1 /*
2 * Copyright (c) 1998, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciMethodData.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/c2_globals.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/opcodes.hpp"
46 #include "opto/predicates.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/vectorization.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "utilities/checkedCast.hpp"
52 #include "utilities/powerOfTwo.hpp"
53
54 //=============================================================================
55 //--------------------------is_cloop_ind_var-----------------------------------
56 // Determine if a node is a counted loop induction variable.
57 // NOTE: The method is declared in "node.hpp".
58 bool Node::is_cloop_ind_var() const {
59 return (is_Phi() &&
60 as_Phi()->region()->is_CountedLoop() &&
61 as_Phi()->region()->as_CountedLoop()->phi() == this);
62 }
63
64 //=============================================================================
65 //------------------------------dump_spec--------------------------------------
66 // Dump special per-node info
67 #ifndef PRODUCT
68 void LoopNode::dump_spec(outputStream *st) const {
69 RegionNode::dump_spec(st);
70 if (is_inner_loop()) st->print( "inner " );
71 if (is_partial_peel_loop()) st->print( "partial_peel " );
72 if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
73 }
74 #endif
75
76 //------------------------------is_valid_counted_loop-------------------------
77 bool LoopNode::is_valid_counted_loop(BasicType bt) const {
78 if (is_BaseCountedLoop() && as_BaseCountedLoop()->bt() == bt) {
79 BaseCountedLoopNode* l = as_BaseCountedLoop();
80 BaseCountedLoopEndNode* le = l->loopexit_or_null();
81 if (le != nullptr &&
82 le->true_proj_or_null() == l->in(LoopNode::LoopBackControl)) {
83 Node* phi = l->phi();
84 IfFalseNode* exit = le->false_proj_or_null();
85 if (exit != nullptr && 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(), "unexpected node");
225 assert(idom(ctl) == parent_ctl, "strange");
226 next = idom(parent_ctl);
227 }
228 } else {
229 next = idom(ctl);
230 }
231 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
232 break;
233 }
234 ctl = next;
235 }
236
237 if (ctl != n->in(0)) {
238 _igvn.replace_input_of(n, 0, ctl);
239 _igvn.hash_insert(n);
240 }
241
242 return ctl;
243 }
244
245
246 //------------------------------set_early_ctrl---------------------------------
247 // Set earliest legal control
248 void PhaseIdealLoop::set_early_ctrl(Node* n, bool update_body) {
249 Node *early = get_early_ctrl(n);
250
251 // Record earliest legal location
252 set_ctrl(n, early);
253 IdealLoopTree *loop = get_loop(early);
254 if (update_body && loop->_child == nullptr) {
255 loop->_body.push(n);
256 }
257 }
258
259 //------------------------------set_subtree_ctrl-------------------------------
260 // set missing _ctrl entries on new nodes
261 void PhaseIdealLoop::set_subtree_ctrl(Node* n, bool update_body) {
262 // Already set? Get out.
263 if (_loop_or_ctrl[n->_idx]) return;
264 // Recursively set _loop_or_ctrl array to indicate where the Node goes
265 uint i;
266 for (i = 0; i < n->req(); ++i) {
267 Node *m = n->in(i);
268 if (m && m != C->root()) {
269 set_subtree_ctrl(m, update_body);
270 }
271 }
272
273 // Fixup self
274 set_early_ctrl(n, update_body);
275 }
276
277 IdealLoopTree* PhaseIdealLoop::insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift) {
278 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
279 IdealLoopTree* parent = loop->_parent;
280 IdealLoopTree* sibling = parent->_child;
281 if (sibling == loop) {
282 parent->_child = outer_ilt;
283 } else {
284 while (sibling->_next != loop) {
285 sibling = sibling->_next;
286 }
287 sibling->_next = outer_ilt;
288 }
289 outer_ilt->_next = loop->_next;
290 outer_ilt->_parent = parent;
291 outer_ilt->_child = loop;
292 outer_ilt->_nest = loop->_nest;
293 loop->_parent = outer_ilt;
294 loop->_next = nullptr;
295 loop->_nest++;
296 assert(loop->_nest <= SHRT_MAX, "sanity");
297 return outer_ilt;
298 }
299
300 // Create a skeleton strip mined outer loop: an OuterStripMinedLoop head before the inner strip mined CountedLoop, a
301 // SafePoint on exit of the inner CountedLoopEnd and an OuterStripMinedLoopEnd test that can't constant fold until loop
302 // optimizations are over. The inner strip mined loop is left as it is. Only once loop optimizations are over, do we
303 // adjust the inner loop exit condition to limit its number of iterations, set the outer loop exit condition and add
304 // Phis to the outer loop head. Some loop optimizations that operate on the inner strip mined loop need to be aware of
305 // the outer strip mined loop: loop unswitching needs to clone the outer loop as well as the inner, unrolling needs to
306 // only clone the inner loop etc. No optimizations need to change the outer strip mined loop as it is only a skeleton.
307 //
308 // Schematically:
309 //
310 // OuterStripMinedLoop -------|
311 // | |
312 // CountedLoop ----------- | |
313 // \- Phi (iv) -| | |
314 // / \ | | |
315 // CmpI AddI --| | |
316 // \ | |
317 // Bool | |
318 // \ | |
319 // CountedLoopEnd | |
320 // / \ | |
321 // IfFalse IfTrue--------| |
322 // | |
323 // SafePoint |
324 // | |
325 // OuterStripMinedLoopEnd |
326 // / \ |
327 // IfFalse IfTrue-----------|
328 // |
329 //
330 //
331 // As loop optimizations transform the inner loop, the outer strip mined loop stays mostly unchanged. The only exception
332 // is nodes referenced from the SafePoint and sunk from the inner loop: they end up in the outer strip mined loop.
333 //
334 // Not adding Phis to the outer loop head from the beginning, and only adding them after loop optimizations does not
335 // conform to C2's IR rules: any variable or memory slice that is mutated in a loop should have a Phi. The main
336 // motivation for such a design that doesn't conform to C2's IR rules is to allow existing loop optimizations to be
337 // mostly unaffected by the outer strip mined loop: the only extra step needed in most cases is to step over the
338 // OuterStripMinedLoop. The main drawback is that once loop optimizations are over, an extra step is needed to finish
339 // constructing the outer loop. This is handled by OuterStripMinedLoopNode::adjust_strip_mined_loop().
340 //
341 // Adding Phis to the outer loop is largely straightforward: there needs to be one Phi in the outer loop for every Phi
342 // in the inner loop. Things may be more complicated for sunk Store nodes: there may not be any inner loop Phi left
343 // after sinking for a particular memory slice but the outer loop needs a Phi. See
344 // OuterStripMinedLoopNode::handle_sunk_stores_when_finishing_construction()
345 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(Node* init_control,
346 IdealLoopTree* loop, float cl_prob, float le_fcnt,
347 Node*& entry_control, Node*& iffalse) {
348 Node* outer_test = intcon(0);
349 Node *orig = iffalse;
350 iffalse = iffalse->clone();
351 _igvn.register_new_node_with_optimizer(iffalse);
352 set_idom(iffalse, idom(orig), dom_depth(orig));
353
354 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
355 Node *outer_ift = new IfTrueNode (outer_le);
356 Node* outer_iff = orig;
357 _igvn.replace_input_of(outer_iff, 0, outer_le);
358
359 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
360 entry_control = outer_l;
361
362 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_l, outer_ift);
363
364 set_loop(iffalse, outer_ilt);
365 // When this code runs, loop bodies have not yet been populated.
366 const bool body_populated = false;
367 register_control(outer_le, outer_ilt, iffalse, body_populated);
368 register_control(outer_ift, outer_ilt, outer_le, body_populated);
369 set_idom(outer_iff, outer_le, dom_depth(outer_le));
370 _igvn.register_new_node_with_optimizer(outer_l);
371 set_loop(outer_l, outer_ilt);
372 set_idom(outer_l, init_control, dom_depth(init_control)+1);
373
374 return outer_ilt;
375 }
376
377 void PhaseIdealLoop::insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj,
378 Node* cmp_limit, Node* bol) {
379 assert(loop_limit_check_parse_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
380 Node* new_predicate_proj = create_new_if_for_predicate(loop_limit_check_parse_proj, nullptr,
381 Deoptimization::Reason_loop_limit_check,
382 Op_If);
383 Node* iff = new_predicate_proj->in(0);
384 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
385 bol = _igvn.register_new_node_with_optimizer(bol);
386 set_subtree_ctrl(bol, false);
387 _igvn.replace_input_of(iff, 1, bol);
388
389 #ifndef PRODUCT
390 // report that the loop predication has been actually performed
391 // for this loop
392 if (TraceLoopLimitCheck) {
393 tty->print_cr("Counted Loop Limit Check generated:");
394 DEBUG_ONLY( bol->dump(2); )
395 }
396 #endif
397 }
398
399 Node* PhaseIdealLoop::loop_exit_control(Node* x, IdealLoopTree* loop) {
400 // Counted loop head must be a good RegionNode with only 3 not null
401 // control input edges: Self, Entry, LoopBack.
402 if (x->in(LoopNode::Self) == nullptr || x->req() != 3 || loop->_irreducible) {
403 return nullptr;
404 }
405 Node *init_control = x->in(LoopNode::EntryControl);
406 Node *back_control = x->in(LoopNode::LoopBackControl);
407 if (init_control == nullptr || back_control == nullptr) { // Partially dead
408 return nullptr;
409 }
410 // Must also check for TOP when looking for a dead loop
411 if (init_control->is_top() || back_control->is_top()) {
412 return nullptr;
413 }
414
415 // Allow funny placement of Safepoint
416 if (back_control->Opcode() == Op_SafePoint) {
417 back_control = back_control->in(TypeFunc::Control);
418 }
419
420 // Controlling test for loop
421 Node *iftrue = back_control;
422 uint iftrue_op = iftrue->Opcode();
423 if (iftrue_op != Op_IfTrue &&
424 iftrue_op != Op_IfFalse) {
425 // I have a weird back-control. Probably the loop-exit test is in
426 // the middle of the loop and I am looking at some trailing control-flow
427 // merge point. To fix this I would have to partially peel the loop.
428 return nullptr; // Obscure back-control
429 }
430
431 // Get boolean guarding loop-back test
432 Node *iff = iftrue->in(0);
433 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) {
434 return nullptr;
435 }
436 return iftrue;
437 }
438
439 Node* PhaseIdealLoop::loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob) {
440 Node* iftrue = back_control;
441 uint iftrue_op = iftrue->Opcode();
442 Node* iff = iftrue->in(0);
443 BoolNode* test = iff->in(1)->as_Bool();
444 bt = test->_test._test;
445 cl_prob = iff->as_If()->_prob;
446 if (iftrue_op == Op_IfFalse) {
447 bt = BoolTest(bt).negate();
448 cl_prob = 1.0 - cl_prob;
449 }
450 // Get backedge compare
451 Node* cmp = test->in(1);
452 if (!cmp->is_Cmp()) {
453 return nullptr;
454 }
455
456 // Find the trip-counter increment & limit. Limit must be loop invariant.
457 incr = cmp->in(1);
458 limit = cmp->in(2);
459
460 // ---------
461 // need 'loop()' test to tell if limit is loop invariant
462 // ---------
463
464 if (!ctrl_is_member(loop, incr)) { // Swapped trip counter and limit?
465 Node* tmp = incr; // Then reverse order into the CmpI
466 incr = limit;
467 limit = tmp;
468 bt = BoolTest(bt).commute(); // And commute the exit test
469 }
470 if (ctrl_is_member(loop, limit)) { // Limit must be loop-invariant
471 return nullptr;
472 }
473 if (!ctrl_is_member(loop, incr)) { // Trip counter must be loop-variant
474 return nullptr;
475 }
476 return cmp;
477 }
478
479 Node* PhaseIdealLoop::loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr) {
480 if (incr->is_Phi()) {
481 if (incr->as_Phi()->region() != x || incr->req() != 3) {
482 return nullptr; // Not simple trip counter expression
483 }
484 phi_incr = incr;
485 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
486 if (!ctrl_is_member(loop, incr)) { // Trip counter must be loop-variant
487 return nullptr;
488 }
489 }
490 return incr;
491 }
492
493 Node* PhaseIdealLoop::loop_iv_stride(Node* incr, Node*& xphi) {
494 assert(incr->Opcode() == Op_AddI || incr->Opcode() == Op_AddL, "caller resp.");
495 // Get merge point
496 xphi = incr->in(1);
497 Node *stride = incr->in(2);
498 if (!stride->is_Con()) { // Oops, swap these
499 if (!xphi->is_Con()) { // Is the other guy a constant?
500 return nullptr; // Nope, unknown stride, bail out
501 }
502 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
503 xphi = stride;
504 stride = tmp;
505 }
506 return stride;
507 }
508
509 PhiNode* PhaseIdealLoop::loop_iv_phi(Node* xphi, Node* phi_incr, Node* x) {
510 if (!xphi->is_Phi()) {
511 return nullptr; // Too much math on the trip counter
512 }
513 if (phi_incr != nullptr && phi_incr != xphi) {
514 return nullptr;
515 }
516 PhiNode *phi = xphi->as_Phi();
517
518 // Phi must be of loop header; backedge must wrap to increment
519 if (phi->region() != x) {
520 return nullptr;
521 }
522 return phi;
523 }
524
525 static int check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt) {
526 if (final_correction > 0) {
527 if (limit_t->lo_as_long() > (max_signed_integer(bt) - final_correction)) {
528 return -1;
529 }
530 if (limit_t->hi_as_long() > (max_signed_integer(bt) - final_correction)) {
531 return 1;
532 }
533 } else {
534 if (limit_t->hi_as_long() < (min_signed_integer(bt) - final_correction)) {
535 return -1;
536 }
537 if (limit_t->lo_as_long() < (min_signed_integer(bt) - final_correction)) {
538 return 1;
539 }
540 }
541 return 0;
542 }
543
544 static bool condition_stride_ok(BoolTest::mask bt, jlong stride_con) {
545 // If the condition is inverted and we will be rolling
546 // through MININT to MAXINT, then bail out.
547 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
548 // Odd stride
549 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
550 // Count down loop rolls through MAXINT
551 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
552 // Count up loop rolls through MININT
553 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
554 return false; // Bail out
555 }
556 return true;
557 }
558
559 Node* PhaseIdealLoop::loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head,
560 BasicType bt) {
561 Node* iv_as_long;
562 if (bt == T_LONG) {
563 iv_as_long = new ConvI2LNode(inner_iv, TypeLong::INT);
564 register_new_node(iv_as_long, inner_head);
565 } else {
566 iv_as_long = inner_iv;
567 }
568 Node* iv_replacement = AddNode::make(outer_phi, iv_as_long, bt);
569 register_new_node(iv_replacement, inner_head);
570 for (DUIterator_Last imin, i = iv_to_replace->last_outs(imin); i >= imin;) {
571 Node* u = iv_to_replace->last_out(i);
572 #ifdef ASSERT
573 if (!is_dominator(inner_head, ctrl_or_self(u))) {
574 assert(u->is_Phi(), "should be a Phi");
575 for (uint j = 1; j < u->req(); j++) {
576 if (u->in(j) == iv_to_replace) {
577 assert(is_dominator(inner_head, u->in(0)->in(j)), "iv use above loop?");
578 }
579 }
580 }
581 #endif
582 _igvn.rehash_node_delayed(u);
583 int nb = u->replace_edge(iv_to_replace, iv_replacement, &_igvn);
584 i -= nb;
585 }
586 return iv_replacement;
587 }
588
589 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
590 void PhaseIdealLoop::add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop,
591 SafePointNode* sfpt) {
592 if (!C->too_many_traps(sfpt->jvms()->method(), sfpt->jvms()->bci(), reason)) {
593 ParsePredicateNode* parse_predicate = new ParsePredicateNode(inner_head->in(LoopNode::EntryControl), reason, &_igvn);
594 register_control(parse_predicate, loop, inner_head->in(LoopNode::EntryControl));
595 Node* if_false = new IfFalseNode(parse_predicate);
596 register_control(if_false, _ltree_root, parse_predicate);
597 Node* if_true = new IfTrueNode(parse_predicate);
598 register_control(if_true, loop, parse_predicate);
599
600 int trap_request = Deoptimization::make_trap_request(reason, Deoptimization::Action_maybe_recompile);
601 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
602 const TypePtr* no_memory_effects = nullptr;
603 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
604 no_memory_effects);
605
606 Node* mem = nullptr;
607 Node* i_o = nullptr;
608 if (sfpt->is_Call()) {
609 mem = sfpt->proj_out(TypeFunc::Memory);
610 i_o = sfpt->proj_out(TypeFunc::I_O);
611 } else {
612 mem = sfpt->memory();
613 i_o = sfpt->i_o();
614 }
615
616 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
617 register_new_node(frame, C->start());
618 Node *ret = new ParmNode(C->start(), TypeFunc::ReturnAdr);
619 register_new_node(ret, C->start());
620
621 unc->init_req(TypeFunc::Control, if_false);
622 unc->init_req(TypeFunc::I_O, i_o);
623 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
624 unc->init_req(TypeFunc::FramePtr, frame);
625 unc->init_req(TypeFunc::ReturnAdr, ret);
626 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
627 unc->set_cnt(PROB_UNLIKELY_MAG(4));
628 unc->copy_call_debug_info(&_igvn, sfpt);
629
630 for (uint i = TypeFunc::Parms; i < unc->req(); i++) {
631 set_subtree_ctrl(unc->in(i), false);
632 }
633 register_control(unc, _ltree_root, if_false);
634
635 Node* ctrl = new ProjNode(unc, TypeFunc::Control);
636 register_control(ctrl, _ltree_root, unc);
637 Node* halt = new HaltNode(ctrl, frame, "uncommon trap returned which should never happen" PRODUCT_ONLY(COMMA /*reachable*/false));
638 register_control(halt, _ltree_root, ctrl);
639 _igvn.add_input_to(C->root(), halt);
640
641 _igvn.replace_input_of(inner_head, LoopNode::EntryControl, if_true);
642 set_idom(inner_head, if_true, dom_depth(inner_head));
643 }
644 }
645
646 // Find a safepoint node that dominates the back edge. We need a
647 // SafePointNode so we can use its jvm state to create empty
648 // predicates.
649 static bool no_side_effect_since_safepoint(Compile* C, Node* x, Node* mem, MergeMemNode* mm, PhaseIdealLoop* phase) {
650 SafePointNode* safepoint = nullptr;
651 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
652 Node* u = x->fast_out(i);
653 if (u->is_memory_phi()) {
654 Node* m = u->in(LoopNode::LoopBackControl);
655 if (u->adr_type() == TypePtr::BOTTOM) {
656 if (m->is_MergeMem() && mem->is_MergeMem()) {
657 if (m != mem DEBUG_ONLY(|| true)) {
658 // MergeMemStream can modify m, for example to adjust the length to mem.
659 // This is unfortunate, and probably unnecessary. But as it is, we need
660 // to add m to the igvn worklist, else we may have a modified node that
661 // is not on the igvn worklist.
662 phase->igvn()._worklist.push(m);
663 for (MergeMemStream mms(m->as_MergeMem(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
664 if (!mms.is_empty()) {
665 if (mms.memory() != mms.memory2()) {
666 return false;
667 }
668 #ifdef ASSERT
669 if (mms.alias_idx() != Compile::AliasIdxBot) {
670 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
671 }
672 #endif
673 }
674 }
675 }
676 } else if (mem->is_MergeMem()) {
677 if (m != mem->as_MergeMem()->base_memory()) {
678 return false;
679 }
680 } else {
681 return false;
682 }
683 } else {
684 if (mem->is_MergeMem()) {
685 if (m != mem->as_MergeMem()->memory_at(C->get_alias_index(u->adr_type()))) {
686 return false;
687 }
688 #ifdef ASSERT
689 mm->set_memory_at(C->get_alias_index(u->adr_type()), mem->as_MergeMem()->base_memory());
690 #endif
691 } else {
692 if (m != mem) {
693 return false;
694 }
695 }
696 }
697 }
698 }
699 return true;
700 }
701
702 SafePointNode* PhaseIdealLoop::find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop) {
703 IfNode* exit_test = back_control->in(0)->as_If();
704 SafePointNode* safepoint = nullptr;
705 if (exit_test->in(0)->is_SafePoint() && exit_test->in(0)->outcnt() == 1) {
706 safepoint = exit_test->in(0)->as_SafePoint();
707 } else {
708 Node* c = back_control;
709 while (c != x && c->Opcode() != Op_SafePoint) {
710 c = idom(c);
711 }
712
713 if (c->Opcode() == Op_SafePoint) {
714 safepoint = c->as_SafePoint();
715 }
716
717 if (safepoint == nullptr) {
718 return nullptr;
719 }
720
721 Node* mem = safepoint->in(TypeFunc::Memory);
722
723 // We can only use that safepoint if there's no side effect between the backedge and the safepoint.
724
725 // mm is the memory state at the safepoint (when it's a MergeMem)
726 // no_side_effect_since_safepoint() goes over the memory state at the backedge. It resets the mm input for each
727 // component of the memory state it encounters so it points to the base memory. Once no_side_effect_since_safepoint()
728 // is done, if no side effect after the safepoint was found, mm should transform to the base memory: the states at
729 // the backedge and safepoint are the same so all components of the memory state at the safepoint should have been
730 // reset.
731 MergeMemNode* mm = nullptr;
732 #ifdef ASSERT
733 if (mem->is_MergeMem()) {
734 mm = mem->clone()->as_MergeMem();
735 _igvn._worklist.push(mm);
736 for (MergeMemStream mms(mem->as_MergeMem()); mms.next_non_empty(); ) {
737 // Loop invariant memory state won't be reset by no_side_effect_since_safepoint(). Do it here.
738 // Escape Analysis can add state to mm that it doesn't add to the backedge memory Phis, breaking verification
739 // code that relies on mm. Clear that extra state here.
740 if (mms.alias_idx() != Compile::AliasIdxBot &&
741 (loop != get_loop(ctrl_or_self(mms.memory())) ||
742 (mms.adr_type()->isa_oop_ptr() && mms.adr_type()->is_known_instance()))) {
743 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
744 }
745 }
746 }
747 #endif
748 if (!no_side_effect_since_safepoint(C, x, mem, mm, this)) {
749 safepoint = nullptr;
750 } else {
751 assert(mm == nullptr|| _igvn.transform(mm) == mem->as_MergeMem()->base_memory(), "all memory state should have been processed");
752 }
753 #ifdef ASSERT
754 if (mm != nullptr) {
755 _igvn.remove_dead_node(mm);
756 }
757 #endif
758 }
759 return safepoint;
760 }
761
762 void PhaseIdealLoop::add_parse_predicates(IdealLoopTree* outer_ilt, LoopNode* inner_head, SafePointNode* cloned_sfpt) {
763 if (ShortRunningLongLoop) {
764 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, inner_head, outer_ilt, cloned_sfpt);
765 }
766 if (UseLoopPredicate) {
767 add_parse_predicate(Deoptimization::Reason_predicate, inner_head, outer_ilt, cloned_sfpt);
768 if (UseProfiledLoopPredicate) {
769 add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
770 }
771 }
772
773 if (UseAutoVectorizationPredicate) {
774 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, inner_head, outer_ilt, cloned_sfpt);
775 }
776
777 add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
778 }
779
780 // If the loop has the shape of a counted loop but with a long
781 // induction variable, transform the loop in a loop nest: an inner
782 // loop that iterates for at most max int iterations with an integer
783 // induction variable and an outer loop that iterates over the full
784 // range of long values from the initial loop in (at most) max int
785 // steps. That is:
786 //
787 // x: for (long phi = init; phi < limit; phi += stride) {
788 // // phi := Phi(L, init, incr)
789 // // incr := AddL(phi, longcon(stride))
790 // long incr = phi + stride;
791 // ... use phi and incr ...
792 // }
793 //
794 // OR:
795 //
796 // x: for (long phi = init; (phi += stride) < limit; ) {
797 // // phi := Phi(L, AddL(init, stride), incr)
798 // // incr := AddL(phi, longcon(stride))
799 // long incr = phi + stride;
800 // ... use phi and (phi + stride) ...
801 // }
802 //
803 // ==transform=>
804 //
805 // const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0
806 // assert(stride <= inner_iters_limit); // else abort transform
807 // assert((extralong)limit + stride <= LONG_MAX); // else deopt
808 // outer_head: for (long outer_phi = init;;) {
809 // // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi)))
810 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi));
811 // long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max);
812 // assert(inner_iters_actual == (int)inner_iters_actual);
813 // int inner_phi, inner_incr;
814 // x: for (inner_phi = 0;; inner_phi = inner_incr) {
815 // // inner_phi := Phi(x, intcon(0), inner_incr)
816 // // inner_incr := AddI(inner_phi, intcon(stride))
817 // inner_incr = inner_phi + stride;
818 // if (inner_incr < inner_iters_actual) {
819 // ... use phi=>(outer_phi+inner_phi) ...
820 // continue;
821 // }
822 // else break;
823 // }
824 // if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit
825 // continue;
826 // else break;
827 // }
828 //
829 // The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int
830 // loops with long range checks transformed to int range checks in the inner loop.
831 bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {
832 Node* x = loop->_head;
833 // Only for inner loops
834 if (loop->_child != nullptr || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) {
835 return false;
836 }
837
838 if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) {
839 return false;
840 }
841
842 BaseCountedLoopNode* head = x->as_BaseCountedLoop();
843 BasicType bt = x->as_BaseCountedLoop()->bt();
844
845 check_counted_loop_shape(loop, x, bt);
846
847 #ifndef PRODUCT
848 if (bt == T_LONG) {
849 AtomicAccess::inc(&_long_loop_candidates);
850 }
851 #endif
852
853 jlong stride_con_long = head->stride_con();
854 assert(stride_con_long != 0, "missed some peephole opt");
855 // We can't iterate for more than max int at a time.
856 if (stride_con_long != (jint)stride_con_long || stride_con_long == min_jint) {
857 assert(bt == T_LONG, "only for long loops");
858 return false;
859 }
860 jint stride_con = checked_cast<jint>(stride_con_long);
861 // The number of iterations for the integer count loop: guarantee no
862 // overflow: max_jint - stride_con max. -1 so there's no need for a
863 // loop limit check if the exit test is <= or >=.
864 int iters_limit = max_jint - ABS(stride_con) - 1;
865 #ifdef ASSERT
866 if (bt == T_LONG && StressLongCountedLoop > 0) {
867 iters_limit = iters_limit / StressLongCountedLoop;
868 }
869 #endif
870 // At least 2 iterations so counted loop construction doesn't fail
871 if (iters_limit/ABS(stride_con) < 2) {
872 return false;
873 }
874
875 assert(iters_limit > 0, "can't be negative");
876
877 PhiNode* phi = head->phi()->as_Phi();
878
879 Node* back_control = head->in(LoopNode::LoopBackControl);
880
881 // data nodes on back branch not supported
882 if (back_control->outcnt() > 1) {
883 return false;
884 }
885
886 Node* limit = head->limit();
887 // We'll need to use the loop limit before the inner loop is entered
888 if (!is_dominator(get_ctrl(limit), x)) {
889 return false;
890 }
891
892 IfNode* exit_test = head->loopexit();
893
894 assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
895
896 Node_List range_checks;
897 iters_limit = extract_long_range_checks(loop, stride_con, iters_limit, phi, range_checks);
898
899 if (bt == T_INT) {
900 // The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further.
901 if (range_checks.size() == 0) {
902 return false;
903 }
904 }
905
906 // Take what we know about the number of iterations of the long counted loop into account when computing the limit of
907 // the inner loop.
908 Node* init = head->init_trip();
909 const TypeInteger* lo = _igvn.type(init)->is_integer(bt);
910 const TypeInteger* hi = _igvn.type(limit)->is_integer(bt);
911 if (stride_con < 0) {
912 swap(lo, hi);
913 }
914 if (hi->hi_as_long() <= lo->lo_as_long()) {
915 // not a loop after all
916 return false;
917 }
918
919 if (range_checks.size() > 0) {
920 // This transformation requires peeling one iteration. Also, if it has range checks and they are eliminated by Loop
921 // Predication, then 2 Hoisted Check Predicates are added for one range check. Finally, transforming a long range
922 // check requires extra logic to be executed before the loop is entered and for the outer loop. As a result, the
923 // transformations can't pay off for a small number of iterations: roughly, if the loop runs for 3 iterations, it's
924 // going to execute as many range checks once transformed with range checks eliminated (1 peeled iteration with
925 // range checks + 2 predicates per range checks) as it would have not transformed. It also has to pay for the extra
926 // logic on loop entry and for the outer loop.
927 loop->compute_trip_count(this, bt);
928 if (head->is_CountedLoop() && head->as_CountedLoop()->has_exact_trip_count()) {
929 if (head->as_CountedLoop()->trip_count() <= 3) {
930 return false;
931 }
932 } else {
933 loop->compute_profile_trip_cnt(this);
934 if (!head->is_profile_trip_failed() && head->profile_trip_cnt() <= 3) {
935 return false;
936 }
937 }
938 }
939
940 if (try_make_short_running_loop(loop, stride_con, range_checks, iters_limit)) {
941 C->set_major_progress();
942 return true;
943 }
944
945 julong orig_iters = (julong)hi->hi_as_long() - lo->lo_as_long();
946 iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
947
948 // We need a safepoint to insert Parse Predicates for the inner loop.
949 SafePointNode* safepoint;
950 if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) {
951 // Loop is strip mined: use the safepoint of the outer strip mined loop
952 OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
953 assert(outer_loop != nullptr, "no outer loop");
954 safepoint = outer_loop->outer_safepoint();
955 outer_loop->transform_to_counted_loop(&_igvn, this);
956 exit_test = head->loopexit();
957 } else {
958 safepoint = find_safepoint(back_control, x, loop);
959 }
960
961 IfFalseNode* exit_branch = exit_test->false_proj();
962 Node* entry_control = head->in(LoopNode::EntryControl);
963
964 // Clone the control flow of the loop to build an outer loop
965 Node* outer_back_branch = back_control->clone();
966 Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
967 Node* inner_exit_branch = exit_branch->clone();
968
969 LoopNode* outer_head = new LoopNode(entry_control, outer_back_branch);
970 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_head, outer_back_branch);
971
972 const bool body_populated = true;
973 register_control(outer_head, outer_ilt, entry_control, body_populated);
974
975 _igvn.register_new_node_with_optimizer(inner_exit_branch);
976 set_loop(inner_exit_branch, outer_ilt);
977 set_idom(inner_exit_branch, exit_test, dom_depth(exit_branch));
978
979 outer_exit_test->set_req(0, inner_exit_branch);
980 register_control(outer_exit_test, outer_ilt, inner_exit_branch, body_populated);
981
982 _igvn.replace_input_of(exit_branch, 0, outer_exit_test);
983 set_idom(exit_branch, outer_exit_test, dom_depth(exit_branch));
984
985 outer_back_branch->set_req(0, outer_exit_test);
986 register_control(outer_back_branch, outer_ilt, outer_exit_test, body_populated);
987
988 _igvn.replace_input_of(x, LoopNode::EntryControl, outer_head);
989 set_idom(x, outer_head, dom_depth(x));
990
991 // add an iv phi to the outer loop and use it to compute the inner
992 // loop iteration limit
993 Node* outer_phi = phi->clone();
994 outer_phi->set_req(0, outer_head);
995 register_new_node(outer_phi, outer_head);
996
997 Node* inner_iters_max = nullptr;
998 if (stride_con > 0) {
999 inner_iters_max = MinMaxNode::max_diff_with_zero(limit, outer_phi, TypeInteger::bottom(bt), _igvn);
1000 } else {
1001 inner_iters_max = MinMaxNode::max_diff_with_zero(outer_phi, limit, TypeInteger::bottom(bt), _igvn);
1002 }
1003
1004 Node* inner_iters_limit = _igvn.integercon(iters_limit, bt);
1005 // inner_iters_max may not fit in a signed integer (iterating from
1006 // Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned
1007 // min.
1008 const TypeInteger* inner_iters_actual_range = TypeInteger::make(0, iters_limit, Type::WidenMin, bt);
1009 Node* inner_iters_actual = MinMaxNode::unsigned_min(inner_iters_max, inner_iters_limit, inner_iters_actual_range, _igvn);
1010
1011 Node* inner_iters_actual_int;
1012 if (bt == T_LONG) {
1013 inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
1014 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1015 // When the inner loop is transformed to a counted loop, a loop limit check is not expected to be needed because
1016 // the loop limit is less or equal to max_jint - stride - 1 (if stride is positive but a similar argument exists for
1017 // a negative stride). We add a CastII here to guarantee that, when the counted loop is created in a subsequent loop
1018 // opts pass, an accurate range of values for the limits is found.
1019 const TypeInt* inner_iters_actual_int_range = TypeInt::make(0, iters_limit, Type::WidenMin);
1020 inner_iters_actual_int = new CastIINode(outer_head, inner_iters_actual_int, inner_iters_actual_int_range, ConstraintCastNode::DependencyType::NonFloatingNonNarrowing);
1021 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1022 } else {
1023 inner_iters_actual_int = inner_iters_actual;
1024 }
1025
1026 Node* int_zero = intcon(0);
1027 if (stride_con < 0) {
1028 inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
1029 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1030 }
1031
1032 // Clone the iv data nodes as an integer iv
1033 Node* int_stride = intcon(stride_con);
1034 Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
1035 Node* inner_incr = new AddINode(inner_phi, int_stride);
1036 Node* inner_cmp = nullptr;
1037 inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
1038 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
1039 inner_phi->set_req(LoopNode::EntryControl, int_zero);
1040 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
1041 register_new_node(inner_phi, x);
1042 register_new_node(inner_incr, x);
1043 register_new_node(inner_cmp, x);
1044 register_new_node(inner_bol, x);
1045
1046 _igvn.replace_input_of(exit_test, 1, inner_bol);
1047
1048 // Clone inner loop phis to outer loop
1049 for (uint i = 0; i < head->outcnt(); i++) {
1050 Node* u = head->raw_out(i);
1051 if (u->is_Phi() && u != inner_phi && u != phi) {
1052 assert(u->in(0) == head, "inconsistent");
1053 Node* clone = u->clone();
1054 clone->set_req(0, outer_head);
1055 register_new_node(clone, outer_head);
1056 _igvn.replace_input_of(u, LoopNode::EntryControl, clone);
1057 }
1058 }
1059
1060 // Replace inner loop long iv phi as inner loop int iv phi + outer
1061 // loop iv phi
1062 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
1063
1064 set_subtree_ctrl(inner_iters_actual_int, body_populated);
1065
1066 LoopNode* inner_head = create_inner_head(loop, head, exit_test);
1067
1068 // Summary of steps from initial loop to loop nest:
1069 //
1070 // == old IR nodes =>
1071 //
1072 // entry_control: {...}
1073 // x:
1074 // for (long phi = init;;) {
1075 // // phi := Phi(x, init, incr)
1076 // // incr := AddL(phi, longcon(stride))
1077 // exit_test:
1078 // if (phi < limit)
1079 // back_control: fallthrough;
1080 // else
1081 // exit_branch: break;
1082 // long incr = phi + stride;
1083 // ... use phi and incr ...
1084 // phi = incr;
1085 // }
1086 //
1087 // == new IR nodes (just before final peel) =>
1088 //
1089 // entry_control: {...}
1090 // long adjusted_limit = limit + stride; //because phi_incr != nullptr
1091 // assert(!limit_check_required || (extralong)limit + stride == adjusted_limit); // else deopt
1092 // ulong inner_iters_limit = max_jint - ABS(stride) - 1; //near 0x7FFFFFF0
1093 // outer_head:
1094 // for (long outer_phi = init;;) {
1095 // // outer_phi := phi->clone(), in(0):=outer_head, => Phi(outer_head, init, incr)
1096 // // REPLACE phi => AddL(outer_phi, I2L(inner_phi))
1097 // // REPLACE incr => AddL(outer_phi, I2L(inner_incr))
1098 // // SO THAT outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_incr)))
1099 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)adjusted_limit - outer_phi) * SGN(stride));
1100 // int inner_iters_actual_int = (int) MIN(inner_iters_limit, inner_iters_max) * SGN(stride);
1101 // inner_head: x: //in(1) := outer_head
1102 // int inner_phi;
1103 // for (inner_phi = 0;;) {
1104 // // inner_phi := Phi(x, intcon(0), inner_phi + stride)
1105 // int inner_incr = inner_phi + stride;
1106 // bool inner_bol = (inner_incr < inner_iters_actual_int);
1107 // exit_test: //exit_test->in(1) := inner_bol;
1108 // if (inner_bol) // WAS (phi < limit)
1109 // back_control: fallthrough;
1110 // else
1111 // inner_exit_branch: break; //exit_branch->clone()
1112 // ... use phi=>(outer_phi+inner_phi) ...
1113 // inner_phi = inner_phi + stride; // inner_incr
1114 // }
1115 // outer_exit_test: //exit_test->clone(), in(0):=inner_exit_branch
1116 // if ((outer_phi+inner_phi) < limit) // WAS (phi < limit)
1117 // outer_back_branch: fallthrough; //back_control->clone(), in(0):=outer_exit_test
1118 // else
1119 // exit_branch: break; //in(0) := outer_exit_test
1120 // }
1121
1122 if (bt == T_INT) {
1123 outer_phi = new ConvI2LNode(outer_phi);
1124 register_new_node(outer_phi, outer_head);
1125 }
1126
1127 transform_long_range_checks(stride_con, range_checks, outer_phi, inner_iters_actual_int,
1128 inner_phi, iv_add, inner_head);
1129 // Peel one iteration of the loop and use the safepoint at the end
1130 // of the peeled iteration to insert Parse Predicates. If no well
1131 // positioned safepoint peel to guarantee a safepoint in the outer
1132 // loop.
1133 if (safepoint != nullptr || !loop->_has_call) {
1134 old_new.clear();
1135 do_peeling(loop, old_new);
1136 } else {
1137 C->set_major_progress();
1138 }
1139
1140 if (safepoint != nullptr) {
1141 SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
1142
1143 add_parse_predicates(outer_ilt, inner_head, cloned_sfpt);
1144 }
1145
1146 #ifndef PRODUCT
1147 if (bt == T_LONG) {
1148 AtomicAccess::inc(&_long_loop_nests);
1149 }
1150 #endif
1151
1152 inner_head->mark_loop_nest_inner_loop();
1153 outer_head->mark_loop_nest_outer_loop();
1154
1155 return true;
1156 }
1157
1158 // Make a copy of Parse/Template Assertion predicates below existing predicates at the loop passed as argument
1159 class CloneShortLoopPredicateVisitor : public PredicateVisitor {
1160 ClonePredicateToTargetLoop _clone_predicate_to_loop;
1161 PhaseIdealLoop* const _phase;
1162 Node* const _new_init;
1163
1164 public:
1165 CloneShortLoopPredicateVisitor(LoopNode* target_loop_head,
1166 Node* new_init,
1167 const NodeInSingleLoopBody &node_in_loop_body,
1168 PhaseIdealLoop* phase)
1169 : _clone_predicate_to_loop(target_loop_head, node_in_loop_body, phase),
1170 _phase(phase),
1171 _new_init(new_init) {
1172 }
1173 NONCOPYABLE(CloneShortLoopPredicateVisitor);
1174
1175 using PredicateVisitor::visit;
1176
1177 void visit(const ParsePredicate& parse_predicate) override {
1178 _clone_predicate_to_loop.clone_parse_predicate(parse_predicate, true);
1179 parse_predicate.kill(_phase->igvn());
1180 }
1181
1182 void visit(const TemplateAssertionPredicate& template_assertion_predicate) override {
1183 _clone_predicate_to_loop.clone_template_assertion_predicate_and_replace_init(template_assertion_predicate, _new_init);
1184 template_assertion_predicate.kill(_phase->igvn());
1185 }
1186 };
1187
1188 // For an int counted loop, try_make_short_running_loop() transforms the loop from:
1189 // for (int = start; i < stop; i+= stride) { ... }
1190 // to
1191 // for (int = 0; i < stop - start; i+= stride) { ... }
1192 // Template Assertion Predicates added so far were with an init value of start. They need to be updated with the new
1193 // init value of 0 (otherwise when a template assertion predicate is turned into an initialized assertion predicate, it
1194 // performs an incorrect check):
1195 // zero
1196 // init |
1197 // | ===> OpaqueLoopInit init
1198 // OpaqueLoopInit \ /
1199 // AddI
1200 //
1201 Node* PhaseIdealLoop::new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero) {
1202 OpaqueLoopInitNode* new_opaque_init = new OpaqueLoopInitNode(C, int_zero);
1203 register_new_node(new_opaque_init, entry_control);
1204 Node* new_init = new AddINode(new_opaque_init, init);
1205 register_new_node(new_init, entry_control);
1206 return new_init;
1207 }
1208
1209 // If the loop is either statically known to run for a small enough number of iterations or if profile data indicates
1210 // that, we don't want an outer loop because the overhead of having an outer loop whose backedge is never taken, has a
1211 // measurable cost. Furthermore, creating the loop nest usually causes one iteration of the loop to be peeled so
1212 // predicates can be set up. If the loop is short running, then it's an extra iteration that's run with range checks
1213 // (compared to an int counted loop with int range checks).
1214 //
1215 // In the short running case, turn the loop into a regular loop again and transform the long range checks:
1216 // - LongCountedLoop: Create LoopNode but keep the loop limit type with a CastLL node to avoid that we later try to
1217 // create a Loop Limit Check when turning the LoopNode into a CountedLoopNode.
1218 // - CountedLoop: Can be reused.
1219 bool PhaseIdealLoop::try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List &range_checks,
1220 const uint iters_limit) {
1221 if (!ShortRunningLongLoop) {
1222 return false;
1223 }
1224 BaseCountedLoopNode* head = loop->_head->as_BaseCountedLoop();
1225 BasicType bt = head->bt();
1226 Node* entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1227
1228 loop->compute_trip_count(this, bt);
1229 // Loop must run for no more than iter_limits as it guarantees no overflow of scale * iv in long range checks (see
1230 // comment above PhaseIdealLoop::transform_long_range_checks()).
1231 // iters_limit / ABS(stride_con) is the largest trip count for which we know it's correct to not create a loop nest:
1232 // it's always beneficial to have a single loop rather than a loop nest, so we try to apply this transformation as
1233 // often as possible.
1234 bool known_short_running_loop = head->trip_count() <= iters_limit / ABS(stride_con);
1235 bool profile_short_running_loop = false;
1236 if (!known_short_running_loop) {
1237 loop->compute_profile_trip_cnt(this);
1238 if (StressShortRunningLongLoop) {
1239 profile_short_running_loop = true;
1240 } else {
1241 profile_short_running_loop = !head->is_profile_trip_failed() && head->profile_trip_cnt() <= iters_limit / ABS(stride_con);
1242 }
1243 }
1244
1245 if (!known_short_running_loop && !profile_short_running_loop) {
1246 return false;
1247 }
1248
1249 Node* limit = head->limit();
1250 Node* init = head->init_trip();
1251
1252 Node* new_limit;
1253 if (stride_con > 0) {
1254 new_limit = SubNode::make(limit, init, bt);
1255 } else {
1256 new_limit = SubNode::make(init, limit, bt);
1257 }
1258 register_new_node(new_limit, entry_control);
1259
1260 Node* int_zero = intcon(0);
1261 PhiNode* phi = head->phi()->as_Phi();
1262 if (profile_short_running_loop) {
1263 // Add a Short Running Long Loop Predicate. It's the first predicate in the predicate chain before entering a loop
1264 // because a cast that's control dependent on the Short Running Long Loop Predicate is added to narrow the limit and
1265 // future predicates may be dependent on the new limit (so have to be between the loop and Short Running Long Loop
1266 // Predicate). The current limit could, itself, be dependent on an existing predicate. Clone parse and template
1267 // assertion predicates below existing predicates to get proper ordering of predicates when walking from the loop
1268 // up: future predicates, Short Running Long Loop Predicate, existing predicates.
1269 //
1270 // Existing Hoisted
1271 // Check Predicates
1272 // |
1273 // New Short Running Long
1274 // Loop Predicate
1275 // |
1276 // Cloned Parse Predicates and
1277 // Template Assertion Predicates
1278 // (future predicates added here)
1279 // |
1280 // Loop
1281 const Predicates predicates_before_cloning(entry_control);
1282 const PredicateBlock* short_running_long_loop_predicate_block = predicates_before_cloning.short_running_long_loop_predicate_block();
1283 if (!short_running_long_loop_predicate_block->has_parse_predicate()) { // already trapped
1284 return false;
1285 }
1286 Node* new_init = new_assertion_predicate_opaque_init(entry_control, init, int_zero);
1287
1288 PredicateIterator predicate_iterator(entry_control);
1289 NodeInSingleLoopBody node_in_short_loop_body(this, loop);
1290 CloneShortLoopPredicateVisitor clone_short_loop_predicates_visitor(head, new_init, node_in_short_loop_body, this);
1291 predicate_iterator.for_each(clone_short_loop_predicates_visitor);
1292
1293 entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1294
1295 const Predicates predicates_after_cloning(entry_control);
1296
1297 ParsePredicateSuccessProj* short_running_loop_predicate_proj = predicates_after_cloning.
1298 short_running_long_loop_predicate_block()->
1299 parse_predicate_success_proj();
1300 assert(short_running_loop_predicate_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
1301
1302 const jlong iters_limit_long = iters_limit;
1303 Node* cmp_limit = CmpNode::make(new_limit, _igvn.integercon(iters_limit_long, bt), bt);
1304 Node* bol = new BoolNode(cmp_limit, BoolTest::le);
1305 Node* new_predicate_proj = create_new_if_for_predicate(short_running_loop_predicate_proj,
1306 nullptr,
1307 Deoptimization::Reason_short_running_long_loop,
1308 Op_If);
1309 Node* iff = new_predicate_proj->in(0);
1310 _igvn.replace_input_of(iff, 1, bol);
1311 register_new_node(cmp_limit, iff->in(0));
1312 register_new_node(bol, iff->in(0));
1313 new_limit = ConstraintCastNode::make_cast_for_basic_type(new_predicate_proj, new_limit,
1314 TypeInteger::make(1, iters_limit_long, Type::WidenMin, bt),
1315 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, bt);
1316 register_new_node(new_limit, new_predicate_proj);
1317
1318 #ifndef PRODUCT
1319 if (TraceLoopLimitCheck) {
1320 tty->print_cr("Short Long Loop Check Predicate generated:");
1321 DEBUG_ONLY(bol->dump(2);)
1322 }
1323 #endif
1324 entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1325 } else if (bt == T_LONG) {
1326 // We're turning a long counted loop into a regular loop that will be converted into an int counted loop. That loop
1327 // won't need loop limit check predicates (iters_limit guarantees that). Add a cast to make sure that, whatever
1328 // transformation happens by the time the counted loop is created (in a subsequent pass of loop opts), C2 knows
1329 // enough about the loop's limit that it doesn't try to add loop limit check predicates.
1330 const Predicates predicates(entry_control);
1331 const TypeLong* new_limit_t = new_limit->Value(&_igvn)->is_long();
1332 new_limit = ConstraintCastNode::make_cast_for_basic_type(predicates.entry(), new_limit,
1333 TypeLong::make(0, new_limit_t->_hi, new_limit_t->_widen),
1334 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, bt);
1335 register_new_node(new_limit, predicates.entry());
1336 } else {
1337 assert(bt == T_INT && known_short_running_loop, "only CountedLoop statically known to be short running");
1338 PredicateIterator predicate_iterator(entry_control);
1339 Node* new_init = new_assertion_predicate_opaque_init(entry_control, init, int_zero);
1340 UpdateInitForTemplateAssertionPredicates update_init_for_template_assertion_predicates(new_init, this);
1341 predicate_iterator.for_each(update_init_for_template_assertion_predicates);
1342 }
1343 IfNode* exit_test = head->loopexit();
1344
1345 if (bt == T_LONG) {
1346 // The loop is short running so new_limit fits into an int: either we determined that statically or added a guard
1347 new_limit = new ConvL2INode(new_limit);
1348 register_new_node(new_limit, entry_control);
1349 }
1350
1351 if (stride_con < 0) {
1352 new_limit = new SubINode(int_zero, new_limit);
1353 register_new_node(new_limit, entry_control);
1354 }
1355
1356 // Clone the iv data nodes as an integer iv
1357 Node* int_stride = intcon(stride_con);
1358 Node* inner_phi = new PhiNode(head, TypeInt::INT);
1359 Node* inner_incr = new AddINode(inner_phi, int_stride);
1360 Node* inner_cmp = new CmpINode(inner_incr, new_limit);
1361 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
1362 inner_phi->set_req(LoopNode::EntryControl, int_zero);
1363 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
1364 register_new_node(inner_phi, head);
1365 register_new_node(inner_incr, head);
1366 register_new_node(inner_cmp, head);
1367 register_new_node(inner_bol, head);
1368
1369 _igvn.replace_input_of(exit_test, 1, inner_bol);
1370
1371 // Replace inner loop long iv phi as inner loop int iv phi + outer
1372 // loop iv phi
1373 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, init, head, bt);
1374
1375 LoopNode* inner_head = head;
1376 if (bt == T_LONG) {
1377 // Turn the loop back to a counted loop
1378 inner_head = create_inner_head(loop, head, exit_test);
1379 } else {
1380 // Use existing counted loop
1381 revert_to_normal_loop(head);
1382 }
1383
1384 if (bt == T_INT) {
1385 init = new ConvI2LNode(init);
1386 register_new_node(init, entry_control);
1387 }
1388
1389 transform_long_range_checks(stride_con, range_checks, init, new_limit,
1390 inner_phi, iv_add, inner_head);
1391
1392 inner_head->mark_loop_nest_inner_loop();
1393
1394 return true;
1395 }
1396
1397 int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1398 Node_List& range_checks) {
1399 const jlong min_iters = 2;
1400 jlong reduced_iters_limit = iters_limit;
1401 jlong original_iters_limit = iters_limit;
1402 for (uint i = 0; i < loop->_body.size(); i++) {
1403 Node* c = loop->_body.at(i);
1404 if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
1405 IfProjNode* if_proj = c->as_IfProj();
1406 CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1407 if (call != nullptr) {
1408 Node* range = nullptr;
1409 Node* offset = nullptr;
1410 jlong scale = 0;
1411 if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1412 loop->is_invariant(range) && loop->is_invariant(offset) &&
1413 scale != min_jlong &&
1414 original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {
1415 assert(scale == (jint)scale, "scale should be an int");
1416 reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1417 range_checks.push(c);
1418 }
1419 }
1420 }
1421 }
1422
1423 return checked_cast<int>(reduced_iters_limit);
1424 }
1425
1426 // One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1427 // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1428 // 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
1429 // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1430
1431 // 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,
1432 // 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)
1433 // 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].
1434
1435 // 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
1436 // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1437 // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1438
1439 // 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
1440 // 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
1441 // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1442 // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1443
1444 // 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
1445 // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1446
1447 // 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].
1448 // 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
1449 // 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
1450 // 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
1451 // down (from C and 0 again).
1452
1453 // 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).
1454 // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1455 // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1456 // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1457 // PhaseIdealLoop::add_constraint.)
1458
1459 // 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
1460 // 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
1461 // 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.
1462 // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1463 // values for the 32-bit counter j within which the range checks can be eliminated.
1464
1465 // 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
1466 // 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
1467 // allowed as an input to the R.C., as long as the R.C. as a whole fails.
1468
1469 // If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1470
1471 // 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
1472 // 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),
1473 // 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)
1474
1475 // 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).
1476 // 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.
1477 // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1478
1479 // Case A: Some Negatives (but no overflow).
1480 // Number line:
1481 // |s64_min . . . 0 . . . s64_max|
1482 // | . Q_min..Q_max . 0 . . . . | s64 negative
1483 // | . . . . R=0 R< R< R< R< | (against R values)
1484 // | . . . Q_min..0..Q_max . . . | small mixed
1485 // | . . . . R R R< R< R< | (against R values)
1486 //
1487 // 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<.
1488 //
1489 // So, if Q_min <s64 0, then use this test:
1490 // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1491 // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1492 // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1493 // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1494
1495 // If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1496
1497 // Case B: No Negatives.
1498 // Number line:
1499 // |s64_min . . . 0 . . . s64_max|
1500 // | . . . . 0 Q_min..Q_max . . | small positive
1501 // | . . . . R> R R R< R< | (against R values)
1502 // | . . . . 0 . Q_min..Q_max . | s64 positive
1503 // | . . . . R> R> R R R< | (against R values)
1504 //
1505 // 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.
1506 // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1507 //
1508 // So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1509 // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1510 // More generally:
1511 // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1512
1513 // Case C: Overflow in the 64-bit domain
1514 // Number line:
1515 // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1516 // | . . . . R> R> R> R> R | (against R values)
1517 //
1518 // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1519 // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1520 // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1521 //
1522 // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1523 // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1524 // More generally:
1525 // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1526 //
1527 // Dropping the bad bound means only Q_min is used to reduce the range of R:
1528 // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1529 //
1530 // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1531 // clamp(X, L, H) := max(L, min(X, H))
1532 // When degenerately L > H, it returns L not H.
1533 //
1534 // All of the formulas above can be merged into a single one:
1535 // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1536 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1537 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1538 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1539 // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1540 // replacement R.C.:
1541 // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1542 // or equivalently:
1543 // j*K + L_2 <u32 R_2
1544 // where
1545 // L_2 = Q_first - L_clamp
1546 // R_2 = R_clamp - L_clamp
1547 //
1548 // Note on why R is never negative:
1549 //
1550 // Various details of this transformation would break badly if R could be negative, so this transformation only
1551 // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1552 // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1553 // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1554 // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1555 // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1556 // R.C., but R values from a.length and inside checkIndex are good to go.
1557 //
1558 void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1559 Node* inner_iters_actual_int, Node* inner_phi,
1560 Node* iv_add, LoopNode* inner_head) {
1561 Node* long_zero = longcon(0);
1562 Node* int_zero = intcon(0);
1563 Node* long_one = longcon(1);
1564 Node* int_stride = intcon(checked_cast<int>(stride_con));
1565
1566 for (uint i = 0; i < range_checks.size(); i++) {
1567 ProjNode* proj = range_checks.at(i)->as_Proj();
1568 RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1569 jlong scale = 0;
1570 Node* offset = nullptr;
1571 Node* rc_bol = rc->in(1);
1572 Node* rc_cmp = rc_bol->in(1);
1573 if (rc_cmp->Opcode() == Op_CmpU) {
1574 // could be shared and have already been taken care of
1575 continue;
1576 }
1577 bool short_scale = false;
1578 bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1579 assert(ok, "inconsistent: was tested before");
1580 Node* range = rc_cmp->in(2);
1581 Node* c = rc->in(0);
1582 Node* entry_control = inner_head->in(LoopNode::EntryControl);
1583
1584 Node* R = range;
1585 Node* K = longcon(scale);
1586
1587 Node* L = offset;
1588
1589 if (short_scale) {
1590 // This converts:
1591 // (int)i*K + L <u64 R
1592 // with K an int into:
1593 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1594 // to protect against an overflow of (int)i*K
1595 //
1596 // Because if (int)i*K overflows, there are K,L where:
1597 // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1598 // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1599 // i*(long)K + L <u64 R is true.
1600 //
1601 // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1602 //
1603 // It's always true that:
1604 // (int)i*K <u64 (long)max_jint + 1
1605 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1606 // As a consequence:
1607 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1608 // is always false in case of overflow of i*K
1609 //
1610 // Note, there are also K,L where i*K overflows and
1611 // i*K + L <u64 R is true, but
1612 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1613 // So this transformation could cause spurious deoptimizations and failed range check elimination
1614 // (but not incorrect execution) for unlikely corner cases with overflow.
1615 // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1616 Node* max_jint_plus_one_long = longcon((jlong)max_jint + 1);
1617 Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1618 register_new_node(max_range, entry_control);
1619 R = MinMaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1620 set_subtree_ctrl(R, true);
1621 }
1622
1623 Node* C = outer_phi;
1624
1625 // Start with 64-bit values:
1626 // i*K + L <u64 R
1627 // (C+j)*K + L <u64 R
1628 // j*K + Q <u64 R where Q = Q_first = C*K+L
1629 Node* Q_first = new MulLNode(C, K);
1630 register_new_node(Q_first, entry_control);
1631 Q_first = new AddLNode(Q_first, L);
1632 register_new_node(Q_first, entry_control);
1633
1634 // Compute endpoints of the range of values j*K + Q.
1635 // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1636 Node* Q_min = Q_first;
1637
1638 // Compute the exact ending value B_2 (which is really A_2 if S < 0)
1639 Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1640 register_new_node(B_2, entry_control);
1641 B_2 = new SubINode(B_2, int_stride);
1642 register_new_node(B_2, entry_control);
1643 B_2 = new ConvI2LNode(B_2);
1644 register_new_node(B_2, entry_control);
1645
1646 Node* Q_max = new MulLNode(B_2, K);
1647 register_new_node(Q_max, entry_control);
1648 Q_max = new AddLNode(Q_max, Q_first);
1649 register_new_node(Q_max, entry_control);
1650
1651 if (scale * stride_con < 0) {
1652 swap(Q_min, Q_max);
1653 }
1654 // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1655
1656 // L_clamp = Q_min < 0 ? 0 : Q_min
1657 Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1658 register_new_node(Q_min_cmp, entry_control);
1659 Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1660 register_new_node(Q_min_bool, entry_control);
1661 Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1662 register_new_node(L_clamp, entry_control);
1663 // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1664
1665 Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1666 register_new_node(Q_max_plus_one, entry_control);
1667
1668 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1669 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1670 Node* max_jlong_long = longcon(max_jlong);
1671 Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1672 register_new_node(Q_max_cmp, entry_control);
1673 Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1674 register_new_node(Q_max_bool, entry_control);
1675 Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1676 register_new_node(H_clamp, entry_control);
1677 // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1678
1679 // R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1680 // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1681 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1682 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1683 // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1684 Node* R_2 = clamp(R, L_clamp, H_clamp);
1685 R_2 = new SubLNode(R_2, L_clamp);
1686 register_new_node(R_2, entry_control);
1687 R_2 = new ConvL2INode(R_2, TypeInt::POS);
1688 register_new_node(R_2, entry_control);
1689
1690 // L_2 = Q_first - L_clamp
1691 // We are subtracting L_clamp from both sides of the <u32 comparison.
1692 // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1693 // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1694 Node* L_2 = new SubLNode(Q_first, L_clamp);
1695 register_new_node(L_2, entry_control);
1696 L_2 = new ConvL2INode(L_2, TypeInt::INT);
1697 register_new_node(L_2, entry_control);
1698
1699 // Transform the range check using the computed values L_2/R_2
1700 // from: i*K + L <u64 R
1701 // to: j*K + L_2 <u32 R_2
1702 // that is:
1703 // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1704 K = intcon(checked_cast<int>(scale));
1705 Node* scaled_iv = new MulINode(inner_phi, K);
1706 register_new_node(scaled_iv, c);
1707 Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1708 register_new_node(scaled_iv_plus_offset, c);
1709
1710 Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1711 register_new_node(new_rc_cmp, c);
1712
1713 _igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1714 }
1715 }
1716
1717 Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {
1718 Node* min = MinMaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1719 set_subtree_ctrl(min, true);
1720 Node* max = MinMaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1721 set_subtree_ctrl(max, true);
1722 return max;
1723 }
1724
1725 LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1726 IfNode* exit_test) {
1727 LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1728 IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1729 _igvn.register_new_node_with_optimizer(new_inner_head);
1730 _igvn.register_new_node_with_optimizer(new_inner_exit);
1731 loop->_body.push(new_inner_head);
1732 loop->_body.push(new_inner_exit);
1733 loop->_body.yank(head);
1734 loop->_body.yank(exit_test);
1735 set_loop(new_inner_head, loop);
1736 set_loop(new_inner_exit, loop);
1737 set_idom(new_inner_head, idom(head), dom_depth(head));
1738 set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1739 replace_node_and_forward_ctrl(head, new_inner_head);
1740 replace_node_and_forward_ctrl(exit_test, new_inner_exit);
1741 loop->_head = new_inner_head;
1742 return new_inner_head;
1743 }
1744
1745 #ifdef ASSERT
1746 void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {
1747 Node* back_control = loop_exit_control(x, loop);
1748 assert(back_control != nullptr, "no back control");
1749
1750 BoolTest::mask mask = BoolTest::illegal;
1751 float cl_prob = 0;
1752 Node* incr = nullptr;
1753 Node* limit = nullptr;
1754
1755 Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1756 assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1757
1758 Node* phi_incr = nullptr;
1759 incr = loop_iv_incr(incr, x, loop, phi_incr);
1760 assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1761
1762 Node* xphi = nullptr;
1763 Node* stride = loop_iv_stride(incr, xphi);
1764
1765 assert(stride != nullptr, "no stride");
1766
1767 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x);
1768
1769 assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1770
1771 jlong stride_con = stride->get_integer_as_long(bt);
1772
1773 assert(condition_stride_ok(mask, stride_con), "illegal condition");
1774
1775 assert(mask != BoolTest::ne, "unexpected condition");
1776 assert(phi_incr == nullptr, "bad loop shape");
1777 assert(cmp->in(1) == incr, "bad exit test shape");
1778
1779 // Safepoint on backedge not supported
1780 assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1781 }
1782 #endif
1783
1784 #ifdef ASSERT
1785 // convert an int counted loop to a long counted to stress handling of
1786 // long counted loops
1787 bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
1788 Unique_Node_List iv_nodes;
1789 Node_List old_new;
1790 iv_nodes.push(cmp);
1791 bool failed = false;
1792
1793 for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
1794 Node* n = iv_nodes.at(i);
1795 switch(n->Opcode()) {
1796 case Op_Phi: {
1797 Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1798 old_new.map(n->_idx, clone);
1799 break;
1800 }
1801 case Op_CmpI: {
1802 Node* clone = new CmpLNode(nullptr, nullptr);
1803 old_new.map(n->_idx, clone);
1804 break;
1805 }
1806 case Op_AddI: {
1807 Node* clone = new AddLNode(nullptr, nullptr);
1808 old_new.map(n->_idx, clone);
1809 break;
1810 }
1811 case Op_CastII: {
1812 failed = true;
1813 break;
1814 }
1815 default:
1816 DEBUG_ONLY(n->dump());
1817 fatal("unexpected");
1818 }
1819
1820 for (uint i = 1; i < n->req(); i++) {
1821 Node* in = n->in(i);
1822 if (in == nullptr) {
1823 continue;
1824 }
1825 if (ctrl_is_member(loop, in)) {
1826 iv_nodes.push(in);
1827 }
1828 }
1829 }
1830
1831 if (failed) {
1832 for (uint i = 0; i < iv_nodes.size(); i++) {
1833 Node* n = iv_nodes.at(i);
1834 Node* clone = old_new[n->_idx];
1835 if (clone != nullptr) {
1836 _igvn.remove_dead_node(clone);
1837 }
1838 }
1839 return false;
1840 }
1841
1842 for (uint i = 0; i < iv_nodes.size(); i++) {
1843 Node* n = iv_nodes.at(i);
1844 Node* clone = old_new[n->_idx];
1845 for (uint i = 1; i < n->req(); i++) {
1846 Node* in = n->in(i);
1847 if (in == nullptr) {
1848 continue;
1849 }
1850 Node* in_clone = old_new[in->_idx];
1851 if (in_clone == nullptr) {
1852 assert(_igvn.type(in)->isa_int(), "");
1853 in_clone = new ConvI2LNode(in);
1854 _igvn.register_new_node_with_optimizer(in_clone);
1855 set_subtree_ctrl(in_clone, false);
1856 }
1857 if (in_clone->in(0) == nullptr) {
1858 in_clone->set_req(0, C->top());
1859 clone->set_req(i, in_clone);
1860 in_clone->set_req(0, nullptr);
1861 } else {
1862 clone->set_req(i, in_clone);
1863 }
1864 }
1865 _igvn.register_new_node_with_optimizer(clone);
1866 }
1867 set_ctrl(old_new[phi->_idx], phi->in(0));
1868
1869 for (uint i = 0; i < iv_nodes.size(); i++) {
1870 Node* n = iv_nodes.at(i);
1871 Node* clone = old_new[n->_idx];
1872 set_subtree_ctrl(clone, false);
1873 Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1874 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1875 Node* u = n->fast_out(i);
1876 if (iv_nodes.member(u)) {
1877 continue;
1878 }
1879 if (m == nullptr) {
1880 m = new ConvL2INode(clone);
1881 _igvn.register_new_node_with_optimizer(m);
1882 set_subtree_ctrl(m, false);
1883 }
1884 _igvn.rehash_node_delayed(u);
1885 int nb = u->replace_edge(n, m, &_igvn);
1886 --i, imax -= nb;
1887 }
1888 }
1889 return true;
1890 }
1891 #endif
1892
1893 //------------------------------is_counted_loop--------------------------------
1894 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop, BasicType iv_bt) {
1895 PhaseGVN *gvn = &_igvn;
1896
1897 Node* back_control = loop_exit_control(x, loop);
1898 if (back_control == nullptr) {
1899 return false;
1900 }
1901
1902 BoolTest::mask bt = BoolTest::illegal;
1903 float cl_prob = 0;
1904 Node* incr = nullptr;
1905 Node* limit = nullptr;
1906 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1907 if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {
1908 return false; // Avoid pointer & float & 64-bit compares
1909 }
1910
1911 // Trip-counter increment must be commutative & associative.
1912 if (incr->Opcode() == Op_Cast(iv_bt)) {
1913 incr = incr->in(1);
1914 }
1915
1916 Node* phi_incr = nullptr;
1917 incr = loop_iv_incr(incr, x, loop, phi_incr);
1918 if (incr == nullptr) {
1919 return false;
1920 }
1921
1922 Node* trunc1 = nullptr;
1923 Node* trunc2 = nullptr;
1924 const TypeInteger* iv_trunc_t = nullptr;
1925 Node* orig_incr = incr;
1926 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {
1927 return false; // Funny increment opcode
1928 }
1929 assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1930
1931 Node* xphi = nullptr;
1932 Node* stride = loop_iv_stride(incr, xphi);
1933
1934 if (stride == nullptr) {
1935 return false;
1936 }
1937
1938 // Iteratively uncast the loop induction variable
1939 // until no more CastII/CastLL nodes are found.
1940 while (xphi->Opcode() == Op_Cast(iv_bt)) {
1941 xphi = xphi->in(1);
1942 }
1943
1944 // Stride must be constant
1945 jlong stride_con = stride->get_integer_as_long(iv_bt);
1946 assert(stride_con != 0, "missed some peephole opt");
1947
1948 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x);
1949
1950 if (phi == nullptr ||
1951 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1952 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
1953 return false;
1954 }
1955
1956 Node* iftrue = back_control;
1957 uint iftrue_op = iftrue->Opcode();
1958 Node* iff = iftrue->in(0);
1959 BoolNode* test = iff->in(1)->as_Bool();
1960
1961 const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1962 if (trunc1 != nullptr) {
1963 // When there is a truncation, we must be sure that after the truncation
1964 // the trip counter will end up higher than the limit, otherwise we are looking
1965 // at an endless loop. Can happen with range checks.
1966
1967 // Example:
1968 // int i = 0;
1969 // while (true)
1970 // sum + = array[i];
1971 // i++;
1972 // i = i && 0x7fff;
1973 // }
1974 //
1975 // If the array is shorter than 0x8000 this exits through a AIOOB
1976 // - Counted loop transformation is ok
1977 // If the array is longer then this is an endless loop
1978 // - No transformation can be done.
1979
1980 const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1981 if (limit_t->hi_as_long() > incr_t->hi_as_long()) {
1982 // if the limit can have a higher value than the increment (before the phi)
1983 return false;
1984 }
1985 }
1986
1987 Node *init_trip = phi->in(LoopNode::EntryControl);
1988
1989 // If iv trunc type is smaller than int, check for possible wrap.
1990 if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
1991 assert(trunc1 != nullptr, "must have found some truncation");
1992
1993 // Get a better type for the phi (filtered thru if's)
1994 const TypeInteger* phi_ft = filtered_type(phi);
1995
1996 // Can iv take on a value that will wrap?
1997 //
1998 // Ensure iv's limit is not within "stride" of the wrap value.
1999 //
2000 // Example for "short" type
2001 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
2002 // If the stride is +10, then the last value of the induction
2003 // variable before the increment (phi_ft->_hi) must be
2004 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
2005 // ensure no truncation occurs after the increment.
2006
2007 if (stride_con > 0) {
2008 if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
2009 iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {
2010 return false; // truncation may occur
2011 }
2012 } else if (stride_con < 0) {
2013 if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
2014 iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {
2015 return false; // truncation may occur
2016 }
2017 }
2018 // No possibility of wrap so truncation can be discarded
2019 // Promote iv type to Int
2020 } else {
2021 assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
2022 }
2023
2024 if (!condition_stride_ok(bt, stride_con)) {
2025 return false;
2026 }
2027
2028 const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
2029
2030 if (stride_con > 0) {
2031 if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {
2032 return false; // cyclic loop
2033 }
2034 } else {
2035 if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {
2036 return false; // cyclic loop
2037 }
2038 }
2039
2040 if (phi_incr != nullptr && bt != BoolTest::ne) {
2041 // check if there is a possibility of IV overflowing after the first increment
2042 if (stride_con > 0) {
2043 if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {
2044 return false;
2045 }
2046 } else {
2047 if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {
2048 return false;
2049 }
2050 }
2051 }
2052
2053 // =================================================
2054 // ---- SUCCESS! Found A Trip-Counted Loop! -----
2055 //
2056
2057 if (x->Opcode() == Op_Region) {
2058 // x has not yet been transformed to Loop or LongCountedLoop.
2059 // This should only happen if we are inside an infinite loop.
2060 // It happens like this:
2061 // build_loop_tree -> do not attach infinite loop and nested loops
2062 // beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
2063 // build_loop_tree -> find and attach infinite and nested loops
2064 // counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
2065 assert(x->as_Region()->is_in_infinite_subgraph(),
2066 "x can only be a Region and not Loop if inside infinite loop");
2067 // Come back later when Region is transformed to LoopNode
2068 return false;
2069 }
2070
2071 assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
2072 C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
2073
2074 // ===================================================
2075 // We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
2076 // This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
2077 // At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
2078 // But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
2079 // to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
2080 // we trap and resume execution before the loop without having executed any iteration of the loop, yet.
2081 //
2082 // These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
2083 // Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
2084 // we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
2085 //
2086 // In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
2087 // for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
2088 //
2089 //
2090 // The loop to be converted does not always need to have the often used shape:
2091 //
2092 // i = init
2093 // i = init loop:
2094 // do { ...
2095 // // ... equivalent i+=stride
2096 // i+=stride <==> if (i < limit)
2097 // } while (i < limit); goto loop
2098 // exit:
2099 // ...
2100 //
2101 // where the loop exit check uses the post-incremented iv phi and a '<'-operator.
2102 //
2103 // We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
2104 // in the loop exit check:
2105 //
2106 // i = init
2107 // loop:
2108 // ...
2109 // if (i <= limit)
2110 // i+=stride
2111 // goto loop
2112 // exit:
2113 // ...
2114 //
2115 // Let's define the following terms:
2116 // - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
2117 // - iv_post_i: The post-incremented iv phi after the i-th iteration.
2118 //
2119 // The iv_pre_i and iv_post_i have the following relation:
2120 // iv_pre_i + stride = iv_post_i
2121 //
2122 // When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
2123 // iv_post_i < adjusted_limit
2124 //
2125 // If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
2126 // (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
2127 // -> adjusted_limit = limit.
2128 // (LE2) iv_post_i <= limit:
2129 // iv_post_i < limit + 1
2130 // -> adjusted limit = limit + 1
2131 // (LE3) iv_pre_i < limit:
2132 // iv_pre_i + stride < limit + stride
2133 // iv_post_i < limit + stride
2134 // -> adjusted_limit = limit + stride
2135 // (LE4) iv_pre_i <= limit:
2136 // iv_pre_i < limit + 1
2137 // iv_pre_i + stride < limit + stride + 1
2138 // iv_post_i < limit + stride + 1
2139 // -> adjusted_limit = limit + stride + 1
2140 //
2141 // Note that:
2142 // (AL) limit <= adjusted_limit.
2143 //
2144 // The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
2145 // loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
2146 // (INV) For i = 1..n, min_int <= iv_post_i <= max_int
2147 //
2148 // To prove (INV), we require the following two conditions/assumptions:
2149 // (i): adjusted_limit - 1 + stride <= max_int
2150 // (ii): init < limit
2151 //
2152 // If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
2153 // induction by assuming (i) and (ii).
2154 //
2155 // Proof by Induction
2156 // ------------------
2157 // > Base case (i = 1): We show that (INV) holds after the first iteration:
2158 // min_int <= iv_post_1 = init + stride <= max_int
2159 // Proof:
2160 // First, we note that (ii) implies
2161 // (iii) init <= limit - 1
2162 // max_int >= adjusted_limit - 1 + stride [using (i)]
2163 // >= limit - 1 + stride [using (AL)]
2164 // >= init + stride [using (iii)]
2165 // >= min_int [using stride > 0, no underflow]
2166 // Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
2167 //
2168 // Note that to prove the base case we need (i) and (ii).
2169 //
2170 // > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
2171 // min_int <= iv_post_j <= max_int
2172 // > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
2173 // min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int
2174 // Proof:
2175 // If iv_post_j >= adjusted_limit:
2176 // We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
2177 // also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.
2178 // If iv_post_j < adjusted_limit:
2179 // First, we note that:
2180 // (iv) iv_post_j <= adjusted_limit - 1
2181 // max_int >= adjusted_limit - 1 + stride [using (i)]
2182 // >= iv_post_j + stride [using (iv)]
2183 // >= min_int [using stride > 0, no underflow]
2184 //
2185 // Note that to prove the step case we only need (i).
2186 //
2187 // Thus, by assuming (i) and (ii), we proved (INV).
2188 //
2189 //
2190 // It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
2191 //
2192 // (1) Loop Limit Check Predicate for (i):
2193 // Using (i): adjusted_limit - 1 + stride <= max_int
2194 //
2195 // This condition is now restated to use limit instead of adjusted_limit:
2196 //
2197 // To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
2198 // max_int - stride + 1 >= adjusted_limit
2199 // We can merge the two constants into
2200 // canonicalized_correction = stride - 1
2201 // which gives us
2202 // max_int - canonicalized_correction >= adjusted_limit
2203 //
2204 // To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
2205 // adjusted_limit = limit + limit_correction
2206 // Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
2207 // max_int - canonicalized_correction - limit_correction >= limit
2208 // Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
2209 // (v) final_correction = canonicalized_correction + limit_correction
2210 //
2211 // which gives us:
2212 //
2213 // Final predicate condition:
2214 // max_int - final_correction >= limit
2215 //
2216 // However, we need to be careful that (v) does not over- or underflow.
2217 // We know that:
2218 // canonicalized_correction = stride - 1
2219 // and
2220 // limit_correction <= stride + 1
2221 // and thus
2222 // canonicalized_correction + limit_correction <= 2 * stride
2223 // To prevent an over- or underflow of (v), we must ensure that
2224 // 2 * stride <= max_int
2225 // which can safely be checked without over- or underflow with
2226 // (vi) stride != min_int AND abs(stride) <= max_int / 2
2227 //
2228 // We could try to further optimize the cases where (vi) does not hold but given that such large strides are
2229 // very uncommon and the loop would only run for a very few iterations anyway, we simply bail out if (vi) fails.
2230 //
2231 // (2) Loop Limit Check Predicate for (ii):
2232 // Using (ii): init < limit
2233 //
2234 // This Loop Limit Check Predicate is not required if we can prove at compile time that either:
2235 // (2.1) type(init) < type(limit)
2236 // In this case, we know:
2237 // all possible values of init < all possible values of limit
2238 // and we can skip the predicate.
2239 //
2240 // (2.2) init < limit is already checked before (i.e. found as a dominating check)
2241 // In this case, we do not need to re-check the condition and can skip the predicate.
2242 // This is often found for while- and for-loops which have the following shape:
2243 //
2244 // if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.
2245 // i = init;
2246 // if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate
2247 // do {
2248 // i += stride;
2249 // } while (i < limit);
2250 // }
2251 //
2252 // (2.3) init + stride <= max_int
2253 // In this case, there is no overflow of the iv phi after the first loop iteration.
2254 // In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
2255 // init < limit
2256 // In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
2257 // compile time that init + stride <= max_int then we have trivially proven the base case and that
2258 // there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
2259 // again and can skip the predicate.
2260
2261 // Check (vi) and bail out if the stride is too big.
2262 if (stride_con == min_signed_integer(iv_bt) || (ABS(stride_con) > max_signed_integer(iv_bt) / 2)) {
2263 return false;
2264 }
2265
2266 // Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
2267 const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
2268
2269 // Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
2270 const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
2271 const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
2272
2273 const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
2274 const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
2275 const jlong final_correction = canonicalized_correction + limit_correction;
2276
2277 int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
2278 Node* init_control = x->in(LoopNode::EntryControl);
2279
2280 // If sov==0, limit's type always satisfies the condition, for
2281 // example, when it is an array length.
2282 if (sov != 0) {
2283 if (sov < 0) {
2284 return false; // Bailout: integer overflow is certain.
2285 }
2286 // (1) Loop Limit Check Predicate is required because we could not statically prove that
2287 // limit + final_correction = adjusted_limit - 1 + stride <= max_int
2288 assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
2289 const Predicates predicates(init_control);
2290 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2291 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2292 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2293 #ifdef ASSERT
2294 if (TraceLoopLimitCheck) {
2295 tty->print("Missing Loop Limit Check Parse Predicate:");
2296 loop->dump_head();
2297 x->dump(1);
2298 }
2299 #endif
2300 return false;
2301 }
2302
2303 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2304 if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {
2305 return false;
2306 }
2307
2308 Node* cmp_limit;
2309 Node* bol;
2310
2311 if (stride_con > 0) {
2312 cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2313 bol = new BoolNode(cmp_limit, BoolTest::le);
2314 } else {
2315 cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2316 bol = new BoolNode(cmp_limit, BoolTest::ge);
2317 }
2318
2319 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2320 }
2321
2322 // (2.3)
2323 const bool init_plus_stride_could_overflow =
2324 (stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2325 (stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2326 // (2.1)
2327 const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2328 (stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2329
2330 if (init_gte_limit && // (2.1)
2331 ((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2332 !has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)
2333 // (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2334 // We use the following condition:
2335 // - stride > 0: init < limit
2336 // - stride < 0: init > limit
2337 //
2338 // This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2339 // a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2340 // check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2341
2342 const Predicates predicates(init_control);
2343 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2344 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2345 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2346 #ifdef ASSERT
2347 if (TraceLoopLimitCheck) {
2348 tty->print("Missing Loop Limit Check Parse Predicate:");
2349 loop->dump_head();
2350 x->dump(1);
2351 }
2352 #endif
2353 return false;
2354 }
2355
2356 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2357 Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2358 if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2359 !is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {
2360 return false;
2361 }
2362
2363 Node* cmp_limit;
2364 Node* bol;
2365
2366 if (stride_con > 0) {
2367 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2368 bol = new BoolNode(cmp_limit, BoolTest::lt);
2369 } else {
2370 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2371 bol = new BoolNode(cmp_limit, BoolTest::gt);
2372 }
2373
2374 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2375 }
2376
2377 if (bt == BoolTest::ne) {
2378 // Now we need to canonicalize the loop condition if it is 'ne'.
2379 assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2380 if (stride_con > 0) {
2381 // 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2382 bt = BoolTest::lt;
2383 } else {
2384 assert(stride_con < 0, "must be");
2385 // 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2386 bt = BoolTest::gt;
2387 }
2388 }
2389
2390 Node* sfpt = nullptr;
2391 if (loop->_child == nullptr) {
2392 sfpt = find_safepoint(back_control, x, loop);
2393 } else {
2394 sfpt = iff->in(0);
2395 if (sfpt->Opcode() != Op_SafePoint) {
2396 sfpt = nullptr;
2397 }
2398 }
2399
2400 if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
2401 Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2402 if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2403 iv_bt == T_LONG) &&
2404 sfpt == nullptr) {
2405 // Leaving the safepoint on the backedge and creating a
2406 // CountedLoop will confuse optimizations. We can't move the
2407 // safepoint around because its jvm state wouldn't match a new
2408 // location. Give up on that loop.
2409 return false;
2410 }
2411 if (is_deleteable_safept(backedge_sfpt)) {
2412 replace_node_and_forward_ctrl(backedge_sfpt, iftrue);
2413 if (loop->_safepts != nullptr) {
2414 loop->_safepts->yank(backedge_sfpt);
2415 }
2416 loop->_tail = iftrue;
2417 }
2418 }
2419
2420
2421 #ifdef ASSERT
2422 if (iv_bt == T_INT &&
2423 !x->as_Loop()->is_loop_nest_inner_loop() &&
2424 StressLongCountedLoop > 0 &&
2425 trunc1 == nullptr &&
2426 convert_to_long_loop(cmp, phi, loop)) {
2427 return false;
2428 }
2429 #endif
2430
2431 Node* adjusted_limit = limit;
2432 if (phi_incr != nullptr) {
2433 // If compare points directly to the phi we need to adjust
2434 // the compare so that it points to the incr. Limit have
2435 // to be adjusted to keep trip count the same and we
2436 // should avoid int overflow.
2437 //
2438 // i = init; do {} while(i++ < limit);
2439 // is converted to
2440 // i = init; do {} while(++i < limit+1);
2441 //
2442 adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2443 }
2444
2445 if (includes_limit) {
2446 // The limit check guaranties that 'limit <= (max_jint - stride)' so
2447 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2448 //
2449 Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2450 adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2451 if (bt == BoolTest::le)
2452 bt = BoolTest::lt;
2453 else if (bt == BoolTest::ge)
2454 bt = BoolTest::gt;
2455 else
2456 ShouldNotReachHere();
2457 }
2458 set_subtree_ctrl(adjusted_limit, false);
2459
2460 // Build a canonical trip test.
2461 // Clone code, as old values may be in use.
2462 incr = incr->clone();
2463 incr->set_req(1,phi);
2464 incr->set_req(2,stride);
2465 incr = _igvn.register_new_node_with_optimizer(incr);
2466 set_early_ctrl(incr, false);
2467 _igvn.rehash_node_delayed(phi);
2468 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2469
2470 // If phi type is more restrictive than Int, raise to
2471 // Int to prevent (almost) infinite recursion in igvn
2472 // which can only handle integer types for constants or minint..maxint.
2473 if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
2474 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2475 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2476 nphi = _igvn.register_new_node_with_optimizer(nphi);
2477 set_ctrl(nphi, get_ctrl(phi));
2478 _igvn.replace_node(phi, nphi);
2479 phi = nphi->as_Phi();
2480 }
2481 cmp = cmp->clone();
2482 cmp->set_req(1,incr);
2483 cmp->set_req(2, adjusted_limit);
2484 cmp = _igvn.register_new_node_with_optimizer(cmp);
2485 set_ctrl(cmp, iff->in(0));
2486
2487 test = test->clone()->as_Bool();
2488 (*(BoolTest*)&test->_test)._test = bt;
2489 test->set_req(1,cmp);
2490 _igvn.register_new_node_with_optimizer(test);
2491 set_ctrl(test, iff->in(0));
2492
2493 // Replace the old IfNode with a new LoopEndNode
2494 Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2495 IfNode *le = lex->as_If();
2496 uint dd = dom_depth(iff);
2497 set_idom(le, le->in(0), dd); // Update dominance for loop exit
2498 set_loop(le, loop);
2499
2500 // Get the loop-exit control
2501 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2502
2503 // Need to swap loop-exit and loop-back control?
2504 if (iftrue_op == Op_IfFalse) {
2505 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2506 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2507
2508 loop->_tail = back_control = ift2;
2509 set_loop(ift2, loop);
2510 set_loop(iff2, get_loop(iffalse));
2511
2512 // Lazy update of 'get_ctrl' mechanism.
2513 replace_node_and_forward_ctrl(iffalse, iff2);
2514 replace_node_and_forward_ctrl(iftrue, ift2);
2515
2516 // Swap names
2517 iffalse = iff2;
2518 iftrue = ift2;
2519 } else {
2520 _igvn.rehash_node_delayed(iffalse);
2521 _igvn.rehash_node_delayed(iftrue);
2522 iffalse->set_req_X( 0, le, &_igvn );
2523 iftrue ->set_req_X( 0, le, &_igvn );
2524 }
2525
2526 set_idom(iftrue, le, dd+1);
2527 set_idom(iffalse, le, dd+1);
2528 assert(iff->outcnt() == 0, "should be dead now");
2529 replace_node_and_forward_ctrl(iff, le); // fix 'get_ctrl'
2530
2531 Node* entry_control = init_control;
2532 bool strip_mine_loop = iv_bt == T_INT &&
2533 loop->_child == nullptr &&
2534 sfpt != nullptr &&
2535 !loop->_has_call &&
2536 is_deleteable_safept(sfpt);
2537 IdealLoopTree* outer_ilt = nullptr;
2538 if (strip_mine_loop) {
2539 outer_ilt = create_outer_strip_mined_loop(init_control, loop, cl_prob, le->_fcnt,
2540 entry_control, iffalse);
2541 }
2542
2543 // Now setup a new CountedLoopNode to replace the existing LoopNode
2544 BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2545 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2546 // The following assert is approximately true, and defines the intention
2547 // of can_be_counted_loop. It fails, however, because phase->type
2548 // is not yet initialized for this loop and its parts.
2549 //assert(l->can_be_counted_loop(this), "sanity");
2550 _igvn.register_new_node_with_optimizer(l);
2551 set_loop(l, loop);
2552 loop->_head = l;
2553 // Fix all data nodes placed at the old loop head.
2554 // Uses the lazy-update mechanism of 'get_ctrl'.
2555 replace_node_and_forward_ctrl(x, l);
2556 set_idom(l, entry_control, dom_depth(entry_control) + 1);
2557
2558 if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {
2559 // Check for immediately preceding SafePoint and remove
2560 if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {
2561 if (strip_mine_loop) {
2562 Node* outer_le = outer_ilt->_tail->in(0);
2563 Node* sfpt_clone = sfpt->clone();
2564 sfpt_clone->set_req(0, iffalse);
2565 outer_le->set_req(0, sfpt_clone);
2566
2567 Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2568 if (polladdr != nullptr && polladdr->is_Load()) {
2569 // Polling load should be pinned outside inner loop.
2570 Node* new_polladdr = polladdr->clone();
2571 new_polladdr->set_req(0, iffalse);
2572 _igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2573 set_ctrl(new_polladdr, iffalse);
2574 sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2575 }
2576 // When this code runs, loop bodies have not yet been populated.
2577 const bool body_populated = false;
2578 register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2579 set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2580 }
2581 replace_node_and_forward_ctrl(sfpt, sfpt->in(TypeFunc::Control));
2582 if (loop->_safepts != nullptr) {
2583 loop->_safepts->yank(sfpt);
2584 }
2585 }
2586 }
2587
2588 #ifdef ASSERT
2589 assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2590 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2591 #endif
2592 #ifndef PRODUCT
2593 if (TraceLoopOpts) {
2594 tty->print("Counted ");
2595 loop->dump_head();
2596 }
2597 #endif
2598
2599 C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2600
2601 // Capture bounds of the loop in the induction variable Phi before
2602 // subsequent transformation (iteration splitting) obscures the
2603 // bounds
2604 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2605
2606 if (strip_mine_loop) {
2607 l->mark_strip_mined();
2608 l->verify_strip_mined(1);
2609 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2610 loop = outer_ilt;
2611 }
2612
2613 #ifndef PRODUCT
2614 if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {
2615 AtomicAccess::inc(&_long_loop_counted_loops);
2616 }
2617 #endif
2618 if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {
2619 l->mark_loop_nest_outer_loop();
2620 }
2621
2622 return true;
2623 }
2624
2625 // Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2626 // If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2627 bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2628 const BasicType iv_bt, Node* loop_entry) {
2629 // Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2630 // successfully find a dominated test with the If node below.
2631 Node* cmp_limit;
2632 Node* bol;
2633 if (stride_con > 0) {
2634 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2635 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2636 } else {
2637 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2638 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2639 }
2640
2641 // Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2642 IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2643 // Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2644 IfFalseNode* if_false = new IfFalseNode(iff);
2645 IfTrueNode* if_true = new IfTrueNode(iff);
2646 Node* dominated_iff = _igvn.transform(iff);
2647 // ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2648 const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2649
2650 // Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2651 _igvn.replace_input_of(iff, 0, C->top());
2652 _igvn.replace_input_of(iff, 1, C->top());
2653 return found_dominating_test;
2654 }
2655
2656 //----------------------exact_limit-------------------------------------------
2657 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
2658 assert(loop->_head->is_CountedLoop(), "");
2659 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2660 assert(cl->is_valid_counted_loop(T_INT), "");
2661
2662 if (cl->stride_con() == 1 ||
2663 cl->stride_con() == -1 ||
2664 cl->limit()->Opcode() == Op_LoopLimit) {
2665 // Old code has exact limit (it could be incorrect in case of int overflow).
2666 // Loop limit is exact with stride == 1. And loop may already have exact limit.
2667 return cl->limit();
2668 }
2669 Node *limit = nullptr;
2670 #ifdef ASSERT
2671 BoolTest::mask bt = cl->loopexit()->test_trip();
2672 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2673 #endif
2674 if (cl->has_exact_trip_count()) {
2675 // Simple case: loop has constant boundaries.
2676 // Use jlongs to avoid integer overflow.
2677 int stride_con = cl->stride_con();
2678 jlong init_con = cl->init_trip()->get_int();
2679 jlong limit_con = cl->limit()->get_int();
2680 julong trip_cnt = cl->trip_count();
2681 jlong final_con = init_con + trip_cnt*stride_con;
2682 int final_int = (int)final_con;
2683 // The final value should be in integer range since the loop
2684 // is counted and the limit was checked for overflow.
2685 assert(final_con == (jlong)final_int, "final value should be integer");
2686 limit = _igvn.intcon(final_int);
2687 } else {
2688 // Create new LoopLimit node to get exact limit (final iv value).
2689 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2690 register_new_node(limit, cl->in(LoopNode::EntryControl));
2691 }
2692 assert(limit != nullptr, "sanity");
2693 return limit;
2694 }
2695
2696 //------------------------------Ideal------------------------------------------
2697 // Return a node which is more "ideal" than the current node.
2698 // Attempt to convert into a counted-loop.
2699 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2700 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
2701 phase->C->set_major_progress();
2702 }
2703 return RegionNode::Ideal(phase, can_reshape);
2704 }
2705
2706 #ifdef ASSERT
2707 void LoopNode::verify_strip_mined(int expect_skeleton) const {
2708 const OuterStripMinedLoopNode* outer = nullptr;
2709 const CountedLoopNode* inner = nullptr;
2710 if (is_strip_mined()) {
2711 if (!is_valid_counted_loop(T_INT)) {
2712 return; // Skip malformed counted loop
2713 }
2714 assert(is_CountedLoop(), "no Loop should be marked strip mined");
2715 inner = as_CountedLoop();
2716 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2717 } else if (is_OuterStripMinedLoop()) {
2718 outer = this->as_OuterStripMinedLoop();
2719 inner = outer->unique_ctrl_out()->as_CountedLoop();
2720 assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2721 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2722 }
2723 if (inner != nullptr || outer != nullptr) {
2724 assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2725 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2726 Node* outer_le = outer_tail->in(0);
2727 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2728 Node* sfpt = outer_le->in(0);
2729 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2730 Node* inner_out = sfpt->in(0);
2731 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2732 assert(cle == inner->loopexit_or_null(), "mismatch");
2733 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2734 if (has_skeleton) {
2735 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2736 assert(outer->outcnt() == 2, "only control nodes");
2737 } else {
2738 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2739 uint phis = 0;
2740 uint be_loads = 0;
2741 Node* be = inner->in(LoopNode::LoopBackControl);
2742 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
2743 Node* u = inner->fast_out(i);
2744 if (u->is_Phi()) {
2745 phis++;
2746 for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
2747 Node* n = be->fast_out(j);
2748 if (n->is_Load()) {
2749 assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2750 do {
2751 n = n->raw_out(0);
2752 } while (!n->is_Phi());
2753 if (n == u) {
2754 be_loads++;
2755 break;
2756 }
2757 }
2758 }
2759 }
2760 }
2761 assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2762 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
2763 Node* u = outer->fast_out(i);
2764 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2765 }
2766 uint stores = 0;
2767 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
2768 Node* u = inner_out->fast_out(i);
2769 if (u->is_Store()) {
2770 stores++;
2771 }
2772 }
2773 // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2774 // not guaranteed to be optimized out.
2775 assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2776 }
2777 assert(sfpt->outcnt() == 1, "no data node");
2778 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2779 }
2780 }
2781 #endif
2782
2783 //=============================================================================
2784 //------------------------------Ideal------------------------------------------
2785 // Return a node which is more "ideal" than the current node.
2786 // Attempt to convert into a counted-loop.
2787 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2788 return RegionNode::Ideal(phase, can_reshape);
2789 }
2790
2791 //------------------------------dump_spec--------------------------------------
2792 // Dump special per-node info
2793 #ifndef PRODUCT
2794 void CountedLoopNode::dump_spec(outputStream *st) const {
2795 LoopNode::dump_spec(st);
2796 if (stride_is_con()) {
2797 st->print("stride: %d ",stride_con());
2798 }
2799 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
2800 if (is_main_loop()) st->print("main of N%d", _idx);
2801 if (is_post_loop()) st->print("post of N%d", _main_idx);
2802 if (is_strip_mined()) st->print(" strip mined");
2803 if (is_multiversion_fast_loop()) { st->print(" multiversion_fast"); }
2804 if (is_multiversion_slow_loop()) { st->print(" multiversion_slow"); }
2805 if (is_multiversion_delayed_slow_loop()) { st->print(" multiversion_delayed_slow"); }
2806 }
2807 #endif
2808
2809 //=============================================================================
2810 jlong BaseCountedLoopEndNode::stride_con() const {
2811 return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2812 }
2813
2814
2815 BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {
2816 if (bt == T_INT) {
2817 return new CountedLoopEndNode(control, test, prob, cnt);
2818 }
2819 assert(bt == T_LONG, "unsupported");
2820 return new LongCountedLoopEndNode(control, test, prob, cnt);
2821 }
2822
2823 //=============================================================================
2824 //------------------------------Value-----------------------------------------
2825 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
2826 const Type* init_t = phase->type(in(Init));
2827 const Type* limit_t = phase->type(in(Limit));
2828 const Type* stride_t = phase->type(in(Stride));
2829 // Either input is TOP ==> the result is TOP
2830 if (init_t == Type::TOP) return Type::TOP;
2831 if (limit_t == Type::TOP) return Type::TOP;
2832 if (stride_t == Type::TOP) return Type::TOP;
2833
2834 int stride_con = stride_t->is_int()->get_con();
2835 if (stride_con == 1)
2836 return bottom_type(); // Identity
2837
2838 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
2839 // Use jlongs to avoid integer overflow.
2840 jlong init_con = init_t->is_int()->get_con();
2841 jlong limit_con = limit_t->is_int()->get_con();
2842 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2843 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2844 jlong final_con = init_con + stride_con*trip_count;
2845 int final_int = (int)final_con;
2846 // The final value should be in integer range in almost all cases,
2847 // since the loop is counted and the limit was checked for overflow.
2848 // There some exceptions, for example:
2849 // - During CCP, there might be a temporary overflow from PhiNodes, see JDK-8309266.
2850 // - During PhaseIdealLoop::split_thru_phi, the LoopLimitNode floats possibly far above
2851 // the loop and its predicates, and we might get constants on one side of the phi that
2852 // would lead to overflows. Such a code path would never lead us to enter the loop
2853 // because of the loop limit overflow check that happens after the LoopLimitNode
2854 // computation with overflow, but before we enter the loop, see JDK-8335747.
2855 if (final_con == (jlong)final_int) {
2856 return TypeInt::make(final_int);
2857 } else {
2858 return bottom_type();
2859 }
2860 }
2861
2862 return bottom_type(); // TypeInt::INT
2863 }
2864
2865 //------------------------------Ideal------------------------------------------
2866 // Return a node which is more "ideal" than the current node.
2867 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2868 if (phase->type(in(Init)) == Type::TOP ||
2869 phase->type(in(Limit)) == Type::TOP ||
2870 phase->type(in(Stride)) == Type::TOP)
2871 return nullptr; // Dead
2872
2873 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2874 if (stride_con == 1)
2875 return nullptr; // Identity
2876
2877 // Delay following optimizations until all loop optimizations
2878 // done to keep Ideal graph simple.
2879 if (!can_reshape || !phase->C->post_loop_opts_phase()) {
2880 phase->C->record_for_post_loop_opts_igvn(this);
2881 return nullptr;
2882 }
2883
2884 const TypeInt* init_t = phase->type(in(Init) )->is_int();
2885 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2886 jlong stride_p;
2887 jlong lim, ini;
2888 julong max;
2889 if (stride_con > 0) {
2890 stride_p = stride_con;
2891 lim = limit_t->_hi;
2892 ini = init_t->_lo;
2893 max = (julong)max_jint;
2894 } else {
2895 stride_p = -(jlong)stride_con;
2896 lim = init_t->_hi;
2897 ini = limit_t->_lo;
2898 max = (julong)(juint)min_jint; // double cast to get 0x0000000080000000, not 0xffffffff80000000
2899 }
2900 julong range = lim - ini + stride_p;
2901 if (range <= max) {
2902 // Convert to integer expression if it is not overflow.
2903 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2904 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2905 Node *bias = phase->transform(new AddINode(range, stride_m));
2906 Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2907 Node *span = phase->transform(new MulINode(trip, in(Stride)));
2908 return new AddINode(span, in(Init)); // exact limit
2909 }
2910
2911 if (is_power_of_2(stride_p) || // divisor is 2^n
2912 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
2913 // Convert to long expression to avoid integer overflow
2914 // and let igvn optimizer convert this division.
2915 //
2916 Node* init = phase->transform( new ConvI2LNode(in(Init)));
2917 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2918 Node* stride = phase->longcon(stride_con);
2919 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2920
2921 Node *range = phase->transform(new SubLNode(limit, init));
2922 Node *bias = phase->transform(new AddLNode(range, stride_m));
2923 Node *span;
2924 if (stride_con > 0 && is_power_of_2(stride_p)) {
2925 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2926 // and avoid generating rounding for division. Zero trip guard should
2927 // guarantee that init < limit but sometimes the guard is missing and
2928 // we can get situation when init > limit. Note, for the empty loop
2929 // optimization zero trip guard is generated explicitly which leaves
2930 // only RCE predicate where exact limit is used and the predicate
2931 // will simply fail forcing recompilation.
2932 Node* neg_stride = phase->longcon(-stride_con);
2933 span = phase->transform(new AndLNode(bias, neg_stride));
2934 } else {
2935 Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2936 span = phase->transform(new MulLNode(trip, stride));
2937 }
2938 // Convert back to int
2939 Node *span_int = phase->transform(new ConvL2INode(span));
2940 return new AddINode(span_int, in(Init)); // exact limit
2941 }
2942
2943 return nullptr; // No progress
2944 }
2945
2946 //------------------------------Identity---------------------------------------
2947 // If stride == 1 return limit node.
2948 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
2949 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2950 if (stride_con == 1 || stride_con == -1)
2951 return in(Limit);
2952 return this;
2953 }
2954
2955 //=============================================================================
2956 //----------------------match_incr_with_optional_truncation--------------------
2957 // Match increment with optional truncation:
2958 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2959 // Return null for failure. Success returns the increment node.
2960 Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2961 const TypeInteger** trunc_type,
2962 BasicType bt) {
2963 // Quick cutouts:
2964 if (expr == nullptr || expr->req() != 3) return nullptr;
2965
2966 Node *t1 = nullptr;
2967 Node *t2 = nullptr;
2968 Node* n1 = expr;
2969 int n1op = n1->Opcode();
2970 const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2971
2972 if (bt == T_INT) {
2973 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
2974 if (n1op == Op_AndI &&
2975 n1->in(2)->is_Con() &&
2976 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
2977 // %%% This check should match any mask of 2**K-1.
2978 t1 = n1;
2979 n1 = t1->in(1);
2980 n1op = n1->Opcode();
2981 trunc_t = TypeInt::CHAR;
2982 } else if (n1op == Op_RShiftI &&
2983 n1->in(1) != nullptr &&
2984 n1->in(1)->Opcode() == Op_LShiftI &&
2985 n1->in(2) == n1->in(1)->in(2) &&
2986 n1->in(2)->is_Con()) {
2987 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2988 // %%% This check should match any shift in [1..31].
2989 if (shift == 16 || shift == 8) {
2990 t1 = n1;
2991 t2 = t1->in(1);
2992 n1 = t2->in(1);
2993 n1op = n1->Opcode();
2994 if (shift == 16) {
2995 trunc_t = TypeInt::SHORT;
2996 } else if (shift == 8) {
2997 trunc_t = TypeInt::BYTE;
2998 }
2999 }
3000 }
3001 }
3002
3003 // If (maybe after stripping) it is an AddI, we won:
3004 if (n1op == Op_Add(bt)) {
3005 *trunc1 = t1;
3006 *trunc2 = t2;
3007 *trunc_type = trunc_t;
3008 return n1;
3009 }
3010
3011 // failed
3012 return nullptr;
3013 }
3014
3015 IfNode* CountedLoopNode::find_multiversion_if_from_multiversion_fast_main_loop() {
3016 assert(is_main_loop() && is_multiversion_fast_loop(), "must be multiversion fast main loop");
3017 CountedLoopEndNode* pre_end = find_pre_loop_end();
3018 if (pre_end == nullptr) { return nullptr; }
3019 Node* pre_entry = pre_end->loopnode()->in(LoopNode::EntryControl);
3020 const Predicates predicates(pre_entry);
3021 IfTrueNode* before_predicates = predicates.entry()->isa_IfTrue();
3022 if (before_predicates != nullptr &&
3023 before_predicates->in(0)->in(1)->is_OpaqueMultiversioning()) {
3024 return before_predicates->in(0)->as_If();
3025 }
3026 return nullptr;
3027 }
3028
3029 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
3030 if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {
3031 verify_strip_mined(expect_skeleton);
3032 return in(EntryControl)->as_Loop();
3033 }
3034 return this;
3035 }
3036
3037 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
3038 assert(is_strip_mined(), "not a strip mined loop");
3039 Node* c = in(EntryControl);
3040 if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {
3041 return nullptr;
3042 }
3043 return c->as_OuterStripMinedLoop();
3044 }
3045
3046 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
3047 Node* c = in(LoopBackControl);
3048 if (c == nullptr || c->is_top()) {
3049 return nullptr;
3050 }
3051 return c->as_IfTrue();
3052 }
3053
3054 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
3055 LoopNode* l = outer_loop();
3056 if (l == nullptr) {
3057 return nullptr;
3058 }
3059 return l->outer_loop_tail();
3060 }
3061
3062 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
3063 IfTrueNode* proj = outer_loop_tail();
3064 if (proj == nullptr) {
3065 return nullptr;
3066 }
3067 Node* c = proj->in(0);
3068 if (c == nullptr || c->is_top() || c->outcnt() != 2) {
3069 return nullptr;
3070 }
3071 return c->as_OuterStripMinedLoopEnd();
3072 }
3073
3074 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
3075 LoopNode* l = outer_loop();
3076 if (l == nullptr) {
3077 return nullptr;
3078 }
3079 return l->outer_loop_end();
3080 }
3081
3082 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
3083 IfNode* le = outer_loop_end();
3084 if (le == nullptr) {
3085 return nullptr;
3086 }
3087 IfFalseNode* c = le->false_proj_or_null();
3088 if (c == nullptr) {
3089 return nullptr;
3090 }
3091 return c->as_IfFalse();
3092 }
3093
3094 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
3095 LoopNode* l = outer_loop();
3096 if (l == nullptr) {
3097 return nullptr;
3098 }
3099 return l->outer_loop_exit();
3100 }
3101
3102 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
3103 IfNode* le = outer_loop_end();
3104 if (le == nullptr) {
3105 return nullptr;
3106 }
3107 Node* c = le->in(0);
3108 if (c == nullptr || c->is_top()) {
3109 return nullptr;
3110 }
3111 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
3112 return c->as_SafePoint();
3113 }
3114
3115 SafePointNode* CountedLoopNode::outer_safepoint() const {
3116 LoopNode* l = outer_loop();
3117 if (l == nullptr) {
3118 return nullptr;
3119 }
3120 return l->outer_safepoint();
3121 }
3122
3123 Node* CountedLoopNode::skip_assertion_predicates_with_halt() {
3124 Node* ctrl = in(LoopNode::EntryControl);
3125 if (ctrl == nullptr) {
3126 // Dying loop.
3127 return nullptr;
3128 }
3129 if (is_main_loop()) {
3130 ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
3131 }
3132 if (is_main_loop() || is_post_loop()) {
3133 AssertionPredicates assertion_predicates(ctrl);
3134 return assertion_predicates.entry();
3135 }
3136 return ctrl;
3137 }
3138
3139
3140 int CountedLoopNode::stride_con() const {
3141 CountedLoopEndNode* cle = loopexit_or_null();
3142 return cle != nullptr ? cle->stride_con() : 0;
3143 }
3144
3145 BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {
3146 if (bt == T_INT) {
3147 return new CountedLoopNode(entry, backedge);
3148 }
3149 assert(bt == T_LONG, "unsupported");
3150 return new LongCountedLoopNode(entry, backedge);
3151 }
3152
3153 void OuterStripMinedLoopNode::fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn,
3154 PhaseIdealLoop* iloop) const {
3155 CountedLoopNode* inner_cl = inner_counted_loop();
3156 IfFalseNode* cle_out = inner_loop_exit();
3157
3158 if (cle_out->outcnt() > 1) {
3159 // Look for chains of stores that were sunk
3160 // out of the inner loop and are in the outer loop
3161 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
3162 Node* u = cle_out->fast_out(i);
3163 if (u->is_Store()) {
3164 int alias_idx = igvn->C->get_alias_index(u->adr_type());
3165 Node* first = u;
3166 for (;;) {
3167 Node* next = first->in(MemNode::Memory);
3168 if (!next->is_Store() || next->in(0) != cle_out) {
3169 break;
3170 }
3171 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
3172 first = next;
3173 }
3174 Node* last = u;
3175 for (;;) {
3176 Node* next = nullptr;
3177 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
3178 Node* uu = last->fast_out(j);
3179 if (uu->is_Store() && uu->in(0) == cle_out) {
3180 assert(next == nullptr, "only one in the outer loop");
3181 next = uu;
3182 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
3183 }
3184 }
3185 if (next == nullptr) {
3186 break;
3187 }
3188 last = next;
3189 }
3190 Node* phi = nullptr;
3191 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
3192 Node* uu = inner_cl->fast_out(j);
3193 if (uu->is_Phi()) {
3194 Node* be = uu->in(LoopNode::LoopBackControl);
3195 if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
3196 assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
3197 }
3198 if (be == last || be == first->in(MemNode::Memory)) {
3199 assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
3200 assert(phi == nullptr, "only one phi");
3201 phi = uu;
3202 }
3203 }
3204 }
3205 #ifdef ASSERT
3206 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
3207 Node* uu = inner_cl->fast_out(j);
3208 if (uu->is_memory_phi()) {
3209 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
3210 assert(phi == uu, "what's that phi?");
3211 } else if (uu->adr_type() == TypePtr::BOTTOM) {
3212 Node* n = uu->in(LoopNode::LoopBackControl);
3213 uint limit = igvn->C->live_nodes();
3214 uint i = 0;
3215 while (n != uu) {
3216 i++;
3217 assert(i < limit, "infinite loop");
3218 if (n->is_Proj()) {
3219 n = n->in(0);
3220 } else if (n->is_SafePoint() || n->is_MemBar()) {
3221 n = n->in(TypeFunc::Memory);
3222 } else if (n->is_Phi()) {
3223 n = n->in(1);
3224 } else if (n->is_MergeMem()) {
3225 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
3226 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
3227 n = n->in(MemNode::Memory);
3228 } else {
3229 n->dump();
3230 ShouldNotReachHere();
3231 }
3232 }
3233 }
3234 }
3235 }
3236 #endif
3237 if (phi == nullptr) {
3238 // If an entire chains was sunk, the
3239 // inner loop has no phi for that memory
3240 // slice, create one for the outer loop
3241 phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
3242 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
3243 phi->set_req(LoopNode::LoopBackControl, last);
3244 phi = register_new_node(phi, inner_cl, igvn, iloop);
3245 igvn->replace_input_of(first, MemNode::Memory, phi);
3246 } else {
3247 // Or fix the outer loop fix to include
3248 // that chain of stores.
3249 Node* be = phi->in(LoopNode::LoopBackControl);
3250 assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
3251 if (be == first->in(MemNode::Memory)) {
3252 if (be == phi->in(LoopNode::LoopBackControl)) {
3253 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
3254 } else {
3255 igvn->replace_input_of(be, MemNode::Memory, last);
3256 }
3257 } else {
3258 #ifdef ASSERT
3259 if (be == phi->in(LoopNode::LoopBackControl)) {
3260 assert(phi->in(LoopNode::LoopBackControl) == last, "");
3261 } else {
3262 assert(be->in(MemNode::Memory) == last, "");
3263 }
3264 #endif
3265 }
3266 }
3267 }
3268 }
3269 }
3270 }
3271
3272 // The outer strip mined loop is initially only partially constructed. In particular Phis are omitted.
3273 // See comment above: PhaseIdealLoop::create_outer_strip_mined_loop()
3274 // We're now in the process of finishing the construction of the outer loop. For each Phi in the inner loop, a Phi in
3275 // the outer loop was just now created. However, Sunk Stores cause an extra challenge:
3276 // 1) If all Stores in the inner loop were sunk for a particular memory slice, there's no Phi left for that memory slice
3277 // in the inner loop anymore, and hence we did not yet add a Phi for the outer loop. So an extra Phi must now be
3278 // added for each chain of sunk Stores for a particular memory slice.
3279 // 2) If some Stores were sunk and some left in the inner loop, a Phi was already created in the outer loop but
3280 // its backedge input wasn't wired correctly to the last Store of the chain: the backedge input was set to the
3281 // backedge of the inner loop Phi instead, but it needs to be the last Store of the chain in the outer loop. We now
3282 // have to fix that too.
3283 void OuterStripMinedLoopNode::handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn) {
3284 IfFalseNode* cle_exit_proj = inner_loop_exit();
3285
3286 // Find Sunk stores: Sunk stores are pinned on the loop exit projection of the inner loop. Indeed, because Sunk Stores
3287 // modify the memory state captured by the SafePoint in the outer strip mined loop, they must be above it. The
3288 // SafePoint's control input is the loop exit projection. It's also the only control out of the inner loop above the
3289 // SafePoint.
3290 #ifdef ASSERT
3291 int stores_in_outer_loop_cnt = 0;
3292 for (DUIterator_Fast imax, i = cle_exit_proj->fast_outs(imax); i < imax; i++) {
3293 Node* u = cle_exit_proj->fast_out(i);
3294 if (u->is_Store()) {
3295 stores_in_outer_loop_cnt++;
3296 }
3297 }
3298 #endif
3299
3300 // Sunk stores are reachable from the memory state of the outer loop safepoint
3301 SafePointNode* safepoint = outer_safepoint();
3302 MergeMemNode* mm = safepoint->in(TypeFunc::Memory)->isa_MergeMem();
3303 if (mm == nullptr) {
3304 // There is no MergeMem, which should only happen if there was no memory node
3305 // sunk out of the loop.
3306 assert(stores_in_outer_loop_cnt == 0, "inconsistent");
3307 return;
3308 }
3309 DEBUG_ONLY(int stores_in_outer_loop_cnt2 = 0);
3310 for (MergeMemStream mms(mm); mms.next_non_empty();) {
3311 Node* mem = mms.memory();
3312 // Traverse up the chain of stores to find the first store pinned
3313 // at the loop exit projection.
3314 Node* last = mem;
3315 Node* first = nullptr;
3316 while (mem->is_Store() && mem->in(0) == cle_exit_proj) {
3317 DEBUG_ONLY(stores_in_outer_loop_cnt2++);
3318 first = mem;
3319 mem = mem->in(MemNode::Memory);
3320 }
3321 if (first != nullptr) {
3322 // Found a chain of Stores that were sunk
3323 // Do we already have a memory Phi for that slice on the outer loop? If that is the case, that Phi was created
3324 // by cloning an inner loop Phi. The inner loop Phi should have mem, the memory state of the first Store out of
3325 // the inner loop, as input on the backedge. So does the outer loop Phi given it's a clone.
3326 Node* phi = nullptr;
3327 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3328 Node* u = mem->fast_out(i);
3329 if (u->is_Phi() && u->in(0) == this && u->in(LoopBackControl) == mem) {
3330 assert(phi == nullptr, "there should be only one");
3331 phi = u;
3332 PRODUCT_ONLY(break);
3333 }
3334 }
3335 if (phi == nullptr) {
3336 // No outer loop Phi? create one
3337 phi = PhiNode::make(this, last);
3338 phi->set_req(EntryControl, mem);
3339 phi = igvn->transform(phi);
3340 igvn->replace_input_of(first, MemNode::Memory, phi);
3341 } else {
3342 // Fix memory state along the backedge: it should be the last sunk Store of the chain
3343 igvn->replace_input_of(phi, LoopBackControl, last);
3344 }
3345 }
3346 }
3347 assert(stores_in_outer_loop_cnt == stores_in_outer_loop_cnt2, "inconsistent");
3348 }
3349
3350 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
3351 verify_strip_mined(1);
3352 // Look for the outer & inner strip mined loop, reduce number of
3353 // iterations of the inner loop, set exit condition of outer loop,
3354 // construct required phi nodes for outer loop.
3355 CountedLoopNode* inner_cl = inner_counted_loop();
3356 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
3357 if (LoopStripMiningIter == 0) {
3358 remove_outer_loop_and_safepoint(igvn);
3359 return;
3360 }
3361 if (LoopStripMiningIter == 1) {
3362 transform_to_counted_loop(igvn, nullptr);
3363 return;
3364 }
3365 Node* inner_iv_phi = inner_cl->phi();
3366 if (inner_iv_phi == nullptr) {
3367 IfNode* outer_le = outer_loop_end();
3368 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3369 igvn->replace_node(outer_le, iff);
3370 inner_cl->clear_strip_mined();
3371 return;
3372 }
3373 CountedLoopEndNode* inner_cle = inner_counted_loop_end();
3374
3375 int stride = inner_cl->stride_con();
3376 // For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
3377 // except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
3378 // stride, the method is guaranteed to return at the next check below.
3379 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
3380 int scaled_iters = (int)scaled_iters_long;
3381 if ((jlong)scaled_iters != scaled_iters_long) {
3382 // Remove outer loop and safepoint (too few iterations)
3383 remove_outer_loop_and_safepoint(igvn);
3384 return;
3385 }
3386 jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
3387 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
3388 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
3389 assert(iter_estimate > 0, "broken");
3390 if (iter_estimate <= short_scaled_iters) {
3391 // Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
3392 remove_outer_loop_and_safepoint(igvn);
3393 return;
3394 }
3395 if (iter_estimate <= scaled_iters_long) {
3396 // We would only go through one iteration of
3397 // the outer loop: drop the outer loop but
3398 // keep the safepoint so we don't run for
3399 // too long without a safepoint
3400 IfNode* outer_le = outer_loop_end();
3401 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3402 igvn->replace_node(outer_le, iff);
3403 inner_cl->clear_strip_mined();
3404 return;
3405 }
3406
3407 IfTrueNode* cle_tail = inner_cle->true_proj();
3408 ResourceMark rm;
3409 Node_List old_new;
3410 if (cle_tail->outcnt() > 1) {
3411 // Look for nodes on backedge of inner loop and clone them
3412 Unique_Node_List backedge_nodes;
3413 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
3414 Node* u = cle_tail->fast_out(i);
3415 if (u != inner_cl) {
3416 assert(!u->is_CFG(), "control flow on the backedge?");
3417 backedge_nodes.push(u);
3418 }
3419 }
3420 uint last = igvn->C->unique();
3421 for (uint next = 0; next < backedge_nodes.size(); next++) {
3422 Node* n = backedge_nodes.at(next);
3423 old_new.map(n->_idx, n->clone());
3424 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3425 Node* u = n->fast_out(i);
3426 assert(!u->is_CFG(), "broken");
3427 if (u->_idx >= last) {
3428 continue;
3429 }
3430 if (!u->is_Phi()) {
3431 backedge_nodes.push(u);
3432 } else {
3433 assert(u->in(0) == inner_cl, "strange phi on the backedge");
3434 }
3435 }
3436 }
3437 // Put the clones on the outer loop backedge
3438 Node* le_tail = outer_loop_tail();
3439 for (uint next = 0; next < backedge_nodes.size(); next++) {
3440 Node *n = old_new[backedge_nodes.at(next)->_idx];
3441 for (uint i = 1; i < n->req(); i++) {
3442 if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {
3443 n->set_req(i, old_new[n->in(i)->_idx]);
3444 }
3445 }
3446 if (n->in(0) != nullptr && n->in(0) == cle_tail) {
3447 n->set_req(0, le_tail);
3448 }
3449 igvn->register_new_node_with_optimizer(n);
3450 }
3451 }
3452
3453 Node* iv_phi = nullptr;
3454 // Make a clone of each phi in the inner loop for the outer loop
3455 // When Stores were Sunk, after this step, a Phi may still be missing or its backedge incorrectly wired. See
3456 // handle_sunk_stores_when_finishing_construction()
3457 for (uint i = 0; i < inner_cl->outcnt(); i++) {
3458 Node* u = inner_cl->raw_out(i);
3459 if (u->is_Phi()) {
3460 assert(u->in(0) == inner_cl, "inconsistent");
3461 Node* phi = u->clone();
3462 phi->set_req(0, this);
3463 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3464 if (be != nullptr) {
3465 phi->set_req(LoopNode::LoopBackControl, be);
3466 }
3467 phi = igvn->transform(phi);
3468 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3469 if (u == inner_iv_phi) {
3470 iv_phi = phi;
3471 }
3472 }
3473 }
3474
3475 handle_sunk_stores_when_finishing_construction(igvn);
3476
3477 if (iv_phi != nullptr) {
3478 // Now adjust the inner loop's exit condition
3479 Node* limit = inner_cl->limit();
3480 // If limit < init for stride > 0 (or limit > init for stride < 0),
3481 // the loop body is run only once. Given limit - init (init - limit resp.)
3482 // would be negative, the unsigned comparison below would cause
3483 // the loop body to be run for LoopStripMiningIter.
3484 Node* max = nullptr;
3485 if (stride > 0) {
3486 max = MinMaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3487 } else {
3488 max = MinMaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3489 }
3490 // sub is positive and can be larger than the max signed int
3491 // value. Use an unsigned min.
3492 Node* const_iters = igvn->intcon(scaled_iters);
3493 Node* min = MinMaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3494 // min is the number of iterations for the next inner loop execution:
3495 // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3496 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3497
3498 Node* new_limit = nullptr;
3499 if (stride > 0) {
3500 new_limit = igvn->transform(new AddINode(min, iv_phi));
3501 } else {
3502 new_limit = igvn->transform(new SubINode(iv_phi, min));
3503 }
3504 Node* inner_cmp = inner_cle->cmp_node();
3505 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3506 Node* outer_bol = inner_bol;
3507 // cmp node for inner loop may be shared
3508 inner_cmp = inner_cmp->clone();
3509 inner_cmp->set_req(2, new_limit);
3510 inner_bol = inner_bol->clone();
3511 inner_bol->set_req(1, igvn->transform(inner_cmp));
3512 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3513 // Set the outer loop's exit condition too
3514 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3515 } else {
3516 assert(false, "should be able to adjust outer loop");
3517 IfNode* outer_le = outer_loop_end();
3518 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3519 igvn->replace_node(outer_le, iff);
3520 inner_cl->clear_strip_mined();
3521 }
3522 }
3523
3524 void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3525 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3526 CountedLoopEndNode* cle = inner_cl->loopexit();
3527 Node* inner_test = cle->in(1);
3528 IfNode* outer_le = outer_loop_end();
3529 Node* safepoint = outer_safepoint();
3530
3531 fix_sunk_stores_when_back_to_counted_loop(igvn, iloop);
3532
3533 // make counted loop exit test always fail
3534 ConINode* zero = igvn->intcon(0);
3535 if (iloop != nullptr) {
3536 iloop->set_root_as_ctrl(zero);
3537 }
3538 igvn->replace_input_of(cle, 1, zero);
3539 // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3540 Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3541 register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3542 if (iloop == nullptr) {
3543 igvn->replace_node(outer_le, new_end);
3544 } else {
3545 iloop->replace_node_and_forward_ctrl(outer_le, new_end);
3546 }
3547 // the backedge of the inner loop must be rewired to the new loop end
3548 IfTrueNode* backedge = cle->true_proj();
3549 igvn->replace_input_of(backedge, 0, new_end);
3550 if (iloop != nullptr) {
3551 iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3552 }
3553 // make the outer loop go away
3554 igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3555 igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3556 inner_cl->clear_strip_mined();
3557 if (iloop != nullptr) {
3558 Unique_Node_List wq;
3559 wq.push(safepoint);
3560
3561 IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3562 IdealLoopTree* loop = iloop->get_loop(inner_cl);
3563
3564 for (uint i = 0; i < wq.size(); i++) {
3565 Node* n = wq.at(i);
3566 for (uint j = 0; j < n->req(); ++j) {
3567 Node* in = n->in(j);
3568 if (in == nullptr || in->is_CFG()) {
3569 continue;
3570 }
3571 if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {
3572 continue;
3573 }
3574 wq.push(in);
3575 }
3576 assert(!loop->_body.contains(n), "Shouldn't append node to body twice");
3577 loop->_body.push(n);
3578 }
3579 iloop->set_loop(safepoint, loop);
3580 loop->_body.push(safepoint);
3581 iloop->set_loop(safepoint->in(0), loop);
3582 loop->_body.push(safepoint->in(0));
3583 outer_loop_ilt->_tail = igvn->C->top();
3584 }
3585 }
3586
3587 void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {
3588 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3589 Node* outer_sfpt = outer_safepoint();
3590 Node* outer_out = outer_loop_exit();
3591 igvn->replace_node(outer_out, outer_sfpt->in(0));
3592 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3593 inner_cl->clear_strip_mined();
3594 }
3595
3596 Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3597 if (iloop == nullptr) {
3598 return igvn->transform(node);
3599 }
3600 iloop->register_new_node(node, ctrl);
3601 return node;
3602 }
3603
3604 Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3605 PhaseIdealLoop* iloop) {
3606 if (iloop == nullptr) {
3607 return igvn->transform(node);
3608 }
3609 iloop->register_control(node, iloop->get_loop(loop), idom);
3610 return node;
3611 }
3612
3613 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
3614 if (!in(0)) return Type::TOP;
3615 if (phase->type(in(0)) == Type::TOP)
3616 return Type::TOP;
3617
3618 // Until expansion, the loop end condition is not set so this should not constant fold.
3619 if (is_expanded(phase)) {
3620 return IfNode::Value(phase);
3621 }
3622
3623 return TypeTuple::IFBOTH;
3624 }
3625
3626 bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {
3627 // The outer strip mined loop head only has Phi uses after expansion
3628 if (phase->is_IterGVN()) {
3629 IfTrueNode* backedge = true_proj_or_null();
3630 if (backedge != nullptr) {
3631 Node* head = backedge->unique_ctrl_out_or_null();
3632 if (head != nullptr && head->is_OuterStripMinedLoop()) {
3633 if (head->find_out_with(Op_Phi) != nullptr) {
3634 return true;
3635 }
3636 }
3637 }
3638 }
3639 return false;
3640 }
3641
3642 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3643 if (remove_dead_region(phase, can_reshape)) return this;
3644
3645 return nullptr;
3646 }
3647
3648 //------------------------------filtered_type--------------------------------
3649 // Return a type based on condition control flow
3650 // A successful return will be a type that is restricted due
3651 // to a series of dominating if-tests, such as:
3652 // if (i < 10) {
3653 // if (i > 0) {
3654 // here: "i" type is [1..10)
3655 // }
3656 // }
3657 // or a control flow merge
3658 // if (i < 10) {
3659 // do {
3660 // phi( , ) -- at top of loop type is [min_int..10)
3661 // i = ?
3662 // } while ( i < 10)
3663 //
3664 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
3665 assert(n && n->bottom_type()->is_int(), "must be int");
3666 const TypeInt* filtered_t = nullptr;
3667 if (!n->is_Phi()) {
3668 assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3669 filtered_t = filtered_type_from_dominators(n, n_ctrl);
3670
3671 } else {
3672 Node* phi = n->as_Phi();
3673 Node* region = phi->in(0);
3674 assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3675 if (region && region != C->top()) {
3676 for (uint i = 1; i < phi->req(); i++) {
3677 Node* val = phi->in(i);
3678 Node* use_c = region->in(i);
3679 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3680 if (val_t != nullptr) {
3681 if (filtered_t == nullptr) {
3682 filtered_t = val_t;
3683 } else {
3684 filtered_t = filtered_t->meet(val_t)->is_int();
3685 }
3686 }
3687 }
3688 }
3689 }
3690 const TypeInt* n_t = _igvn.type(n)->is_int();
3691 if (filtered_t != nullptr) {
3692 n_t = n_t->join(filtered_t)->is_int();
3693 }
3694 return n_t;
3695 }
3696
3697
3698 //------------------------------filtered_type_from_dominators--------------------------------
3699 // Return a possibly more restrictive type for val based on condition control flow of dominators
3700 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
3701 if (val->is_Con()) {
3702 return val->bottom_type()->is_int();
3703 }
3704 uint if_limit = 10; // Max number of dominating if's visited
3705 const TypeInt* rtn_t = nullptr;
3706
3707 if (use_ctrl && use_ctrl != C->top()) {
3708 Node* val_ctrl = get_ctrl(val);
3709 uint val_dom_depth = dom_depth(val_ctrl);
3710 Node* pred = use_ctrl;
3711 uint if_cnt = 0;
3712 while (if_cnt < if_limit) {
3713 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
3714 if_cnt++;
3715 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3716 if (if_t != nullptr) {
3717 if (rtn_t == nullptr) {
3718 rtn_t = if_t;
3719 } else {
3720 rtn_t = rtn_t->join(if_t)->is_int();
3721 }
3722 }
3723 }
3724 pred = idom(pred);
3725 if (pred == nullptr || pred == C->top()) {
3726 break;
3727 }
3728 // Stop if going beyond definition block of val
3729 if (dom_depth(pred) < val_dom_depth) {
3730 break;
3731 }
3732 }
3733 }
3734 return rtn_t;
3735 }
3736
3737
3738 //------------------------------dump_spec--------------------------------------
3739 // Dump special per-node info
3740 #ifndef PRODUCT
3741 void CountedLoopEndNode::dump_spec(outputStream *st) const {
3742 if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {
3743 BoolTest bt( test_trip()); // Added this for g++.
3744
3745 st->print("[");
3746 bt.dump_on(st);
3747 st->print("]");
3748 }
3749 st->print(" ");
3750 IfNode::dump_spec(st);
3751 }
3752 #endif
3753
3754 IdealLoopTree::IdealLoopTree(PhaseIdealLoop* phase, Node* head, Node* tail): _parent(nullptr), _next(nullptr), _child(nullptr),
3755 _head(head), _tail(tail),
3756 _phase(phase),
3757 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
3758 _body(phase->arena()),
3759 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
3760 _has_range_checks(0), _has_range_checks_computed(0),
3761 _safepts(nullptr),
3762 _required_safept(nullptr),
3763 _allow_optimizations(true) {
3764 precond(_head != nullptr);
3765 precond(_tail != nullptr);
3766 }
3767
3768 //=============================================================================
3769 //------------------------------is_member--------------------------------------
3770 // Is 'l' a member of 'this'?
3771 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
3772 while( l->_nest > _nest ) l = l->_parent;
3773 return l == this;
3774 }
3775
3776 //------------------------------set_nest---------------------------------------
3777 // Set loop tree nesting depth. Accumulate _has_call bits.
3778 int IdealLoopTree::set_nest( uint depth ) {
3779 assert(depth <= SHRT_MAX, "sanity");
3780 _nest = depth;
3781 int bits = _has_call;
3782 if( _child ) bits |= _child->set_nest(depth+1);
3783 if( bits ) _has_call = 1;
3784 if( _next ) bits |= _next ->set_nest(depth );
3785 return bits;
3786 }
3787
3788 //------------------------------split_fall_in----------------------------------
3789 // Split out multiple fall-in edges from the loop header. Move them to a
3790 // private RegionNode before the loop. This becomes the loop landing pad.
3791 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
3792 PhaseIterGVN &igvn = phase->_igvn;
3793 uint i;
3794
3795 // Make a new RegionNode to be the landing pad.
3796 RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3797 phase->set_loop(landing_pad,_parent);
3798 // If _head was irreducible loop entry, landing_pad may now be too
3799 landing_pad->set_loop_status(_head->as_Region()->loop_status());
3800 // Gather all the fall-in control paths into the landing pad
3801 uint icnt = fall_in_cnt;
3802 uint oreq = _head->req();
3803 for( i = oreq-1; i>0; i-- )
3804 if( !phase->is_member( this, _head->in(i) ) )
3805 landing_pad->set_req(icnt--,_head->in(i));
3806
3807 // Peel off PhiNode edges as well
3808 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3809 Node *oj = _head->fast_out(j);
3810 if( oj->is_Phi() ) {
3811 PhiNode* old_phi = oj->as_Phi();
3812 assert( old_phi->region() == _head, "" );
3813 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3814 Node *p = PhiNode::make_blank(landing_pad, old_phi);
3815 uint icnt = fall_in_cnt;
3816 for( i = oreq-1; i>0; i-- ) {
3817 if( !phase->is_member( this, _head->in(i) ) ) {
3818 p->init_req(icnt--, old_phi->in(i));
3819 // Go ahead and clean out old edges from old phi
3820 old_phi->del_req(i);
3821 }
3822 }
3823 // Search for CSE's here, because ZKM.jar does a lot of
3824 // loop hackery and we need to be a little incremental
3825 // with the CSE to avoid O(N^2) node blow-up.
3826 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3827 if( p2 ) { // Found CSE
3828 p->destruct(&igvn); // Recover useless new node
3829 p = p2; // Use old node
3830 } else {
3831 igvn.register_new_node_with_optimizer(p, old_phi);
3832 }
3833 // Make old Phi refer to new Phi.
3834 old_phi->add_req(p);
3835 // Check for the special case of making the old phi useless and
3836 // disappear it. In JavaGrande I have a case where this useless
3837 // Phi is the loop limit and prevents recognizing a CountedLoop
3838 // which in turn prevents removing an empty loop.
3839 Node *id_old_phi = old_phi->Identity(&igvn);
3840 if( id_old_phi != old_phi ) { // Found a simple identity?
3841 // Note that I cannot call 'replace_node' here, because
3842 // that will yank the edge from old_phi to the Region and
3843 // I'm mid-iteration over the Region's uses.
3844 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
3845 Node* use = old_phi->last_out(i);
3846 igvn.rehash_node_delayed(use);
3847 uint uses_found = 0;
3848 for (uint j = 0; j < use->len(); j++) {
3849 if (use->in(j) == old_phi) {
3850 if (j < use->req()) use->set_req (j, id_old_phi);
3851 else use->set_prec(j, id_old_phi);
3852 uses_found++;
3853 }
3854 }
3855 i -= uses_found; // we deleted 1 or more copies of this edge
3856 }
3857 }
3858 igvn._worklist.push(old_phi);
3859 }
3860 }
3861 // Finally clean out the fall-in edges from the RegionNode
3862 for( i = oreq-1; i>0; i-- ) {
3863 if( !phase->is_member( this, _head->in(i) ) ) {
3864 _head->del_req(i);
3865 }
3866 }
3867 igvn.rehash_node_delayed(_head);
3868 // Transform landing pad
3869 igvn.register_new_node_with_optimizer(landing_pad, _head);
3870 // Insert landing pad into the header
3871 _head->add_req(landing_pad);
3872 }
3873
3874 //------------------------------split_outer_loop-------------------------------
3875 // Split out the outermost loop from this shared header.
3876 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
3877 PhaseIterGVN &igvn = phase->_igvn;
3878
3879 // Find index of outermost loop; it should also be my tail.
3880 uint outer_idx = 1;
3881 while( _head->in(outer_idx) != _tail ) outer_idx++;
3882
3883 // Make a LoopNode for the outermost loop.
3884 Node *ctl = _head->in(LoopNode::EntryControl);
3885 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3886 outer = igvn.register_new_node_with_optimizer(outer, _head);
3887 phase->set_created_loop_node();
3888
3889 // Outermost loop falls into '_head' loop
3890 _head->set_req(LoopNode::EntryControl, outer);
3891 _head->del_req(outer_idx);
3892 // Split all the Phis up between '_head' loop and 'outer' loop.
3893 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3894 Node *out = _head->fast_out(j);
3895 if( out->is_Phi() ) {
3896 PhiNode *old_phi = out->as_Phi();
3897 assert( old_phi->region() == _head, "" );
3898 Node *phi = PhiNode::make_blank(outer, old_phi);
3899 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3900 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3901 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3902 // Make old Phi point to new Phi on the fall-in path
3903 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3904 old_phi->del_req(outer_idx);
3905 }
3906 }
3907
3908 // Use the new loop head instead of the old shared one
3909 _head = outer;
3910 phase->set_loop(_head, this);
3911 }
3912
3913 //------------------------------fix_parent-------------------------------------
3914 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
3915 loop->_parent = parent;
3916 if( loop->_child ) fix_parent( loop->_child, loop );
3917 if( loop->_next ) fix_parent( loop->_next , parent );
3918 }
3919
3920 //------------------------------estimate_path_freq-----------------------------
3921 static float estimate_path_freq( Node *n ) {
3922 // Try to extract some path frequency info
3923 IfNode *iff;
3924 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
3925 uint nop = n->Opcode();
3926 if( nop == Op_SafePoint ) { // Skip any safepoint
3927 n = n->in(0);
3928 continue;
3929 }
3930 if( nop == Op_CatchProj ) { // Get count from a prior call
3931 // Assume call does not always throw exceptions: means the call-site
3932 // count is also the frequency of the fall-through path.
3933 assert( n->is_CatchProj(), "" );
3934 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3935 return 0.0f; // Assume call exception path is rare
3936 Node *call = n->in(0)->in(0)->in(0);
3937 assert( call->is_Call(), "expect a call here" );
3938 const JVMState *jvms = ((CallNode*)call)->jvms();
3939 ciMethodData* methodData = jvms->method()->method_data();
3940 if (!methodData->is_mature()) return 0.0f; // No call-site data
3941 ciProfileData* data = methodData->bci_to_data(jvms->bci());
3942 if ((data == nullptr) || !data->is_CounterData()) {
3943 // no call profile available, try call's control input
3944 n = n->in(0);
3945 continue;
3946 }
3947 return data->as_CounterData()->count()/FreqCountInvocations;
3948 }
3949 // See if there's a gating IF test
3950 Node *n_c = n->in(0);
3951 if( !n_c->is_If() ) break; // No estimate available
3952 iff = n_c->as_If();
3953 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3954 // Compute how much count comes on this path
3955 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3956 // Have no count info. Skip dull uncommon-trap like branches.
3957 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3958 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3959 break;
3960 // Skip through never-taken branch; look for a real loop exit.
3961 n = iff->in(0);
3962 }
3963 return 0.0f; // No estimate available
3964 }
3965
3966 //------------------------------merge_many_backedges---------------------------
3967 // Merge all the backedges from the shared header into a private Region.
3968 // Feed that region as the one backedge to this loop.
3969 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
3970 uint i;
3971
3972 // Scan for the top 2 hottest backedges
3973 float hotcnt = 0.0f;
3974 float warmcnt = 0.0f;
3975 uint hot_idx = 0;
3976 // Loop starts at 2 because slot 1 is the fall-in path
3977 for( i = 2; i < _head->req(); i++ ) {
3978 float cnt = estimate_path_freq(_head->in(i));
3979 if( cnt > hotcnt ) { // Grab hottest path
3980 warmcnt = hotcnt;
3981 hotcnt = cnt;
3982 hot_idx = i;
3983 } else if( cnt > warmcnt ) { // And 2nd hottest path
3984 warmcnt = cnt;
3985 }
3986 }
3987
3988 // See if the hottest backedge is worthy of being an inner loop
3989 // by being much hotter than the next hottest backedge.
3990 if( hotcnt <= 0.0001 ||
3991 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3992
3993 // Peel out the backedges into a private merge point; peel
3994 // them all except optionally hot_idx.
3995 PhaseIterGVN &igvn = phase->_igvn;
3996
3997 Node *hot_tail = nullptr;
3998 // Make a Region for the merge point
3999 Node *r = new RegionNode(1);
4000 for( i = 2; i < _head->req(); i++ ) {
4001 if( i != hot_idx )
4002 r->add_req( _head->in(i) );
4003 else hot_tail = _head->in(i);
4004 }
4005 igvn.register_new_node_with_optimizer(r, _head);
4006 // Plug region into end of loop _head, followed by hot_tail
4007 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
4008 igvn.replace_input_of(_head, 2, r);
4009 if( hot_idx ) _head->add_req(hot_tail);
4010
4011 // Split all the Phis up between '_head' loop and the Region 'r'
4012 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
4013 Node *out = _head->fast_out(j);
4014 if( out->is_Phi() ) {
4015 PhiNode* n = out->as_Phi();
4016 igvn.hash_delete(n); // Delete from hash before hacking edges
4017 Node *hot_phi = nullptr;
4018 Node *phi = new PhiNode(r, n->type(), n->adr_type());
4019 // Check all inputs for the ones to peel out
4020 uint j = 1;
4021 for( uint i = 2; i < n->req(); i++ ) {
4022 if( i != hot_idx )
4023 phi->set_req( j++, n->in(i) );
4024 else hot_phi = n->in(i);
4025 }
4026 // Register the phi but do not transform until whole place transforms
4027 igvn.register_new_node_with_optimizer(phi, n);
4028 // Add the merge phi to the old Phi
4029 while( n->req() > 3 ) n->del_req( n->req()-1 );
4030 igvn.replace_input_of(n, 2, phi);
4031 if( hot_idx ) n->add_req(hot_phi);
4032 }
4033 }
4034
4035
4036 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
4037 // of self loop tree. Turn self into a loop headed by _head and with
4038 // tail being the new merge point.
4039 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
4040 phase->set_loop(_tail,ilt); // Adjust tail
4041 _tail = r; // Self's tail is new merge point
4042 phase->set_loop(r,this);
4043 ilt->_child = _child; // New guy has my children
4044 _child = ilt; // Self has new guy as only child
4045 ilt->_parent = this; // new guy has self for parent
4046 ilt->_nest = _nest; // Same nesting depth (for now)
4047
4048 // Starting with 'ilt', look for child loop trees using the same shared
4049 // header. Flatten these out; they will no longer be loops in the end.
4050 IdealLoopTree **pilt = &_child;
4051 while( ilt ) {
4052 if( ilt->_head == _head ) {
4053 uint i;
4054 for( i = 2; i < _head->req(); i++ )
4055 if( _head->in(i) == ilt->_tail )
4056 break; // Still a loop
4057 if( i == _head->req() ) { // No longer a loop
4058 // Flatten ilt. Hang ilt's "_next" list from the end of
4059 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
4060 IdealLoopTree **cp = &ilt->_child;
4061 while( *cp ) cp = &(*cp)->_next; // Find end of child list
4062 *cp = ilt->_next; // Hang next list at end of child list
4063 *pilt = ilt->_child; // Move child up to replace ilt
4064 ilt->_head = nullptr; // Flag as a loop UNIONED into parent
4065 ilt = ilt->_child; // Repeat using new ilt
4066 continue; // do not advance over ilt->_child
4067 }
4068 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
4069 phase->set_loop(_head,ilt);
4070 }
4071 pilt = &ilt->_child; // Advance to next
4072 ilt = *pilt;
4073 }
4074
4075 if( _child ) fix_parent( _child, this );
4076 }
4077
4078 //------------------------------beautify_loops---------------------------------
4079 // Split shared headers and insert loop landing pads.
4080 // Insert a LoopNode to replace the RegionNode.
4081 // Return TRUE if loop tree is structurally changed.
4082 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
4083 bool result = false;
4084 // Cache parts in locals for easy
4085 PhaseIterGVN &igvn = phase->_igvn;
4086
4087 igvn.hash_delete(_head); // Yank from hash before hacking edges
4088
4089 // Check for multiple fall-in paths. Peel off a landing pad if need be.
4090 int fall_in_cnt = 0;
4091 for( uint i = 1; i < _head->req(); i++ )
4092 if( !phase->is_member( this, _head->in(i) ) )
4093 fall_in_cnt++;
4094 assert( fall_in_cnt, "at least 1 fall-in path" );
4095 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
4096 split_fall_in( phase, fall_in_cnt );
4097
4098 // Swap inputs to the _head and all Phis to move the fall-in edge to
4099 // the left.
4100 fall_in_cnt = 1;
4101 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
4102 fall_in_cnt++;
4103 if( fall_in_cnt > 1 ) {
4104 // Since I am just swapping inputs I do not need to update def-use info
4105 Node *tmp = _head->in(1);
4106 igvn.rehash_node_delayed(_head);
4107 _head->set_req( 1, _head->in(fall_in_cnt) );
4108 _head->set_req( fall_in_cnt, tmp );
4109 // Swap also all Phis
4110 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
4111 Node* phi = _head->fast_out(i);
4112 if( phi->is_Phi() ) {
4113 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
4114 tmp = phi->in(1);
4115 phi->set_req( 1, phi->in(fall_in_cnt) );
4116 phi->set_req( fall_in_cnt, tmp );
4117 }
4118 }
4119 }
4120 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
4121 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
4122
4123 // If I am a shared header (multiple backedges), peel off the many
4124 // backedges into a private merge point and use the merge point as
4125 // the one true backedge.
4126 if (_head->req() > 3) {
4127 // Merge the many backedges into a single backedge but leave
4128 // the hottest backedge as separate edge for the following peel.
4129 if (!_irreducible) {
4130 merge_many_backedges( phase );
4131 }
4132
4133 // When recursively beautify my children, split_fall_in can change
4134 // loop tree structure when I am an irreducible loop. Then the head
4135 // of my children has a req() not bigger than 3. Here we need to set
4136 // result to true to catch that case in order to tell the caller to
4137 // rebuild loop tree. See issue JDK-8244407 for details.
4138 result = true;
4139 }
4140
4141 // If I have one hot backedge, peel off myself loop.
4142 // I better be the outermost loop.
4143 if (_head->req() > 3 && !_irreducible) {
4144 split_outer_loop( phase );
4145 result = true;
4146
4147 } else if (!_head->is_Loop() && !_irreducible) {
4148 // Make a new LoopNode to replace the old loop head
4149 Node *l = new LoopNode( _head->in(1), _head->in(2) );
4150 l = igvn.register_new_node_with_optimizer(l, _head);
4151 phase->set_created_loop_node();
4152 // Go ahead and replace _head
4153 phase->_igvn.replace_node( _head, l );
4154 _head = l;
4155 phase->set_loop(_head, this);
4156 }
4157
4158 // Now recursively beautify nested loops
4159 if( _child ) result |= _child->beautify_loops( phase );
4160 if( _next ) result |= _next ->beautify_loops( phase );
4161 return result;
4162 }
4163
4164 //------------------------------allpaths_check_safepts----------------------------
4165 // Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
4166 // safepoint encountered. Helper for check_safepts.
4167 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
4168 assert(stack.size() == 0, "empty stack");
4169 stack.push(_head);
4170 visited.clear();
4171 visited.set(_head->_idx);
4172 while (stack.size() > 0) {
4173 Node* n = stack.pop();
4174 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
4175 // Terminate this path
4176 } else if (n->Opcode() == Op_SafePoint) {
4177 if (_phase->get_loop(n) != this) {
4178 if (_required_safept == nullptr) _required_safept = new Node_List();
4179 // save the first we run into on that path: closest to the tail if the head has a single backedge
4180 _required_safept->push(n);
4181 }
4182 // Terminate this path
4183 } else {
4184 uint start = n->is_Region() ? 1 : 0;
4185 uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
4186 for (uint i = start; i < end; i++) {
4187 Node* in = n->in(i);
4188 assert(in->is_CFG(), "must be");
4189 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
4190 stack.push(in);
4191 }
4192 }
4193 }
4194 }
4195 }
4196
4197 //------------------------------check_safepts----------------------------
4198 // Given dominators, try to find loops with calls that must always be
4199 // executed (call dominates loop tail). These loops do not need non-call
4200 // safepoints (ncsfpt).
4201 //
4202 // A complication is that a safepoint in a inner loop may be needed
4203 // by an outer loop. In the following, the inner loop sees it has a
4204 // call (block 3) on every path from the head (block 2) to the
4205 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
4206 // in block 2, _but_ this leaves the outer loop without a safepoint.
4207 //
4208 // entry 0
4209 // |
4210 // v
4211 // outer 1,2,4 +-> 1
4212 // | \
4213 // | v
4214 // inner 2,3 | 2 <---+ ncsfpt in 2
4215 // | / \ |
4216 // | v v |
4217 // | 4 3 | call in 3
4218 // |_/ \ \_|
4219 // |
4220 // v
4221 // exit 5
4222 //
4223 // This method maintains a list (_required_safept) of ncsfpts that must
4224 // be protected for each loop. It only marks ncsfpts for prevervation,
4225 // and does not actually delete any of them.
4226 //
4227 // If some other method needs to delete a ncsfpt later, it will make sure
4228 // the ncsfpt is not in the list of all outer loops of the current loop.
4229 // See `PhaseIdealLoop::is_deleteable_safept`.
4230 //
4231 // The insights into the problem:
4232 // A) Counted loops are okay (i.e. do not need to preserve ncsfpts),
4233 // they will be handled in `IdealLoopTree::counted_loop`
4234 // B) Innermost loops are okay because there's no inner loops that can
4235 // delete their ncsfpts. Only outer loops need to mark safepoints for
4236 // protection, because only loops further in can accidentally delete
4237 // their ncsfpts
4238 // C) If an outer loop has a call that's guaranteed to execute (on the
4239 // idom-path), then that loop is okay. Because the call will always
4240 // perform a safepoint poll, regardless of what safepoints are deleted
4241 // from its inner loops
4242 // D) Similarly, if an outer loop has a ncsfpt on the idom-path that isn't
4243 // inside any nested loop, then that loop is okay
4244 // E) Otherwise, if an outer loop's ncsfpt on the idom-path is nested in
4245 // an inner loop, we need to prevent the inner loop from deleting it
4246 //
4247 // There are two analyses:
4248 // 1) The first, and cheaper one, scans the loop body from
4249 // tail to head following the idom (immediate dominator)
4250 // chain, looking for the cases (C,D,E) above.
4251 // Since inner loops are scanned before outer loops, there is summary
4252 // information about inner loops. Inner loops can be skipped over
4253 // when the tail of an inner loop is encountered.
4254 //
4255 // 2) The second, invoked if the first fails to find a call or ncsfpt on
4256 // the idom path (which is rare), scans all predecessor control paths
4257 // from the tail to the head, terminating a path when a call or sfpt
4258 // is encountered, to find the ncsfpt's that are closest to the tail.
4259 //
4260 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
4261 // Bottom up traversal
4262 IdealLoopTree* ch = _child;
4263 if (_child) _child->check_safepts(visited, stack);
4264 if (_next) _next ->check_safepts(visited, stack);
4265
4266 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {
4267 bool has_call = false; // call on dom-path
4268 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
4269 Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
4270 if (!_irreducible) {
4271 // Scan the dom-path nodes from tail to head
4272 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
4273 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
4274 has_call = true;
4275 _has_sfpt = 1; // Then no need for a safept!
4276 break;
4277 } else if (n->Opcode() == Op_SafePoint) {
4278 if (_phase->get_loop(n) == this) {
4279 // We found a local ncsfpt.
4280 // Continue searching for a call that is guaranteed to be a safepoint.
4281 has_local_ncsfpt = true;
4282 } else if (nonlocal_ncsfpt == nullptr) {
4283 nonlocal_ncsfpt = n; // save the one closest to the tail
4284 }
4285 } else {
4286 IdealLoopTree* nlpt = _phase->get_loop(n);
4287 if (this != nlpt) {
4288 // If at an inner loop tail, see if the inner loop has already
4289 // recorded seeing a call on the dom-path (and stop.) If not,
4290 // jump to the head of the inner loop.
4291 assert(is_member(nlpt), "nested loop");
4292 Node* tail = nlpt->_tail;
4293 if (tail->in(0)->is_If()) tail = tail->in(0);
4294 if (n == tail) {
4295 // If inner loop has call on dom-path, so does outer loop
4296 if (nlpt->_has_sfpt) {
4297 has_call = true;
4298 _has_sfpt = 1;
4299 break;
4300 }
4301 // Skip to head of inner loop
4302 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
4303 n = nlpt->_head;
4304 if (_head == n) {
4305 // this and nlpt (inner loop) have the same loop head. This should not happen because
4306 // during beautify_loops we call merge_many_backedges. However, infinite loops may not
4307 // have been attached to the loop-tree during build_loop_tree before beautify_loops,
4308 // but then attached in the build_loop_tree afterwards, and so still have unmerged
4309 // backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
4310 // since we have reached the loop head of this.
4311 assert(_head->as_Region()->is_in_infinite_subgraph(),
4312 "only expect unmerged backedges in infinite loops");
4313 break;
4314 }
4315 }
4316 }
4317 }
4318 }
4319 }
4320 // Record safept's that this loop needs preserved when an
4321 // inner loop attempts to delete it's safepoints.
4322 if (_child != nullptr && !has_call && !has_local_ncsfpt) {
4323 if (nonlocal_ncsfpt != nullptr) {
4324 if (_required_safept == nullptr) _required_safept = new Node_List();
4325 _required_safept->push(nonlocal_ncsfpt);
4326 } else {
4327 // Failed to find a suitable safept on the dom-path. Now use
4328 // an all paths walk from tail to head, looking for safepoints to preserve.
4329 allpaths_check_safepts(visited, stack);
4330 }
4331 }
4332 }
4333 }
4334
4335 //---------------------------is_deleteable_safept----------------------------
4336 // Is safept not required by an outer loop?
4337 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
4338 assert(sfpt->Opcode() == Op_SafePoint, "");
4339 IdealLoopTree* lp = get_loop(sfpt)->_parent;
4340 while (lp != nullptr) {
4341 Node_List* sfpts = lp->_required_safept;
4342 if (sfpts != nullptr) {
4343 for (uint i = 0; i < sfpts->size(); i++) {
4344 if (sfpt == sfpts->at(i))
4345 return false;
4346 }
4347 }
4348 lp = lp->_parent;
4349 }
4350 return true;
4351 }
4352
4353 //---------------------------replace_parallel_iv-------------------------------
4354 // Replace parallel induction variable (parallel to trip counter)
4355 // This optimization looks for patterns similar to:
4356 //
4357 // int a = init2;
4358 // for (int iv = init; iv < limit; iv += stride_con) {
4359 // a += stride_con2;
4360 // }
4361 //
4362 // and transforms it to:
4363 //
4364 // int iv2 = init2
4365 // int iv = init
4366 // loop:
4367 // if (iv >= limit) goto exit
4368 // iv += stride_con
4369 // iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
4370 // goto loop
4371 // exit:
4372 // ...
4373 //
4374 // Such transformation introduces more optimization opportunities. In this
4375 // particular example, the loop can be eliminated entirely given that
4376 // `stride_con2 / stride_con` is exact (i.e., no remainder). Checks are in
4377 // place to only perform this optimization if such a division is exact. This
4378 // example will be transformed into its semantic equivalence:
4379 //
4380 // int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
4381 //
4382 // which corresponds to the structure of transformed subgraph.
4383 //
4384 // However, if there is a mismatch between types of the loop and the parallel
4385 // induction variable (e.g., a long-typed IV in an int-typed loop), type
4386 // conversions are required:
4387 //
4388 // long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
4389 //
4390 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
4391 assert(loop->_head->is_CountedLoop(), "");
4392 CountedLoopNode *cl = loop->_head->as_CountedLoop();
4393 if (!cl->is_valid_counted_loop(T_INT)) {
4394 return; // skip malformed counted loop
4395 }
4396 Node *incr = cl->incr();
4397 if (incr == nullptr) {
4398 return; // Dead loop?
4399 }
4400 Node *init = cl->init_trip();
4401 Node *phi = cl->phi();
4402 jlong stride_con = cl->stride_con();
4403
4404 // Visit all children, looking for Phis
4405 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
4406 Node *out = cl->out(i);
4407 // Look for other phis (secondary IVs). Skip dead ones
4408 if (!out->is_Phi() || out == phi || !has_node(out)) {
4409 continue;
4410 }
4411
4412 PhiNode* phi2 = out->as_Phi();
4413 Node* incr2 = phi2->in(LoopNode::LoopBackControl);
4414 // Look for induction variables of the form: X += constant
4415 if (phi2->region() != loop->_head ||
4416 incr2->req() != 3 ||
4417 incr2->in(1)->uncast() != phi2 ||
4418 incr2 == incr ||
4419 (incr2->Opcode() != Op_AddI && incr2->Opcode() != Op_AddL) ||
4420 !incr2->in(2)->is_Con()) {
4421 continue;
4422 }
4423
4424 if (incr2->in(1)->is_ConstraintCast() &&
4425 !(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {
4426 // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
4427 continue;
4428 }
4429 // Check for parallel induction variable (parallel to trip counter)
4430 // via an affine function. In particular, count-down loops with
4431 // count-up array indices are common. We only RCE references off
4432 // the trip-counter, so we need to convert all these to trip-counter
4433 // expressions.
4434 Node* init2 = phi2->in(LoopNode::EntryControl);
4435
4436 // Determine the basic type of the stride constant (and the iv being incremented).
4437 BasicType stride_con2_bt = incr2->Opcode() == Op_AddI ? T_INT : T_LONG;
4438 jlong stride_con2 = incr2->in(2)->get_integer_as_long(stride_con2_bt);
4439
4440 // The ratio of the two strides cannot be represented as an int
4441 // if stride_con2 is min_jint (or min_jlong, respectively) and
4442 // stride_con is -1.
4443 if (stride_con2 == min_signed_integer(stride_con2_bt) && stride_con == -1) {
4444 continue;
4445 }
4446
4447 // The general case here gets a little tricky. We want to find the
4448 // GCD of all possible parallel IV's and make a new IV using this
4449 // GCD for the loop. Then all possible IVs are simple multiples of
4450 // the GCD. In practice, this will cover very few extra loops.
4451 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
4452 // where +/-1 is the common case, but other integer multiples are
4453 // also easy to handle.
4454 jlong ratio_con = stride_con2 / stride_con;
4455
4456 if ((ratio_con * stride_con) != stride_con2) { // Check for exact (no remainder)
4457 continue;
4458 }
4459
4460 #ifndef PRODUCT
4461 if (TraceLoopOpts) {
4462 tty->print("Parallel IV: %d ", phi2->_idx);
4463 loop->dump_head();
4464 }
4465 #endif
4466
4467 // Convert to using the trip counter. The parallel induction
4468 // variable differs from the trip counter by a loop-invariant
4469 // amount, the difference between their respective initial values.
4470 // It is scaled by the 'ratio_con'.
4471 Node* ratio = integercon(ratio_con, stride_con2_bt);
4472
4473 Node* init_converted = insert_convert_node_if_needed(stride_con2_bt, init);
4474 Node* phi_converted = insert_convert_node_if_needed(stride_con2_bt, phi);
4475
4476 Node* ratio_init = MulNode::make(init_converted, ratio, stride_con2_bt);
4477 _igvn.register_new_node_with_optimizer(ratio_init, init_converted);
4478 set_early_ctrl(ratio_init, false);
4479
4480 Node* diff = SubNode::make(init2, ratio_init, stride_con2_bt);
4481 _igvn.register_new_node_with_optimizer(diff, init2);
4482 set_early_ctrl(diff, false);
4483
4484 Node* ratio_idx = MulNode::make(phi_converted, ratio, stride_con2_bt);
4485 _igvn.register_new_node_with_optimizer(ratio_idx, phi_converted);
4486 set_ctrl(ratio_idx, cl);
4487
4488 Node* add = AddNode::make(ratio_idx, diff, stride_con2_bt);
4489 _igvn.register_new_node_with_optimizer(add);
4490 set_ctrl(add, cl);
4491
4492 _igvn.replace_node( phi2, add );
4493 // Sometimes an induction variable is unused
4494 if (add->outcnt() == 0) {
4495 _igvn.remove_dead_node(add);
4496 }
4497 --i; // deleted this phi; rescan starting with next position
4498 }
4499 }
4500
4501 Node* PhaseIdealLoop::insert_convert_node_if_needed(BasicType target, Node* input) {
4502 BasicType source = _igvn.type(input)->basic_type();
4503 if (source == target) {
4504 return input;
4505 }
4506
4507 Node* converted = ConvertNode::create_convert(source, target, input);
4508 _igvn.register_new_node_with_optimizer(converted, input);
4509 set_early_ctrl(converted, false);
4510
4511 return converted;
4512 }
4513
4514 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
4515 Node* keep = nullptr;
4516 if (keep_one) {
4517 // Look for a safepoint on the idom-path.
4518 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
4519 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
4520 keep = i;
4521 break; // Found one
4522 }
4523 }
4524 }
4525
4526 // Don't remove any safepoints if it is requested to keep a single safepoint and
4527 // no safepoint was found on idom-path. It is not safe to remove any safepoint
4528 // in this case since there's no safepoint dominating all paths in the loop body.
4529 bool prune = !keep_one || keep != nullptr;
4530
4531 // Delete other safepoints in this loop.
4532 Node_List* sfpts = _safepts;
4533 if (prune && sfpts != nullptr) {
4534 assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4535 for (uint i = 0; i < sfpts->size(); i++) {
4536 Node* n = sfpts->at(i);
4537 assert(phase->get_loop(n) == this, "");
4538 if (n != keep && phase->is_deleteable_safept(n)) {
4539 phase->replace_node_and_forward_ctrl(n, n->in(TypeFunc::Control));
4540 }
4541 }
4542 }
4543 }
4544
4545 //------------------------------counted_loop-----------------------------------
4546 // Convert to counted loops where possible
4547 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
4548
4549 // For grins, set the inner-loop flag here
4550 if (!_child) {
4551 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4552 }
4553
4554 IdealLoopTree* loop = this;
4555 if (_head->is_CountedLoop() ||
4556 phase->is_counted_loop(_head, loop, T_INT)) {
4557
4558 if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {
4559 // Indicate we do not need a safepoint here
4560 _has_sfpt = 1;
4561 }
4562
4563 // Remove safepoints
4564 bool keep_one_sfpt = !(_has_call || _has_sfpt);
4565 remove_safepoints(phase, keep_one_sfpt);
4566
4567 // Look for induction variables
4568 phase->replace_parallel_iv(this);
4569 } else if (_head->is_LongCountedLoop() ||
4570 phase->is_counted_loop(_head, loop, T_LONG)) {
4571 remove_safepoints(phase, true);
4572 } else {
4573 assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4574 if (_parent != nullptr && !_irreducible) {
4575 // Not a counted loop. Keep one safepoint.
4576 bool keep_one_sfpt = true;
4577 remove_safepoints(phase, keep_one_sfpt);
4578 }
4579 }
4580
4581 // Recursively
4582 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4583 assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4584 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4585 if (loop->_next) loop->_next ->counted_loop(phase);
4586 }
4587
4588
4589 // The Estimated Loop Clone Size:
4590 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4591 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4592 // respectively, used to ensure that the node usage estimates made are on the
4593 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
4594 // possible additional/excessive nodes generated due to data and control flow
4595 // merging, for edges reaching outside the loop.
4596 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
4597
4598 precond(0 < factor && factor < 16);
4599
4600 uint const bc = 13;
4601 uint const cc = 17;
4602 uint const sz = _body.size() + (_body.size() + 7) / 2;
4603 uint estimate = factor * (sz + bc) + cc;
4604
4605 assert((estimate - cc) / factor == sz + bc, "overflow");
4606
4607 return estimate + est_loop_flow_merge_sz();
4608 }
4609
4610 // The Estimated Loop (full-) Unroll Size:
4611 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4612 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4613 // node usage estimates made are on the safe side, for the most part. This is
4614 // a "light" version of the loop clone size calculation (above), based on the
4615 // assumption that most of the loop-construct overhead will be unraveled when
4616 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4617 // including an overflow check and returning UINT_MAX in case of an overflow.
4618 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
4619
4620 precond(factor > 0);
4621
4622 // Take into account that after unroll conjoined heads and tails will fold.
4623 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4624 uint const cc = 7;
4625 uint const sz = b0 + (b0 + 15) / 16;
4626 uint estimate = factor * sz + cc;
4627
4628 if ((estimate - cc) / factor != sz) {
4629 return UINT_MAX;
4630 }
4631
4632 return estimate + est_loop_flow_merge_sz();
4633 }
4634
4635 // Estimate the growth effect (in nodes) of merging control and data flow when
4636 // cloning a loop body, based on the amount of control and data flow reaching
4637 // outside of the (current) loop body.
4638 uint IdealLoopTree::est_loop_flow_merge_sz() const {
4639
4640 uint ctrl_edge_out_cnt = 0;
4641 uint data_edge_out_cnt = 0;
4642
4643 for (uint i = 0; i < _body.size(); i++) {
4644 Node* node = _body.at(i);
4645 uint outcnt = node->outcnt();
4646
4647 for (uint k = 0; k < outcnt; k++) {
4648 Node* out = node->raw_out(k);
4649 if (out == nullptr) continue;
4650 if (out->is_CFG()) {
4651 if (!is_member(_phase->get_loop(out))) {
4652 ctrl_edge_out_cnt++;
4653 }
4654 } else if (_phase->has_ctrl(out)) {
4655 Node* ctrl = _phase->get_ctrl(out);
4656 assert(ctrl != nullptr, "must be");
4657 assert(ctrl->is_CFG(), "must be");
4658 if (!is_member(_phase->get_loop(ctrl))) {
4659 data_edge_out_cnt++;
4660 }
4661 }
4662 }
4663 }
4664 // Use data and control count (x2.0) in estimate iff both are > 0. This is
4665 // a rather pessimistic estimate for the most part, in particular for some
4666 // complex loops, but still not enough to capture all loops.
4667 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
4668 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4669 }
4670 return 0;
4671 }
4672
4673 #ifndef PRODUCT
4674 //------------------------------dump_head--------------------------------------
4675 // Dump 1 liner for loop header info
4676 void IdealLoopTree::dump_head() {
4677 tty->sp(2 * _nest);
4678 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
4679 if (_irreducible) tty->print(" IRREDUCIBLE");
4680 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4681 : _head->in(LoopNode::EntryControl);
4682 const Predicates predicates(entry);
4683 if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {
4684 tty->print(" limit_check");
4685 }
4686 if (predicates.short_running_long_loop_predicate_block()->is_non_empty()) {
4687 tty->print(" short_running");
4688 }
4689 if (UseLoopPredicate) {
4690 if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {
4691 tty->print(" profile_predicated");
4692 }
4693 if (predicates.loop_predicate_block()->is_non_empty()) {
4694 tty->print(" predicated");
4695 }
4696 }
4697 if (UseAutoVectorizationPredicate && predicates.auto_vectorization_check_block()->is_non_empty()) {
4698 tty->print(" auto_vectorization_check_predicate");
4699 }
4700 if (_head->is_CountedLoop()) {
4701 CountedLoopNode *cl = _head->as_CountedLoop();
4702 tty->print(" counted");
4703
4704 Node* init_n = cl->init_trip();
4705 if (init_n != nullptr && init_n->is_Con())
4706 tty->print(" [%d,", cl->init_trip()->get_int());
4707 else
4708 tty->print(" [int,");
4709 Node* limit_n = cl->limit();
4710 if (limit_n != nullptr && limit_n->is_Con())
4711 tty->print("%d),", cl->limit()->get_int());
4712 else
4713 tty->print("int),");
4714 int stride_con = cl->stride_con();
4715 if (stride_con > 0) tty->print("+");
4716 tty->print("%d", stride_con);
4717
4718 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
4719
4720 if (cl->is_pre_loop ()) tty->print(" pre" );
4721 if (cl->is_main_loop()) tty->print(" main");
4722 if (cl->is_post_loop()) tty->print(" post");
4723 if (cl->is_vectorized_loop()) tty->print(" vector");
4724 if (range_checks_present()) tty->print(" rc ");
4725 if (cl->is_multiversion_fast_loop()) { tty->print(" multiversion_fast"); }
4726 if (cl->is_multiversion_slow_loop()) { tty->print(" multiversion_slow"); }
4727 if (cl->is_multiversion_delayed_slow_loop()) { tty->print(" multiversion_delayed_slow"); }
4728 }
4729 if (_has_call) tty->print(" has_call");
4730 if (_has_sfpt) tty->print(" has_sfpt");
4731 if (_rce_candidate) tty->print(" rce");
4732 if (_safepts != nullptr && _safepts->size() > 0) {
4733 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
4734 }
4735 if (_required_safept != nullptr && _required_safept->size() > 0) {
4736 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
4737 }
4738 if (Verbose) {
4739 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
4740 }
4741 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
4742 tty->print(" strip_mined");
4743 }
4744 tty->cr();
4745 }
4746
4747 //------------------------------dump-------------------------------------------
4748 // Dump loops by loop tree
4749 void IdealLoopTree::dump() {
4750 dump_head();
4751 if (_child) _child->dump();
4752 if (_next) _next ->dump();
4753 }
4754
4755 #endif
4756
4757 static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
4758 if (loop == root) {
4759 if (loop->_child != nullptr) {
4760 log->begin_head("loop_tree");
4761 log->end_head();
4762 log_loop_tree_helper(root, loop->_child, log);
4763 log->tail("loop_tree");
4764 assert(loop->_next == nullptr, "what?");
4765 }
4766 } else if (loop != nullptr) {
4767 Node* head = loop->_head;
4768 log->begin_head("loop");
4769 log->print(" idx='%d' ", head->_idx);
4770 if (loop->_irreducible) log->print("irreducible='1' ");
4771 if (head->is_Loop()) {
4772 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
4773 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
4774 } else if (head->is_CountedLoop()) {
4775 CountedLoopNode* cl = head->as_CountedLoop();
4776 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
4777 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
4778 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
4779 }
4780 log->end_head();
4781 log_loop_tree_helper(root, loop->_child, log);
4782 log->tail("loop");
4783 log_loop_tree_helper(root, loop->_next, log);
4784 }
4785 }
4786
4787 void PhaseIdealLoop::log_loop_tree() {
4788 if (C->log() != nullptr) {
4789 log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4790 }
4791 }
4792
4793 // Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4794 // predicates will be removed during the next round of IGVN.
4795 void PhaseIdealLoop::eliminate_useless_predicates() const {
4796 if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {
4797 return; // No predicates left.
4798 }
4799
4800 EliminateUselessPredicates eliminate_useless_predicates(_igvn, _ltree_root);
4801 eliminate_useless_predicates.eliminate();
4802 }
4803
4804 // If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4805 void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {
4806 if (_zero_trip_guard_opaque_nodes.size() == 0) {
4807 return;
4808 }
4809 Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4810 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4811 IdealLoopTree* lpt = iter.current();
4812 if (lpt->_child == nullptr && lpt->is_counted()) {
4813 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4814 Node* opaque = head->is_canonical_loop_entry();
4815 if (opaque != nullptr) {
4816 useful_zero_trip_guard_opaques_nodes.push(opaque);
4817 #ifdef ASSERT
4818 // See PhaseIdealLoop::do_unroll
4819 // This property is required in do_unroll, but it may not hold after cloning a loop.
4820 // In such a case, we bail out from unrolling, and rely on IGVN to clean up the graph.
4821 // We are here before loop cloning (before iteration_split), so if this property
4822 // does not hold, it must come from the previous round of loop optimizations, meaning
4823 // that IGVN failed to clean it: we will catch that here.
4824 // On the other hand, if this assert passes, a bailout in do_unroll means that
4825 // this property was broken in the current round of loop optimization (between here
4826 // and do_unroll), so we give a chance to IGVN to make the property true again.
4827 if (head->is_main_loop()) {
4828 assert(opaque->outcnt() == 1, "opaque node should not be shared");
4829 assert(opaque->in(1) == head->limit(), "After IGVN cleanup, input of opaque node must be the limit.");
4830 }
4831 if (head->is_post_loop()) {
4832 assert(opaque->outcnt() == 1, "opaque node should not be shared");
4833 }
4834 #endif
4835 }
4836 }
4837 }
4838 for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {
4839 OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4840 DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4841 if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {
4842 IfNode* iff = opaque->if_node();
4843 IdealLoopTree* loop = get_loop(iff);
4844 while (loop != _ltree_root && loop != nullptr) {
4845 loop = loop->_parent;
4846 }
4847 if (loop == nullptr) {
4848 // unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4849 // We can't tell if the opaque node is useful or not
4850 assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4851 } else {
4852 assert(guarded_loop == nullptr, "");
4853 this->_igvn.replace_node(opaque, opaque->in(1));
4854 }
4855 } else {
4856 assert(guarded_loop != nullptr, "");
4857 }
4858 }
4859 }
4860
4861 void PhaseIdealLoop::eliminate_useless_multiversion_if() {
4862 if (_multiversion_opaque_nodes.size() == 0) {
4863 return;
4864 }
4865
4866 ResourceMark rm;
4867 Unique_Node_List useful_multiversioning_opaque_nodes;
4868
4869 // The OpaqueMultiversioning is only used from the fast main loop in AutoVectorization, to add
4870 // speculative runtime-checks to the multiversion_if. Thus, a OpaqueMultiversioning is only
4871 // useful if it can be found from a fast main loop. If it can not be found from a fast main loop,
4872 // then we cannot ever use that multiversion_if to add more speculative runtime-checks, and hence
4873 // it is useless. If it is still in delayed mode, i.e. has not yet had any runtime-checks added,
4874 // then we can let it constant fold towards the fast loop.
4875 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4876 IdealLoopTree* lpt = iter.current();
4877 if (lpt->_child == nullptr && lpt->is_counted()) {
4878 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4879 if (head->is_main_loop() && head->is_multiversion_fast_loop()) {
4880 // There are fast_loop pre/main/post loops, but the finding traversal starts at the main
4881 // loop, and traverses via the fast pre loop to the multiversion_if.
4882 IfNode* multiversion_if = head->find_multiversion_if_from_multiversion_fast_main_loop();
4883 if (multiversion_if != nullptr) {
4884 useful_multiversioning_opaque_nodes.push(multiversion_if->in(1)->as_OpaqueMultiversioning());
4885 } else {
4886 // We could not find the multiversion_if, and would never find it again. Remove the
4887 // multiversion marking for consistency.
4888 head->set_no_multiversion();
4889 }
4890 }
4891 }
4892 }
4893
4894 for (uint i = 0; i < _multiversion_opaque_nodes.size(); i++) {
4895 OpaqueMultiversioningNode* opaque = _multiversion_opaque_nodes.at(i)->as_OpaqueMultiversioning();
4896 if (!useful_multiversioning_opaque_nodes.member(opaque)) {
4897 if (opaque->is_delayed_slow_loop()) {
4898 // We cannot hack the node directly, otherwise the slow_loop will complain that it cannot
4899 // find the multiversioning opaque node. Instead, we mark the opaque node as useless, and
4900 // it can be constant folded during IGVN.
4901 opaque->mark_useless(_igvn);
4902 }
4903 }
4904 }
4905 }
4906
4907 //------------------------process_expensive_nodes-----------------------------
4908 // Expensive nodes have their control input set to prevent the GVN
4909 // from commoning them and as a result forcing the resulting node to
4910 // be in a more frequent path. Use CFG information here, to change the
4911 // control inputs so that some expensive nodes can be commoned while
4912 // not executed more frequently.
4913 bool PhaseIdealLoop::process_expensive_nodes() {
4914 assert(OptimizeExpensiveOps, "optimization off?");
4915
4916 // Sort nodes to bring similar nodes together
4917 C->sort_expensive_nodes();
4918
4919 bool progress = false;
4920
4921 for (int i = 0; i < C->expensive_count(); ) {
4922 Node* n = C->expensive_node(i);
4923 int start = i;
4924 // Find nodes similar to n
4925 i++;
4926 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4927 int end = i;
4928 // And compare them two by two
4929 for (int j = start; j < end; j++) {
4930 Node* n1 = C->expensive_node(j);
4931 if (is_node_unreachable(n1)) {
4932 continue;
4933 }
4934 for (int k = j+1; k < end; k++) {
4935 Node* n2 = C->expensive_node(k);
4936 if (is_node_unreachable(n2)) {
4937 continue;
4938 }
4939
4940 assert(n1 != n2, "should be pair of nodes");
4941
4942 Node* c1 = n1->in(0);
4943 Node* c2 = n2->in(0);
4944
4945 Node* parent_c1 = c1;
4946 Node* parent_c2 = c2;
4947
4948 // The call to get_early_ctrl_for_expensive() moves the
4949 // expensive nodes up but stops at loops that are in a if
4950 // branch. See whether we can exit the loop and move above the
4951 // If.
4952 if (c1->is_Loop()) {
4953 parent_c1 = c1->in(1);
4954 }
4955 if (c2->is_Loop()) {
4956 parent_c2 = c2->in(1);
4957 }
4958
4959 if (parent_c1 == parent_c2) {
4960 _igvn._worklist.push(n1);
4961 _igvn._worklist.push(n2);
4962 continue;
4963 }
4964
4965 // Look for identical expensive node up the dominator chain.
4966 if (is_dominator(c1, c2)) {
4967 c2 = c1;
4968 } else if (is_dominator(c2, c1)) {
4969 c1 = c2;
4970 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4971 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
4972 // Both branches have the same expensive node so move it up
4973 // before the if.
4974 c1 = c2 = idom(parent_c1->in(0));
4975 }
4976 // Do the actual moves
4977 if (n1->in(0) != c1) {
4978 _igvn.replace_input_of(n1, 0, c1);
4979 progress = true;
4980 }
4981 if (n2->in(0) != c2) {
4982 _igvn.replace_input_of(n2, 0, c2);
4983 progress = true;
4984 }
4985 }
4986 }
4987 }
4988
4989 return progress;
4990 }
4991
4992 //=============================================================================
4993 //----------------------------build_and_optimize-------------------------------
4994 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4995 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4996 void PhaseIdealLoop::build_and_optimize() {
4997 assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4998
4999 bool do_split_ifs = (_mode == LoopOptsDefault);
5000 bool skip_loop_opts = (_mode == LoopOptsNone);
5001 bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
5002
5003
5004 bool old_progress = C->major_progress();
5005 uint orig_worklist_size = _igvn._worklist.size();
5006
5007 // Reset major-progress flag for the driver's heuristics
5008 C->clear_major_progress();
5009
5010 #ifndef PRODUCT
5011 // Capture for later assert
5012 uint unique = C->unique();
5013 _loop_invokes++;
5014 _loop_work += unique;
5015 #endif
5016
5017 // True if the method has at least 1 irreducible loop
5018 _has_irreducible_loops = false;
5019
5020 _created_loop_node = false;
5021
5022 VectorSet visited;
5023 // Pre-grow the mapping from Nodes to IdealLoopTrees.
5024 _loop_or_ctrl.map(C->unique(), nullptr);
5025 memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
5026
5027 // Pre-build the top-level outermost loop tree entry
5028 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
5029 // Do not need a safepoint at the top level
5030 _ltree_root->_has_sfpt = 1;
5031
5032 // Initialize Dominators.
5033 // Checked in clone_loop_predicate() during beautify_loops().
5034 _idom_size = 0;
5035 _idom = nullptr;
5036 _dom_depth = nullptr;
5037 _dom_stk = nullptr;
5038
5039 // Empty pre-order array
5040 allocate_preorders();
5041
5042 // Build a loop tree on the fly. Build a mapping from CFG nodes to
5043 // IdealLoopTree entries. Data nodes are NOT walked.
5044 build_loop_tree();
5045 // Check for bailout, and return
5046 if (C->failing()) {
5047 return;
5048 }
5049
5050 // Verify that the has_loops() flag set at parse time is consistent with the just built loop tree. When the back edge
5051 // is an exception edge, parsing doesn't set has_loops().
5052 assert(_ltree_root->_child == nullptr || C->has_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
5053 // No loops after all
5054 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
5055
5056 // There should always be an outer loop containing the Root and Return nodes.
5057 // If not, we have a degenerate empty program. Bail out in this case.
5058 if (!has_node(C->root())) {
5059 if (!_verify_only) {
5060 C->clear_major_progress();
5061 assert(false, "empty program detected during loop optimization");
5062 C->record_method_not_compilable("empty program detected during loop optimization");
5063 }
5064 return;
5065 }
5066
5067 // Nothing to do, so get out
5068 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
5069 !_verify_only;
5070 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
5071 if (stop_early && !do_expensive_nodes) {
5072 return;
5073 }
5074
5075 // Set loop nesting depth
5076 _ltree_root->set_nest( 0 );
5077
5078 // Split shared headers and insert loop landing pads.
5079 // Do not bother doing this on the Root loop of course.
5080 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
5081 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
5082 if( _ltree_root->_child->beautify_loops( this ) ) {
5083 // Re-build loop tree!
5084 _ltree_root->_child = nullptr;
5085 _loop_or_ctrl.clear();
5086 reallocate_preorders();
5087 build_loop_tree();
5088 // Check for bailout, and return
5089 if (C->failing()) {
5090 return;
5091 }
5092 // Reset loop nesting depth
5093 _ltree_root->set_nest( 0 );
5094
5095 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
5096 }
5097 }
5098
5099 // Build Dominators for elision of null checks & loop finding.
5100 // Since nodes do not have a slot for immediate dominator, make
5101 // a persistent side array for that info indexed on node->_idx.
5102 _idom_size = C->unique();
5103 _idom = NEW_ARENA_ARRAY(&_arena, Node*, _idom_size);
5104 _dom_depth = NEW_ARENA_ARRAY(&_arena, uint, _idom_size);
5105 _dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
5106 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
5107
5108 Dominators();
5109
5110 if (!_verify_only) {
5111 // As a side effect, Dominators removed any unreachable CFG paths
5112 // into RegionNodes. It doesn't do this test against Root, so
5113 // we do it here.
5114 for( uint i = 1; i < C->root()->req(); i++ ) {
5115 if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?
5116 _igvn.delete_input_of(C->root(), i);
5117 i--; // Rerun same iteration on compressed edges
5118 }
5119 }
5120
5121 // Given dominators, try to find inner loops with calls that must
5122 // always be executed (call dominates loop tail). These loops do
5123 // not need a separate safepoint.
5124 Node_List cisstack;
5125 _ltree_root->check_safepts(visited, cisstack);
5126 }
5127
5128 // Walk the DATA nodes and place into loops. Find earliest control
5129 // node. For CFG nodes, the _loop_or_ctrl array starts out and remains
5130 // holding the associated IdealLoopTree pointer. For DATA nodes, the
5131 // _loop_or_ctrl array holds the earliest legal controlling CFG node.
5132
5133 // Allocate stack with enough space to avoid frequent realloc
5134 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
5135 Node_Stack nstack(stack_size);
5136
5137 visited.clear();
5138 Node_List worklist;
5139 // Don't need C->root() on worklist since
5140 // it will be processed among C->top() inputs
5141 worklist.push(C->top());
5142 visited.set(C->top()->_idx); // Set C->top() as visited now
5143 build_loop_early( visited, worklist, nstack );
5144
5145 // Given early legal placement, try finding counted loops. This placement
5146 // is good enough to discover most loop invariants.
5147 if (!_verify_me && !_verify_only) {
5148 _ltree_root->counted_loop( this );
5149 }
5150
5151 // Find latest loop placement. Find ideal loop placement.
5152 visited.clear();
5153 init_dom_lca_tags();
5154 // Need C->root() on worklist when processing outs
5155 worklist.push(C->root());
5156 NOT_PRODUCT( C->verify_graph_edges(); )
5157 worklist.push(C->top());
5158 build_loop_late( visited, worklist, nstack );
5159 if (C->failing()) { return; }
5160
5161 if (_verify_only) {
5162 C->restore_major_progress(old_progress);
5163 assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
5164 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
5165 return;
5166 }
5167
5168 // clear out the dead code after build_loop_late
5169 while (_deadlist.size()) {
5170 _igvn.remove_globally_dead_node(_deadlist.pop());
5171 }
5172
5173 eliminate_useless_zero_trip_guard();
5174 eliminate_useless_multiversion_if();
5175
5176 if (stop_early) {
5177 assert(do_expensive_nodes, "why are we here?");
5178 if (process_expensive_nodes()) {
5179 // If we made some progress when processing expensive nodes then
5180 // the IGVN may modify the graph in a way that will allow us to
5181 // make some more progress: we need to try processing expensive
5182 // nodes again.
5183 C->set_major_progress();
5184 }
5185 return;
5186 }
5187
5188 // Some parser-inserted loop predicates could never be used by loop
5189 // predication or they were moved away from loop during some optimizations.
5190 // For example, peeling. Eliminate them before next loop optimizations.
5191 eliminate_useless_predicates();
5192
5193 #ifndef PRODUCT
5194 C->verify_graph_edges();
5195 if (_verify_me) { // Nested verify pass?
5196 // Check to see if the verify _mode is broken
5197 assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
5198 return;
5199 }
5200 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
5201 if (TraceLoopOpts && C->has_loops()) {
5202 _ltree_root->dump();
5203 }
5204 #endif
5205
5206 if (skip_loop_opts) {
5207 C->restore_major_progress(old_progress);
5208 return;
5209 }
5210
5211 if (do_max_unroll) {
5212 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5213 IdealLoopTree* lpt = iter.current();
5214 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
5215 lpt->compute_trip_count(this, T_INT);
5216 if (!lpt->do_one_iteration_loop(this) &&
5217 !lpt->do_remove_empty_loop(this)) {
5218 AutoNodeBudget node_budget(this);
5219 if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
5220 lpt->policy_maximally_unroll(this)) {
5221 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
5222 do_maximally_unroll(lpt, worklist);
5223 }
5224 }
5225 }
5226 }
5227
5228 C->restore_major_progress(old_progress);
5229 return;
5230 }
5231
5232 if (ReassociateInvariants && !C->major_progress()) {
5233 // Reassociate invariants and prep for split_thru_phi
5234 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5235 IdealLoopTree* lpt = iter.current();
5236 if (!lpt->is_loop()) {
5237 continue;
5238 }
5239 Node* head = lpt->_head;
5240 if (!lpt->is_innermost()) continue;
5241
5242 // check for vectorized loops, any reassociation of invariants was already done
5243 if (head->is_CountedLoop() && head->as_CountedLoop()->is_unroll_only()) {
5244 continue;
5245 } else {
5246 AutoNodeBudget node_budget(this);
5247 lpt->reassociate_invariants(this);
5248 }
5249 // Because RCE opportunities can be masked by split_thru_phi,
5250 // look for RCE candidates and inhibit split_thru_phi
5251 // on just their loop-phi's for this pass of loop opts
5252 if (SplitIfBlocks && do_split_ifs &&
5253 head->is_BaseCountedLoop() &&
5254 head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
5255 (lpt->policy_range_check(this, true, T_LONG) ||
5256 (head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
5257 lpt->_rce_candidate = 1; // = true
5258 }
5259 }
5260 }
5261
5262 // Check for aggressive application of split-if and other transforms
5263 // that require basic-block info (like cloning through Phi's)
5264 if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
5265 visited.clear();
5266 split_if_with_blocks(visited, nstack);
5267 if (C->failing()) {
5268 return;
5269 }
5270 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
5271 }
5272
5273 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
5274 C->set_major_progress();
5275 }
5276
5277 // Perform loop predication before iteration splitting
5278 if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {
5279 _ltree_root->_child->loop_predication(this);
5280 }
5281
5282 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
5283 if (do_intrinsify_fill()) {
5284 C->set_major_progress();
5285 }
5286 }
5287
5288 // Perform iteration-splitting on inner loops. Split iterations to avoid
5289 // range checks or one-shot null checks.
5290
5291 // If split-if's didn't hack the graph too bad (no CFG changes)
5292 // then do loop opts.
5293 if (C->has_loops() && !C->major_progress()) {
5294 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
5295 _ltree_root->_child->iteration_split( this, worklist );
5296 // No verify after peeling! GCM has hoisted code out of the loop.
5297 // After peeling, the hoisted code could sink inside the peeled area.
5298 // The peeling code does not try to recompute the best location for
5299 // all the code before the peeled area, so the verify pass will always
5300 // complain about it.
5301 }
5302
5303 // Check for bailout, and return
5304 if (C->failing()) {
5305 return;
5306 }
5307
5308 // Do verify graph edges in any case
5309 NOT_PRODUCT( C->verify_graph_edges(); );
5310
5311 if (!do_split_ifs) {
5312 // We saw major progress in Split-If to get here. We forced a
5313 // pass with unrolling and not split-if, however more split-if's
5314 // might make progress. If the unrolling didn't make progress
5315 // then the major-progress flag got cleared and we won't try
5316 // another round of Split-If. In particular the ever-common
5317 // instance-of/check-cast pattern requires at least 2 rounds of
5318 // Split-If to clear out.
5319 C->set_major_progress();
5320 }
5321
5322 // Repeat loop optimizations if new loops were seen
5323 if (created_loop_node()) {
5324 C->set_major_progress();
5325 }
5326
5327 // Auto-vectorize main-loop
5328 if (C->do_superword() && C->has_loops() && !C->major_progress()) {
5329 Compile::TracePhase tp(_t_autoVectorize);
5330
5331 // Shared data structures for all AutoVectorizations, to reduce allocations
5332 // of large arrays.
5333 VSharedData vshared;
5334 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5335 IdealLoopTree* lpt = iter.current();
5336 AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
5337
5338 if (status == AutoVectorizeStatus::TriedAndFailed) {
5339 // We tried vectorization, but failed. From now on only unroll the loop.
5340 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
5341 if (cl->has_passed_slp()) {
5342 C->set_major_progress();
5343 cl->set_notpassed_slp();
5344 cl->mark_do_unroll_only();
5345 }
5346 }
5347 }
5348 }
5349
5350 // Keep loop predicates and perform optimizations with them
5351 // until no more loop optimizations could be done.
5352 // After that switch predicates off and do more loop optimizations.
5353 if (!C->major_progress() && (C->parse_predicate_count() > 0)) {
5354 C->mark_parse_predicate_nodes_useless(_igvn);
5355 assert(C->parse_predicate_count() == 0, "should be zero now");
5356 if (TraceLoopOpts) {
5357 tty->print_cr("PredicatesOff");
5358 }
5359 C->set_major_progress();
5360 }
5361 }
5362
5363 #ifndef PRODUCT
5364 //------------------------------print_statistics-------------------------------
5365 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
5366 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
5367 volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
5368 volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
5369 volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
5370 void PhaseIdealLoop::print_statistics() {
5371 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);
5372 }
5373 #endif
5374
5375 #ifdef ASSERT
5376 // Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
5377 void PhaseIdealLoop::verify() const {
5378 ResourceMark rm;
5379 bool old_progress = C->major_progress();
5380 bool success = true;
5381
5382 PhaseIdealLoop phase_verify(_igvn, this);
5383 if (C->failing_internal()) {
5384 return;
5385 }
5386
5387 // Verify ctrl and idom of every node.
5388 success &= verify_idom_and_nodes(C->root(), &phase_verify);
5389
5390 // Verify loop-tree.
5391 success &= _ltree_root->verify_tree(phase_verify._ltree_root);
5392
5393 assert(success, "VerifyLoopOptimizations failed");
5394
5395 // Major progress was cleared by creating a verify version of PhaseIdealLoop.
5396 C->restore_major_progress(old_progress);
5397 }
5398
5399 // Perform a BFS starting at n, through all inputs.
5400 // Call verify_idom and verify_node on all nodes of BFS traversal.
5401 bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {
5402 Unique_Node_List worklist;
5403 worklist.push(root);
5404 bool success = true;
5405 for (uint i = 0; i < worklist.size(); i++) {
5406 Node* n = worklist.at(i);
5407 // process node
5408 success &= verify_idom(n, phase_verify);
5409 success &= verify_loop_ctrl(n, phase_verify);
5410 // visit inputs
5411 for (uint j = 0; j < n->req(); j++) {
5412 if (n->in(j) != nullptr) {
5413 worklist.push(n->in(j));
5414 }
5415 }
5416 }
5417 return success;
5418 }
5419
5420 // Verify dominator structure (IDOM).
5421 bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {
5422 // Verify IDOM for all CFG nodes (except root).
5423 if (!n->is_CFG() || n->is_Root()) {
5424 return true; // pass
5425 }
5426
5427 if (n->_idx >= _idom_size) {
5428 tty->print("CFG Node with no idom: ");
5429 n->dump();
5430 return false; // fail
5431 }
5432
5433 Node* id = idom_no_update(n);
5434 Node* id_verify = phase_verify->idom_no_update(n);
5435 if (id != id_verify) {
5436 tty->print("Mismatching idom for node: ");
5437 n->dump();
5438 tty->print(" We have idom: ");
5439 id->dump();
5440 tty->print(" Verify has idom: ");
5441 id_verify->dump();
5442 tty->cr();
5443 return false; // fail
5444 }
5445 return true; // pass
5446 }
5447
5448 // Verify "_loop_or_ctrl": control and loop membership.
5449 // (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5450 // (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5451 // (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5452 // (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5453 bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {
5454 const uint i = n->_idx;
5455 // The loop-tree was built from def to use (top-down).
5456 // The verification happens from use to def (bottom-up).
5457 // We may thus find nodes during verification that are not in the loop-tree.
5458 if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {
5459 if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {
5460 tty->print_cr("Was reachable in only one. this %d, verify %d.",
5461 _loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5462 n->dump();
5463 return false; // fail
5464 }
5465 // Not reachable for both.
5466 return true; // pass
5467 }
5468
5469 if (n->is_CFG() == has_ctrl(n)) {
5470 tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));
5471 n->dump();
5472 return false; // fail
5473 }
5474
5475 if (has_ctrl(n) != phase_verify->has_ctrl(n)) {
5476 tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));
5477 n->dump();
5478 return false; // fail
5479 } else if (has_ctrl(n)) {
5480 assert(phase_verify->has_ctrl(n), "sanity");
5481 // n is a data node.
5482 // Verify that its ctrl is the same.
5483
5484 // Broken part of VerifyLoopOptimizations (A)
5485 // Reason:
5486 // BUG, wrong control set for example in
5487 // PhaseIdealLoop::split_if_with_blocks
5488 // at "set_ctrl(x, new_ctrl);"
5489 /*
5490 if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5491 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
5492 tty->print("Mismatched control setting for: ");
5493 n->dump();
5494 if( fail++ > 10 ) return;
5495 Node *c = get_ctrl_no_update(n);
5496 tty->print("We have it as: ");
5497 if( c->in(0) ) c->dump();
5498 else tty->print_cr("N%d",c->_idx);
5499 tty->print("Verify thinks: ");
5500 if( loop_verify->has_ctrl(n) )
5501 loop_verify->get_ctrl_no_update(n)->dump();
5502 else
5503 loop_verify->get_loop_idx(n)->dump();
5504 tty->cr();
5505 }
5506 */
5507 return true; // pass
5508 } else {
5509 assert(!phase_verify->has_ctrl(n), "sanity");
5510 // n is a ctrl node.
5511 // Verify that not has_ctrl, and that get_loop_idx is the same.
5512
5513 // Broken part of VerifyLoopOptimizations (B)
5514 // Reason:
5515 // NeverBranch node for example is added to loop outside its scope.
5516 // Once we run build_loop_tree again, it is added to the correct loop.
5517 /*
5518 if (!C->major_progress()) {
5519 // Loop selection can be messed up if we did a major progress
5520 // operation, like split-if. Do not verify in that case.
5521 IdealLoopTree *us = get_loop_idx(n);
5522 IdealLoopTree *them = loop_verify->get_loop_idx(n);
5523 if( us->_head != them->_head || us->_tail != them->_tail ) {
5524 tty->print("Unequals loops for: ");
5525 n->dump();
5526 if( fail++ > 10 ) return;
5527 tty->print("We have it as: ");
5528 us->dump();
5529 tty->print("Verify thinks: ");
5530 them->dump();
5531 tty->cr();
5532 }
5533 }
5534 */
5535 return true; // pass
5536 }
5537 }
5538
5539 static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {
5540 assert(a != nullptr && b != nullptr, "must be");
5541 return a->_head->_idx - b->_head->_idx;
5542 }
5543
5544 GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {
5545 GrowableArray<IdealLoopTree*> children;
5546 IdealLoopTree* child = _child;
5547 while (child != nullptr) {
5548 assert(child->_parent == this, "all must be children of this");
5549 children.insert_sorted<compare_tree>(child);
5550 child = child->_next;
5551 }
5552 return children;
5553 }
5554
5555 // Verify that tree structures match. Because the CFG can change, siblings
5556 // within the loop tree can be reordered. We attempt to deal with that by
5557 // reordering the verify's loop tree if possible.
5558 bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {
5559 assert(_head == loop_verify->_head, "mismatched loop head");
5560 assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5561
5562 // Collect the children
5563 GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5564 GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5565
5566 bool success = true;
5567
5568 // Compare the two children lists
5569 for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {
5570 IdealLoopTree* child = nullptr;
5571 IdealLoopTree* child_verify = nullptr;
5572 // Read from both lists, if possible.
5573 if (i < children.length()) {
5574 child = children.at(i);
5575 }
5576 if (j < children_verify.length()) {
5577 child_verify = children_verify.at(j);
5578 }
5579 assert(child != nullptr || child_verify != nullptr, "must find at least one");
5580 if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {
5581 // We found two non-equal children. Select the smaller one.
5582 if (child->_head->_idx < child_verify->_head->_idx) {
5583 child_verify = nullptr;
5584 } else {
5585 child = nullptr;
5586 }
5587 }
5588 // Process the two children, or potentially log the failure if we only found one.
5589 if (child_verify == nullptr) {
5590 if (child->_irreducible && Compile::current()->major_progress()) {
5591 // Irreducible loops can pick a different header (one of its entries).
5592 } else {
5593 tty->print_cr("We have a loop that verify does not have");
5594 child->dump();
5595 success = false;
5596 }
5597 i++; // step for this
5598 } else if (child == nullptr) {
5599 if (child_verify->_irreducible && Compile::current()->major_progress()) {
5600 // Irreducible loops can pick a different header (one of its entries).
5601 } else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {
5602 // Infinite loops do not get attached to the loop-tree on their first visit.
5603 // "this" runs before "loop_verify". It is thus possible that we find the
5604 // infinite loop only for "child_verify". Only finding it with "child" would
5605 // mean that we lost it, which is not ok.
5606 } else {
5607 tty->print_cr("Verify has a loop that we do not have");
5608 child_verify->dump();
5609 success = false;
5610 }
5611 j++; // step for verify
5612 } else {
5613 assert(child->_head == child_verify->_head, "We have both and they are equal");
5614 success &= child->verify_tree(child_verify); // Recursion
5615 i++; // step for this
5616 j++; // step for verify
5617 }
5618 }
5619
5620 // Broken part of VerifyLoopOptimizations (D)
5621 // Reason:
5622 // split_if has to update the _tail, if it is modified. But that is done by
5623 // checking to what loop the iff belongs to. That info can be wrong, and then
5624 // we do not update the _tail correctly.
5625 /*
5626 Node *tail = _tail; // Inline a non-updating version of
5627 while( !tail->in(0) ) // the 'tail()' call.
5628 tail = tail->in(1);
5629 assert( tail == loop->_tail, "mismatched loop tail" );
5630 */
5631
5632 if (_head->is_CountedLoop()) {
5633 CountedLoopNode *cl = _head->as_CountedLoop();
5634
5635 Node* ctrl = cl->init_control();
5636 Node* back = cl->back_control();
5637 assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5638 assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5639 cl->loopexit(); // assert implied
5640 }
5641
5642 // Broken part of VerifyLoopOptimizations (E)
5643 // Reason:
5644 // PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5645 // to the loop body. Or maybe they are not added to the correct loop.
5646 // at "Node* x = n->clone();"
5647 /*
5648 // Innermost loops need to verify loop bodies,
5649 // but only if no 'major_progress'
5650 int fail = 0;
5651 if (!Compile::current()->major_progress() && _child == nullptr) {
5652 for( uint i = 0; i < _body.size(); i++ ) {
5653 Node *n = _body.at(i);
5654 if (n->outcnt() == 0) continue; // Ignore dead
5655 uint j;
5656 for( j = 0; j < loop->_body.size(); j++ )
5657 if( loop->_body.at(j) == n )
5658 break;
5659 if( j == loop->_body.size() ) { // Not found in loop body
5660 // Last ditch effort to avoid assertion: Its possible that we
5661 // have some users (so outcnt not zero) but are still dead.
5662 // Try to find from root.
5663 if (Compile::current()->root()->find(n->_idx)) {
5664 fail++;
5665 tty->print("We have that verify does not: ");
5666 n->dump();
5667 }
5668 }
5669 }
5670 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
5671 Node *n = loop->_body.at(i2);
5672 if (n->outcnt() == 0) continue; // Ignore dead
5673 uint j;
5674 for( j = 0; j < _body.size(); j++ )
5675 if( _body.at(j) == n )
5676 break;
5677 if( j == _body.size() ) { // Not found in loop body
5678 // Last ditch effort to avoid assertion: Its possible that we
5679 // have some users (so outcnt not zero) but are still dead.
5680 // Try to find from root.
5681 if (Compile::current()->root()->find(n->_idx)) {
5682 fail++;
5683 tty->print("Verify has that we do not: ");
5684 n->dump();
5685 }
5686 }
5687 }
5688 assert( !fail, "loop body mismatch" );
5689 }
5690 */
5691 return success;
5692 }
5693 #endif
5694
5695 //------------------------------set_idom---------------------------------------
5696 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
5697 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
5698 uint idx = d->_idx;
5699 if (idx >= _idom_size) {
5700 uint newsize = next_power_of_2(idx);
5701 _idom = REALLOC_ARENA_ARRAY(&_arena, Node*, _idom,_idom_size,newsize);
5702 _dom_depth = REALLOC_ARENA_ARRAY(&_arena, uint, _dom_depth,_idom_size,newsize);
5703 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5704 _idom_size = newsize;
5705 }
5706 _idom[idx] = n;
5707 _dom_depth[idx] = dom_depth;
5708 }
5709
5710 //------------------------------recompute_dom_depth---------------------------------------
5711 // The dominator tree is constructed with only parent pointers.
5712 // This recomputes the depth in the tree by first tagging all
5713 // nodes as "no depth yet" marker. The next pass then runs up
5714 // the dom tree from each node marked "no depth yet", and computes
5715 // the depth on the way back down.
5716 void PhaseIdealLoop::recompute_dom_depth() {
5717 uint no_depth_marker = C->unique();
5718 uint i;
5719 // Initialize depth to "no depth yet" and realize all lazy updates
5720 for (i = 0; i < _idom_size; i++) {
5721 // Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5722 if (_dom_depth[i] > 0 && _idom[i] != nullptr) {
5723 _dom_depth[i] = no_depth_marker;
5724
5725 // heal _idom if it has a fwd mapping in _loop_or_ctrl
5726 if (_idom[i]->in(0) == nullptr) {
5727 idom(i);
5728 }
5729 }
5730 }
5731 if (_dom_stk == nullptr) {
5732 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5733 if (init_size < 10) init_size = 10;
5734 _dom_stk = new GrowableArray<uint>(init_size);
5735 }
5736 // Compute new depth for each node.
5737 for (i = 0; i < _idom_size; i++) {
5738 uint j = i;
5739 // Run up the dom tree to find a node with a depth
5740 while (_dom_depth[j] == no_depth_marker) {
5741 _dom_stk->push(j);
5742 j = _idom[j]->_idx;
5743 }
5744 // Compute the depth on the way back down this tree branch
5745 uint dd = _dom_depth[j] + 1;
5746 while (_dom_stk->length() > 0) {
5747 uint j = _dom_stk->pop();
5748 _dom_depth[j] = dd;
5749 dd++;
5750 }
5751 }
5752 }
5753
5754 //------------------------------sort-------------------------------------------
5755 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5756 // loop tree, not the root.
5757 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
5758 if( !innermost ) return loop; // New innermost loop
5759
5760 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5761 assert( loop_preorder, "not yet post-walked loop" );
5762 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5763 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5764
5765 // Insert at start of list
5766 while( l ) { // Insertion sort based on pre-order
5767 if( l == loop ) return innermost; // Already on list!
5768 int l_preorder = get_preorder(l->_head); // Cache pre-order number
5769 assert( l_preorder, "not yet post-walked l" );
5770 // Check header pre-order number to figure proper nesting
5771 if( loop_preorder > l_preorder )
5772 break; // End of insertion
5773 // If headers tie (e.g., shared headers) check tail pre-order numbers.
5774 // Since I split shared headers, you'd think this could not happen.
5775 // BUT: I must first do the preorder numbering before I can discover I
5776 // have shared headers, so the split headers all get the same preorder
5777 // number as the RegionNode they split from.
5778 if( loop_preorder == l_preorder &&
5779 get_preorder(loop->_tail) < get_preorder(l->_tail) )
5780 break; // Also check for shared headers (same pre#)
5781 pp = &l->_parent; // Chain up list
5782 l = *pp;
5783 }
5784 // Link into list
5785 // Point predecessor to me
5786 *pp = loop;
5787 // Point me to successor
5788 IdealLoopTree *p = loop->_parent;
5789 loop->_parent = l; // Point me to successor
5790 if( p ) sort( p, innermost ); // Insert my parents into list as well
5791 return innermost;
5792 }
5793
5794 //------------------------------build_loop_tree--------------------------------
5795 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5796 // bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5797 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5798 // tightest enclosing IdealLoopTree for post-walked.
5799 //
5800 // During my forward walk I do a short 1-layer lookahead to see if I can find
5801 // a loop backedge with that doesn't have any work on the backedge. This
5802 // helps me construct nested loops with shared headers better.
5803 //
5804 // Once I've done the forward recursion, I do the post-work. For each child
5805 // I check to see if there is a backedge. Backedges define a loop! I
5806 // insert an IdealLoopTree at the target of the backedge.
5807 //
5808 // During the post-work I also check to see if I have several children
5809 // belonging to different loops. If so, then this Node is a decision point
5810 // where control flow can choose to change loop nests. It is at this
5811 // decision point where I can figure out how loops are nested. At this
5812 // time I can properly order the different loop nests from my children.
5813 // Note that there may not be any backedges at the decision point!
5814 //
5815 // Since the decision point can be far removed from the backedges, I can't
5816 // order my loops at the time I discover them. Thus at the decision point
5817 // I need to inspect loop header pre-order numbers to properly nest my
5818 // loops. This means I need to sort my childrens' loops by pre-order.
5819 // The sort is of size number-of-control-children, which generally limits
5820 // it to size 2 (i.e., I just choose between my 2 target loops).
5821 void PhaseIdealLoop::build_loop_tree() {
5822 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5823 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5824 Node *n = C->root();
5825 bltstack.push(n);
5826 int pre_order = 1;
5827 int stack_size;
5828
5829 while ( ( stack_size = bltstack.length() ) != 0 ) {
5830 n = bltstack.top(); // Leave node on stack
5831 if ( !is_visited(n) ) {
5832 // ---- Pre-pass Work ----
5833 // Pre-walked but not post-walked nodes need a pre_order number.
5834
5835 set_preorder_visited( n, pre_order ); // set as visited
5836
5837 // ---- Scan over children ----
5838 // Scan first over control projections that lead to loop headers.
5839 // This helps us find inner-to-outer loops with shared headers better.
5840
5841 // Scan children's children for loop headers.
5842 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
5843 Node* m = n->raw_out(i); // Child
5844 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
5845 // Scan over children's children to find loop
5846 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5847 Node* l = m->fast_out(j);
5848 if( is_visited(l) && // Been visited?
5849 !is_postvisited(l) && // But not post-visited
5850 get_preorder(l) < pre_order ) { // And smaller pre-order
5851 // Found! Scan the DFS down this path before doing other paths
5852 bltstack.push(m);
5853 break;
5854 }
5855 }
5856 }
5857 }
5858 pre_order++;
5859 }
5860 else if ( !is_postvisited(n) ) {
5861 // Note: build_loop_tree_impl() adds out edges on rare occasions,
5862 // such as com.sun.rsasign.am::a.
5863 // For non-recursive version, first, process current children.
5864 // On next iteration, check if additional children were added.
5865 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
5866 Node* u = n->raw_out(k);
5867 if ( u->is_CFG() && !is_visited(u) ) {
5868 bltstack.push(u);
5869 }
5870 }
5871 if ( bltstack.length() == stack_size ) {
5872 // There were no additional children, post visit node now
5873 (void)bltstack.pop(); // Remove node from stack
5874 pre_order = build_loop_tree_impl(n, pre_order);
5875 // Check for bailout
5876 if (C->failing()) {
5877 return;
5878 }
5879 // Check to grow _preorders[] array for the case when
5880 // build_loop_tree_impl() adds new nodes.
5881 check_grow_preorders();
5882 }
5883 }
5884 else {
5885 (void)bltstack.pop(); // Remove post-visited node from stack
5886 }
5887 }
5888 DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5889 }
5890
5891 //------------------------------build_loop_tree_impl---------------------------
5892 int PhaseIdealLoop::build_loop_tree_impl(Node* n, int pre_order) {
5893 // ---- Post-pass Work ----
5894 // Pre-walked but not post-walked nodes need a pre_order number.
5895
5896 // Tightest enclosing loop for this Node
5897 IdealLoopTree *innermost = nullptr;
5898
5899 // For all children, see if any edge is a backedge. If so, make a loop
5900 // for it. Then find the tightest enclosing loop for the self Node.
5901 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5902 Node* m = n->fast_out(i); // Child
5903 if (n == m) continue; // Ignore control self-cycles
5904 if (!m->is_CFG()) continue;// Ignore non-CFG edges
5905
5906 IdealLoopTree *l; // Child's loop
5907 if (!is_postvisited(m)) { // Child visited but not post-visited?
5908 // Found a backedge
5909 assert(get_preorder(m) < pre_order, "should be backedge");
5910 // Check for the RootNode, which is already a LoopNode and is allowed
5911 // to have multiple "backedges".
5912 if (m == C->root()) { // Found the root?
5913 l = _ltree_root; // Root is the outermost LoopNode
5914 } else { // Else found a nested loop
5915 // Insert a LoopNode to mark this loop.
5916 l = new IdealLoopTree(this, m, n);
5917 } // End of Else found a nested loop
5918 if (!has_loop(m)) { // If 'm' does not already have a loop set
5919 set_loop(m, l); // Set loop header to loop now
5920 }
5921 } else { // Else not a nested loop
5922 if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5923 IdealLoopTree* m_loop = get_loop(m);
5924 l = m_loop; // Get previously determined loop
5925 // If successor is header of a loop (nest), move up-loop till it
5926 // is a member of some outer enclosing loop. Since there are no
5927 // shared headers (I've split them already) I only need to go up
5928 // at most 1 level.
5929 while (l && l->_head == m) { // Successor heads loop?
5930 l = l->_parent; // Move up 1 for me
5931 }
5932 // If this loop is not properly parented, then this loop
5933 // has no exit path out, i.e. its an infinite loop.
5934 if (!l) {
5935 // Make loop "reachable" from root so the CFG is reachable. Basically
5936 // insert a bogus loop exit that is never taken. 'm', the loop head,
5937 // points to 'n', one (of possibly many) fall-in paths. There may be
5938 // many backedges as well.
5939
5940 if (!_verify_only) {
5941 // Insert the NeverBranch between 'm' and it's control user.
5942 NeverBranchNode *iff = new NeverBranchNode( m );
5943 _igvn.register_new_node_with_optimizer(iff);
5944 set_loop(iff, m_loop);
5945 Node *if_t = new CProjNode( iff, 0 );
5946 _igvn.register_new_node_with_optimizer(if_t);
5947 set_loop(if_t, m_loop);
5948
5949 Node* cfg = nullptr; // Find the One True Control User of m
5950 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5951 Node* x = m->fast_out(j);
5952 if (x->is_CFG() && x != m && x != iff)
5953 { cfg = x; break; }
5954 }
5955 assert(cfg != nullptr, "must find the control user of m");
5956 uint k = 0; // Probably cfg->in(0)
5957 while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5958 _igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5959
5960 // Now create the never-taken loop exit
5961 Node *if_f = new CProjNode( iff, 1 );
5962 _igvn.register_new_node_with_optimizer(if_f);
5963 set_loop(if_f, _ltree_root);
5964 // Find frame ptr for Halt. Relies on the optimizer
5965 // V-N'ing. Easier and quicker than searching through
5966 // the program structure.
5967 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5968 _igvn.register_new_node_with_optimizer(frame);
5969 // Halt & Catch Fire
5970 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5971 _igvn.register_new_node_with_optimizer(halt);
5972 set_loop(halt, _ltree_root);
5973 _igvn.add_input_to(C->root(), halt);
5974 }
5975 set_loop(C->root(), _ltree_root);
5976 // move to outer most loop with same header
5977 l = m_loop;
5978 while (true) {
5979 IdealLoopTree* next = l->_parent;
5980 if (next == nullptr || next->_head != m) {
5981 break;
5982 }
5983 l = next;
5984 }
5985 // properly insert infinite loop in loop tree
5986 sort(_ltree_root, l);
5987 // fix child link from parent
5988 IdealLoopTree* p = l->_parent;
5989 l->_next = p->_child;
5990 p->_child = l;
5991 // code below needs enclosing loop
5992 l = l->_parent;
5993 }
5994 }
5995 if (is_postvisited(l->_head)) {
5996 // We are currently visiting l, but its head has already been post-visited.
5997 // l is irreducible: we just found a second entry m.
5998 _has_irreducible_loops = true;
5999 RegionNode* secondary_entry = m->as_Region();
6000
6001 if (!secondary_entry->can_be_irreducible_entry()) {
6002 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
6003 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
6004 return pre_order;
6005 }
6006
6007 // Walk up the loop-tree, mark all loops that are already post-visited as irreducible
6008 // Since m is a secondary entry to them all.
6009 while( is_postvisited(l->_head) ) {
6010 l->_irreducible = 1; // = true
6011 RegionNode* head = l->_head->as_Region();
6012 if (!head->can_be_irreducible_entry()) {
6013 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
6014 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
6015 return pre_order;
6016 }
6017 l = l->_parent;
6018 // Check for bad CFG here to prevent crash, and bailout of compile
6019 if (l == nullptr) {
6020 #ifndef PRODUCT
6021 if (TraceLoopOpts) {
6022 tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");
6023 m->dump();
6024 }
6025 #endif
6026 // This is a rare case that we do not want to handle in C2.
6027 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
6028 return pre_order;
6029 }
6030 }
6031 }
6032 if (!_verify_only) {
6033 C->set_has_irreducible_loop(_has_irreducible_loops);
6034 }
6035
6036 // This Node might be a decision point for loops. It is only if
6037 // it's children belong to several different loops. The sort call
6038 // does a trivial amount of work if there is only 1 child or all
6039 // children belong to the same loop. If however, the children
6040 // belong to different loops, the sort call will properly set the
6041 // _parent pointers to show how the loops nest.
6042 //
6043 // In any case, it returns the tightest enclosing loop.
6044 innermost = sort( l, innermost );
6045 }
6046
6047 // Def-use info will have some dead stuff; dead stuff will have no
6048 // loop decided on.
6049
6050 // Am I a loop header? If so fix up my parent's child and next ptrs.
6051 if( innermost && innermost->_head == n ) {
6052 assert( get_loop(n) == innermost, "" );
6053 IdealLoopTree *p = innermost->_parent;
6054 IdealLoopTree *l = innermost;
6055 while (p && l->_head == n) {
6056 l->_next = p->_child; // Put self on parents 'next child'
6057 p->_child = l; // Make self as first child of parent
6058 l = p; // Now walk up the parent chain
6059 p = l->_parent;
6060 }
6061 } else {
6062 // Note that it is possible for a LoopNode to reach here, if the
6063 // backedge has been made unreachable (hence the LoopNode no longer
6064 // denotes a Loop, and will eventually be removed).
6065
6066 // Record tightest enclosing loop for self. Mark as post-visited.
6067 set_loop(n, innermost);
6068 // Also record has_call flag early on
6069 if (innermost) {
6070 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
6071 // Do not count uncommon calls
6072 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
6073 Node *iff = n->in(0)->in(0);
6074 // No any calls for vectorized loops.
6075 if (C->do_superword() ||
6076 !iff->is_If() ||
6077 (n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
6078 iff->as_If()->_prob >= 0.01) {
6079 innermost->_has_call = 1;
6080 }
6081 }
6082 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
6083 // Disable loop optimizations if the loop has a scalar replaceable
6084 // allocation. This disabling may cause a potential performance lost
6085 // if the allocation is not eliminated for some reason.
6086 innermost->_allow_optimizations = false;
6087 innermost->_has_call = 1; // = true
6088 } else if (n->Opcode() == Op_SafePoint) {
6089 // Record all safepoints in this loop.
6090 if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
6091 innermost->_safepts->push(n);
6092 }
6093 }
6094 }
6095
6096 // Flag as post-visited now
6097 set_postvisited(n);
6098 return pre_order;
6099 }
6100
6101 #ifdef ASSERT
6102 //--------------------------verify_regions_in_irreducible_loops----------------
6103 // Iterate down from Root through CFG, verify for every region:
6104 // if it is in an irreducible loop it must be marked as such
6105 void PhaseIdealLoop::verify_regions_in_irreducible_loops() {
6106 ResourceMark rm;
6107 if (!_has_irreducible_loops) {
6108 // last build_loop_tree has not found any irreducible loops
6109 // hence no region has to be marked is_in_irreduible_loop
6110 return;
6111 }
6112
6113 RootNode* root = C->root();
6114 Unique_Node_List worklist; // visit all nodes once
6115 worklist.push(root);
6116 bool failure = false;
6117 for (uint i = 0; i < worklist.size(); i++) {
6118 Node* n = worklist.at(i);
6119 if (n->is_Region()) {
6120 RegionNode* region = n->as_Region();
6121 if (is_in_irreducible_loop(region) &&
6122 region->loop_status() == RegionNode::LoopStatus::Reducible) {
6123 failure = true;
6124 tty->print("irreducible! ");
6125 region->dump();
6126 }
6127 }
6128 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
6129 Node* use = n->fast_out(j);
6130 if (use->is_CFG()) {
6131 worklist.push(use); // push if was not pushed before
6132 }
6133 }
6134 }
6135 assert(!failure, "region in irreducible loop was marked as reducible");
6136 }
6137
6138 //---------------------------is_in_irreducible_loop-------------------------
6139 // Analogous to ciTypeFlow::Block::is_in_irreducible_loop
6140 bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {
6141 if (!_has_irreducible_loops) {
6142 return false; // no irreducible loop in graph
6143 }
6144 IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
6145 do {
6146 if (l->_irreducible) {
6147 return true; // found it
6148 }
6149 if (l == _ltree_root) {
6150 return false; // reached root, terimnate
6151 }
6152 l = l->_parent;
6153 } while (l != nullptr);
6154 assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
6155 // We have "l->_parent == nullptr", which happens only for infinite loops,
6156 // where no parent is attached to the loop. We did not find any irreducible
6157 // loop from this block out to lp. Thus lp only has one entry, and no exit
6158 // (it is infinite and reducible). We can always rewrite an infinite loop
6159 // that is nested inside other loops:
6160 // while(condition) { infinite_loop; }
6161 // with an equivalent program where the infinite loop is an outermost loop
6162 // that is not nested in any loop:
6163 // while(condition) { break; } infinite_loop;
6164 // Thus, we can understand lp as an outermost loop, and can terminate and
6165 // conclude: this block is in no irreducible loop.
6166 return false;
6167 }
6168 #endif
6169
6170 //------------------------------build_loop_early-------------------------------
6171 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6172 // First pass computes the earliest controlling node possible. This is the
6173 // controlling input with the deepest dominating depth.
6174 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6175 while (worklist.size() != 0) {
6176 // Use local variables nstack_top_n & nstack_top_i to cache values
6177 // on nstack's top.
6178 Node *nstack_top_n = worklist.pop();
6179 uint nstack_top_i = 0;
6180 //while_nstack_nonempty:
6181 while (true) {
6182 // Get parent node and next input's index from stack's top.
6183 Node *n = nstack_top_n;
6184 uint i = nstack_top_i;
6185 uint cnt = n->req(); // Count of inputs
6186 if (i == 0) { // Pre-process the node.
6187 if( has_node(n) && // Have either loop or control already?
6188 !has_ctrl(n) ) { // Have loop picked out already?
6189 // During "merge_many_backedges" we fold up several nested loops
6190 // into a single loop. This makes the members of the original
6191 // loop bodies pointing to dead loops; they need to move up
6192 // to the new UNION'd larger loop. I set the _head field of these
6193 // dead loops to null and the _parent field points to the owning
6194 // loop. Shades of UNION-FIND algorithm.
6195 IdealLoopTree *ilt;
6196 while( !(ilt = get_loop(n))->_head ) {
6197 // Normally I would use a set_loop here. But in this one special
6198 // case, it is legal (and expected) to change what loop a Node
6199 // belongs to.
6200 _loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
6201 }
6202 // Remove safepoints ONLY if I've already seen I don't need one.
6203 // (the old code here would yank a 2nd safepoint after seeing a
6204 // first one, even though the 1st did not dominate in the loop body
6205 // and thus could be avoided indefinitely)
6206 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
6207 is_deleteable_safept(n)) {
6208 Node *in = n->in(TypeFunc::Control);
6209 replace_node_and_forward_ctrl(n, in); // Pull safepoint now
6210 if (ilt->_safepts != nullptr) {
6211 ilt->_safepts->yank(n);
6212 }
6213 // Carry on with the recursion "as if" we are walking
6214 // only the control input
6215 if( !visited.test_set( in->_idx ) ) {
6216 worklist.push(in); // Visit this guy later, using worklist
6217 }
6218 // Get next node from nstack:
6219 // - skip n's inputs processing by setting i > cnt;
6220 // - we also will not call set_early_ctrl(n) since
6221 // has_node(n) == true (see the condition above).
6222 i = cnt + 1;
6223 }
6224 }
6225 } // if (i == 0)
6226
6227 // Visit all inputs
6228 bool done = true; // Assume all n's inputs will be processed
6229 while (i < cnt) {
6230 Node *in = n->in(i);
6231 ++i;
6232 if (in == nullptr) continue;
6233 if (in->pinned() && !in->is_CFG())
6234 set_ctrl(in, in->in(0));
6235 int is_visited = visited.test_set( in->_idx );
6236 if (!has_node(in)) { // No controlling input yet?
6237 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
6238 assert( !is_visited, "visit only once" );
6239 nstack.push(n, i); // Save parent node and next input's index.
6240 nstack_top_n = in; // Process current input now.
6241 nstack_top_i = 0;
6242 done = false; // Not all n's inputs processed.
6243 break; // continue while_nstack_nonempty;
6244 } else if (!is_visited) {
6245 // This guy has a location picked out for him, but has not yet
6246 // been visited. Happens to all CFG nodes, for instance.
6247 // Visit him using the worklist instead of recursion, to break
6248 // cycles. Since he has a location already we do not need to
6249 // find his location before proceeding with the current Node.
6250 worklist.push(in); // Visit this guy later, using worklist
6251 }
6252 }
6253 if (done) {
6254 // All of n's inputs have been processed, complete post-processing.
6255
6256 // Compute earliest point this Node can go.
6257 // CFG, Phi, pinned nodes already know their controlling input.
6258 if (!has_node(n)) {
6259 // Record earliest legal location
6260 set_early_ctrl(n, false);
6261 }
6262 if (nstack.is_empty()) {
6263 // Finished all nodes on stack.
6264 // Process next node on the worklist.
6265 break;
6266 }
6267 // Get saved parent node and next input's index.
6268 nstack_top_n = nstack.node();
6269 nstack_top_i = nstack.index();
6270 nstack.pop();
6271 }
6272 } // while (true)
6273 }
6274 }
6275
6276 //------------------------------dom_lca_internal--------------------------------
6277 // Pair-wise LCA
6278 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
6279 if( !n1 ) return n2; // Handle null original LCA
6280 assert( n1->is_CFG(), "" );
6281 assert( n2->is_CFG(), "" );
6282 // find LCA of all uses
6283 uint d1 = dom_depth(n1);
6284 uint d2 = dom_depth(n2);
6285 while (n1 != n2) {
6286 if (d1 > d2) {
6287 n1 = idom(n1);
6288 d1 = dom_depth(n1);
6289 } else if (d1 < d2) {
6290 n2 = idom(n2);
6291 d2 = dom_depth(n2);
6292 } else {
6293 // Here d1 == d2. Due to edits of the dominator-tree, sections
6294 // of the tree might have the same depth. These sections have
6295 // to be searched more carefully.
6296
6297 // Scan up all the n1's with equal depth, looking for n2.
6298 Node *t1 = idom(n1);
6299 while (dom_depth(t1) == d1) {
6300 if (t1 == n2) return n2;
6301 t1 = idom(t1);
6302 }
6303 // Scan up all the n2's with equal depth, looking for n1.
6304 Node *t2 = idom(n2);
6305 while (dom_depth(t2) == d2) {
6306 if (t2 == n1) return n1;
6307 t2 = idom(t2);
6308 }
6309 // Move up to a new dominator-depth value as well as up the dom-tree.
6310 n1 = t1;
6311 n2 = t2;
6312 d1 = dom_depth(n1);
6313 d2 = dom_depth(n2);
6314 }
6315 }
6316 return n1;
6317 }
6318
6319 //------------------------------compute_idom-----------------------------------
6320 // Locally compute IDOM using dom_lca call. Correct only if the incoming
6321 // IDOMs are correct.
6322 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
6323 assert( region->is_Region(), "" );
6324 Node *LCA = nullptr;
6325 for( uint i = 1; i < region->req(); i++ ) {
6326 if( region->in(i) != C->top() )
6327 LCA = dom_lca( LCA, region->in(i) );
6328 }
6329 return LCA;
6330 }
6331
6332 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
6333 bool had_error = false;
6334 #ifdef ASSERT
6335 if (early != C->root()) {
6336 // Make sure that there's a dominance path from LCA to early
6337 Node* d = LCA;
6338 while (d != early) {
6339 if (d == C->root()) {
6340 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
6341 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
6342 had_error = true;
6343 break;
6344 }
6345 d = idom(d);
6346 }
6347 }
6348 #endif
6349 return had_error;
6350 }
6351
6352
6353 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
6354 // Compute LCA over list of uses
6355 bool had_error = false;
6356 Node *LCA = nullptr;
6357 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
6358 Node* c = n->fast_out(i);
6359 if (_loop_or_ctrl[c->_idx] == nullptr)
6360 continue; // Skip the occasional dead node
6361 if( c->is_Phi() ) { // For Phis, we must land above on the path
6362 for( uint j=1; j<c->req(); j++ ) {// For all inputs
6363 if( c->in(j) == n ) { // Found matching input?
6364 Node *use = c->in(0)->in(j);
6365 if (_verify_only && use->is_top()) continue;
6366 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6367 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6368 }
6369 }
6370 } else {
6371 // For CFG data-users, use is in the block just prior
6372 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
6373 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6374 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6375 }
6376 }
6377 assert(!had_error, "bad dominance");
6378 return LCA;
6379 }
6380
6381 // Check the shape of the graph at the loop entry. In some cases,
6382 // the shape of the graph does not match the shape outlined below.
6383 // That is caused by the Opaque1 node "protecting" the shape of
6384 // the graph being removed by, for example, the IGVN performed
6385 // in PhaseIdealLoop::build_and_optimize().
6386 //
6387 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
6388 // loop unswitching, and IGVN, or a combination of them) can freely change
6389 // the graph's shape. As a result, the graph shape outlined below cannot
6390 // be guaranteed anymore.
6391 Node* CountedLoopNode::is_canonical_loop_entry() {
6392 if (!is_main_loop() && !is_post_loop()) {
6393 return nullptr;
6394 }
6395 Node* ctrl = skip_assertion_predicates_with_halt();
6396
6397 if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
6398 return nullptr;
6399 }
6400 Node* iffm = ctrl->in(0);
6401 if (iffm == nullptr || iffm->Opcode() != Op_If) {
6402 return nullptr;
6403 }
6404 Node* bolzm = iffm->in(1);
6405 if (bolzm == nullptr || !bolzm->is_Bool()) {
6406 return nullptr;
6407 }
6408 Node* cmpzm = bolzm->in(1);
6409 if (cmpzm == nullptr || !cmpzm->is_Cmp()) {
6410 return nullptr;
6411 }
6412
6413 uint input = is_main_loop() ? 2 : 1;
6414 if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {
6415 return nullptr;
6416 }
6417 bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
6418 #ifdef ASSERT
6419 bool found_opaque = false;
6420 for (uint i = 1; i < cmpzm->req(); i++) {
6421 Node* opnd = cmpzm->in(i);
6422 if (opnd && opnd->is_Opaque1()) {
6423 found_opaque = true;
6424 break;
6425 }
6426 }
6427 assert(found_opaque == res, "wrong pattern");
6428 #endif
6429 return res ? cmpzm->in(input) : nullptr;
6430 }
6431
6432 // Find pre loop end from main loop. Returns nullptr if none.
6433 CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {
6434 assert(is_main_loop(), "Can only find pre-loop from main-loop");
6435 // The loop cannot be optimized if the graph shape at the loop entry is
6436 // inappropriate.
6437 if (is_canonical_loop_entry() == nullptr) {
6438 return nullptr;
6439 }
6440
6441 Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
6442 if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {
6443 return nullptr;
6444 }
6445 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
6446 CountedLoopNode* loop_node = pre_end->loopnode();
6447 if (loop_node == nullptr || !loop_node->is_pre_loop()) {
6448 return nullptr;
6449 }
6450 return pre_end;
6451 }
6452
6453 Node* CountedLoopNode::uncasted_init_trip(bool uncast) {
6454 Node* init = init_trip();
6455 if (uncast && init->is_CastII()) {
6456 // skip over the cast added by PhaseIdealLoop::cast_incr_before_loop() when pre/post/main loops are created because
6457 // it can get in the way of type propagation. For instance, the index tested by an Assertion Predicate, if the cast
6458 // is not skipped over, could be (1):
6459 // (AddI (CastII (AddI pre_loop_iv -2) int) 1)
6460 // while without the cast, it is (2):
6461 // (AddI (AddI pre_loop_iv -2) 1)
6462 // which is be transformed to (3):
6463 // (AddI pre_loop_iv -1)
6464 // The compiler may be able to constant fold the Assertion Predicate condition for (3) but not (1)
6465 assert(init->as_CastII()->carry_dependency() && skip_assertion_predicates_with_halt() == init->in(0), "casted iv phi from pre loop expected");
6466 init = init->in(1);
6467 }
6468 return init;
6469 }
6470
6471 //------------------------------get_late_ctrl----------------------------------
6472 // Compute latest legal control.
6473 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
6474 assert(early != nullptr, "early control should not be null");
6475
6476 Node* LCA = compute_lca_of_uses(n, early);
6477 #ifdef ASSERT
6478 if (LCA == C->root() && LCA != early) {
6479 // def doesn't dominate uses so print some useful debugging output
6480 compute_lca_of_uses(n, early, true);
6481 }
6482 #endif
6483
6484 if (n->is_Load() && LCA != early) {
6485 LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6486 }
6487
6488 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6489 return LCA;
6490 }
6491
6492 // if this is a load, check for anti-dependent stores
6493 // We use a conservative algorithm to identify potential interfering
6494 // instructions and for rescheduling the load. The users of the memory
6495 // input of this load are examined. Any use which is not a load and is
6496 // dominated by early is considered a potentially interfering store.
6497 // This can produce false positives.
6498 Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {
6499 int load_alias_idx = C->get_alias_index(n->adr_type());
6500 if (C->alias_type(load_alias_idx)->is_rewritable()) {
6501 Unique_Node_List worklist;
6502
6503 Node* mem = n->in(MemNode::Memory);
6504 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
6505 Node* s = mem->fast_out(i);
6506 worklist.push(s);
6507 }
6508 for (uint i = 0; i < worklist.size() && LCA != early; i++) {
6509 Node* s = worklist.at(i);
6510 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6511 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6512 s->is_Phi()) {
6513 continue;
6514 } else if (s->is_MergeMem()) {
6515 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6516 Node* s1 = s->fast_out(i);
6517 worklist.push(s1);
6518 }
6519 } else {
6520 Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6521 assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6522 if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {
6523 const TypePtr* adr_type = s->adr_type();
6524 if (s->is_ArrayCopy()) {
6525 // Copy to known instance needs destination type to test for aliasing
6526 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6527 if (dest_type != TypeOopPtr::BOTTOM) {
6528 adr_type = dest_type;
6529 }
6530 }
6531 if (C->can_alias(adr_type, load_alias_idx)) {
6532 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6533 } else if (s->is_CFG() && s->is_Multi()) {
6534 // Look for the memory use of s (that is the use of its memory projection)
6535 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6536 Node* s1 = s->fast_out(i);
6537 assert(s1->is_Proj(), "projection expected");
6538 if (_igvn.type(s1) == Type::MEMORY) {
6539 for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
6540 Node* s2 = s1->fast_out(j);
6541 worklist.push(s2);
6542 }
6543 }
6544 }
6545 }
6546 }
6547 }
6548 }
6549 // For Phis only consider Region's inputs that were reached by following the memory edges
6550 if (LCA != early) {
6551 for (uint i = 0; i < worklist.size(); i++) {
6552 Node* s = worklist.at(i);
6553 if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
6554 Node* r = s->in(0);
6555 for (uint j = 1; j < s->req(); j++) {
6556 Node* in = s->in(j);
6557 Node* r_in = r->in(j);
6558 // We can't reach any node from a Phi because we don't enqueue Phi's uses above
6559 if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
6560 LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6561 }
6562 }
6563 }
6564 }
6565 }
6566 }
6567 return LCA;
6568 }
6569
6570 // Is CFG node 'dominator' dominating node 'n'?
6571 bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {
6572 if (dominator == n) {
6573 return true;
6574 }
6575 assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6576 uint dd = dom_depth(dominator);
6577 while (dom_depth(n) >= dd) {
6578 if (n == dominator) {
6579 return true;
6580 }
6581 n = idom(n);
6582 }
6583 return false;
6584 }
6585
6586 // Is CFG node 'dominator' strictly dominating node 'n'?
6587 bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {
6588 return dominator != n && is_dominator(dominator, n);
6589 }
6590
6591 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
6592 // Pair-wise LCA with tags.
6593 // Tag each index with the node 'tag' currently being processed
6594 // before advancing up the dominator chain using idom().
6595 // Later calls that find a match to 'tag' know that this path has already
6596 // been considered in the current LCA (which is input 'n1' by convention).
6597 // Since get_late_ctrl() is only called once for each node, the tag array
6598 // does not need to be cleared between calls to get_late_ctrl().
6599 // Algorithm trades a larger constant factor for better asymptotic behavior
6600 //
6601 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
6602 uint d1 = dom_depth(n1);
6603 uint d2 = dom_depth(n2);
6604 jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6605
6606 do {
6607 if (d1 > d2) {
6608 // current lca is deeper than n2
6609 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6610 n1 = idom(n1);
6611 d1 = dom_depth(n1);
6612 } else if (d1 < d2) {
6613 // n2 is deeper than current lca
6614 jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6615 if (memo == tag) {
6616 return n1; // Return the current LCA
6617 }
6618 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6619 n2 = idom(n2);
6620 d2 = dom_depth(n2);
6621 } else {
6622 // Here d1 == d2. Due to edits of the dominator-tree, sections
6623 // of the tree might have the same depth. These sections have
6624 // to be searched more carefully.
6625
6626 // Scan up all the n1's with equal depth, looking for n2.
6627 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6628 Node *t1 = idom(n1);
6629 while (dom_depth(t1) == d1) {
6630 if (t1 == n2) return n2;
6631 _dom_lca_tags.at_put_grow(t1->_idx, tag);
6632 t1 = idom(t1);
6633 }
6634 // Scan up all the n2's with equal depth, looking for n1.
6635 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6636 Node *t2 = idom(n2);
6637 while (dom_depth(t2) == d2) {
6638 if (t2 == n1) return n1;
6639 _dom_lca_tags.at_put_grow(t2->_idx, tag);
6640 t2 = idom(t2);
6641 }
6642 // Move up to a new dominator-depth value as well as up the dom-tree.
6643 n1 = t1;
6644 n2 = t2;
6645 d1 = dom_depth(n1);
6646 d2 = dom_depth(n2);
6647 }
6648 } while (n1 != n2);
6649 return n1;
6650 }
6651
6652 //------------------------------init_dom_lca_tags------------------------------
6653 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
6654 // Intended use does not involve any growth for the array, so it could
6655 // be of fixed size.
6656 void PhaseIdealLoop::init_dom_lca_tags() {
6657 uint limit = C->unique() + 1;
6658 _dom_lca_tags.at_grow(limit, 0);
6659 _dom_lca_tags_round = 0;
6660 #ifdef ASSERT
6661 for (uint i = 0; i < limit; ++i) {
6662 assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6663 }
6664 #endif // ASSERT
6665 }
6666
6667 //------------------------------build_loop_late--------------------------------
6668 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6669 // Second pass finds latest legal placement, and ideal loop placement.
6670 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6671 while (worklist.size() != 0) {
6672 Node *n = worklist.pop();
6673 // Only visit once
6674 if (visited.test_set(n->_idx)) continue;
6675 uint cnt = n->outcnt();
6676 uint i = 0;
6677 while (true) {
6678 assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6679 // Visit all children
6680 if (i < cnt) {
6681 Node* use = n->raw_out(i);
6682 ++i;
6683 // Check for dead uses. Aggressively prune such junk. It might be
6684 // dead in the global sense, but still have local uses so I cannot
6685 // easily call 'remove_dead_node'.
6686 if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?
6687 // Due to cycles, we might not hit the same fixed point in the verify
6688 // pass as we do in the regular pass. Instead, visit such phis as
6689 // simple uses of the loop head.
6690 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
6691 if( !visited.test(use->_idx) )
6692 worklist.push(use);
6693 } else if( !visited.test_set(use->_idx) ) {
6694 nstack.push(n, i); // Save parent and next use's index.
6695 n = use; // Process all children of current use.
6696 cnt = use->outcnt();
6697 i = 0;
6698 }
6699 } else {
6700 // Do not visit around the backedge of loops via data edges.
6701 // push dead code onto a worklist
6702 _deadlist.push(use);
6703 }
6704 } else {
6705 // All of n's children have been processed, complete post-processing.
6706 build_loop_late_post(n);
6707 if (C->failing()) { return; }
6708 if (nstack.is_empty()) {
6709 // Finished all nodes on stack.
6710 // Process next node on the worklist.
6711 break;
6712 }
6713 // Get saved parent node and next use's index. Visit the rest of uses.
6714 n = nstack.node();
6715 cnt = n->outcnt();
6716 i = nstack.index();
6717 nstack.pop();
6718 }
6719 }
6720 }
6721 }
6722
6723 // Verify that no data node is scheduled in the outer loop of a strip
6724 // mined loop.
6725 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
6726 #ifdef ASSERT
6727 if (get_loop(least)->_nest == 0) {
6728 return;
6729 }
6730 IdealLoopTree* loop = get_loop(least);
6731 Node* head = loop->_head;
6732 if (head->is_OuterStripMinedLoop() &&
6733 // Verification can't be applied to fully built strip mined loops
6734 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
6735 Node* sfpt = head->as_Loop()->outer_safepoint();
6736 ResourceMark rm;
6737 Unique_Node_List wq;
6738 wq.push(sfpt);
6739 for (uint i = 0; i < wq.size(); i++) {
6740 Node *m = wq.at(i);
6741 for (uint i = 1; i < m->req(); i++) {
6742 Node* nn = m->in(i);
6743 if (nn == n) {
6744 return;
6745 }
6746 if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
6747 wq.push(nn);
6748 }
6749 }
6750 }
6751 ShouldNotReachHere();
6752 }
6753 #endif
6754 }
6755
6756
6757 //------------------------------build_loop_late_post---------------------------
6758 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6759 // Second pass finds latest legal placement, and ideal loop placement.
6760 void PhaseIdealLoop::build_loop_late_post(Node *n) {
6761 build_loop_late_post_work(n, true);
6762 }
6763
6764 // Class to visit all predicates in a predicate chain to find out which are dominated by a given node. Keeps track of
6765 // the entry to the earliest predicate that is still dominated by the given dominator. This class is used when trying to
6766 // legally skip all predicates when figuring out the latest placement such that a node does not interfere with Loop
6767 // Predication or creating a Loop Limit Check Predicate later.
6768 class DominatedPredicates : public UnifiedPredicateVisitor {
6769 Node* const _dominator;
6770 Node* _earliest_dominated_predicate_entry;
6771 bool _should_continue;
6772 PhaseIdealLoop* const _phase;
6773
6774 public:
6775 DominatedPredicates(Node* dominator, Node* start_node, PhaseIdealLoop* phase)
6776 : _dominator(dominator),
6777 _earliest_dominated_predicate_entry(start_node),
6778 _should_continue(true),
6779 _phase(phase) {}
6780 NONCOPYABLE(DominatedPredicates);
6781
6782 bool should_continue() const override {
6783 return _should_continue;
6784 }
6785
6786 // Returns the entry to the earliest predicate that is still dominated by the given dominator (all could be dominated).
6787 Node* earliest_dominated_predicate_entry() const {
6788 return _earliest_dominated_predicate_entry;
6789 }
6790
6791 void visit_predicate(const Predicate& predicate) override {
6792 Node* entry = predicate.entry();
6793 if (_phase->is_strict_dominator(entry, _dominator)) {
6794 _should_continue = false;
6795 } else {
6796 _earliest_dominated_predicate_entry = entry;
6797 }
6798 }
6799 };
6800
6801 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
6802
6803 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
6804 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6805 }
6806
6807 #ifdef ASSERT
6808 if (_verify_only && !n->is_CFG()) {
6809 // Check def-use domination.
6810 // We would like to expose this check in product but it appears to be expensive.
6811 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
6812 }
6813 #endif
6814
6815 // CFG and pinned nodes already handled
6816 if( n->in(0) ) {
6817 if( n->in(0)->is_top() ) return; // Dead?
6818
6819 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
6820 // _must_ be pinned (they have to observe their control edge of course).
6821 // Unlike Stores (which modify an unallocable resource, the memory
6822 // state), Mods/Loads can float around. So free them up.
6823 switch( n->Opcode() ) {
6824 case Op_DivI:
6825 case Op_DivF:
6826 case Op_DivD:
6827 case Op_ModI:
6828 case Op_LoadB: // Same with Loads; they can sink
6829 case Op_LoadUB: // during loop optimizations.
6830 case Op_LoadUS:
6831 case Op_LoadD:
6832 case Op_LoadF:
6833 case Op_LoadI:
6834 case Op_LoadKlass:
6835 case Op_LoadNKlass:
6836 case Op_LoadL:
6837 case Op_LoadS:
6838 case Op_LoadP:
6839 case Op_LoadN:
6840 case Op_LoadRange:
6841 case Op_LoadD_unaligned:
6842 case Op_LoadL_unaligned:
6843 case Op_StrComp: // Does a bunch of load-like effects
6844 case Op_StrEquals:
6845 case Op_StrIndexOf:
6846 case Op_StrIndexOfChar:
6847 case Op_AryEq:
6848 case Op_VectorizedHashCode:
6849 case Op_CountPositives:
6850 pinned = false;
6851 }
6852 if (n->is_CMove() || n->is_ConstraintCast()) {
6853 pinned = false;
6854 }
6855 if( pinned ) {
6856 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6857 if( !chosen_loop->_child ) // Inner loop?
6858 chosen_loop->_body.push(n); // Collect inner loops
6859 return;
6860 }
6861 } else { // No slot zero
6862 if( n->is_CFG() ) { // CFG with no slot 0 is dead
6863 _loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6864 return;
6865 }
6866 assert(!n->is_CFG() || n->outcnt() == 0, "");
6867 }
6868
6869 // Do I have a "safe range" I can select over?
6870 Node *early = get_ctrl(n);// Early location already computed
6871
6872 // Compute latest point this Node can go
6873 Node *LCA = get_late_ctrl( n, early );
6874 // LCA is null due to uses being dead
6875 if( LCA == nullptr ) {
6876 #ifdef ASSERT
6877 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
6878 assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6879 }
6880 #endif
6881 _loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6882 _deadlist.push(n);
6883 return;
6884 }
6885 assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6886
6887 Node *legal = LCA; // Walk 'legal' up the IDOM chain
6888 Node *least = legal; // Best legal position so far
6889 while( early != legal ) { // While not at earliest legal
6890 if (legal->is_Start() && !early->is_Root()) {
6891 #ifdef ASSERT
6892 // Bad graph. Print idom path and fail.
6893 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
6894 assert(false, "Bad graph detected in build_loop_late");
6895 #endif
6896 C->record_method_not_compilable("Bad graph detected in build_loop_late");
6897 return;
6898 }
6899 // Find least loop nesting depth
6900 legal = idom(legal); // Bump up the IDOM tree
6901 // Check for lower nesting depth
6902 if( get_loop(legal)->_nest < get_loop(least)->_nest )
6903 least = legal;
6904 }
6905 assert(early == legal || legal != C->root(), "bad dominance of inputs");
6906
6907 if (least != early) {
6908 // Move the node above predicates as far up as possible so a
6909 // following pass of Loop Predication doesn't hoist a predicate
6910 // that depends on it above that node.
6911 const PredicateIterator predicate_iterator(least);
6912 DominatedPredicates dominated_predicates(early, least, this);
6913 predicate_iterator.for_each(dominated_predicates);
6914 least = dominated_predicates.earliest_dominated_predicate_entry();
6915 }
6916 // Try not to place code on a loop entry projection
6917 // which can inhibit range check elimination.
6918 if (least != early) {
6919 Node* ctrl_out = least->unique_ctrl_out_or_null();
6920 if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6921 least == ctrl_out->in(LoopNode::EntryControl) &&
6922 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
6923 Node* least_dom = idom(least);
6924 if (get_loop(least_dom)->is_member(get_loop(least))) {
6925 least = least_dom;
6926 }
6927 }
6928 }
6929 // Don't extend live ranges of raw oops
6930 if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6931 !n->bottom_type()->isa_rawptr()) {
6932 least = early;
6933 }
6934
6935 #ifdef ASSERT
6936 // Broken part of VerifyLoopOptimizations (F)
6937 // Reason:
6938 // _verify_me->get_ctrl_no_update(n) seems to return wrong result
6939 /*
6940 // If verifying, verify that 'verify_me' has a legal location
6941 // and choose it as our location.
6942 if( _verify_me ) {
6943 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6944 Node *legal = LCA;
6945 while( early != legal ) { // While not at earliest legal
6946 if( legal == v_ctrl ) break; // Check for prior good location
6947 legal = idom(legal) ;// Bump up the IDOM tree
6948 }
6949 // Check for prior good location
6950 if( legal == v_ctrl ) least = legal; // Keep prior if found
6951 }
6952 */
6953 #endif
6954
6955 // Assign discovered "here or above" point
6956 least = find_non_split_ctrl(least);
6957 verify_strip_mined_scheduling(n, least);
6958 set_ctrl(n, least);
6959
6960 // Collect inner loop bodies
6961 IdealLoopTree *chosen_loop = get_loop(least);
6962 if( !chosen_loop->_child ) // Inner loop?
6963 chosen_loop->_body.push(n);// Collect inner loops
6964
6965 if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {
6966 _zero_trip_guard_opaque_nodes.push(n);
6967 }
6968
6969 if (!_verify_only && n->Opcode() == Op_OpaqueMultiversioning) {
6970 _multiversion_opaque_nodes.push(n);
6971 }
6972 }
6973
6974 #ifdef ASSERT
6975 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
6976 tty->print_cr("%s", msg);
6977 tty->print("n: "); n->dump();
6978 tty->print("early(n): "); early->dump();
6979 if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6980 n->in(0) != early && !n->in(0)->is_Root()) {
6981 tty->print("n->in(0): "); n->in(0)->dump();
6982 }
6983 for (uint i = 1; i < n->req(); i++) {
6984 Node* in1 = n->in(i);
6985 if (in1 != nullptr && in1 != n && !in1->is_top()) {
6986 tty->print("n->in(%d): ", i); in1->dump();
6987 Node* in1_early = get_ctrl(in1);
6988 tty->print("early(n->in(%d)): ", i); in1_early->dump();
6989 if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6990 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
6991 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
6992 }
6993 for (uint j = 1; j < in1->req(); j++) {
6994 Node* in2 = in1->in(j);
6995 if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {
6996 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
6997 Node* in2_early = get_ctrl(in2);
6998 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
6999 if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
7000 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
7001 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
7002 }
7003 }
7004 }
7005 }
7006 }
7007 tty->cr();
7008 tty->print("LCA(n): "); LCA->dump();
7009 for (uint i = 0; i < n->outcnt(); i++) {
7010 Node* u1 = n->raw_out(i);
7011 if (u1 == n)
7012 continue;
7013 tty->print("n->out(%d): ", i); u1->dump();
7014 if (u1->is_CFG()) {
7015 for (uint j = 0; j < u1->outcnt(); j++) {
7016 Node* u2 = u1->raw_out(j);
7017 if (u2 != u1 && u2 != n && u2->is_CFG()) {
7018 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
7019 }
7020 }
7021 } else {
7022 Node* u1_later = get_ctrl(u1);
7023 tty->print("later(n->out(%d)): ", i); u1_later->dump();
7024 if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
7025 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
7026 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
7027 }
7028 for (uint j = 0; j < u1->outcnt(); j++) {
7029 Node* u2 = u1->raw_out(j);
7030 if (u2 == n || u2 == u1)
7031 continue;
7032 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
7033 if (!u2->is_CFG()) {
7034 Node* u2_later = get_ctrl(u2);
7035 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
7036 if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
7037 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
7038 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
7039 }
7040 }
7041 }
7042 }
7043 }
7044 dump_idoms(early, LCA);
7045 tty->cr();
7046 }
7047
7048 // Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
7049 class RealLCA {
7050 const PhaseIdealLoop* _phase;
7051 Node* _early;
7052 Node* _wrong_lca;
7053 uint _early_index;
7054 int _wrong_lca_index;
7055
7056 // Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
7057 // is different: Return the previously visited node which must be the real LCA.
7058 // The node lists also contain _early and _wrong_lca, respectively.
7059 Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {
7060 int early_index = early_with_idoms.size() - 1;
7061 int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
7062 bool found_difference = false;
7063 do {
7064 if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {
7065 // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
7066 found_difference = true;
7067 break;
7068 }
7069 early_index--;
7070 wrong_lca_index--;
7071 } while (wrong_lca_index >= 0);
7072
7073 assert(early_index >= 0, "must always find an LCA - cannot be early");
7074 _early_index = early_index;
7075 _wrong_lca_index = wrong_lca_index;
7076 Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
7077 assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
7078 return real_lca;
7079 }
7080
7081 void dump(Node* real_lca) {
7082 tty->cr();
7083 tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
7084 _phase->dump_idom(_early, _early_index + 1);
7085
7086 tty->cr();
7087 tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());
7088 _phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
7089
7090 tty->cr();
7091 tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",
7092 _early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
7093 if (_wrong_lca_index >= 0) {
7094 tty->print(" (idom[%d])", _wrong_lca_index);
7095 }
7096 tty->print_cr(":");
7097 real_lca->dump();
7098 }
7099
7100 public:
7101 RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
7102 : _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {
7103 assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
7104 }
7105
7106 void compute_and_dump() {
7107 ResourceMark rm;
7108 Unique_Node_List early_with_idoms;
7109 Unique_Node_List wrong_lca_with_idoms;
7110 early_with_idoms.push(_early);
7111 wrong_lca_with_idoms.push(_wrong_lca);
7112 _phase->get_idoms(_early, 10000, early_with_idoms);
7113 _phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
7114 Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
7115 dump(real_lca);
7116 }
7117 };
7118
7119 // Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
7120 void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
7121 assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
7122 assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
7123
7124 RealLCA real_lca(this, early, wrong_lca);
7125 real_lca.compute_and_dump();
7126 }
7127 #endif // ASSERT
7128
7129 #ifndef PRODUCT
7130 //------------------------------dump-------------------------------------------
7131 void PhaseIdealLoop::dump() const {
7132 ResourceMark rm;
7133 Node_Stack stack(C->live_nodes() >> 2);
7134 Node_List rpo_list;
7135 VectorSet visited;
7136 visited.set(C->top()->_idx);
7137 rpo(C->root(), stack, visited, rpo_list);
7138 // Dump root loop indexed by last element in PO order
7139 dump(_ltree_root, rpo_list.size(), rpo_list);
7140 }
7141
7142 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
7143 loop->dump_head();
7144
7145 // Now scan for CFG nodes in the same loop
7146 for (uint j = idx; j > 0; j--) {
7147 Node* n = rpo_list[j-1];
7148 if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
7149 continue;
7150
7151 if (get_loop(n) != loop) { // Wrong loop nest
7152 if (get_loop(n)->_head == n && // Found nested loop?
7153 get_loop(n)->_parent == loop)
7154 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
7155 continue;
7156 }
7157
7158 // Dump controlling node
7159 tty->sp(2 * loop->_nest);
7160 tty->print("C");
7161 if (n == C->root()) {
7162 n->dump();
7163 } else {
7164 Node* cached_idom = idom_no_update(n);
7165 Node* computed_idom = n->in(0);
7166 if (n->is_Region()) {
7167 computed_idom = compute_idom(n);
7168 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
7169 // any MultiBranch ctrl node), so apply a similar transform to
7170 // the cached idom returned from idom_no_update.
7171 cached_idom = find_non_split_ctrl(cached_idom);
7172 }
7173 tty->print(" ID:%d", computed_idom->_idx);
7174 n->dump();
7175 if (cached_idom != computed_idom) {
7176 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
7177 computed_idom->_idx, cached_idom->_idx);
7178 }
7179 }
7180 // Dump nodes it controls
7181 for (uint k = 0; k < _loop_or_ctrl.max(); k++) {
7182 // (k < C->unique() && get_ctrl(find(k)) == n)
7183 if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {
7184 Node* m = C->root()->find(k);
7185 if (m && m->outcnt() > 0) {
7186 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
7187 tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",
7188 _loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
7189 }
7190 tty->sp(2 * loop->_nest + 1);
7191 m->dump();
7192 }
7193 }
7194 }
7195 }
7196 }
7197
7198 void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {
7199 if (has_ctrl(n)) {
7200 tty->print_cr("No idom for data nodes");
7201 } else {
7202 ResourceMark rm;
7203 Unique_Node_List idoms;
7204 get_idoms(n, count, idoms);
7205 dump_idoms_in_reverse(n, idoms);
7206 }
7207 }
7208
7209 void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
7210 Node* next = n;
7211 for (uint i = 0; !next->is_Start() && i < count; i++) {
7212 next = idom(next);
7213 assert(!idoms.member(next), "duplicated idom is not possible");
7214 idoms.push(next);
7215 }
7216 }
7217
7218 void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
7219 Node* next;
7220 uint padding = 3;
7221 uint node_index_padding_width = (C->unique() == 0 ? 0 : static_cast<int>(log10(static_cast<double>(C->unique())))) + 1;
7222 for (int i = idom_list.size() - 1; i >= 0; i--) {
7223 if (i == 9 || i == 99) {
7224 padding++;
7225 }
7226 next = idom_list[i];
7227 tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
7228 }
7229 tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
7230 }
7231 #endif // NOT PRODUCT
7232
7233 // Collect a R-P-O for the whole CFG.
7234 // Result list is in post-order (scan backwards for RPO)
7235 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
7236 stk.push(start, 0);
7237 visited.set(start->_idx);
7238
7239 while (stk.is_nonempty()) {
7240 Node* m = stk.node();
7241 uint idx = stk.index();
7242 if (idx < m->outcnt()) {
7243 stk.set_index(idx + 1);
7244 Node* n = m->raw_out(idx);
7245 if (n->is_CFG() && !visited.test_set(n->_idx)) {
7246 stk.push(n, 0);
7247 }
7248 } else {
7249 rpo_list.push(m);
7250 stk.pop();
7251 }
7252 }
7253 }
7254
7255 ConINode* PhaseIdealLoop::intcon(jint i) {
7256 ConINode* node = _igvn.intcon(i);
7257 set_root_as_ctrl(node);
7258 return node;
7259 }
7260
7261 ConLNode* PhaseIdealLoop::longcon(jlong i) {
7262 ConLNode* node = _igvn.longcon(i);
7263 set_root_as_ctrl(node);
7264 return node;
7265 }
7266
7267 ConNode* PhaseIdealLoop::makecon(const Type* t) {
7268 ConNode* node = _igvn.makecon(t);
7269 set_root_as_ctrl(node);
7270 return node;
7271 }
7272
7273 ConNode* PhaseIdealLoop::integercon(jlong l, BasicType bt) {
7274 ConNode* node = _igvn.integercon(l, bt);
7275 set_root_as_ctrl(node);
7276 return node;
7277 }
7278
7279 ConNode* PhaseIdealLoop::zerocon(BasicType bt) {
7280 ConNode* node = _igvn.zerocon(bt);
7281 set_root_as_ctrl(node);
7282 return node;
7283 }
7284
7285
7286 //=============================================================================
7287 //------------------------------LoopTreeIterator-------------------------------
7288
7289 // Advance to next loop tree using a preorder, left-to-right traversal.
7290 void LoopTreeIterator::next() {
7291 assert(!done(), "must not be done.");
7292 if (_curnt->_child != nullptr) {
7293 _curnt = _curnt->_child;
7294 } else if (_curnt->_next != nullptr) {
7295 _curnt = _curnt->_next;
7296 } else {
7297 while (_curnt != _root && _curnt->_next == nullptr) {
7298 _curnt = _curnt->_parent;
7299 }
7300 if (_curnt == _root) {
7301 _curnt = nullptr;
7302 assert(done(), "must be done.");
7303 } else {
7304 assert(_curnt->_next != nullptr, "must be more to do");
7305 _curnt = _curnt->_next;
7306 }
7307 }
7308 }