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() ||
225 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
226 assert(idom(ctl) == parent_ctl, "strange");
227 next = idom(parent_ctl);
228 }
229 } else {
230 next = idom(ctl);
231 }
232 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
233 break;
234 }
235 ctl = next;
236 }
237
238 if (ctl != n->in(0)) {
239 _igvn.replace_input_of(n, 0, ctl);
240 _igvn.hash_insert(n);
241 }
242
243 return ctl;
244 }
245
246
247 //------------------------------set_early_ctrl---------------------------------
248 // Set earliest legal control
249 void PhaseIdealLoop::set_early_ctrl(Node* n, bool update_body) {
250 Node *early = get_early_ctrl(n);
251
252 // Record earliest legal location
253 set_ctrl(n, early);
254 IdealLoopTree *loop = get_loop(early);
255 if (update_body && loop->_child == nullptr) {
256 loop->_body.push(n);
257 }
258 }
259
260 //------------------------------set_subtree_ctrl-------------------------------
261 // set missing _ctrl entries on new nodes
262 void PhaseIdealLoop::set_subtree_ctrl(Node* n, bool update_body) {
263 // Already set? Get out.
264 if (_loop_or_ctrl[n->_idx]) return;
265 // Recursively set _loop_or_ctrl array to indicate where the Node goes
266 uint i;
267 for (i = 0; i < n->req(); ++i) {
268 Node *m = n->in(i);
269 if (m && m != C->root()) {
270 set_subtree_ctrl(m, update_body);
271 }
272 }
273
274 // Fixup self
275 set_early_ctrl(n, update_body);
276 }
277
278 IdealLoopTree* PhaseIdealLoop::insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift) {
279 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
280 IdealLoopTree* parent = loop->_parent;
281 IdealLoopTree* sibling = parent->_child;
282 if (sibling == loop) {
283 parent->_child = outer_ilt;
284 } else {
285 while (sibling->_next != loop) {
286 sibling = sibling->_next;
287 }
288 sibling->_next = outer_ilt;
289 }
290 outer_ilt->_next = loop->_next;
291 outer_ilt->_parent = parent;
292 outer_ilt->_child = loop;
293 outer_ilt->_nest = loop->_nest;
294 loop->_parent = outer_ilt;
295 loop->_next = nullptr;
296 loop->_nest++;
297 assert(loop->_nest <= SHRT_MAX, "sanity");
298 return outer_ilt;
299 }
300
301 // Create a skeleton strip mined outer loop: an OuterStripMinedLoop head before the inner strip mined CountedLoop, a
302 // SafePoint on exit of the inner CountedLoopEnd and an OuterStripMinedLoopEnd test that can't constant fold until loop
303 // optimizations are over. The inner strip mined loop is left as it is. Only once loop optimizations are over, do we
304 // adjust the inner loop exit condition to limit its number of iterations, set the outer loop exit condition and add
305 // Phis to the outer loop head. Some loop optimizations that operate on the inner strip mined loop need to be aware of
306 // the outer strip mined loop: loop unswitching needs to clone the outer loop as well as the inner, unrolling needs to
307 // only clone the inner loop etc. No optimizations need to change the outer strip mined loop as it is only a skeleton.
308 //
309 // Schematically:
310 //
311 // OuterStripMinedLoop -------|
312 // | |
313 // CountedLoop ----------- | |
314 // \- Phi (iv) -| | |
315 // / \ | | |
316 // CmpI AddI --| | |
317 // \ | |
318 // Bool | |
319 // \ | |
320 // CountedLoopEnd | |
321 // / \ | |
322 // IfFalse IfTrue--------| |
323 // | |
324 // SafePoint |
325 // | |
326 // OuterStripMinedLoopEnd |
327 // / \ |
328 // IfFalse IfTrue-----------|
329 // |
330 //
331 //
332 // As loop optimizations transform the inner loop, the outer strip mined loop stays mostly unchanged. The only exception
333 // is nodes referenced from the SafePoint and sunk from the inner loop: they end up in the outer strip mined loop.
334 //
335 // Not adding Phis to the outer loop head from the beginning, and only adding them after loop optimizations does not
336 // conform to C2's IR rules: any variable or memory slice that is mutated in a loop should have a Phi. The main
337 // motivation for such a design that doesn't conform to C2's IR rules is to allow existing loop optimizations to be
338 // mostly unaffected by the outer strip mined loop: the only extra step needed in most cases is to step over the
339 // OuterStripMinedLoop. The main drawback is that once loop optimizations are over, an extra step is needed to finish
340 // constructing the outer loop. This is handled by OuterStripMinedLoopNode::adjust_strip_mined_loop().
341 //
342 // Adding Phis to the outer loop is largely straightforward: there needs to be one Phi in the outer loop for every Phi
343 // in the inner loop. Things may be more complicated for sunk Store nodes: there may not be any inner loop Phi left
344 // after sinking for a particular memory slice but the outer loop needs a Phi. See
345 // OuterStripMinedLoopNode::handle_sunk_stores_when_finishing_construction()
346 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(Node* init_control,
347 IdealLoopTree* loop, float cl_prob, float le_fcnt,
348 Node*& entry_control, Node*& iffalse) {
349 Node* outer_test = intcon(0);
350 Node *orig = iffalse;
351 iffalse = iffalse->clone();
352 _igvn.register_new_node_with_optimizer(iffalse);
353 set_idom(iffalse, idom(orig), dom_depth(orig));
354
355 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
356 Node *outer_ift = new IfTrueNode (outer_le);
357 Node* outer_iff = orig;
358 _igvn.replace_input_of(outer_iff, 0, outer_le);
359
360 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
361 entry_control = outer_l;
362
363 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_l, outer_ift);
364
365 set_loop(iffalse, outer_ilt);
366 // When this code runs, loop bodies have not yet been populated.
367 const bool body_populated = false;
368 register_control(outer_le, outer_ilt, iffalse, body_populated);
369 register_control(outer_ift, outer_ilt, outer_le, body_populated);
370 set_idom(outer_iff, outer_le, dom_depth(outer_le));
371 _igvn.register_new_node_with_optimizer(outer_l);
372 set_loop(outer_l, outer_ilt);
373 set_idom(outer_l, init_control, dom_depth(init_control)+1);
374
375 return outer_ilt;
376 }
377
378 void PhaseIdealLoop::insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj,
379 Node* cmp_limit, Node* bol) {
380 assert(loop_limit_check_parse_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
381 Node* new_predicate_proj = create_new_if_for_predicate(loop_limit_check_parse_proj, nullptr,
382 Deoptimization::Reason_loop_limit_check,
383 Op_If);
384 Node* iff = new_predicate_proj->in(0);
385 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
386 bol = _igvn.register_new_node_with_optimizer(bol);
387 set_subtree_ctrl(bol, false);
388 _igvn.replace_input_of(iff, 1, bol);
389
390 #ifndef PRODUCT
391 // report that the loop predication has been actually performed
392 // for this loop
393 if (TraceLoopLimitCheck) {
394 tty->print_cr("Counted Loop Limit Check generated:");
395 DEBUG_ONLY( bol->dump(2); )
396 }
397 #endif
398 }
399
400 Node* PhaseIdealLoop::loop_exit_control(Node* x, IdealLoopTree* loop) {
401 // Counted loop head must be a good RegionNode with only 3 not null
402 // control input edges: Self, Entry, LoopBack.
403 if (x->in(LoopNode::Self) == nullptr || x->req() != 3 || loop->_irreducible) {
404 return nullptr;
405 }
406 Node *init_control = x->in(LoopNode::EntryControl);
407 Node *back_control = x->in(LoopNode::LoopBackControl);
408 if (init_control == nullptr || back_control == nullptr) { // Partially dead
409 return nullptr;
410 }
411 // Must also check for TOP when looking for a dead loop
412 if (init_control->is_top() || back_control->is_top()) {
413 return nullptr;
414 }
415
416 // Allow funny placement of Safepoint
417 if (back_control->Opcode() == Op_SafePoint) {
418 back_control = back_control->in(TypeFunc::Control);
419 }
420
421 // Controlling test for loop
422 Node *iftrue = back_control;
423 uint iftrue_op = iftrue->Opcode();
424 if (iftrue_op != Op_IfTrue &&
425 iftrue_op != Op_IfFalse) {
426 // I have a weird back-control. Probably the loop-exit test is in
427 // the middle of the loop and I am looking at some trailing control-flow
428 // merge point. To fix this I would have to partially peel the loop.
429 return nullptr; // Obscure back-control
430 }
431
432 // Get boolean guarding loop-back test
433 Node *iff = iftrue->in(0);
434 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) {
435 return nullptr;
436 }
437 return iftrue;
438 }
439
440 Node* PhaseIdealLoop::loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob) {
441 Node* iftrue = back_control;
442 uint iftrue_op = iftrue->Opcode();
443 Node* iff = iftrue->in(0);
444 BoolNode* test = iff->in(1)->as_Bool();
445 bt = test->_test._test;
446 cl_prob = iff->as_If()->_prob;
447 if (iftrue_op == Op_IfFalse) {
448 bt = BoolTest(bt).negate();
449 cl_prob = 1.0 - cl_prob;
450 }
451 // Get backedge compare
452 Node* cmp = test->in(1);
453 if (!cmp->is_Cmp()) {
454 return nullptr;
455 }
456
457 // Find the trip-counter increment & limit. Limit must be loop invariant.
458 incr = cmp->in(1);
459 limit = cmp->in(2);
460
461 // ---------
462 // need 'loop()' test to tell if limit is loop invariant
463 // ---------
464
465 if (!ctrl_is_member(loop, incr)) { // Swapped trip counter and limit?
466 Node* tmp = incr; // Then reverse order into the CmpI
467 incr = limit;
468 limit = tmp;
469 bt = BoolTest(bt).commute(); // And commute the exit test
470 }
471 if (ctrl_is_member(loop, limit)) { // Limit must be loop-invariant
472 return nullptr;
473 }
474 if (!ctrl_is_member(loop, incr)) { // Trip counter must be loop-variant
475 return nullptr;
476 }
477 return cmp;
478 }
479
480 Node* PhaseIdealLoop::loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr) {
481 if (incr->is_Phi()) {
482 if (incr->as_Phi()->region() != x || incr->req() != 3) {
483 return nullptr; // Not simple trip counter expression
484 }
485 phi_incr = incr;
486 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
487 if (!ctrl_is_member(loop, incr)) { // Trip counter must be loop-variant
488 return nullptr;
489 }
490 }
491 return incr;
492 }
493
494 Node* PhaseIdealLoop::loop_iv_stride(Node* incr, Node*& xphi) {
495 assert(incr->Opcode() == Op_AddI || incr->Opcode() == Op_AddL, "caller resp.");
496 // Get merge point
497 xphi = incr->in(1);
498 Node *stride = incr->in(2);
499 if (!stride->is_Con()) { // Oops, swap these
500 if (!xphi->is_Con()) { // Is the other guy a constant?
501 return nullptr; // Nope, unknown stride, bail out
502 }
503 Node *tmp = xphi; // 'incr' is commutative, so ok to swap
504 xphi = stride;
505 stride = tmp;
506 }
507 return stride;
508 }
509
510 PhiNode* PhaseIdealLoop::loop_iv_phi(Node* xphi, Node* phi_incr, Node* x) {
511 if (!xphi->is_Phi()) {
512 return nullptr; // Too much math on the trip counter
513 }
514 if (phi_incr != nullptr && phi_incr != xphi) {
515 return nullptr;
516 }
517 PhiNode *phi = xphi->as_Phi();
518
519 // Phi must be of loop header; backedge must wrap to increment
520 if (phi->region() != x) {
521 return nullptr;
522 }
523 return phi;
524 }
525
526 static int check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt) {
527 if (final_correction > 0) {
528 if (limit_t->lo_as_long() > (max_signed_integer(bt) - final_correction)) {
529 return -1;
530 }
531 if (limit_t->hi_as_long() > (max_signed_integer(bt) - final_correction)) {
532 return 1;
533 }
534 } else {
535 if (limit_t->hi_as_long() < (min_signed_integer(bt) - final_correction)) {
536 return -1;
537 }
538 if (limit_t->lo_as_long() < (min_signed_integer(bt) - final_correction)) {
539 return 1;
540 }
541 }
542 return 0;
543 }
544
545 static bool condition_stride_ok(BoolTest::mask bt, jlong stride_con) {
546 // If the condition is inverted and we will be rolling
547 // through MININT to MAXINT, then bail out.
548 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
549 // Odd stride
550 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
551 // Count down loop rolls through MAXINT
552 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
553 // Count up loop rolls through MININT
554 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
555 return false; // Bail out
556 }
557 return true;
558 }
559
560 Node* PhaseIdealLoop::loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head,
561 BasicType bt) {
562 Node* iv_as_long;
563 if (bt == T_LONG) {
564 iv_as_long = new ConvI2LNode(inner_iv, TypeLong::INT);
565 register_new_node(iv_as_long, inner_head);
566 } else {
567 iv_as_long = inner_iv;
568 }
569 Node* iv_replacement = AddNode::make(outer_phi, iv_as_long, bt);
570 register_new_node(iv_replacement, inner_head);
571 for (DUIterator_Last imin, i = iv_to_replace->last_outs(imin); i >= imin;) {
572 Node* u = iv_to_replace->last_out(i);
573 #ifdef ASSERT
574 if (!is_dominator(inner_head, ctrl_or_self(u))) {
575 assert(u->is_Phi(), "should be a Phi");
576 for (uint j = 1; j < u->req(); j++) {
577 if (u->in(j) == iv_to_replace) {
578 assert(is_dominator(inner_head, u->in(0)->in(j)), "iv use above loop?");
579 }
580 }
581 }
582 #endif
583 _igvn.rehash_node_delayed(u);
584 int nb = u->replace_edge(iv_to_replace, iv_replacement, &_igvn);
585 i -= nb;
586 }
587 return iv_replacement;
588 }
589
590 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
591 void PhaseIdealLoop::add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop,
592 SafePointNode* sfpt) {
593 if (!C->too_many_traps(sfpt->jvms()->method(), sfpt->jvms()->bci(), reason)) {
594 ParsePredicateNode* parse_predicate = new ParsePredicateNode(inner_head->in(LoopNode::EntryControl), reason, &_igvn);
595 register_control(parse_predicate, loop, inner_head->in(LoopNode::EntryControl));
596 Node* if_false = new IfFalseNode(parse_predicate);
597 register_control(if_false, _ltree_root, parse_predicate);
598 Node* if_true = new IfTrueNode(parse_predicate);
599 register_control(if_true, loop, parse_predicate);
600
601 int trap_request = Deoptimization::make_trap_request(reason, Deoptimization::Action_maybe_recompile);
602 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
603 const TypePtr* no_memory_effects = nullptr;
604 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
605 no_memory_effects);
606
607 Node* mem = nullptr;
608 Node* i_o = nullptr;
609 if (sfpt->is_Call()) {
610 mem = sfpt->proj_out(TypeFunc::Memory);
611 i_o = sfpt->proj_out(TypeFunc::I_O);
612 } else {
613 mem = sfpt->memory();
614 i_o = sfpt->i_o();
615 }
616
617 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
618 register_new_node(frame, C->start());
619 Node *ret = new ParmNode(C->start(), TypeFunc::ReturnAdr);
620 register_new_node(ret, C->start());
621
622 unc->init_req(TypeFunc::Control, if_false);
623 unc->init_req(TypeFunc::I_O, i_o);
624 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
625 unc->init_req(TypeFunc::FramePtr, frame);
626 unc->init_req(TypeFunc::ReturnAdr, ret);
627 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
628 unc->set_cnt(PROB_UNLIKELY_MAG(4));
629 unc->copy_call_debug_info(&_igvn, sfpt);
630
631 for (uint i = TypeFunc::Parms; i < unc->req(); i++) {
632 set_subtree_ctrl(unc->in(i), false);
633 }
634 register_control(unc, _ltree_root, if_false);
635
636 Node* ctrl = new ProjNode(unc, TypeFunc::Control);
637 register_control(ctrl, _ltree_root, unc);
638 Node* halt = new HaltNode(ctrl, frame, "uncommon trap returned which should never happen" PRODUCT_ONLY(COMMA /*reachable*/false));
639 register_control(halt, _ltree_root, ctrl);
640 _igvn.add_input_to(C->root(), halt);
641
642 _igvn.replace_input_of(inner_head, LoopNode::EntryControl, if_true);
643 set_idom(inner_head, if_true, dom_depth(inner_head));
644 }
645 }
646
647 // Find a safepoint node that dominates the back edge. We need a
648 // SafePointNode so we can use its jvm state to create empty
649 // predicates.
650 static bool no_side_effect_since_safepoint(Compile* C, Node* x, Node* mem, MergeMemNode* mm, PhaseIdealLoop* phase) {
651 SafePointNode* safepoint = nullptr;
652 for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {
653 Node* u = x->fast_out(i);
654 if (u->is_memory_phi()) {
655 Node* m = u->in(LoopNode::LoopBackControl);
656 if (u->adr_type() == TypePtr::BOTTOM) {
657 if (m->is_MergeMem() && mem->is_MergeMem()) {
658 if (m != mem DEBUG_ONLY(|| true)) {
659 // MergeMemStream can modify m, for example to adjust the length to mem.
660 // This is unfortunate, and probably unnecessary. But as it is, we need
661 // to add m to the igvn worklist, else we may have a modified node that
662 // is not on the igvn worklist.
663 phase->igvn()._worklist.push(m);
664 for (MergeMemStream mms(m->as_MergeMem(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
665 if (!mms.is_empty()) {
666 if (mms.memory() != mms.memory2()) {
667 return false;
668 }
669 #ifdef ASSERT
670 if (mms.alias_idx() != Compile::AliasIdxBot) {
671 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
672 }
673 #endif
674 }
675 }
676 }
677 } else if (mem->is_MergeMem()) {
678 if (m != mem->as_MergeMem()->base_memory()) {
679 return false;
680 }
681 } else {
682 return false;
683 }
684 } else {
685 if (mem->is_MergeMem()) {
686 if (m != mem->as_MergeMem()->memory_at(C->get_alias_index(u->adr_type()))) {
687 return false;
688 }
689 #ifdef ASSERT
690 mm->set_memory_at(C->get_alias_index(u->adr_type()), mem->as_MergeMem()->base_memory());
691 #endif
692 } else {
693 if (m != mem) {
694 return false;
695 }
696 }
697 }
698 }
699 }
700 return true;
701 }
702
703 SafePointNode* PhaseIdealLoop::find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop) {
704 IfNode* exit_test = back_control->in(0)->as_If();
705 SafePointNode* safepoint = nullptr;
706 if (exit_test->in(0)->is_SafePoint() && exit_test->in(0)->outcnt() == 1) {
707 safepoint = exit_test->in(0)->as_SafePoint();
708 } else {
709 Node* c = back_control;
710 while (c != x && c->Opcode() != Op_SafePoint) {
711 c = idom(c);
712 }
713
714 if (c->Opcode() == Op_SafePoint) {
715 safepoint = c->as_SafePoint();
716 }
717
718 if (safepoint == nullptr) {
719 return nullptr;
720 }
721
722 Node* mem = safepoint->in(TypeFunc::Memory);
723
724 // We can only use that safepoint if there's no side effect between the backedge and the safepoint.
725
726 // mm is the memory state at the safepoint (when it's a MergeMem)
727 // no_side_effect_since_safepoint() goes over the memory state at the backedge. It resets the mm input for each
728 // component of the memory state it encounters so it points to the base memory. Once no_side_effect_since_safepoint()
729 // is done, if no side effect after the safepoint was found, mm should transform to the base memory: the states at
730 // the backedge and safepoint are the same so all components of the memory state at the safepoint should have been
731 // reset.
732 MergeMemNode* mm = nullptr;
733 #ifdef ASSERT
734 if (mem->is_MergeMem()) {
735 mm = mem->clone()->as_MergeMem();
736 _igvn._worklist.push(mm);
737 for (MergeMemStream mms(mem->as_MergeMem()); mms.next_non_empty(); ) {
738 // Loop invariant memory state won't be reset by no_side_effect_since_safepoint(). Do it here.
739 // Escape Analysis can add state to mm that it doesn't add to the backedge memory Phis, breaking verification
740 // code that relies on mm. Clear that extra state here.
741 if (mms.alias_idx() != Compile::AliasIdxBot &&
742 (loop != get_loop(ctrl_or_self(mms.memory())) ||
743 (mms.adr_type()->isa_oop_ptr() && mms.adr_type()->is_known_instance()))) {
744 mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
745 }
746 }
747 }
748 #endif
749 if (!no_side_effect_since_safepoint(C, x, mem, mm, this)) {
750 safepoint = nullptr;
751 } else {
752 assert(mm == nullptr|| _igvn.transform(mm) == mem->as_MergeMem()->base_memory(), "all memory state should have been processed");
753 }
754 #ifdef ASSERT
755 if (mm != nullptr) {
756 _igvn.remove_dead_node(mm);
757 }
758 #endif
759 }
760 return safepoint;
761 }
762
763 void PhaseIdealLoop::add_parse_predicates(IdealLoopTree* outer_ilt, LoopNode* inner_head, SafePointNode* cloned_sfpt) {
764 if (ShortRunningLongLoop) {
765 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, inner_head, outer_ilt, cloned_sfpt);
766 }
767 if (UseLoopPredicate) {
768 add_parse_predicate(Deoptimization::Reason_predicate, inner_head, outer_ilt, cloned_sfpt);
769 if (UseProfiledLoopPredicate) {
770 add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
771 }
772 }
773
774 if (UseAutoVectorizationPredicate) {
775 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, inner_head, outer_ilt, cloned_sfpt);
776 }
777
778 add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
779 }
780
781 // If the loop has the shape of a counted loop but with a long
782 // induction variable, transform the loop in a loop nest: an inner
783 // loop that iterates for at most max int iterations with an integer
784 // induction variable and an outer loop that iterates over the full
785 // range of long values from the initial loop in (at most) max int
786 // steps. That is:
787 //
788 // x: for (long phi = init; phi < limit; phi += stride) {
789 // // phi := Phi(L, init, incr)
790 // // incr := AddL(phi, longcon(stride))
791 // long incr = phi + stride;
792 // ... use phi and incr ...
793 // }
794 //
795 // OR:
796 //
797 // x: for (long phi = init; (phi += stride) < limit; ) {
798 // // phi := Phi(L, AddL(init, stride), incr)
799 // // incr := AddL(phi, longcon(stride))
800 // long incr = phi + stride;
801 // ... use phi and (phi + stride) ...
802 // }
803 //
804 // ==transform=>
805 //
806 // const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0
807 // assert(stride <= inner_iters_limit); // else abort transform
808 // assert((extralong)limit + stride <= LONG_MAX); // else deopt
809 // outer_head: for (long outer_phi = init;;) {
810 // // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi)))
811 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi));
812 // long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max);
813 // assert(inner_iters_actual == (int)inner_iters_actual);
814 // int inner_phi, inner_incr;
815 // x: for (inner_phi = 0;; inner_phi = inner_incr) {
816 // // inner_phi := Phi(x, intcon(0), inner_incr)
817 // // inner_incr := AddI(inner_phi, intcon(stride))
818 // inner_incr = inner_phi + stride;
819 // if (inner_incr < inner_iters_actual) {
820 // ... use phi=>(outer_phi+inner_phi) ...
821 // continue;
822 // }
823 // else break;
824 // }
825 // if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit
826 // continue;
827 // else break;
828 // }
829 //
830 // The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int
831 // loops with long range checks transformed to int range checks in the inner loop.
832 bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {
833 Node* x = loop->_head;
834 // Only for inner loops
835 if (loop->_child != nullptr || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) {
836 return false;
837 }
838
839 if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) {
840 return false;
841 }
842
843 BaseCountedLoopNode* head = x->as_BaseCountedLoop();
844 BasicType bt = x->as_BaseCountedLoop()->bt();
845
846 check_counted_loop_shape(loop, x, bt);
847
848 #ifndef PRODUCT
849 if (bt == T_LONG) {
850 AtomicAccess::inc(&_long_loop_candidates);
851 }
852 #endif
853
854 jlong stride_con_long = head->stride_con();
855 assert(stride_con_long != 0, "missed some peephole opt");
856 // We can't iterate for more than max int at a time.
857 if (stride_con_long != (jint)stride_con_long || stride_con_long == min_jint) {
858 assert(bt == T_LONG, "only for long loops");
859 return false;
860 }
861 jint stride_con = checked_cast<jint>(stride_con_long);
862 // The number of iterations for the integer count loop: guarantee no
863 // overflow: max_jint - stride_con max. -1 so there's no need for a
864 // loop limit check if the exit test is <= or >=.
865 int iters_limit = max_jint - ABS(stride_con) - 1;
866 #ifdef ASSERT
867 if (bt == T_LONG && StressLongCountedLoop > 0) {
868 iters_limit = iters_limit / StressLongCountedLoop;
869 }
870 #endif
871 // At least 2 iterations so counted loop construction doesn't fail
872 if (iters_limit/ABS(stride_con) < 2) {
873 return false;
874 }
875
876 assert(iters_limit > 0, "can't be negative");
877
878 PhiNode* phi = head->phi()->as_Phi();
879
880 Node* back_control = head->in(LoopNode::LoopBackControl);
881
882 // data nodes on back branch not supported
883 if (back_control->outcnt() > 1) {
884 return false;
885 }
886
887 Node* limit = head->limit();
888 // We'll need to use the loop limit before the inner loop is entered
889 if (!is_dominator(get_ctrl(limit), x)) {
890 return false;
891 }
892
893 IfNode* exit_test = head->loopexit();
894
895 assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
896
897 Node_List range_checks;
898 iters_limit = extract_long_range_checks(loop, stride_con, iters_limit, phi, range_checks);
899
900 if (bt == T_INT) {
901 // The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further.
902 if (range_checks.size() == 0) {
903 return false;
904 }
905 }
906
907 // Take what we know about the number of iterations of the long counted loop into account when computing the limit of
908 // the inner loop.
909 Node* init = head->init_trip();
910 const TypeInteger* lo = _igvn.type(init)->is_integer(bt);
911 const TypeInteger* hi = _igvn.type(limit)->is_integer(bt);
912 if (stride_con < 0) {
913 swap(lo, hi);
914 }
915 if (hi->hi_as_long() <= lo->lo_as_long()) {
916 // not a loop after all
917 return false;
918 }
919
920 if (range_checks.size() > 0) {
921 // This transformation requires peeling one iteration. Also, if it has range checks and they are eliminated by Loop
922 // Predication, then 2 Hoisted Check Predicates are added for one range check. Finally, transforming a long range
923 // check requires extra logic to be executed before the loop is entered and for the outer loop. As a result, the
924 // transformations can't pay off for a small number of iterations: roughly, if the loop runs for 3 iterations, it's
925 // going to execute as many range checks once transformed with range checks eliminated (1 peeled iteration with
926 // range checks + 2 predicates per range checks) as it would have not transformed. It also has to pay for the extra
927 // logic on loop entry and for the outer loop.
928 loop->compute_trip_count(this, bt);
929 if (head->is_CountedLoop() && head->as_CountedLoop()->has_exact_trip_count()) {
930 if (head->as_CountedLoop()->trip_count() <= 3) {
931 return false;
932 }
933 } else {
934 loop->compute_profile_trip_cnt(this);
935 if (!head->is_profile_trip_failed() && head->profile_trip_cnt() <= 3) {
936 return false;
937 }
938 }
939 }
940
941 if (try_make_short_running_loop(loop, stride_con, range_checks, iters_limit)) {
942 C->set_major_progress();
943 return true;
944 }
945
946 julong orig_iters = (julong)hi->hi_as_long() - lo->lo_as_long();
947 iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
948
949 // We need a safepoint to insert Parse Predicates for the inner loop.
950 SafePointNode* safepoint;
951 if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) {
952 // Loop is strip mined: use the safepoint of the outer strip mined loop
953 OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
954 assert(outer_loop != nullptr, "no outer loop");
955 safepoint = outer_loop->outer_safepoint();
956 outer_loop->transform_to_counted_loop(&_igvn, this);
957 exit_test = head->loopexit();
958 } else {
959 safepoint = find_safepoint(back_control, x, loop);
960 }
961
962 IfFalseNode* exit_branch = exit_test->false_proj();
963 Node* entry_control = head->in(LoopNode::EntryControl);
964
965 // Clone the control flow of the loop to build an outer loop
966 Node* outer_back_branch = back_control->clone();
967 Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
968 Node* inner_exit_branch = exit_branch->clone();
969
970 LoopNode* outer_head = new LoopNode(entry_control, outer_back_branch);
971 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_head, outer_back_branch);
972
973 const bool body_populated = true;
974 register_control(outer_head, outer_ilt, entry_control, body_populated);
975
976 _igvn.register_new_node_with_optimizer(inner_exit_branch);
977 set_loop(inner_exit_branch, outer_ilt);
978 set_idom(inner_exit_branch, exit_test, dom_depth(exit_branch));
979
980 outer_exit_test->set_req(0, inner_exit_branch);
981 register_control(outer_exit_test, outer_ilt, inner_exit_branch, body_populated);
982
983 _igvn.replace_input_of(exit_branch, 0, outer_exit_test);
984 set_idom(exit_branch, outer_exit_test, dom_depth(exit_branch));
985
986 outer_back_branch->set_req(0, outer_exit_test);
987 register_control(outer_back_branch, outer_ilt, outer_exit_test, body_populated);
988
989 _igvn.replace_input_of(x, LoopNode::EntryControl, outer_head);
990 set_idom(x, outer_head, dom_depth(x));
991
992 // add an iv phi to the outer loop and use it to compute the inner
993 // loop iteration limit
994 Node* outer_phi = phi->clone();
995 outer_phi->set_req(0, outer_head);
996 register_new_node(outer_phi, outer_head);
997
998 Node* inner_iters_max = nullptr;
999 if (stride_con > 0) {
1000 inner_iters_max = MinMaxNode::max_diff_with_zero(limit, outer_phi, TypeInteger::bottom(bt), _igvn);
1001 } else {
1002 inner_iters_max = MinMaxNode::max_diff_with_zero(outer_phi, limit, TypeInteger::bottom(bt), _igvn);
1003 }
1004
1005 Node* inner_iters_limit = _igvn.integercon(iters_limit, bt);
1006 // inner_iters_max may not fit in a signed integer (iterating from
1007 // Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned
1008 // min.
1009 const TypeInteger* inner_iters_actual_range = TypeInteger::make(0, iters_limit, Type::WidenMin, bt);
1010 Node* inner_iters_actual = MinMaxNode::unsigned_min(inner_iters_max, inner_iters_limit, inner_iters_actual_range, _igvn);
1011
1012 Node* inner_iters_actual_int;
1013 if (bt == T_LONG) {
1014 inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
1015 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1016 // When the inner loop is transformed to a counted loop, a loop limit check is not expected to be needed because
1017 // the loop limit is less or equal to max_jint - stride - 1 (if stride is positive but a similar argument exists for
1018 // a negative stride). We add a CastII here to guarantee that, when the counted loop is created in a subsequent loop
1019 // opts pass, an accurate range of values for the limits is found.
1020 const TypeInt* inner_iters_actual_int_range = TypeInt::make(0, iters_limit, Type::WidenMin);
1021 inner_iters_actual_int = new CastIINode(outer_head, inner_iters_actual_int, inner_iters_actual_int_range, ConstraintCastNode::DependencyType::NonFloatingNonNarrowing);
1022 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1023 } else {
1024 inner_iters_actual_int = inner_iters_actual;
1025 }
1026
1027 Node* int_zero = intcon(0);
1028 if (stride_con < 0) {
1029 inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
1030 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1031 }
1032
1033 // Clone the iv data nodes as an integer iv
1034 Node* int_stride = intcon(stride_con);
1035 Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
1036 Node* inner_incr = new AddINode(inner_phi, int_stride);
1037 Node* inner_cmp = nullptr;
1038 inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
1039 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
1040 inner_phi->set_req(LoopNode::EntryControl, int_zero);
1041 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
1042 register_new_node(inner_phi, x);
1043 register_new_node(inner_incr, x);
1044 register_new_node(inner_cmp, x);
1045 register_new_node(inner_bol, x);
1046
1047 _igvn.replace_input_of(exit_test, 1, inner_bol);
1048
1049 // Clone inner loop phis to outer loop
1050 for (uint i = 0; i < head->outcnt(); i++) {
1051 Node* u = head->raw_out(i);
1052 if (u->is_Phi() && u != inner_phi && u != phi) {
1053 assert(u->in(0) == head, "inconsistent");
1054 Node* clone = u->clone();
1055 clone->set_req(0, outer_head);
1056 register_new_node(clone, outer_head);
1057 _igvn.replace_input_of(u, LoopNode::EntryControl, clone);
1058 }
1059 }
1060
1061 // Replace inner loop long iv phi as inner loop int iv phi + outer
1062 // loop iv phi
1063 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
1064
1065 set_subtree_ctrl(inner_iters_actual_int, body_populated);
1066
1067 LoopNode* inner_head = create_inner_head(loop, head, exit_test);
1068
1069 // Summary of steps from initial loop to loop nest:
1070 //
1071 // == old IR nodes =>
1072 //
1073 // entry_control: {...}
1074 // x:
1075 // for (long phi = init;;) {
1076 // // phi := Phi(x, init, incr)
1077 // // incr := AddL(phi, longcon(stride))
1078 // exit_test:
1079 // if (phi < limit)
1080 // back_control: fallthrough;
1081 // else
1082 // exit_branch: break;
1083 // long incr = phi + stride;
1084 // ... use phi and incr ...
1085 // phi = incr;
1086 // }
1087 //
1088 // == new IR nodes (just before final peel) =>
1089 //
1090 // entry_control: {...}
1091 // long adjusted_limit = limit + stride; //because phi_incr != nullptr
1092 // assert(!limit_check_required || (extralong)limit + stride == adjusted_limit); // else deopt
1093 // ulong inner_iters_limit = max_jint - ABS(stride) - 1; //near 0x7FFFFFF0
1094 // outer_head:
1095 // for (long outer_phi = init;;) {
1096 // // outer_phi := phi->clone(), in(0):=outer_head, => Phi(outer_head, init, incr)
1097 // // REPLACE phi => AddL(outer_phi, I2L(inner_phi))
1098 // // REPLACE incr => AddL(outer_phi, I2L(inner_incr))
1099 // // SO THAT outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_incr)))
1100 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)adjusted_limit - outer_phi) * SGN(stride));
1101 // int inner_iters_actual_int = (int) MIN(inner_iters_limit, inner_iters_max) * SGN(stride);
1102 // inner_head: x: //in(1) := outer_head
1103 // int inner_phi;
1104 // for (inner_phi = 0;;) {
1105 // // inner_phi := Phi(x, intcon(0), inner_phi + stride)
1106 // int inner_incr = inner_phi + stride;
1107 // bool inner_bol = (inner_incr < inner_iters_actual_int);
1108 // exit_test: //exit_test->in(1) := inner_bol;
1109 // if (inner_bol) // WAS (phi < limit)
1110 // back_control: fallthrough;
1111 // else
1112 // inner_exit_branch: break; //exit_branch->clone()
1113 // ... use phi=>(outer_phi+inner_phi) ...
1114 // inner_phi = inner_phi + stride; // inner_incr
1115 // }
1116 // outer_exit_test: //exit_test->clone(), in(0):=inner_exit_branch
1117 // if ((outer_phi+inner_phi) < limit) // WAS (phi < limit)
1118 // outer_back_branch: fallthrough; //back_control->clone(), in(0):=outer_exit_test
1119 // else
1120 // exit_branch: break; //in(0) := outer_exit_test
1121 // }
1122
1123 if (bt == T_INT) {
1124 outer_phi = new ConvI2LNode(outer_phi);
1125 register_new_node(outer_phi, outer_head);
1126 }
1127
1128 transform_long_range_checks(stride_con, range_checks, outer_phi, inner_iters_actual_int,
1129 inner_phi, iv_add, inner_head);
1130 // Peel one iteration of the loop and use the safepoint at the end
1131 // of the peeled iteration to insert Parse Predicates. If no well
1132 // positioned safepoint peel to guarantee a safepoint in the outer
1133 // loop.
1134 if (safepoint != nullptr || !loop->_has_call) {
1135 old_new.clear();
1136 do_peeling(loop, old_new);
1137 } else {
1138 C->set_major_progress();
1139 }
1140
1141 if (safepoint != nullptr) {
1142 SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
1143
1144 add_parse_predicates(outer_ilt, inner_head, cloned_sfpt);
1145 }
1146
1147 #ifndef PRODUCT
1148 if (bt == T_LONG) {
1149 AtomicAccess::inc(&_long_loop_nests);
1150 }
1151 #endif
1152
1153 inner_head->mark_loop_nest_inner_loop();
1154 outer_head->mark_loop_nest_outer_loop();
1155
1156 return true;
1157 }
1158
1159 // Make a copy of Parse/Template Assertion predicates below existing predicates at the loop passed as argument
1160 class CloneShortLoopPredicateVisitor : public PredicateVisitor {
1161 ClonePredicateToTargetLoop _clone_predicate_to_loop;
1162 PhaseIdealLoop* const _phase;
1163 Node* const _new_init;
1164
1165 public:
1166 CloneShortLoopPredicateVisitor(LoopNode* target_loop_head,
1167 Node* new_init,
1168 const NodeInSingleLoopBody &node_in_loop_body,
1169 PhaseIdealLoop* phase)
1170 : _clone_predicate_to_loop(target_loop_head, node_in_loop_body, phase),
1171 _phase(phase),
1172 _new_init(new_init) {
1173 }
1174 NONCOPYABLE(CloneShortLoopPredicateVisitor);
1175
1176 using PredicateVisitor::visit;
1177
1178 void visit(const ParsePredicate& parse_predicate) override {
1179 _clone_predicate_to_loop.clone_parse_predicate(parse_predicate, true);
1180 parse_predicate.kill(_phase->igvn());
1181 }
1182
1183 void visit(const TemplateAssertionPredicate& template_assertion_predicate) override {
1184 _clone_predicate_to_loop.clone_template_assertion_predicate_and_replace_init(template_assertion_predicate, _new_init);
1185 template_assertion_predicate.kill(_phase->igvn());
1186 }
1187 };
1188
1189 // For an int counted loop, try_make_short_running_loop() transforms the loop from:
1190 // for (int = start; i < stop; i+= stride) { ... }
1191 // to
1192 // for (int = 0; i < stop - start; i+= stride) { ... }
1193 // Template Assertion Predicates added so far were with an init value of start. They need to be updated with the new
1194 // init value of 0 (otherwise when a template assertion predicate is turned into an initialized assertion predicate, it
1195 // performs an incorrect check):
1196 // zero
1197 // init |
1198 // | ===> OpaqueLoopInit init
1199 // OpaqueLoopInit \ /
1200 // AddI
1201 //
1202 Node* PhaseIdealLoop::new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero) {
1203 OpaqueLoopInitNode* new_opaque_init = new OpaqueLoopInitNode(C, int_zero);
1204 register_new_node(new_opaque_init, entry_control);
1205 Node* new_init = new AddINode(new_opaque_init, init);
1206 register_new_node(new_init, entry_control);
1207 return new_init;
1208 }
1209
1210 // If the loop is either statically known to run for a small enough number of iterations or if profile data indicates
1211 // that, we don't want an outer loop because the overhead of having an outer loop whose backedge is never taken, has a
1212 // measurable cost. Furthermore, creating the loop nest usually causes one iteration of the loop to be peeled so
1213 // predicates can be set up. If the loop is short running, then it's an extra iteration that's run with range checks
1214 // (compared to an int counted loop with int range checks).
1215 //
1216 // In the short running case, turn the loop into a regular loop again and transform the long range checks:
1217 // - LongCountedLoop: Create LoopNode but keep the loop limit type with a CastLL node to avoid that we later try to
1218 // create a Loop Limit Check when turning the LoopNode into a CountedLoopNode.
1219 // - CountedLoop: Can be reused.
1220 bool PhaseIdealLoop::try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List &range_checks,
1221 const uint iters_limit) {
1222 if (!ShortRunningLongLoop) {
1223 return false;
1224 }
1225 BaseCountedLoopNode* head = loop->_head->as_BaseCountedLoop();
1226 BasicType bt = head->bt();
1227 Node* entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1228
1229 loop->compute_trip_count(this, bt);
1230 // Loop must run for no more than iter_limits as it guarantees no overflow of scale * iv in long range checks (see
1231 // comment above PhaseIdealLoop::transform_long_range_checks()).
1232 // iters_limit / ABS(stride_con) is the largest trip count for which we know it's correct to not create a loop nest:
1233 // it's always beneficial to have a single loop rather than a loop nest, so we try to apply this transformation as
1234 // often as possible.
1235 bool known_short_running_loop = head->trip_count() <= iters_limit / ABS(stride_con);
1236 bool profile_short_running_loop = false;
1237 if (!known_short_running_loop) {
1238 loop->compute_profile_trip_cnt(this);
1239 if (StressShortRunningLongLoop) {
1240 profile_short_running_loop = true;
1241 } else {
1242 profile_short_running_loop = !head->is_profile_trip_failed() && head->profile_trip_cnt() <= iters_limit / ABS(stride_con);
1243 }
1244 }
1245
1246 if (!known_short_running_loop && !profile_short_running_loop) {
1247 return false;
1248 }
1249
1250 Node* limit = head->limit();
1251 Node* init = head->init_trip();
1252
1253 Node* new_limit;
1254 if (stride_con > 0) {
1255 new_limit = SubNode::make(limit, init, bt);
1256 } else {
1257 new_limit = SubNode::make(init, limit, bt);
1258 }
1259 register_new_node(new_limit, entry_control);
1260
1261 Node* int_zero = intcon(0);
1262 PhiNode* phi = head->phi()->as_Phi();
1263 if (profile_short_running_loop) {
1264 // Add a Short Running Long Loop Predicate. It's the first predicate in the predicate chain before entering a loop
1265 // because a cast that's control dependent on the Short Running Long Loop Predicate is added to narrow the limit and
1266 // future predicates may be dependent on the new limit (so have to be between the loop and Short Running Long Loop
1267 // Predicate). The current limit could, itself, be dependent on an existing predicate. Clone parse and template
1268 // assertion predicates below existing predicates to get proper ordering of predicates when walking from the loop
1269 // up: future predicates, Short Running Long Loop Predicate, existing predicates.
1270 //
1271 // Existing Hoisted
1272 // Check Predicates
1273 // |
1274 // New Short Running Long
1275 // Loop Predicate
1276 // |
1277 // Cloned Parse Predicates and
1278 // Template Assertion Predicates
1279 // (future predicates added here)
1280 // |
1281 // Loop
1282 const Predicates predicates_before_cloning(entry_control);
1283 const PredicateBlock* short_running_long_loop_predicate_block = predicates_before_cloning.short_running_long_loop_predicate_block();
1284 if (!short_running_long_loop_predicate_block->has_parse_predicate()) { // already trapped
1285 return false;
1286 }
1287 Node* new_init = new_assertion_predicate_opaque_init(entry_control, init, int_zero);
1288
1289 PredicateIterator predicate_iterator(entry_control);
1290 NodeInSingleLoopBody node_in_short_loop_body(this, loop);
1291 CloneShortLoopPredicateVisitor clone_short_loop_predicates_visitor(head, new_init, node_in_short_loop_body, this);
1292 predicate_iterator.for_each(clone_short_loop_predicates_visitor);
1293
1294 entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1295
1296 const Predicates predicates_after_cloning(entry_control);
1297
1298 ParsePredicateSuccessProj* short_running_loop_predicate_proj = predicates_after_cloning.
1299 short_running_long_loop_predicate_block()->
1300 parse_predicate_success_proj();
1301 assert(short_running_loop_predicate_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
1302
1303 const jlong iters_limit_long = iters_limit;
1304 Node* cmp_limit = CmpNode::make(new_limit, _igvn.integercon(iters_limit_long, bt), bt);
1305 Node* bol = new BoolNode(cmp_limit, BoolTest::le);
1306 Node* new_predicate_proj = create_new_if_for_predicate(short_running_loop_predicate_proj,
1307 nullptr,
1308 Deoptimization::Reason_short_running_long_loop,
1309 Op_If);
1310 Node* iff = new_predicate_proj->in(0);
1311 _igvn.replace_input_of(iff, 1, bol);
1312 register_new_node(cmp_limit, iff->in(0));
1313 register_new_node(bol, iff->in(0));
1314 new_limit = ConstraintCastNode::make_cast_for_basic_type(new_predicate_proj, new_limit,
1315 TypeInteger::make(1, iters_limit_long, Type::WidenMin, bt),
1316 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, bt);
1317 register_new_node(new_limit, new_predicate_proj);
1318
1319 #ifndef PRODUCT
1320 if (TraceLoopLimitCheck) {
1321 tty->print_cr("Short Long Loop Check Predicate generated:");
1322 DEBUG_ONLY(bol->dump(2);)
1323 }
1324 #endif
1325 entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1326 } else if (bt == T_LONG) {
1327 // We're turning a long counted loop into a regular loop that will be converted into an int counted loop. That loop
1328 // won't need loop limit check predicates (iters_limit guarantees that). Add a cast to make sure that, whatever
1329 // transformation happens by the time the counted loop is created (in a subsequent pass of loop opts), C2 knows
1330 // enough about the loop's limit that it doesn't try to add loop limit check predicates.
1331 const Predicates predicates(entry_control);
1332 const TypeLong* new_limit_t = new_limit->Value(&_igvn)->is_long();
1333 new_limit = ConstraintCastNode::make_cast_for_basic_type(predicates.entry(), new_limit,
1334 TypeLong::make(0, new_limit_t->_hi, new_limit_t->_widen),
1335 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, bt);
1336 register_new_node(new_limit, predicates.entry());
1337 } else {
1338 assert(bt == T_INT && known_short_running_loop, "only CountedLoop statically known to be short running");
1339 PredicateIterator predicate_iterator(entry_control);
1340 Node* new_init = new_assertion_predicate_opaque_init(entry_control, init, int_zero);
1341 UpdateInitForTemplateAssertionPredicates update_init_for_template_assertion_predicates(new_init, this);
1342 predicate_iterator.for_each(update_init_for_template_assertion_predicates);
1343 }
1344 IfNode* exit_test = head->loopexit();
1345
1346 if (bt == T_LONG) {
1347 // The loop is short running so new_limit fits into an int: either we determined that statically or added a guard
1348 new_limit = new ConvL2INode(new_limit);
1349 register_new_node(new_limit, entry_control);
1350 }
1351
1352 if (stride_con < 0) {
1353 new_limit = new SubINode(int_zero, new_limit);
1354 register_new_node(new_limit, entry_control);
1355 }
1356
1357 // Clone the iv data nodes as an integer iv
1358 Node* int_stride = intcon(stride_con);
1359 Node* inner_phi = new PhiNode(head, TypeInt::INT);
1360 Node* inner_incr = new AddINode(inner_phi, int_stride);
1361 Node* inner_cmp = new CmpINode(inner_incr, new_limit);
1362 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
1363 inner_phi->set_req(LoopNode::EntryControl, int_zero);
1364 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
1365 register_new_node(inner_phi, head);
1366 register_new_node(inner_incr, head);
1367 register_new_node(inner_cmp, head);
1368 register_new_node(inner_bol, head);
1369
1370 _igvn.replace_input_of(exit_test, 1, inner_bol);
1371
1372 // Replace inner loop long iv phi as inner loop int iv phi + outer
1373 // loop iv phi
1374 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, init, head, bt);
1375
1376 LoopNode* inner_head = head;
1377 if (bt == T_LONG) {
1378 // Turn the loop back to a counted loop
1379 inner_head = create_inner_head(loop, head, exit_test);
1380 } else {
1381 // Use existing counted loop
1382 revert_to_normal_loop(head);
1383 }
1384
1385 if (bt == T_INT) {
1386 init = new ConvI2LNode(init);
1387 register_new_node(init, entry_control);
1388 }
1389
1390 transform_long_range_checks(stride_con, range_checks, init, new_limit,
1391 inner_phi, iv_add, inner_head);
1392
1393 inner_head->mark_loop_nest_inner_loop();
1394
1395 return true;
1396 }
1397
1398 int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1399 Node_List& range_checks) {
1400 const jlong min_iters = 2;
1401 jlong reduced_iters_limit = iters_limit;
1402 jlong original_iters_limit = iters_limit;
1403 for (uint i = 0; i < loop->_body.size(); i++) {
1404 Node* c = loop->_body.at(i);
1405 if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
1406 IfProjNode* if_proj = c->as_IfProj();
1407 CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1408 if (call != nullptr) {
1409 Node* range = nullptr;
1410 Node* offset = nullptr;
1411 jlong scale = 0;
1412 if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1413 loop->is_invariant(range) && loop->is_invariant(offset) &&
1414 scale != min_jlong &&
1415 original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {
1416 assert(scale == (jint)scale, "scale should be an int");
1417 reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1418 range_checks.push(c);
1419 }
1420 }
1421 }
1422 }
1423
1424 return checked_cast<int>(reduced_iters_limit);
1425 }
1426
1427 // One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1428 // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1429 // 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
1430 // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1431
1432 // 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,
1433 // 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)
1434 // 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].
1435
1436 // 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
1437 // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1438 // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1439
1440 // 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
1441 // 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
1442 // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1443 // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1444
1445 // 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
1446 // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1447
1448 // 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].
1449 // 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
1450 // 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
1451 // 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
1452 // down (from C and 0 again).
1453
1454 // 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).
1455 // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1456 // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1457 // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1458 // PhaseIdealLoop::add_constraint.)
1459
1460 // 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
1461 // 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
1462 // 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.
1463 // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1464 // values for the 32-bit counter j within which the range checks can be eliminated.
1465
1466 // 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
1467 // 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
1468 // allowed as an input to the R.C., as long as the R.C. as a whole fails.
1469
1470 // If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1471
1472 // 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
1473 // 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),
1474 // 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)
1475
1476 // 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).
1477 // 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.
1478 // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1479
1480 // Case A: Some Negatives (but no overflow).
1481 // Number line:
1482 // |s64_min . . . 0 . . . s64_max|
1483 // | . Q_min..Q_max . 0 . . . . | s64 negative
1484 // | . . . . R=0 R< R< R< R< | (against R values)
1485 // | . . . Q_min..0..Q_max . . . | small mixed
1486 // | . . . . R R R< R< R< | (against R values)
1487 //
1488 // 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<.
1489 //
1490 // So, if Q_min <s64 0, then use this test:
1491 // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1492 // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1493 // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1494 // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1495
1496 // If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1497
1498 // Case B: No Negatives.
1499 // Number line:
1500 // |s64_min . . . 0 . . . s64_max|
1501 // | . . . . 0 Q_min..Q_max . . | small positive
1502 // | . . . . R> R R R< R< | (against R values)
1503 // | . . . . 0 . Q_min..Q_max . | s64 positive
1504 // | . . . . R> R> R R R< | (against R values)
1505 //
1506 // 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.
1507 // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1508 //
1509 // So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1510 // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1511 // More generally:
1512 // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1513
1514 // Case C: Overflow in the 64-bit domain
1515 // Number line:
1516 // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1517 // | . . . . R> R> R> R> R | (against R values)
1518 //
1519 // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1520 // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1521 // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1522 //
1523 // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1524 // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1525 // More generally:
1526 // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1527 //
1528 // Dropping the bad bound means only Q_min is used to reduce the range of R:
1529 // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1530 //
1531 // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1532 // clamp(X, L, H) := max(L, min(X, H))
1533 // When degenerately L > H, it returns L not H.
1534 //
1535 // All of the formulas above can be merged into a single one:
1536 // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1537 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1538 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1539 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1540 // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1541 // replacement R.C.:
1542 // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1543 // or equivalently:
1544 // j*K + L_2 <u32 R_2
1545 // where
1546 // L_2 = Q_first - L_clamp
1547 // R_2 = R_clamp - L_clamp
1548 //
1549 // Note on why R is never negative:
1550 //
1551 // Various details of this transformation would break badly if R could be negative, so this transformation only
1552 // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1553 // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1554 // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1555 // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1556 // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1557 // R.C., but R values from a.length and inside checkIndex are good to go.
1558 //
1559 void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1560 Node* inner_iters_actual_int, Node* inner_phi,
1561 Node* iv_add, LoopNode* inner_head) {
1562 Node* long_zero = longcon(0);
1563 Node* int_zero = intcon(0);
1564 Node* long_one = longcon(1);
1565 Node* int_stride = intcon(checked_cast<int>(stride_con));
1566
1567 for (uint i = 0; i < range_checks.size(); i++) {
1568 ProjNode* proj = range_checks.at(i)->as_Proj();
1569 RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1570 jlong scale = 0;
1571 Node* offset = nullptr;
1572 Node* rc_bol = rc->in(1);
1573 Node* rc_cmp = rc_bol->in(1);
1574 if (rc_cmp->Opcode() == Op_CmpU) {
1575 // could be shared and have already been taken care of
1576 continue;
1577 }
1578 bool short_scale = false;
1579 bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1580 assert(ok, "inconsistent: was tested before");
1581 Node* range = rc_cmp->in(2);
1582 Node* c = rc->in(0);
1583 Node* entry_control = inner_head->in(LoopNode::EntryControl);
1584
1585 Node* R = range;
1586 Node* K = longcon(scale);
1587
1588 Node* L = offset;
1589
1590 if (short_scale) {
1591 // This converts:
1592 // (int)i*K + L <u64 R
1593 // with K an int into:
1594 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1595 // to protect against an overflow of (int)i*K
1596 //
1597 // Because if (int)i*K overflows, there are K,L where:
1598 // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1599 // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1600 // i*(long)K + L <u64 R is true.
1601 //
1602 // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1603 //
1604 // It's always true that:
1605 // (int)i*K <u64 (long)max_jint + 1
1606 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1607 // As a consequence:
1608 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1609 // is always false in case of overflow of i*K
1610 //
1611 // Note, there are also K,L where i*K overflows and
1612 // i*K + L <u64 R is true, but
1613 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1614 // So this transformation could cause spurious deoptimizations and failed range check elimination
1615 // (but not incorrect execution) for unlikely corner cases with overflow.
1616 // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1617 Node* max_jint_plus_one_long = longcon((jlong)max_jint + 1);
1618 Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1619 register_new_node(max_range, entry_control);
1620 R = MinMaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1621 set_subtree_ctrl(R, true);
1622 }
1623
1624 Node* C = outer_phi;
1625
1626 // Start with 64-bit values:
1627 // i*K + L <u64 R
1628 // (C+j)*K + L <u64 R
1629 // j*K + Q <u64 R where Q = Q_first = C*K+L
1630 Node* Q_first = new MulLNode(C, K);
1631 register_new_node(Q_first, entry_control);
1632 Q_first = new AddLNode(Q_first, L);
1633 register_new_node(Q_first, entry_control);
1634
1635 // Compute endpoints of the range of values j*K + Q.
1636 // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1637 Node* Q_min = Q_first;
1638
1639 // Compute the exact ending value B_2 (which is really A_2 if S < 0)
1640 Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1641 register_new_node(B_2, entry_control);
1642 B_2 = new SubINode(B_2, int_stride);
1643 register_new_node(B_2, entry_control);
1644 B_2 = new ConvI2LNode(B_2);
1645 register_new_node(B_2, entry_control);
1646
1647 Node* Q_max = new MulLNode(B_2, K);
1648 register_new_node(Q_max, entry_control);
1649 Q_max = new AddLNode(Q_max, Q_first);
1650 register_new_node(Q_max, entry_control);
1651
1652 if (scale * stride_con < 0) {
1653 swap(Q_min, Q_max);
1654 }
1655 // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1656
1657 // L_clamp = Q_min < 0 ? 0 : Q_min
1658 Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1659 register_new_node(Q_min_cmp, entry_control);
1660 Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1661 register_new_node(Q_min_bool, entry_control);
1662 Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1663 register_new_node(L_clamp, entry_control);
1664 // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1665
1666 Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1667 register_new_node(Q_max_plus_one, entry_control);
1668
1669 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1670 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1671 Node* max_jlong_long = longcon(max_jlong);
1672 Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1673 register_new_node(Q_max_cmp, entry_control);
1674 Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1675 register_new_node(Q_max_bool, entry_control);
1676 Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1677 register_new_node(H_clamp, entry_control);
1678 // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1679
1680 // R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1681 // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1682 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1683 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1684 // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1685 Node* R_2 = clamp(R, L_clamp, H_clamp);
1686 R_2 = new SubLNode(R_2, L_clamp);
1687 register_new_node(R_2, entry_control);
1688 R_2 = new ConvL2INode(R_2, TypeInt::POS);
1689 register_new_node(R_2, entry_control);
1690
1691 // L_2 = Q_first - L_clamp
1692 // We are subtracting L_clamp from both sides of the <u32 comparison.
1693 // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1694 // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1695 Node* L_2 = new SubLNode(Q_first, L_clamp);
1696 register_new_node(L_2, entry_control);
1697 L_2 = new ConvL2INode(L_2, TypeInt::INT);
1698 register_new_node(L_2, entry_control);
1699
1700 // Transform the range check using the computed values L_2/R_2
1701 // from: i*K + L <u64 R
1702 // to: j*K + L_2 <u32 R_2
1703 // that is:
1704 // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1705 K = intcon(checked_cast<int>(scale));
1706 Node* scaled_iv = new MulINode(inner_phi, K);
1707 register_new_node(scaled_iv, c);
1708 Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1709 register_new_node(scaled_iv_plus_offset, c);
1710
1711 Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1712 register_new_node(new_rc_cmp, c);
1713
1714 _igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1715 }
1716 }
1717
1718 Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {
1719 Node* min = MinMaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1720 set_subtree_ctrl(min, true);
1721 Node* max = MinMaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1722 set_subtree_ctrl(max, true);
1723 return max;
1724 }
1725
1726 LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1727 IfNode* exit_test) {
1728 LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1729 IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1730 _igvn.register_new_node_with_optimizer(new_inner_head);
1731 _igvn.register_new_node_with_optimizer(new_inner_exit);
1732 loop->_body.push(new_inner_head);
1733 loop->_body.push(new_inner_exit);
1734 loop->_body.yank(head);
1735 loop->_body.yank(exit_test);
1736 set_loop(new_inner_head, loop);
1737 set_loop(new_inner_exit, loop);
1738 set_idom(new_inner_head, idom(head), dom_depth(head));
1739 set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1740 replace_node_and_forward_ctrl(head, new_inner_head);
1741 replace_node_and_forward_ctrl(exit_test, new_inner_exit);
1742 loop->_head = new_inner_head;
1743 return new_inner_head;
1744 }
1745
1746 #ifdef ASSERT
1747 void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {
1748 Node* back_control = loop_exit_control(x, loop);
1749 assert(back_control != nullptr, "no back control");
1750
1751 BoolTest::mask mask = BoolTest::illegal;
1752 float cl_prob = 0;
1753 Node* incr = nullptr;
1754 Node* limit = nullptr;
1755
1756 Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1757 assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1758
1759 Node* phi_incr = nullptr;
1760 incr = loop_iv_incr(incr, x, loop, phi_incr);
1761 assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1762
1763 Node* xphi = nullptr;
1764 Node* stride = loop_iv_stride(incr, xphi);
1765
1766 assert(stride != nullptr, "no stride");
1767
1768 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x);
1769
1770 assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1771
1772 jlong stride_con = stride->get_integer_as_long(bt);
1773
1774 assert(condition_stride_ok(mask, stride_con), "illegal condition");
1775
1776 assert(mask != BoolTest::ne, "unexpected condition");
1777 assert(phi_incr == nullptr, "bad loop shape");
1778 assert(cmp->in(1) == incr, "bad exit test shape");
1779
1780 // Safepoint on backedge not supported
1781 assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1782 }
1783 #endif
1784
1785 #ifdef ASSERT
1786 // convert an int counted loop to a long counted to stress handling of
1787 // long counted loops
1788 bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
1789 Unique_Node_List iv_nodes;
1790 Node_List old_new;
1791 iv_nodes.push(cmp);
1792 bool failed = false;
1793
1794 for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
1795 Node* n = iv_nodes.at(i);
1796 switch(n->Opcode()) {
1797 case Op_Phi: {
1798 Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1799 old_new.map(n->_idx, clone);
1800 break;
1801 }
1802 case Op_CmpI: {
1803 Node* clone = new CmpLNode(nullptr, nullptr);
1804 old_new.map(n->_idx, clone);
1805 break;
1806 }
1807 case Op_AddI: {
1808 Node* clone = new AddLNode(nullptr, nullptr);
1809 old_new.map(n->_idx, clone);
1810 break;
1811 }
1812 case Op_CastII: {
1813 failed = true;
1814 break;
1815 }
1816 default:
1817 DEBUG_ONLY(n->dump());
1818 fatal("unexpected");
1819 }
1820
1821 for (uint i = 1; i < n->req(); i++) {
1822 Node* in = n->in(i);
1823 if (in == nullptr) {
1824 continue;
1825 }
1826 if (ctrl_is_member(loop, in)) {
1827 iv_nodes.push(in);
1828 }
1829 }
1830 }
1831
1832 if (failed) {
1833 for (uint i = 0; i < iv_nodes.size(); i++) {
1834 Node* n = iv_nodes.at(i);
1835 Node* clone = old_new[n->_idx];
1836 if (clone != nullptr) {
1837 _igvn.remove_dead_node(clone);
1838 }
1839 }
1840 return false;
1841 }
1842
1843 for (uint i = 0; i < iv_nodes.size(); i++) {
1844 Node* n = iv_nodes.at(i);
1845 Node* clone = old_new[n->_idx];
1846 for (uint i = 1; i < n->req(); i++) {
1847 Node* in = n->in(i);
1848 if (in == nullptr) {
1849 continue;
1850 }
1851 Node* in_clone = old_new[in->_idx];
1852 if (in_clone == nullptr) {
1853 assert(_igvn.type(in)->isa_int(), "");
1854 in_clone = new ConvI2LNode(in);
1855 _igvn.register_new_node_with_optimizer(in_clone);
1856 set_subtree_ctrl(in_clone, false);
1857 }
1858 if (in_clone->in(0) == nullptr) {
1859 in_clone->set_req(0, C->top());
1860 clone->set_req(i, in_clone);
1861 in_clone->set_req(0, nullptr);
1862 } else {
1863 clone->set_req(i, in_clone);
1864 }
1865 }
1866 _igvn.register_new_node_with_optimizer(clone);
1867 }
1868 set_ctrl(old_new[phi->_idx], phi->in(0));
1869
1870 for (uint i = 0; i < iv_nodes.size(); i++) {
1871 Node* n = iv_nodes.at(i);
1872 Node* clone = old_new[n->_idx];
1873 set_subtree_ctrl(clone, false);
1874 Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1875 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1876 Node* u = n->fast_out(i);
1877 if (iv_nodes.member(u)) {
1878 continue;
1879 }
1880 if (m == nullptr) {
1881 m = new ConvL2INode(clone);
1882 _igvn.register_new_node_with_optimizer(m);
1883 set_subtree_ctrl(m, false);
1884 }
1885 _igvn.rehash_node_delayed(u);
1886 int nb = u->replace_edge(n, m, &_igvn);
1887 --i, imax -= nb;
1888 }
1889 }
1890 return true;
1891 }
1892 #endif
1893
1894 //------------------------------is_counted_loop--------------------------------
1895 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop, BasicType iv_bt) {
1896 PhaseGVN *gvn = &_igvn;
1897
1898 Node* back_control = loop_exit_control(x, loop);
1899 if (back_control == nullptr) {
1900 return false;
1901 }
1902
1903 BoolTest::mask bt = BoolTest::illegal;
1904 float cl_prob = 0;
1905 Node* incr = nullptr;
1906 Node* limit = nullptr;
1907 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1908 if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {
1909 return false; // Avoid pointer & float & 64-bit compares
1910 }
1911
1912 // Trip-counter increment must be commutative & associative.
1913 if (incr->Opcode() == Op_Cast(iv_bt)) {
1914 incr = incr->in(1);
1915 }
1916
1917 Node* phi_incr = nullptr;
1918 incr = loop_iv_incr(incr, x, loop, phi_incr);
1919 if (incr == nullptr) {
1920 return false;
1921 }
1922
1923 Node* trunc1 = nullptr;
1924 Node* trunc2 = nullptr;
1925 const TypeInteger* iv_trunc_t = nullptr;
1926 Node* orig_incr = incr;
1927 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {
1928 return false; // Funny increment opcode
1929 }
1930 assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1931
1932 Node* xphi = nullptr;
1933 Node* stride = loop_iv_stride(incr, xphi);
1934
1935 if (stride == nullptr) {
1936 return false;
1937 }
1938
1939 // Iteratively uncast the loop induction variable
1940 // until no more CastII/CastLL nodes are found.
1941 while (xphi->Opcode() == Op_Cast(iv_bt)) {
1942 xphi = xphi->in(1);
1943 }
1944
1945 // Stride must be constant
1946 jlong stride_con = stride->get_integer_as_long(iv_bt);
1947 assert(stride_con != 0, "missed some peephole opt");
1948
1949 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x);
1950
1951 if (phi == nullptr ||
1952 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1953 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
1954 return false;
1955 }
1956
1957 Node* iftrue = back_control;
1958 uint iftrue_op = iftrue->Opcode();
1959 Node* iff = iftrue->in(0);
1960 BoolNode* test = iff->in(1)->as_Bool();
1961
1962 const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1963 if (trunc1 != nullptr) {
1964 // When there is a truncation, we must be sure that after the truncation
1965 // the trip counter will end up higher than the limit, otherwise we are looking
1966 // at an endless loop. Can happen with range checks.
1967
1968 // Example:
1969 // int i = 0;
1970 // while (true)
1971 // sum + = array[i];
1972 // i++;
1973 // i = i && 0x7fff;
1974 // }
1975 //
1976 // If the array is shorter than 0x8000 this exits through a AIOOB
1977 // - Counted loop transformation is ok
1978 // If the array is longer then this is an endless loop
1979 // - No transformation can be done.
1980
1981 const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1982 if (limit_t->hi_as_long() > incr_t->hi_as_long()) {
1983 // if the limit can have a higher value than the increment (before the phi)
1984 return false;
1985 }
1986 }
1987
1988 Node *init_trip = phi->in(LoopNode::EntryControl);
1989
1990 // If iv trunc type is smaller than int, check for possible wrap.
1991 if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
1992 assert(trunc1 != nullptr, "must have found some truncation");
1993
1994 // Get a better type for the phi (filtered thru if's)
1995 const TypeInteger* phi_ft = filtered_type(phi);
1996
1997 // Can iv take on a value that will wrap?
1998 //
1999 // Ensure iv's limit is not within "stride" of the wrap value.
2000 //
2001 // Example for "short" type
2002 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
2003 // If the stride is +10, then the last value of the induction
2004 // variable before the increment (phi_ft->_hi) must be
2005 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
2006 // ensure no truncation occurs after the increment.
2007
2008 if (stride_con > 0) {
2009 if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
2010 iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {
2011 return false; // truncation may occur
2012 }
2013 } else if (stride_con < 0) {
2014 if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
2015 iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {
2016 return false; // truncation may occur
2017 }
2018 }
2019 // No possibility of wrap so truncation can be discarded
2020 // Promote iv type to Int
2021 } else {
2022 assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
2023 }
2024
2025 if (!condition_stride_ok(bt, stride_con)) {
2026 return false;
2027 }
2028
2029 const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
2030
2031 if (stride_con > 0) {
2032 if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {
2033 return false; // cyclic loop
2034 }
2035 } else {
2036 if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {
2037 return false; // cyclic loop
2038 }
2039 }
2040
2041 if (phi_incr != nullptr && bt != BoolTest::ne) {
2042 // check if there is a possibility of IV overflowing after the first increment
2043 if (stride_con > 0) {
2044 if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {
2045 return false;
2046 }
2047 } else {
2048 if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {
2049 return false;
2050 }
2051 }
2052 }
2053
2054 // =================================================
2055 // ---- SUCCESS! Found A Trip-Counted Loop! -----
2056 //
2057
2058 if (x->Opcode() == Op_Region) {
2059 // x has not yet been transformed to Loop or LongCountedLoop.
2060 // This should only happen if we are inside an infinite loop.
2061 // It happens like this:
2062 // build_loop_tree -> do not attach infinite loop and nested loops
2063 // beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
2064 // build_loop_tree -> find and attach infinite and nested loops
2065 // counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
2066 assert(x->as_Region()->is_in_infinite_subgraph(),
2067 "x can only be a Region and not Loop if inside infinite loop");
2068 // Come back later when Region is transformed to LoopNode
2069 return false;
2070 }
2071
2072 assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
2073 C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
2074
2075 // ===================================================
2076 // We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
2077 // This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
2078 // At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
2079 // But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
2080 // to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
2081 // we trap and resume execution before the loop without having executed any iteration of the loop, yet.
2082 //
2083 // These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
2084 // Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
2085 // we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
2086 //
2087 // In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
2088 // for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
2089 //
2090 //
2091 // The loop to be converted does not always need to have the often used shape:
2092 //
2093 // i = init
2094 // i = init loop:
2095 // do { ...
2096 // // ... equivalent i+=stride
2097 // i+=stride <==> if (i < limit)
2098 // } while (i < limit); goto loop
2099 // exit:
2100 // ...
2101 //
2102 // where the loop exit check uses the post-incremented iv phi and a '<'-operator.
2103 //
2104 // We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
2105 // in the loop exit check:
2106 //
2107 // i = init
2108 // loop:
2109 // ...
2110 // if (i <= limit)
2111 // i+=stride
2112 // goto loop
2113 // exit:
2114 // ...
2115 //
2116 // Let's define the following terms:
2117 // - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
2118 // - iv_post_i: The post-incremented iv phi after the i-th iteration.
2119 //
2120 // The iv_pre_i and iv_post_i have the following relation:
2121 // iv_pre_i + stride = iv_post_i
2122 //
2123 // When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
2124 // iv_post_i < adjusted_limit
2125 //
2126 // If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
2127 // (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
2128 // -> adjusted_limit = limit.
2129 // (LE2) iv_post_i <= limit:
2130 // iv_post_i < limit + 1
2131 // -> adjusted limit = limit + 1
2132 // (LE3) iv_pre_i < limit:
2133 // iv_pre_i + stride < limit + stride
2134 // iv_post_i < limit + stride
2135 // -> adjusted_limit = limit + stride
2136 // (LE4) iv_pre_i <= limit:
2137 // iv_pre_i < limit + 1
2138 // iv_pre_i + stride < limit + stride + 1
2139 // iv_post_i < limit + stride + 1
2140 // -> adjusted_limit = limit + stride + 1
2141 //
2142 // Note that:
2143 // (AL) limit <= adjusted_limit.
2144 //
2145 // The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
2146 // loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
2147 // (INV) For i = 1..n, min_int <= iv_post_i <= max_int
2148 //
2149 // To prove (INV), we require the following two conditions/assumptions:
2150 // (i): adjusted_limit - 1 + stride <= max_int
2151 // (ii): init < limit
2152 //
2153 // If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
2154 // induction by assuming (i) and (ii).
2155 //
2156 // Proof by Induction
2157 // ------------------
2158 // > Base case (i = 1): We show that (INV) holds after the first iteration:
2159 // min_int <= iv_post_1 = init + stride <= max_int
2160 // Proof:
2161 // First, we note that (ii) implies
2162 // (iii) init <= limit - 1
2163 // max_int >= adjusted_limit - 1 + stride [using (i)]
2164 // >= limit - 1 + stride [using (AL)]
2165 // >= init + stride [using (iii)]
2166 // >= min_int [using stride > 0, no underflow]
2167 // Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
2168 //
2169 // Note that to prove the base case we need (i) and (ii).
2170 //
2171 // > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
2172 // min_int <= iv_post_j <= max_int
2173 // > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
2174 // min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int
2175 // Proof:
2176 // If iv_post_j >= adjusted_limit:
2177 // We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
2178 // also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.
2179 // If iv_post_j < adjusted_limit:
2180 // First, we note that:
2181 // (iv) iv_post_j <= adjusted_limit - 1
2182 // max_int >= adjusted_limit - 1 + stride [using (i)]
2183 // >= iv_post_j + stride [using (iv)]
2184 // >= min_int [using stride > 0, no underflow]
2185 //
2186 // Note that to prove the step case we only need (i).
2187 //
2188 // Thus, by assuming (i) and (ii), we proved (INV).
2189 //
2190 //
2191 // It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
2192 //
2193 // (1) Loop Limit Check Predicate for (i):
2194 // Using (i): adjusted_limit - 1 + stride <= max_int
2195 //
2196 // This condition is now restated to use limit instead of adjusted_limit:
2197 //
2198 // To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
2199 // max_int - stride + 1 >= adjusted_limit
2200 // We can merge the two constants into
2201 // canonicalized_correction = stride - 1
2202 // which gives us
2203 // max_int - canonicalized_correction >= adjusted_limit
2204 //
2205 // To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
2206 // adjusted_limit = limit + limit_correction
2207 // Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
2208 // max_int - canonicalized_correction - limit_correction >= limit
2209 // Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
2210 // (v) final_correction = canonicalized_correction + limit_correction
2211 //
2212 // which gives us:
2213 //
2214 // Final predicate condition:
2215 // max_int - final_correction >= limit
2216 //
2217 // However, we need to be careful that (v) does not over- or underflow.
2218 // We know that:
2219 // canonicalized_correction = stride - 1
2220 // and
2221 // limit_correction <= stride + 1
2222 // and thus
2223 // canonicalized_correction + limit_correction <= 2 * stride
2224 // To prevent an over- or underflow of (v), we must ensure that
2225 // 2 * stride <= max_int
2226 // which can safely be checked without over- or underflow with
2227 // (vi) stride != min_int AND abs(stride) <= max_int / 2
2228 //
2229 // We could try to further optimize the cases where (vi) does not hold but given that such large strides are
2230 // very uncommon and the loop would only run for a very few iterations anyway, we simply bail out if (vi) fails.
2231 //
2232 // (2) Loop Limit Check Predicate for (ii):
2233 // Using (ii): init < limit
2234 //
2235 // This Loop Limit Check Predicate is not required if we can prove at compile time that either:
2236 // (2.1) type(init) < type(limit)
2237 // In this case, we know:
2238 // all possible values of init < all possible values of limit
2239 // and we can skip the predicate.
2240 //
2241 // (2.2) init < limit is already checked before (i.e. found as a dominating check)
2242 // In this case, we do not need to re-check the condition and can skip the predicate.
2243 // This is often found for while- and for-loops which have the following shape:
2244 //
2245 // if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.
2246 // i = init;
2247 // if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate
2248 // do {
2249 // i += stride;
2250 // } while (i < limit);
2251 // }
2252 //
2253 // (2.3) init + stride <= max_int
2254 // In this case, there is no overflow of the iv phi after the first loop iteration.
2255 // In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
2256 // init < limit
2257 // In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
2258 // compile time that init + stride <= max_int then we have trivially proven the base case and that
2259 // there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
2260 // again and can skip the predicate.
2261
2262 // Check (vi) and bail out if the stride is too big.
2263 if (stride_con == min_signed_integer(iv_bt) || (ABS(stride_con) > max_signed_integer(iv_bt) / 2)) {
2264 return false;
2265 }
2266
2267 // Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
2268 const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
2269
2270 // Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
2271 const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
2272 const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
2273
2274 const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
2275 const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
2276 const jlong final_correction = canonicalized_correction + limit_correction;
2277
2278 int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
2279 Node* init_control = x->in(LoopNode::EntryControl);
2280
2281 // If sov==0, limit's type always satisfies the condition, for
2282 // example, when it is an array length.
2283 if (sov != 0) {
2284 if (sov < 0) {
2285 return false; // Bailout: integer overflow is certain.
2286 }
2287 // (1) Loop Limit Check Predicate is required because we could not statically prove that
2288 // limit + final_correction = adjusted_limit - 1 + stride <= max_int
2289 assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
2290 const Predicates predicates(init_control);
2291 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2292 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2293 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2294 #ifdef ASSERT
2295 if (TraceLoopLimitCheck) {
2296 tty->print("Missing Loop Limit Check Parse Predicate:");
2297 loop->dump_head();
2298 x->dump(1);
2299 }
2300 #endif
2301 return false;
2302 }
2303
2304 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2305 if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {
2306 return false;
2307 }
2308
2309 Node* cmp_limit;
2310 Node* bol;
2311
2312 if (stride_con > 0) {
2313 cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2314 bol = new BoolNode(cmp_limit, BoolTest::le);
2315 } else {
2316 cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2317 bol = new BoolNode(cmp_limit, BoolTest::ge);
2318 }
2319
2320 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2321 }
2322
2323 // (2.3)
2324 const bool init_plus_stride_could_overflow =
2325 (stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2326 (stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2327 // (2.1)
2328 const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2329 (stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2330
2331 if (init_gte_limit && // (2.1)
2332 ((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2333 !has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)
2334 // (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2335 // We use the following condition:
2336 // - stride > 0: init < limit
2337 // - stride < 0: init > limit
2338 //
2339 // This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2340 // a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2341 // check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2342
2343 const Predicates predicates(init_control);
2344 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2345 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2346 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2347 #ifdef ASSERT
2348 if (TraceLoopLimitCheck) {
2349 tty->print("Missing Loop Limit Check Parse Predicate:");
2350 loop->dump_head();
2351 x->dump(1);
2352 }
2353 #endif
2354 return false;
2355 }
2356
2357 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2358 Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2359 if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2360 !is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {
2361 return false;
2362 }
2363
2364 Node* cmp_limit;
2365 Node* bol;
2366
2367 if (stride_con > 0) {
2368 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2369 bol = new BoolNode(cmp_limit, BoolTest::lt);
2370 } else {
2371 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2372 bol = new BoolNode(cmp_limit, BoolTest::gt);
2373 }
2374
2375 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2376 }
2377
2378 if (bt == BoolTest::ne) {
2379 // Now we need to canonicalize the loop condition if it is 'ne'.
2380 assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2381 if (stride_con > 0) {
2382 // 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2383 bt = BoolTest::lt;
2384 } else {
2385 assert(stride_con < 0, "must be");
2386 // 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2387 bt = BoolTest::gt;
2388 }
2389 }
2390
2391 Node* sfpt = nullptr;
2392 if (loop->_child == nullptr) {
2393 sfpt = find_safepoint(back_control, x, loop);
2394 } else {
2395 sfpt = iff->in(0);
2396 if (sfpt->Opcode() != Op_SafePoint) {
2397 sfpt = nullptr;
2398 }
2399 }
2400
2401 if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
2402 Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2403 if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2404 iv_bt == T_LONG) &&
2405 sfpt == nullptr) {
2406 // Leaving the safepoint on the backedge and creating a
2407 // CountedLoop will confuse optimizations. We can't move the
2408 // safepoint around because its jvm state wouldn't match a new
2409 // location. Give up on that loop.
2410 return false;
2411 }
2412 if (is_deleteable_safept(backedge_sfpt)) {
2413 replace_node_and_forward_ctrl(backedge_sfpt, iftrue);
2414 if (loop->_safepts != nullptr) {
2415 loop->_safepts->yank(backedge_sfpt);
2416 }
2417 loop->_tail = iftrue;
2418 }
2419 }
2420
2421
2422 #ifdef ASSERT
2423 if (iv_bt == T_INT &&
2424 !x->as_Loop()->is_loop_nest_inner_loop() &&
2425 StressLongCountedLoop > 0 &&
2426 trunc1 == nullptr &&
2427 convert_to_long_loop(cmp, phi, loop)) {
2428 return false;
2429 }
2430 #endif
2431
2432 Node* adjusted_limit = limit;
2433 if (phi_incr != nullptr) {
2434 // If compare points directly to the phi we need to adjust
2435 // the compare so that it points to the incr. Limit have
2436 // to be adjusted to keep trip count the same and we
2437 // should avoid int overflow.
2438 //
2439 // i = init; do {} while(i++ < limit);
2440 // is converted to
2441 // i = init; do {} while(++i < limit+1);
2442 //
2443 adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2444 }
2445
2446 if (includes_limit) {
2447 // The limit check guaranties that 'limit <= (max_jint - stride)' so
2448 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2449 //
2450 Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2451 adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2452 if (bt == BoolTest::le)
2453 bt = BoolTest::lt;
2454 else if (bt == BoolTest::ge)
2455 bt = BoolTest::gt;
2456 else
2457 ShouldNotReachHere();
2458 }
2459 set_subtree_ctrl(adjusted_limit, false);
2460
2461 // Build a canonical trip test.
2462 // Clone code, as old values may be in use.
2463 incr = incr->clone();
2464 incr->set_req(1,phi);
2465 incr->set_req(2,stride);
2466 incr = _igvn.register_new_node_with_optimizer(incr);
2467 set_early_ctrl(incr, false);
2468 _igvn.rehash_node_delayed(phi);
2469 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2470
2471 // If phi type is more restrictive than Int, raise to
2472 // Int to prevent (almost) infinite recursion in igvn
2473 // which can only handle integer types for constants or minint..maxint.
2474 if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
2475 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2476 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2477 nphi = _igvn.register_new_node_with_optimizer(nphi);
2478 set_ctrl(nphi, get_ctrl(phi));
2479 _igvn.replace_node(phi, nphi);
2480 phi = nphi->as_Phi();
2481 }
2482 cmp = cmp->clone();
2483 cmp->set_req(1,incr);
2484 cmp->set_req(2, adjusted_limit);
2485 cmp = _igvn.register_new_node_with_optimizer(cmp);
2486 set_ctrl(cmp, iff->in(0));
2487
2488 test = test->clone()->as_Bool();
2489 (*(BoolTest*)&test->_test)._test = bt;
2490 test->set_req(1,cmp);
2491 _igvn.register_new_node_with_optimizer(test);
2492 set_ctrl(test, iff->in(0));
2493
2494 // Replace the old IfNode with a new LoopEndNode
2495 Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2496 IfNode *le = lex->as_If();
2497 uint dd = dom_depth(iff);
2498 set_idom(le, le->in(0), dd); // Update dominance for loop exit
2499 set_loop(le, loop);
2500
2501 // Get the loop-exit control
2502 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2503
2504 // Need to swap loop-exit and loop-back control?
2505 if (iftrue_op == Op_IfFalse) {
2506 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2507 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2508
2509 loop->_tail = back_control = ift2;
2510 set_loop(ift2, loop);
2511 set_loop(iff2, get_loop(iffalse));
2512
2513 // Lazy update of 'get_ctrl' mechanism.
2514 replace_node_and_forward_ctrl(iffalse, iff2);
2515 replace_node_and_forward_ctrl(iftrue, ift2);
2516
2517 // Swap names
2518 iffalse = iff2;
2519 iftrue = ift2;
2520 } else {
2521 _igvn.rehash_node_delayed(iffalse);
2522 _igvn.rehash_node_delayed(iftrue);
2523 iffalse->set_req_X( 0, le, &_igvn );
2524 iftrue ->set_req_X( 0, le, &_igvn );
2525 }
2526
2527 set_idom(iftrue, le, dd+1);
2528 set_idom(iffalse, le, dd+1);
2529 assert(iff->outcnt() == 0, "should be dead now");
2530 replace_node_and_forward_ctrl(iff, le); // fix 'get_ctrl'
2531
2532 Node* entry_control = init_control;
2533 bool strip_mine_loop = iv_bt == T_INT &&
2534 loop->_child == nullptr &&
2535 sfpt != nullptr &&
2536 !loop->_has_call &&
2537 is_deleteable_safept(sfpt);
2538 IdealLoopTree* outer_ilt = nullptr;
2539 if (strip_mine_loop) {
2540 outer_ilt = create_outer_strip_mined_loop(init_control, loop, cl_prob, le->_fcnt,
2541 entry_control, iffalse);
2542 }
2543
2544 // Now setup a new CountedLoopNode to replace the existing LoopNode
2545 BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2546 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2547 // The following assert is approximately true, and defines the intention
2548 // of can_be_counted_loop. It fails, however, because phase->type
2549 // is not yet initialized for this loop and its parts.
2550 //assert(l->can_be_counted_loop(this), "sanity");
2551 _igvn.register_new_node_with_optimizer(l);
2552 set_loop(l, loop);
2553 loop->_head = l;
2554 // Fix all data nodes placed at the old loop head.
2555 // Uses the lazy-update mechanism of 'get_ctrl'.
2556 replace_node_and_forward_ctrl(x, l);
2557 set_idom(l, entry_control, dom_depth(entry_control) + 1);
2558
2559 if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {
2560 // Check for immediately preceding SafePoint and remove
2561 if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {
2562 if (strip_mine_loop) {
2563 Node* outer_le = outer_ilt->_tail->in(0);
2564 Node* sfpt_clone = sfpt->clone();
2565 sfpt_clone->set_req(0, iffalse);
2566 outer_le->set_req(0, sfpt_clone);
2567
2568 Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2569 if (polladdr != nullptr && polladdr->is_Load()) {
2570 // Polling load should be pinned outside inner loop.
2571 Node* new_polladdr = polladdr->clone();
2572 new_polladdr->set_req(0, iffalse);
2573 _igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2574 set_ctrl(new_polladdr, iffalse);
2575 sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2576 }
2577 // When this code runs, loop bodies have not yet been populated.
2578 const bool body_populated = false;
2579 register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2580 set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2581 }
2582 replace_node_and_forward_ctrl(sfpt, sfpt->in(TypeFunc::Control));
2583 if (loop->_safepts != nullptr) {
2584 loop->_safepts->yank(sfpt);
2585 }
2586 }
2587 }
2588
2589 #ifdef ASSERT
2590 assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2591 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2592 #endif
2593 #ifndef PRODUCT
2594 if (TraceLoopOpts) {
2595 tty->print("Counted ");
2596 loop->dump_head();
2597 }
2598 #endif
2599
2600 C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2601
2602 // Capture bounds of the loop in the induction variable Phi before
2603 // subsequent transformation (iteration splitting) obscures the
2604 // bounds
2605 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2606
2607 if (strip_mine_loop) {
2608 l->mark_strip_mined();
2609 l->verify_strip_mined(1);
2610 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2611 loop = outer_ilt;
2612 }
2613
2614 #ifndef PRODUCT
2615 if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {
2616 AtomicAccess::inc(&_long_loop_counted_loops);
2617 }
2618 #endif
2619 if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {
2620 l->mark_loop_nest_outer_loop();
2621 }
2622
2623 return true;
2624 }
2625
2626 // Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2627 // If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2628 bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2629 const BasicType iv_bt, Node* loop_entry) {
2630 // Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2631 // successfully find a dominated test with the If node below.
2632 Node* cmp_limit;
2633 Node* bol;
2634 if (stride_con > 0) {
2635 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2636 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2637 } else {
2638 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2639 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2640 }
2641
2642 // Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2643 IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2644 // Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2645 IfFalseNode* if_false = new IfFalseNode(iff);
2646 IfTrueNode* if_true = new IfTrueNode(iff);
2647 Node* dominated_iff = _igvn.transform(iff);
2648 // ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2649 const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2650
2651 // Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2652 _igvn.replace_input_of(iff, 0, C->top());
2653 _igvn.replace_input_of(iff, 1, C->top());
2654 return found_dominating_test;
2655 }
2656
2657 //----------------------exact_limit-------------------------------------------
2658 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
2659 assert(loop->_head->is_CountedLoop(), "");
2660 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2661 assert(cl->is_valid_counted_loop(T_INT), "");
2662
2663 if (cl->stride_con() == 1 ||
2664 cl->stride_con() == -1 ||
2665 cl->limit()->Opcode() == Op_LoopLimit) {
2666 // Old code has exact limit (it could be incorrect in case of int overflow).
2667 // Loop limit is exact with stride == 1. And loop may already have exact limit.
2668 return cl->limit();
2669 }
2670 Node *limit = nullptr;
2671 #ifdef ASSERT
2672 BoolTest::mask bt = cl->loopexit()->test_trip();
2673 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2674 #endif
2675 if (cl->has_exact_trip_count()) {
2676 // Simple case: loop has constant boundaries.
2677 // Use jlongs to avoid integer overflow.
2678 int stride_con = cl->stride_con();
2679 jlong init_con = cl->init_trip()->get_int();
2680 jlong limit_con = cl->limit()->get_int();
2681 julong trip_cnt = cl->trip_count();
2682 jlong final_con = init_con + trip_cnt*stride_con;
2683 int final_int = (int)final_con;
2684 // The final value should be in integer range since the loop
2685 // is counted and the limit was checked for overflow.
2686 assert(final_con == (jlong)final_int, "final value should be integer");
2687 limit = _igvn.intcon(final_int);
2688 } else {
2689 // Create new LoopLimit node to get exact limit (final iv value).
2690 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2691 register_new_node(limit, cl->in(LoopNode::EntryControl));
2692 }
2693 assert(limit != nullptr, "sanity");
2694 return limit;
2695 }
2696
2697 //------------------------------Ideal------------------------------------------
2698 // Return a node which is more "ideal" than the current node.
2699 // Attempt to convert into a counted-loop.
2700 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2701 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
2702 phase->C->set_major_progress();
2703 }
2704 return RegionNode::Ideal(phase, can_reshape);
2705 }
2706
2707 #ifdef ASSERT
2708 void LoopNode::verify_strip_mined(int expect_skeleton) const {
2709 const OuterStripMinedLoopNode* outer = nullptr;
2710 const CountedLoopNode* inner = nullptr;
2711 if (is_strip_mined()) {
2712 if (!is_valid_counted_loop(T_INT)) {
2713 return; // Skip malformed counted loop
2714 }
2715 assert(is_CountedLoop(), "no Loop should be marked strip mined");
2716 inner = as_CountedLoop();
2717 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2718 } else if (is_OuterStripMinedLoop()) {
2719 outer = this->as_OuterStripMinedLoop();
2720 inner = outer->unique_ctrl_out()->as_CountedLoop();
2721 assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2722 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2723 }
2724 if (inner != nullptr || outer != nullptr) {
2725 assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2726 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2727 Node* outer_le = outer_tail->in(0);
2728 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2729 Node* sfpt = outer_le->in(0);
2730 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2731 Node* inner_out = sfpt->in(0);
2732 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2733 assert(cle == inner->loopexit_or_null(), "mismatch");
2734 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2735 if (has_skeleton) {
2736 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2737 assert(outer->outcnt() == 2, "only control nodes");
2738 } else {
2739 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2740 uint phis = 0;
2741 uint be_loads = 0;
2742 Node* be = inner->in(LoopNode::LoopBackControl);
2743 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
2744 Node* u = inner->fast_out(i);
2745 if (u->is_Phi()) {
2746 phis++;
2747 for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
2748 Node* n = be->fast_out(j);
2749 if (n->is_Load()) {
2750 assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2751 do {
2752 n = n->raw_out(0);
2753 } while (!n->is_Phi());
2754 if (n == u) {
2755 be_loads++;
2756 break;
2757 }
2758 }
2759 }
2760 }
2761 }
2762 assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2763 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
2764 Node* u = outer->fast_out(i);
2765 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2766 }
2767 uint stores = 0;
2768 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
2769 Node* u = inner_out->fast_out(i);
2770 if (u->is_Store()) {
2771 stores++;
2772 }
2773 }
2774 // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2775 // not guaranteed to be optimized out.
2776 assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2777 }
2778 assert(sfpt->outcnt() == 1, "no data node");
2779 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2780 }
2781 }
2782 #endif
2783
2784 //=============================================================================
2785 //------------------------------Ideal------------------------------------------
2786 // Return a node which is more "ideal" than the current node.
2787 // Attempt to convert into a counted-loop.
2788 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2789 return RegionNode::Ideal(phase, can_reshape);
2790 }
2791
2792 //------------------------------dump_spec--------------------------------------
2793 // Dump special per-node info
2794 #ifndef PRODUCT
2795 void CountedLoopNode::dump_spec(outputStream *st) const {
2796 LoopNode::dump_spec(st);
2797 if (stride_is_con()) {
2798 st->print("stride: %d ",stride_con());
2799 }
2800 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
2801 if (is_main_loop()) st->print("main of N%d", _idx);
2802 if (is_post_loop()) st->print("post of N%d", _main_idx);
2803 if (is_strip_mined()) st->print(" strip mined");
2804 if (is_multiversion_fast_loop()) { st->print(" multiversion_fast"); }
2805 if (is_multiversion_slow_loop()) { st->print(" multiversion_slow"); }
2806 if (is_multiversion_delayed_slow_loop()) { st->print(" multiversion_delayed_slow"); }
2807 }
2808 #endif
2809
2810 //=============================================================================
2811 jlong BaseCountedLoopEndNode::stride_con() const {
2812 return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2813 }
2814
2815
2816 BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {
2817 if (bt == T_INT) {
2818 return new CountedLoopEndNode(control, test, prob, cnt);
2819 }
2820 assert(bt == T_LONG, "unsupported");
2821 return new LongCountedLoopEndNode(control, test, prob, cnt);
2822 }
2823
2824 //=============================================================================
2825 //------------------------------Value-----------------------------------------
2826 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
2827 const Type* init_t = phase->type(in(Init));
2828 const Type* limit_t = phase->type(in(Limit));
2829 const Type* stride_t = phase->type(in(Stride));
2830 // Either input is TOP ==> the result is TOP
2831 if (init_t == Type::TOP) return Type::TOP;
2832 if (limit_t == Type::TOP) return Type::TOP;
2833 if (stride_t == Type::TOP) return Type::TOP;
2834
2835 int stride_con = stride_t->is_int()->get_con();
2836 if (stride_con == 1)
2837 return bottom_type(); // Identity
2838
2839 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
2840 // Use jlongs to avoid integer overflow.
2841 jlong init_con = init_t->is_int()->get_con();
2842 jlong limit_con = limit_t->is_int()->get_con();
2843 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2844 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2845 jlong final_con = init_con + stride_con*trip_count;
2846 int final_int = (int)final_con;
2847 // The final value should be in integer range in almost all cases,
2848 // since the loop is counted and the limit was checked for overflow.
2849 // There some exceptions, for example:
2850 // - During CCP, there might be a temporary overflow from PhiNodes, see JDK-8309266.
2851 // - During PhaseIdealLoop::split_thru_phi, the LoopLimitNode floats possibly far above
2852 // the loop and its predicates, and we might get constants on one side of the phi that
2853 // would lead to overflows. Such a code path would never lead us to enter the loop
2854 // because of the loop limit overflow check that happens after the LoopLimitNode
2855 // computation with overflow, but before we enter the loop, see JDK-8335747.
2856 if (final_con == (jlong)final_int) {
2857 return TypeInt::make(final_int);
2858 } else {
2859 return bottom_type();
2860 }
2861 }
2862
2863 return bottom_type(); // TypeInt::INT
2864 }
2865
2866 //------------------------------Ideal------------------------------------------
2867 // Return a node which is more "ideal" than the current node.
2868 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2869 if (phase->type(in(Init)) == Type::TOP ||
2870 phase->type(in(Limit)) == Type::TOP ||
2871 phase->type(in(Stride)) == Type::TOP)
2872 return nullptr; // Dead
2873
2874 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2875 if (stride_con == 1)
2876 return nullptr; // Identity
2877
2878 // Delay following optimizations until all loop optimizations
2879 // done to keep Ideal graph simple.
2880 if (!can_reshape || !phase->C->post_loop_opts_phase()) {
2881 phase->C->record_for_post_loop_opts_igvn(this);
2882 return nullptr;
2883 }
2884
2885 const TypeInt* init_t = phase->type(in(Init) )->is_int();
2886 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2887 jlong stride_p;
2888 jlong lim, ini;
2889 julong max;
2890 if (stride_con > 0) {
2891 stride_p = stride_con;
2892 lim = limit_t->_hi;
2893 ini = init_t->_lo;
2894 max = (julong)max_jint;
2895 } else {
2896 stride_p = -(jlong)stride_con;
2897 lim = init_t->_hi;
2898 ini = limit_t->_lo;
2899 max = (julong)(juint)min_jint; // double cast to get 0x0000000080000000, not 0xffffffff80000000
2900 }
2901 julong range = lim - ini + stride_p;
2902 if (range <= max) {
2903 // Convert to integer expression if it is not overflow.
2904 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2905 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2906 Node *bias = phase->transform(new AddINode(range, stride_m));
2907 Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2908 Node *span = phase->transform(new MulINode(trip, in(Stride)));
2909 return new AddINode(span, in(Init)); // exact limit
2910 }
2911
2912 if (is_power_of_2(stride_p) || // divisor is 2^n
2913 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
2914 // Convert to long expression to avoid integer overflow
2915 // and let igvn optimizer convert this division.
2916 //
2917 Node* init = phase->transform( new ConvI2LNode(in(Init)));
2918 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2919 Node* stride = phase->longcon(stride_con);
2920 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2921
2922 Node *range = phase->transform(new SubLNode(limit, init));
2923 Node *bias = phase->transform(new AddLNode(range, stride_m));
2924 Node *span;
2925 if (stride_con > 0 && is_power_of_2(stride_p)) {
2926 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2927 // and avoid generating rounding for division. Zero trip guard should
2928 // guarantee that init < limit but sometimes the guard is missing and
2929 // we can get situation when init > limit. Note, for the empty loop
2930 // optimization zero trip guard is generated explicitly which leaves
2931 // only RCE predicate where exact limit is used and the predicate
2932 // will simply fail forcing recompilation.
2933 Node* neg_stride = phase->longcon(-stride_con);
2934 span = phase->transform(new AndLNode(bias, neg_stride));
2935 } else {
2936 Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2937 span = phase->transform(new MulLNode(trip, stride));
2938 }
2939 // Convert back to int
2940 Node *span_int = phase->transform(new ConvL2INode(span));
2941 return new AddINode(span_int, in(Init)); // exact limit
2942 }
2943
2944 return nullptr; // No progress
2945 }
2946
2947 //------------------------------Identity---------------------------------------
2948 // If stride == 1 return limit node.
2949 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
2950 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2951 if (stride_con == 1 || stride_con == -1)
2952 return in(Limit);
2953 return this;
2954 }
2955
2956 //=============================================================================
2957 //----------------------match_incr_with_optional_truncation--------------------
2958 // Match increment with optional truncation:
2959 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2960 // Return null for failure. Success returns the increment node.
2961 Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2962 const TypeInteger** trunc_type,
2963 BasicType bt) {
2964 // Quick cutouts:
2965 if (expr == nullptr || expr->req() != 3) return nullptr;
2966
2967 Node *t1 = nullptr;
2968 Node *t2 = nullptr;
2969 Node* n1 = expr;
2970 int n1op = n1->Opcode();
2971 const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2972
2973 if (bt == T_INT) {
2974 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
2975 if (n1op == Op_AndI &&
2976 n1->in(2)->is_Con() &&
2977 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
2978 // %%% This check should match any mask of 2**K-1.
2979 t1 = n1;
2980 n1 = t1->in(1);
2981 n1op = n1->Opcode();
2982 trunc_t = TypeInt::CHAR;
2983 } else if (n1op == Op_RShiftI &&
2984 n1->in(1) != nullptr &&
2985 n1->in(1)->Opcode() == Op_LShiftI &&
2986 n1->in(2) == n1->in(1)->in(2) &&
2987 n1->in(2)->is_Con()) {
2988 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2989 // %%% This check should match any shift in [1..31].
2990 if (shift == 16 || shift == 8) {
2991 t1 = n1;
2992 t2 = t1->in(1);
2993 n1 = t2->in(1);
2994 n1op = n1->Opcode();
2995 if (shift == 16) {
2996 trunc_t = TypeInt::SHORT;
2997 } else if (shift == 8) {
2998 trunc_t = TypeInt::BYTE;
2999 }
3000 }
3001 }
3002 }
3003
3004 // If (maybe after stripping) it is an AddI, we won:
3005 if (n1op == Op_Add(bt)) {
3006 *trunc1 = t1;
3007 *trunc2 = t2;
3008 *trunc_type = trunc_t;
3009 return n1;
3010 }
3011
3012 // failed
3013 return nullptr;
3014 }
3015
3016 IfNode* CountedLoopNode::find_multiversion_if_from_multiversion_fast_main_loop() {
3017 assert(is_main_loop() && is_multiversion_fast_loop(), "must be multiversion fast main loop");
3018 CountedLoopEndNode* pre_end = find_pre_loop_end();
3019 if (pre_end == nullptr) { return nullptr; }
3020 Node* pre_entry = pre_end->loopnode()->in(LoopNode::EntryControl);
3021 const Predicates predicates(pre_entry);
3022 IfTrueNode* before_predicates = predicates.entry()->isa_IfTrue();
3023 if (before_predicates != nullptr &&
3024 before_predicates->in(0)->in(1)->is_OpaqueMultiversioning()) {
3025 return before_predicates->in(0)->as_If();
3026 }
3027 return nullptr;
3028 }
3029
3030 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
3031 if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {
3032 verify_strip_mined(expect_skeleton);
3033 return in(EntryControl)->as_Loop();
3034 }
3035 return this;
3036 }
3037
3038 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
3039 assert(is_strip_mined(), "not a strip mined loop");
3040 Node* c = in(EntryControl);
3041 if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {
3042 return nullptr;
3043 }
3044 return c->as_OuterStripMinedLoop();
3045 }
3046
3047 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
3048 Node* c = in(LoopBackControl);
3049 if (c == nullptr || c->is_top()) {
3050 return nullptr;
3051 }
3052 return c->as_IfTrue();
3053 }
3054
3055 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
3056 LoopNode* l = outer_loop();
3057 if (l == nullptr) {
3058 return nullptr;
3059 }
3060 return l->outer_loop_tail();
3061 }
3062
3063 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
3064 IfTrueNode* proj = outer_loop_tail();
3065 if (proj == nullptr) {
3066 return nullptr;
3067 }
3068 Node* c = proj->in(0);
3069 if (c == nullptr || c->is_top() || c->outcnt() != 2) {
3070 return nullptr;
3071 }
3072 return c->as_OuterStripMinedLoopEnd();
3073 }
3074
3075 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
3076 LoopNode* l = outer_loop();
3077 if (l == nullptr) {
3078 return nullptr;
3079 }
3080 return l->outer_loop_end();
3081 }
3082
3083 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
3084 IfNode* le = outer_loop_end();
3085 if (le == nullptr) {
3086 return nullptr;
3087 }
3088 IfFalseNode* c = le->false_proj_or_null();
3089 if (c == nullptr) {
3090 return nullptr;
3091 }
3092 return c->as_IfFalse();
3093 }
3094
3095 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
3096 LoopNode* l = outer_loop();
3097 if (l == nullptr) {
3098 return nullptr;
3099 }
3100 return l->outer_loop_exit();
3101 }
3102
3103 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
3104 IfNode* le = outer_loop_end();
3105 if (le == nullptr) {
3106 return nullptr;
3107 }
3108 Node* c = le->in(0);
3109 if (c == nullptr || c->is_top()) {
3110 return nullptr;
3111 }
3112 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
3113 return c->as_SafePoint();
3114 }
3115
3116 SafePointNode* CountedLoopNode::outer_safepoint() const {
3117 LoopNode* l = outer_loop();
3118 if (l == nullptr) {
3119 return nullptr;
3120 }
3121 return l->outer_safepoint();
3122 }
3123
3124 Node* CountedLoopNode::skip_assertion_predicates_with_halt() {
3125 Node* ctrl = in(LoopNode::EntryControl);
3126 if (ctrl == nullptr) {
3127 // Dying loop.
3128 return nullptr;
3129 }
3130 if (is_main_loop()) {
3131 ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
3132 }
3133 if (is_main_loop() || is_post_loop()) {
3134 AssertionPredicates assertion_predicates(ctrl);
3135 return assertion_predicates.entry();
3136 }
3137 return ctrl;
3138 }
3139
3140
3141 int CountedLoopNode::stride_con() const {
3142 CountedLoopEndNode* cle = loopexit_or_null();
3143 return cle != nullptr ? cle->stride_con() : 0;
3144 }
3145
3146 BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {
3147 if (bt == T_INT) {
3148 return new CountedLoopNode(entry, backedge);
3149 }
3150 assert(bt == T_LONG, "unsupported");
3151 return new LongCountedLoopNode(entry, backedge);
3152 }
3153
3154 void OuterStripMinedLoopNode::fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn,
3155 PhaseIdealLoop* iloop) const {
3156 CountedLoopNode* inner_cl = inner_counted_loop();
3157 IfFalseNode* cle_out = inner_loop_exit();
3158
3159 if (cle_out->outcnt() > 1) {
3160 // Look for chains of stores that were sunk
3161 // out of the inner loop and are in the outer loop
3162 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
3163 Node* u = cle_out->fast_out(i);
3164 if (u->is_Store()) {
3165 int alias_idx = igvn->C->get_alias_index(u->adr_type());
3166 Node* first = u;
3167 for (;;) {
3168 Node* next = first->in(MemNode::Memory);
3169 if (!next->is_Store() || next->in(0) != cle_out) {
3170 break;
3171 }
3172 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
3173 first = next;
3174 }
3175 Node* last = u;
3176 for (;;) {
3177 Node* next = nullptr;
3178 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
3179 Node* uu = last->fast_out(j);
3180 if (uu->is_Store() && uu->in(0) == cle_out) {
3181 assert(next == nullptr, "only one in the outer loop");
3182 next = uu;
3183 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
3184 }
3185 }
3186 if (next == nullptr) {
3187 break;
3188 }
3189 last = next;
3190 }
3191 Node* phi = nullptr;
3192 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
3193 Node* uu = inner_cl->fast_out(j);
3194 if (uu->is_Phi()) {
3195 Node* be = uu->in(LoopNode::LoopBackControl);
3196 if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
3197 assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
3198 }
3199 if (be == last || be == first->in(MemNode::Memory)) {
3200 assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
3201 assert(phi == nullptr, "only one phi");
3202 phi = uu;
3203 }
3204 }
3205 }
3206 #ifdef ASSERT
3207 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
3208 Node* uu = inner_cl->fast_out(j);
3209 if (uu->is_memory_phi()) {
3210 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
3211 assert(phi == uu, "what's that phi?");
3212 } else if (uu->adr_type() == TypePtr::BOTTOM) {
3213 Node* n = uu->in(LoopNode::LoopBackControl);
3214 uint limit = igvn->C->live_nodes();
3215 uint i = 0;
3216 while (n != uu) {
3217 i++;
3218 assert(i < limit, "infinite loop");
3219 if (n->is_Proj()) {
3220 n = n->in(0);
3221 } else if (n->is_SafePoint() || n->is_MemBar()) {
3222 n = n->in(TypeFunc::Memory);
3223 } else if (n->is_Phi()) {
3224 n = n->in(1);
3225 } else if (n->is_MergeMem()) {
3226 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
3227 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
3228 n = n->in(MemNode::Memory);
3229 } else {
3230 n->dump();
3231 ShouldNotReachHere();
3232 }
3233 }
3234 }
3235 }
3236 }
3237 #endif
3238 if (phi == nullptr) {
3239 // If an entire chains was sunk, the
3240 // inner loop has no phi for that memory
3241 // slice, create one for the outer loop
3242 phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
3243 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
3244 phi->set_req(LoopNode::LoopBackControl, last);
3245 phi = register_new_node(phi, inner_cl, igvn, iloop);
3246 igvn->replace_input_of(first, MemNode::Memory, phi);
3247 } else {
3248 // Or fix the outer loop fix to include
3249 // that chain of stores.
3250 Node* be = phi->in(LoopNode::LoopBackControl);
3251 assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
3252 if (be == first->in(MemNode::Memory)) {
3253 if (be == phi->in(LoopNode::LoopBackControl)) {
3254 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
3255 } else {
3256 igvn->replace_input_of(be, MemNode::Memory, last);
3257 }
3258 } else {
3259 #ifdef ASSERT
3260 if (be == phi->in(LoopNode::LoopBackControl)) {
3261 assert(phi->in(LoopNode::LoopBackControl) == last, "");
3262 } else {
3263 assert(be->in(MemNode::Memory) == last, "");
3264 }
3265 #endif
3266 }
3267 }
3268 }
3269 }
3270 }
3271 }
3272
3273 // The outer strip mined loop is initially only partially constructed. In particular Phis are omitted.
3274 // See comment above: PhaseIdealLoop::create_outer_strip_mined_loop()
3275 // We're now in the process of finishing the construction of the outer loop. For each Phi in the inner loop, a Phi in
3276 // the outer loop was just now created. However, Sunk Stores cause an extra challenge:
3277 // 1) If all Stores in the inner loop were sunk for a particular memory slice, there's no Phi left for that memory slice
3278 // 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
3279 // added for each chain of sunk Stores for a particular memory slice.
3280 // 2) If some Stores were sunk and some left in the inner loop, a Phi was already created in the outer loop but
3281 // its backedge input wasn't wired correctly to the last Store of the chain: the backedge input was set to the
3282 // backedge of the inner loop Phi instead, but it needs to be the last Store of the chain in the outer loop. We now
3283 // have to fix that too.
3284 void OuterStripMinedLoopNode::handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn) {
3285 IfFalseNode* cle_exit_proj = inner_loop_exit();
3286
3287 // Find Sunk stores: Sunk stores are pinned on the loop exit projection of the inner loop. Indeed, because Sunk Stores
3288 // modify the memory state captured by the SafePoint in the outer strip mined loop, they must be above it. The
3289 // SafePoint's control input is the loop exit projection. It's also the only control out of the inner loop above the
3290 // SafePoint.
3291 #ifdef ASSERT
3292 int stores_in_outer_loop_cnt = 0;
3293 for (DUIterator_Fast imax, i = cle_exit_proj->fast_outs(imax); i < imax; i++) {
3294 Node* u = cle_exit_proj->fast_out(i);
3295 if (u->is_Store()) {
3296 stores_in_outer_loop_cnt++;
3297 }
3298 }
3299 #endif
3300
3301 // Sunk stores are reachable from the memory state of the outer loop safepoint
3302 SafePointNode* safepoint = outer_safepoint();
3303 MergeMemNode* mm = safepoint->in(TypeFunc::Memory)->isa_MergeMem();
3304 if (mm == nullptr) {
3305 // There is no MergeMem, which should only happen if there was no memory node
3306 // sunk out of the loop.
3307 assert(stores_in_outer_loop_cnt == 0, "inconsistent");
3308 return;
3309 }
3310 DEBUG_ONLY(int stores_in_outer_loop_cnt2 = 0);
3311 for (MergeMemStream mms(mm); mms.next_non_empty();) {
3312 Node* mem = mms.memory();
3313 // Traverse up the chain of stores to find the first store pinned
3314 // at the loop exit projection.
3315 Node* last = mem;
3316 Node* first = nullptr;
3317 while (mem->is_Store() && mem->in(0) == cle_exit_proj) {
3318 DEBUG_ONLY(stores_in_outer_loop_cnt2++);
3319 first = mem;
3320 mem = mem->in(MemNode::Memory);
3321 }
3322 if (first != nullptr) {
3323 // Found a chain of Stores that were sunk
3324 // Do we already have a memory Phi for that slice on the outer loop? If that is the case, that Phi was created
3325 // by cloning an inner loop Phi. The inner loop Phi should have mem, the memory state of the first Store out of
3326 // the inner loop, as input on the backedge. So does the outer loop Phi given it's a clone.
3327 Node* phi = nullptr;
3328 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3329 Node* u = mem->fast_out(i);
3330 if (u->is_Phi() && u->in(0) == this && u->in(LoopBackControl) == mem) {
3331 assert(phi == nullptr, "there should be only one");
3332 phi = u;
3333 PRODUCT_ONLY(break);
3334 }
3335 }
3336 if (phi == nullptr) {
3337 // No outer loop Phi? create one
3338 phi = PhiNode::make(this, last);
3339 phi->set_req(EntryControl, mem);
3340 phi = igvn->transform(phi);
3341 igvn->replace_input_of(first, MemNode::Memory, phi);
3342 } else {
3343 // Fix memory state along the backedge: it should be the last sunk Store of the chain
3344 igvn->replace_input_of(phi, LoopBackControl, last);
3345 }
3346 }
3347 }
3348 assert(stores_in_outer_loop_cnt == stores_in_outer_loop_cnt2, "inconsistent");
3349 }
3350
3351 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
3352 verify_strip_mined(1);
3353 // Look for the outer & inner strip mined loop, reduce number of
3354 // iterations of the inner loop, set exit condition of outer loop,
3355 // construct required phi nodes for outer loop.
3356 CountedLoopNode* inner_cl = inner_counted_loop();
3357 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
3358 if (LoopStripMiningIter == 0) {
3359 remove_outer_loop_and_safepoint(igvn);
3360 return;
3361 }
3362 if (LoopStripMiningIter == 1) {
3363 transform_to_counted_loop(igvn, nullptr);
3364 return;
3365 }
3366 Node* inner_iv_phi = inner_cl->phi();
3367 if (inner_iv_phi == nullptr) {
3368 IfNode* outer_le = outer_loop_end();
3369 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3370 igvn->replace_node(outer_le, iff);
3371 inner_cl->clear_strip_mined();
3372 return;
3373 }
3374 CountedLoopEndNode* inner_cle = inner_counted_loop_end();
3375
3376 int stride = inner_cl->stride_con();
3377 // For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
3378 // except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
3379 // stride, the method is guaranteed to return at the next check below.
3380 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
3381 int scaled_iters = (int)scaled_iters_long;
3382 if ((jlong)scaled_iters != scaled_iters_long) {
3383 // Remove outer loop and safepoint (too few iterations)
3384 remove_outer_loop_and_safepoint(igvn);
3385 return;
3386 }
3387 jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
3388 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
3389 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
3390 assert(iter_estimate > 0, "broken");
3391 if (iter_estimate <= short_scaled_iters) {
3392 // Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
3393 remove_outer_loop_and_safepoint(igvn);
3394 return;
3395 }
3396 if (iter_estimate <= scaled_iters_long) {
3397 // We would only go through one iteration of
3398 // the outer loop: drop the outer loop but
3399 // keep the safepoint so we don't run for
3400 // too long without a safepoint
3401 IfNode* outer_le = outer_loop_end();
3402 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3403 igvn->replace_node(outer_le, iff);
3404 inner_cl->clear_strip_mined();
3405 return;
3406 }
3407
3408 IfTrueNode* cle_tail = inner_cle->true_proj();
3409 ResourceMark rm;
3410 Node_List old_new;
3411 if (cle_tail->outcnt() > 1) {
3412 // Look for nodes on backedge of inner loop and clone them
3413 Unique_Node_List backedge_nodes;
3414 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
3415 Node* u = cle_tail->fast_out(i);
3416 if (u != inner_cl) {
3417 assert(!u->is_CFG(), "control flow on the backedge?");
3418 backedge_nodes.push(u);
3419 }
3420 }
3421 uint last = igvn->C->unique();
3422 for (uint next = 0; next < backedge_nodes.size(); next++) {
3423 Node* n = backedge_nodes.at(next);
3424 old_new.map(n->_idx, n->clone());
3425 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3426 Node* u = n->fast_out(i);
3427 assert(!u->is_CFG(), "broken");
3428 if (u->_idx >= last) {
3429 continue;
3430 }
3431 if (!u->is_Phi()) {
3432 backedge_nodes.push(u);
3433 } else {
3434 assert(u->in(0) == inner_cl, "strange phi on the backedge");
3435 }
3436 }
3437 }
3438 // Put the clones on the outer loop backedge
3439 Node* le_tail = outer_loop_tail();
3440 for (uint next = 0; next < backedge_nodes.size(); next++) {
3441 Node *n = old_new[backedge_nodes.at(next)->_idx];
3442 for (uint i = 1; i < n->req(); i++) {
3443 if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {
3444 n->set_req(i, old_new[n->in(i)->_idx]);
3445 }
3446 }
3447 if (n->in(0) != nullptr && n->in(0) == cle_tail) {
3448 n->set_req(0, le_tail);
3449 }
3450 igvn->register_new_node_with_optimizer(n);
3451 }
3452 }
3453
3454 Node* iv_phi = nullptr;
3455 // Make a clone of each phi in the inner loop for the outer loop
3456 // When Stores were Sunk, after this step, a Phi may still be missing or its backedge incorrectly wired. See
3457 // handle_sunk_stores_when_finishing_construction()
3458 for (uint i = 0; i < inner_cl->outcnt(); i++) {
3459 Node* u = inner_cl->raw_out(i);
3460 if (u->is_Phi()) {
3461 assert(u->in(0) == inner_cl, "inconsistent");
3462 Node* phi = u->clone();
3463 phi->set_req(0, this);
3464 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3465 if (be != nullptr) {
3466 phi->set_req(LoopNode::LoopBackControl, be);
3467 }
3468 phi = igvn->transform(phi);
3469 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3470 if (u == inner_iv_phi) {
3471 iv_phi = phi;
3472 }
3473 }
3474 }
3475
3476 handle_sunk_stores_when_finishing_construction(igvn);
3477
3478 if (iv_phi != nullptr) {
3479 // Now adjust the inner loop's exit condition
3480 Node* limit = inner_cl->limit();
3481 // If limit < init for stride > 0 (or limit > init for stride < 0),
3482 // the loop body is run only once. Given limit - init (init - limit resp.)
3483 // would be negative, the unsigned comparison below would cause
3484 // the loop body to be run for LoopStripMiningIter.
3485 Node* max = nullptr;
3486 if (stride > 0) {
3487 max = MinMaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3488 } else {
3489 max = MinMaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3490 }
3491 // sub is positive and can be larger than the max signed int
3492 // value. Use an unsigned min.
3493 Node* const_iters = igvn->intcon(scaled_iters);
3494 Node* min = MinMaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3495 // min is the number of iterations for the next inner loop execution:
3496 // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3497 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3498
3499 Node* new_limit = nullptr;
3500 if (stride > 0) {
3501 new_limit = igvn->transform(new AddINode(min, iv_phi));
3502 } else {
3503 new_limit = igvn->transform(new SubINode(iv_phi, min));
3504 }
3505 Node* inner_cmp = inner_cle->cmp_node();
3506 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3507 Node* outer_bol = inner_bol;
3508 // cmp node for inner loop may be shared
3509 inner_cmp = inner_cmp->clone();
3510 inner_cmp->set_req(2, new_limit);
3511 inner_bol = inner_bol->clone();
3512 inner_bol->set_req(1, igvn->transform(inner_cmp));
3513 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3514 // Set the outer loop's exit condition too
3515 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3516 } else {
3517 assert(false, "should be able to adjust outer loop");
3518 IfNode* outer_le = outer_loop_end();
3519 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3520 igvn->replace_node(outer_le, iff);
3521 inner_cl->clear_strip_mined();
3522 }
3523 }
3524
3525 void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3526 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3527 CountedLoopEndNode* cle = inner_cl->loopexit();
3528 Node* inner_test = cle->in(1);
3529 IfNode* outer_le = outer_loop_end();
3530 Node* safepoint = outer_safepoint();
3531
3532 fix_sunk_stores_when_back_to_counted_loop(igvn, iloop);
3533
3534 // make counted loop exit test always fail
3535 ConINode* zero = igvn->intcon(0);
3536 if (iloop != nullptr) {
3537 iloop->set_root_as_ctrl(zero);
3538 }
3539 igvn->replace_input_of(cle, 1, zero);
3540 // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3541 Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3542 register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3543 if (iloop == nullptr) {
3544 igvn->replace_node(outer_le, new_end);
3545 } else {
3546 iloop->replace_node_and_forward_ctrl(outer_le, new_end);
3547 }
3548 // the backedge of the inner loop must be rewired to the new loop end
3549 IfTrueNode* backedge = cle->true_proj();
3550 igvn->replace_input_of(backedge, 0, new_end);
3551 if (iloop != nullptr) {
3552 iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3553 }
3554 // make the outer loop go away
3555 igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3556 igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3557 inner_cl->clear_strip_mined();
3558 if (iloop != nullptr) {
3559 Unique_Node_List wq;
3560 wq.push(safepoint);
3561
3562 IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3563 IdealLoopTree* loop = iloop->get_loop(inner_cl);
3564
3565 for (uint i = 0; i < wq.size(); i++) {
3566 Node* n = wq.at(i);
3567 for (uint j = 0; j < n->req(); ++j) {
3568 Node* in = n->in(j);
3569 if (in == nullptr || in->is_CFG()) {
3570 continue;
3571 }
3572 if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {
3573 continue;
3574 }
3575 wq.push(in);
3576 }
3577 assert(!loop->_body.contains(n), "Shouldn't append node to body twice");
3578 loop->_body.push(n);
3579 }
3580 iloop->set_loop(safepoint, loop);
3581 loop->_body.push(safepoint);
3582 iloop->set_loop(safepoint->in(0), loop);
3583 loop->_body.push(safepoint->in(0));
3584 outer_loop_ilt->_tail = igvn->C->top();
3585 }
3586 }
3587
3588 void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {
3589 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3590 Node* outer_sfpt = outer_safepoint();
3591 Node* outer_out = outer_loop_exit();
3592 igvn->replace_node(outer_out, outer_sfpt->in(0));
3593 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3594 inner_cl->clear_strip_mined();
3595 }
3596
3597 Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3598 if (iloop == nullptr) {
3599 return igvn->transform(node);
3600 }
3601 iloop->register_new_node(node, ctrl);
3602 return node;
3603 }
3604
3605 Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3606 PhaseIdealLoop* iloop) {
3607 if (iloop == nullptr) {
3608 return igvn->transform(node);
3609 }
3610 iloop->register_control(node, iloop->get_loop(loop), idom);
3611 return node;
3612 }
3613
3614 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
3615 if (!in(0)) return Type::TOP;
3616 if (phase->type(in(0)) == Type::TOP)
3617 return Type::TOP;
3618
3619 // Until expansion, the loop end condition is not set so this should not constant fold.
3620 if (is_expanded(phase)) {
3621 return IfNode::Value(phase);
3622 }
3623
3624 return TypeTuple::IFBOTH;
3625 }
3626
3627 bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {
3628 // The outer strip mined loop head only has Phi uses after expansion
3629 if (phase->is_IterGVN()) {
3630 IfTrueNode* backedge = true_proj_or_null();
3631 if (backedge != nullptr) {
3632 Node* head = backedge->unique_ctrl_out_or_null();
3633 if (head != nullptr && head->is_OuterStripMinedLoop()) {
3634 if (head->find_out_with(Op_Phi) != nullptr) {
3635 return true;
3636 }
3637 }
3638 }
3639 }
3640 return false;
3641 }
3642
3643 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3644 if (remove_dead_region(phase, can_reshape)) return this;
3645
3646 return nullptr;
3647 }
3648
3649 //------------------------------filtered_type--------------------------------
3650 // Return a type based on condition control flow
3651 // A successful return will be a type that is restricted due
3652 // to a series of dominating if-tests, such as:
3653 // if (i < 10) {
3654 // if (i > 0) {
3655 // here: "i" type is [1..10)
3656 // }
3657 // }
3658 // or a control flow merge
3659 // if (i < 10) {
3660 // do {
3661 // phi( , ) -- at top of loop type is [min_int..10)
3662 // i = ?
3663 // } while ( i < 10)
3664 //
3665 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
3666 assert(n && n->bottom_type()->is_int(), "must be int");
3667 const TypeInt* filtered_t = nullptr;
3668 if (!n->is_Phi()) {
3669 assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3670 filtered_t = filtered_type_from_dominators(n, n_ctrl);
3671
3672 } else {
3673 Node* phi = n->as_Phi();
3674 Node* region = phi->in(0);
3675 assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3676 if (region && region != C->top()) {
3677 for (uint i = 1; i < phi->req(); i++) {
3678 Node* val = phi->in(i);
3679 Node* use_c = region->in(i);
3680 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3681 if (val_t != nullptr) {
3682 if (filtered_t == nullptr) {
3683 filtered_t = val_t;
3684 } else {
3685 filtered_t = filtered_t->meet(val_t)->is_int();
3686 }
3687 }
3688 }
3689 }
3690 }
3691 const TypeInt* n_t = _igvn.type(n)->is_int();
3692 if (filtered_t != nullptr) {
3693 n_t = n_t->join(filtered_t)->is_int();
3694 }
3695 return n_t;
3696 }
3697
3698
3699 //------------------------------filtered_type_from_dominators--------------------------------
3700 // Return a possibly more restrictive type for val based on condition control flow of dominators
3701 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
3702 if (val->is_Con()) {
3703 return val->bottom_type()->is_int();
3704 }
3705 uint if_limit = 10; // Max number of dominating if's visited
3706 const TypeInt* rtn_t = nullptr;
3707
3708 if (use_ctrl && use_ctrl != C->top()) {
3709 Node* val_ctrl = get_ctrl(val);
3710 uint val_dom_depth = dom_depth(val_ctrl);
3711 Node* pred = use_ctrl;
3712 uint if_cnt = 0;
3713 while (if_cnt < if_limit) {
3714 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
3715 if_cnt++;
3716 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3717 if (if_t != nullptr) {
3718 if (rtn_t == nullptr) {
3719 rtn_t = if_t;
3720 } else {
3721 rtn_t = rtn_t->join(if_t)->is_int();
3722 }
3723 }
3724 }
3725 pred = idom(pred);
3726 if (pred == nullptr || pred == C->top()) {
3727 break;
3728 }
3729 // Stop if going beyond definition block of val
3730 if (dom_depth(pred) < val_dom_depth) {
3731 break;
3732 }
3733 }
3734 }
3735 return rtn_t;
3736 }
3737
3738
3739 //------------------------------dump_spec--------------------------------------
3740 // Dump special per-node info
3741 #ifndef PRODUCT
3742 void CountedLoopEndNode::dump_spec(outputStream *st) const {
3743 if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {
3744 BoolTest bt( test_trip()); // Added this for g++.
3745
3746 st->print("[");
3747 bt.dump_on(st);
3748 st->print("]");
3749 }
3750 st->print(" ");
3751 IfNode::dump_spec(st);
3752 }
3753 #endif
3754
3755 IdealLoopTree::IdealLoopTree(PhaseIdealLoop* phase, Node* head, Node* tail): _parent(nullptr), _next(nullptr), _child(nullptr),
3756 _head(head), _tail(tail),
3757 _phase(phase),
3758 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
3759 _body(phase->arena()),
3760 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
3761 _has_range_checks(0), _has_range_checks_computed(0),
3762 _safepts(nullptr),
3763 _required_safept(nullptr),
3764 _allow_optimizations(true) {
3765 precond(_head != nullptr);
3766 precond(_tail != nullptr);
3767 }
3768
3769 //=============================================================================
3770 //------------------------------is_member--------------------------------------
3771 // Is 'l' a member of 'this'?
3772 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
3773 while( l->_nest > _nest ) l = l->_parent;
3774 return l == this;
3775 }
3776
3777 //------------------------------set_nest---------------------------------------
3778 // Set loop tree nesting depth. Accumulate _has_call bits.
3779 int IdealLoopTree::set_nest( uint depth ) {
3780 assert(depth <= SHRT_MAX, "sanity");
3781 _nest = depth;
3782 int bits = _has_call;
3783 if( _child ) bits |= _child->set_nest(depth+1);
3784 if( bits ) _has_call = 1;
3785 if( _next ) bits |= _next ->set_nest(depth );
3786 return bits;
3787 }
3788
3789 //------------------------------split_fall_in----------------------------------
3790 // Split out multiple fall-in edges from the loop header. Move them to a
3791 // private RegionNode before the loop. This becomes the loop landing pad.
3792 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
3793 PhaseIterGVN &igvn = phase->_igvn;
3794 uint i;
3795
3796 // Make a new RegionNode to be the landing pad.
3797 RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3798 phase->set_loop(landing_pad,_parent);
3799 // If _head was irreducible loop entry, landing_pad may now be too
3800 landing_pad->set_loop_status(_head->as_Region()->loop_status());
3801 // Gather all the fall-in control paths into the landing pad
3802 uint icnt = fall_in_cnt;
3803 uint oreq = _head->req();
3804 for( i = oreq-1; i>0; i-- )
3805 if( !phase->is_member( this, _head->in(i) ) )
3806 landing_pad->set_req(icnt--,_head->in(i));
3807
3808 // Peel off PhiNode edges as well
3809 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3810 Node *oj = _head->fast_out(j);
3811 if( oj->is_Phi() ) {
3812 PhiNode* old_phi = oj->as_Phi();
3813 assert( old_phi->region() == _head, "" );
3814 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3815 Node *p = PhiNode::make_blank(landing_pad, old_phi);
3816 uint icnt = fall_in_cnt;
3817 for( i = oreq-1; i>0; i-- ) {
3818 if( !phase->is_member( this, _head->in(i) ) ) {
3819 p->init_req(icnt--, old_phi->in(i));
3820 // Go ahead and clean out old edges from old phi
3821 old_phi->del_req(i);
3822 }
3823 }
3824 // Search for CSE's here, because ZKM.jar does a lot of
3825 // loop hackery and we need to be a little incremental
3826 // with the CSE to avoid O(N^2) node blow-up.
3827 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3828 if( p2 ) { // Found CSE
3829 p->destruct(&igvn); // Recover useless new node
3830 p = p2; // Use old node
3831 } else {
3832 igvn.register_new_node_with_optimizer(p, old_phi);
3833 }
3834 // Make old Phi refer to new Phi.
3835 old_phi->add_req(p);
3836 // Check for the special case of making the old phi useless and
3837 // disappear it. In JavaGrande I have a case where this useless
3838 // Phi is the loop limit and prevents recognizing a CountedLoop
3839 // which in turn prevents removing an empty loop.
3840 Node *id_old_phi = old_phi->Identity(&igvn);
3841 if( id_old_phi != old_phi ) { // Found a simple identity?
3842 // Note that I cannot call 'replace_node' here, because
3843 // that will yank the edge from old_phi to the Region and
3844 // I'm mid-iteration over the Region's uses.
3845 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
3846 Node* use = old_phi->last_out(i);
3847 igvn.rehash_node_delayed(use);
3848 uint uses_found = 0;
3849 for (uint j = 0; j < use->len(); j++) {
3850 if (use->in(j) == old_phi) {
3851 if (j < use->req()) use->set_req (j, id_old_phi);
3852 else use->set_prec(j, id_old_phi);
3853 uses_found++;
3854 }
3855 }
3856 i -= uses_found; // we deleted 1 or more copies of this edge
3857 }
3858 }
3859 igvn._worklist.push(old_phi);
3860 }
3861 }
3862 // Finally clean out the fall-in edges from the RegionNode
3863 for( i = oreq-1; i>0; i-- ) {
3864 if( !phase->is_member( this, _head->in(i) ) ) {
3865 _head->del_req(i);
3866 }
3867 }
3868 igvn.rehash_node_delayed(_head);
3869 // Transform landing pad
3870 igvn.register_new_node_with_optimizer(landing_pad, _head);
3871 // Insert landing pad into the header
3872 _head->add_req(landing_pad);
3873 }
3874
3875 //------------------------------split_outer_loop-------------------------------
3876 // Split out the outermost loop from this shared header.
3877 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
3878 PhaseIterGVN &igvn = phase->_igvn;
3879
3880 // Find index of outermost loop; it should also be my tail.
3881 uint outer_idx = 1;
3882 while( _head->in(outer_idx) != _tail ) outer_idx++;
3883
3884 // Make a LoopNode for the outermost loop.
3885 Node *ctl = _head->in(LoopNode::EntryControl);
3886 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3887 outer = igvn.register_new_node_with_optimizer(outer, _head);
3888 phase->set_created_loop_node();
3889
3890 // Outermost loop falls into '_head' loop
3891 _head->set_req(LoopNode::EntryControl, outer);
3892 _head->del_req(outer_idx);
3893 // Split all the Phis up between '_head' loop and 'outer' loop.
3894 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3895 Node *out = _head->fast_out(j);
3896 if( out->is_Phi() ) {
3897 PhiNode *old_phi = out->as_Phi();
3898 assert( old_phi->region() == _head, "" );
3899 Node *phi = PhiNode::make_blank(outer, old_phi);
3900 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3901 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3902 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3903 // Make old Phi point to new Phi on the fall-in path
3904 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3905 old_phi->del_req(outer_idx);
3906 }
3907 }
3908
3909 // Use the new loop head instead of the old shared one
3910 _head = outer;
3911 phase->set_loop(_head, this);
3912 }
3913
3914 //------------------------------fix_parent-------------------------------------
3915 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
3916 loop->_parent = parent;
3917 if( loop->_child ) fix_parent( loop->_child, loop );
3918 if( loop->_next ) fix_parent( loop->_next , parent );
3919 }
3920
3921 //------------------------------estimate_path_freq-----------------------------
3922 static float estimate_path_freq( Node *n ) {
3923 // Try to extract some path frequency info
3924 IfNode *iff;
3925 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
3926 uint nop = n->Opcode();
3927 if( nop == Op_SafePoint ) { // Skip any safepoint
3928 n = n->in(0);
3929 continue;
3930 }
3931 if( nop == Op_CatchProj ) { // Get count from a prior call
3932 // Assume call does not always throw exceptions: means the call-site
3933 // count is also the frequency of the fall-through path.
3934 assert( n->is_CatchProj(), "" );
3935 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3936 return 0.0f; // Assume call exception path is rare
3937 Node *call = n->in(0)->in(0)->in(0);
3938 assert( call->is_Call(), "expect a call here" );
3939 const JVMState *jvms = ((CallNode*)call)->jvms();
3940 ciMethodData* methodData = jvms->method()->method_data();
3941 if (!methodData->is_mature()) return 0.0f; // No call-site data
3942 ciProfileData* data = methodData->bci_to_data(jvms->bci());
3943 if ((data == nullptr) || !data->is_CounterData()) {
3944 // no call profile available, try call's control input
3945 n = n->in(0);
3946 continue;
3947 }
3948 return data->as_CounterData()->count()/FreqCountInvocations;
3949 }
3950 // See if there's a gating IF test
3951 Node *n_c = n->in(0);
3952 if( !n_c->is_If() ) break; // No estimate available
3953 iff = n_c->as_If();
3954 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3955 // Compute how much count comes on this path
3956 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3957 // Have no count info. Skip dull uncommon-trap like branches.
3958 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3959 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3960 break;
3961 // Skip through never-taken branch; look for a real loop exit.
3962 n = iff->in(0);
3963 }
3964 return 0.0f; // No estimate available
3965 }
3966
3967 //------------------------------merge_many_backedges---------------------------
3968 // Merge all the backedges from the shared header into a private Region.
3969 // Feed that region as the one backedge to this loop.
3970 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
3971 uint i;
3972
3973 // Scan for the top 2 hottest backedges
3974 float hotcnt = 0.0f;
3975 float warmcnt = 0.0f;
3976 uint hot_idx = 0;
3977 // Loop starts at 2 because slot 1 is the fall-in path
3978 for( i = 2; i < _head->req(); i++ ) {
3979 float cnt = estimate_path_freq(_head->in(i));
3980 if( cnt > hotcnt ) { // Grab hottest path
3981 warmcnt = hotcnt;
3982 hotcnt = cnt;
3983 hot_idx = i;
3984 } else if( cnt > warmcnt ) { // And 2nd hottest path
3985 warmcnt = cnt;
3986 }
3987 }
3988
3989 // See if the hottest backedge is worthy of being an inner loop
3990 // by being much hotter than the next hottest backedge.
3991 if( hotcnt <= 0.0001 ||
3992 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3993
3994 // Peel out the backedges into a private merge point; peel
3995 // them all except optionally hot_idx.
3996 PhaseIterGVN &igvn = phase->_igvn;
3997
3998 Node *hot_tail = nullptr;
3999 // Make a Region for the merge point
4000 Node *r = new RegionNode(1);
4001 for( i = 2; i < _head->req(); i++ ) {
4002 if( i != hot_idx )
4003 r->add_req( _head->in(i) );
4004 else hot_tail = _head->in(i);
4005 }
4006 igvn.register_new_node_with_optimizer(r, _head);
4007 // Plug region into end of loop _head, followed by hot_tail
4008 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
4009 igvn.replace_input_of(_head, 2, r);
4010 if( hot_idx ) _head->add_req(hot_tail);
4011
4012 // Split all the Phis up between '_head' loop and the Region 'r'
4013 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
4014 Node *out = _head->fast_out(j);
4015 if( out->is_Phi() ) {
4016 PhiNode* n = out->as_Phi();
4017 igvn.hash_delete(n); // Delete from hash before hacking edges
4018 Node *hot_phi = nullptr;
4019 Node *phi = new PhiNode(r, n->type(), n->adr_type());
4020 // Check all inputs for the ones to peel out
4021 uint j = 1;
4022 for( uint i = 2; i < n->req(); i++ ) {
4023 if( i != hot_idx )
4024 phi->set_req( j++, n->in(i) );
4025 else hot_phi = n->in(i);
4026 }
4027 // Register the phi but do not transform until whole place transforms
4028 igvn.register_new_node_with_optimizer(phi, n);
4029 // Add the merge phi to the old Phi
4030 while( n->req() > 3 ) n->del_req( n->req()-1 );
4031 igvn.replace_input_of(n, 2, phi);
4032 if( hot_idx ) n->add_req(hot_phi);
4033 }
4034 }
4035
4036
4037 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
4038 // of self loop tree. Turn self into a loop headed by _head and with
4039 // tail being the new merge point.
4040 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
4041 phase->set_loop(_tail,ilt); // Adjust tail
4042 _tail = r; // Self's tail is new merge point
4043 phase->set_loop(r,this);
4044 ilt->_child = _child; // New guy has my children
4045 _child = ilt; // Self has new guy as only child
4046 ilt->_parent = this; // new guy has self for parent
4047 ilt->_nest = _nest; // Same nesting depth (for now)
4048
4049 // Starting with 'ilt', look for child loop trees using the same shared
4050 // header. Flatten these out; they will no longer be loops in the end.
4051 IdealLoopTree **pilt = &_child;
4052 while( ilt ) {
4053 if( ilt->_head == _head ) {
4054 uint i;
4055 for( i = 2; i < _head->req(); i++ )
4056 if( _head->in(i) == ilt->_tail )
4057 break; // Still a loop
4058 if( i == _head->req() ) { // No longer a loop
4059 // Flatten ilt. Hang ilt's "_next" list from the end of
4060 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
4061 IdealLoopTree **cp = &ilt->_child;
4062 while( *cp ) cp = &(*cp)->_next; // Find end of child list
4063 *cp = ilt->_next; // Hang next list at end of child list
4064 *pilt = ilt->_child; // Move child up to replace ilt
4065 ilt->_head = nullptr; // Flag as a loop UNIONED into parent
4066 ilt = ilt->_child; // Repeat using new ilt
4067 continue; // do not advance over ilt->_child
4068 }
4069 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
4070 phase->set_loop(_head,ilt);
4071 }
4072 pilt = &ilt->_child; // Advance to next
4073 ilt = *pilt;
4074 }
4075
4076 if( _child ) fix_parent( _child, this );
4077 }
4078
4079 //------------------------------beautify_loops---------------------------------
4080 // Split shared headers and insert loop landing pads.
4081 // Insert a LoopNode to replace the RegionNode.
4082 // Return TRUE if loop tree is structurally changed.
4083 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
4084 bool result = false;
4085 // Cache parts in locals for easy
4086 PhaseIterGVN &igvn = phase->_igvn;
4087
4088 igvn.hash_delete(_head); // Yank from hash before hacking edges
4089
4090 // Check for multiple fall-in paths. Peel off a landing pad if need be.
4091 int fall_in_cnt = 0;
4092 for( uint i = 1; i < _head->req(); i++ )
4093 if( !phase->is_member( this, _head->in(i) ) )
4094 fall_in_cnt++;
4095 assert( fall_in_cnt, "at least 1 fall-in path" );
4096 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
4097 split_fall_in( phase, fall_in_cnt );
4098
4099 // Swap inputs to the _head and all Phis to move the fall-in edge to
4100 // the left.
4101 fall_in_cnt = 1;
4102 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
4103 fall_in_cnt++;
4104 if( fall_in_cnt > 1 ) {
4105 // Since I am just swapping inputs I do not need to update def-use info
4106 Node *tmp = _head->in(1);
4107 igvn.rehash_node_delayed(_head);
4108 _head->set_req( 1, _head->in(fall_in_cnt) );
4109 _head->set_req( fall_in_cnt, tmp );
4110 // Swap also all Phis
4111 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
4112 Node* phi = _head->fast_out(i);
4113 if( phi->is_Phi() ) {
4114 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
4115 tmp = phi->in(1);
4116 phi->set_req( 1, phi->in(fall_in_cnt) );
4117 phi->set_req( fall_in_cnt, tmp );
4118 }
4119 }
4120 }
4121 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
4122 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
4123
4124 // If I am a shared header (multiple backedges), peel off the many
4125 // backedges into a private merge point and use the merge point as
4126 // the one true backedge.
4127 if (_head->req() > 3) {
4128 // Merge the many backedges into a single backedge but leave
4129 // the hottest backedge as separate edge for the following peel.
4130 if (!_irreducible) {
4131 merge_many_backedges( phase );
4132 }
4133
4134 // When recursively beautify my children, split_fall_in can change
4135 // loop tree structure when I am an irreducible loop. Then the head
4136 // of my children has a req() not bigger than 3. Here we need to set
4137 // result to true to catch that case in order to tell the caller to
4138 // rebuild loop tree. See issue JDK-8244407 for details.
4139 result = true;
4140 }
4141
4142 // If I have one hot backedge, peel off myself loop.
4143 // I better be the outermost loop.
4144 if (_head->req() > 3 && !_irreducible) {
4145 split_outer_loop( phase );
4146 result = true;
4147
4148 } else if (!_head->is_Loop() && !_irreducible) {
4149 // Make a new LoopNode to replace the old loop head
4150 Node *l = new LoopNode( _head->in(1), _head->in(2) );
4151 l = igvn.register_new_node_with_optimizer(l, _head);
4152 phase->set_created_loop_node();
4153 // Go ahead and replace _head
4154 phase->_igvn.replace_node( _head, l );
4155 _head = l;
4156 phase->set_loop(_head, this);
4157 }
4158
4159 // Now recursively beautify nested loops
4160 if( _child ) result |= _child->beautify_loops( phase );
4161 if( _next ) result |= _next ->beautify_loops( phase );
4162 return result;
4163 }
4164
4165 //------------------------------allpaths_check_safepts----------------------------
4166 // Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
4167 // safepoint encountered. Helper for check_safepts.
4168 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
4169 assert(stack.size() == 0, "empty stack");
4170 stack.push(_head);
4171 visited.clear();
4172 visited.set(_head->_idx);
4173 while (stack.size() > 0) {
4174 Node* n = stack.pop();
4175 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
4176 // Terminate this path
4177 } else if (n->Opcode() == Op_SafePoint) {
4178 if (_phase->get_loop(n) != this) {
4179 if (_required_safept == nullptr) _required_safept = new Node_List();
4180 // save the first we run into on that path: closest to the tail if the head has a single backedge
4181 _required_safept->push(n);
4182 }
4183 // Terminate this path
4184 } else {
4185 uint start = n->is_Region() ? 1 : 0;
4186 uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
4187 for (uint i = start; i < end; i++) {
4188 Node* in = n->in(i);
4189 assert(in->is_CFG(), "must be");
4190 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
4191 stack.push(in);
4192 }
4193 }
4194 }
4195 }
4196 }
4197
4198 //------------------------------check_safepts----------------------------
4199 // Given dominators, try to find loops with calls that must always be
4200 // executed (call dominates loop tail). These loops do not need non-call
4201 // safepoints (ncsfpt).
4202 //
4203 // A complication is that a safepoint in a inner loop may be needed
4204 // by an outer loop. In the following, the inner loop sees it has a
4205 // call (block 3) on every path from the head (block 2) to the
4206 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
4207 // in block 2, _but_ this leaves the outer loop without a safepoint.
4208 //
4209 // entry 0
4210 // |
4211 // v
4212 // outer 1,2,4 +-> 1
4213 // | \
4214 // | v
4215 // inner 2,3 | 2 <---+ ncsfpt in 2
4216 // | / \ |
4217 // | v v |
4218 // | 4 3 | call in 3
4219 // |_/ \ \_|
4220 // |
4221 // v
4222 // exit 5
4223 //
4224 // This method maintains a list (_required_safept) of ncsfpts that must
4225 // be protected for each loop. It only marks ncsfpts for prevervation,
4226 // and does not actually delete any of them.
4227 //
4228 // If some other method needs to delete a ncsfpt later, it will make sure
4229 // the ncsfpt is not in the list of all outer loops of the current loop.
4230 // See `PhaseIdealLoop::is_deleteable_safept`.
4231 //
4232 // The insights into the problem:
4233 // A) Counted loops are okay (i.e. do not need to preserve ncsfpts),
4234 // they will be handled in `IdealLoopTree::counted_loop`
4235 // B) Innermost loops are okay because there's no inner loops that can
4236 // delete their ncsfpts. Only outer loops need to mark safepoints for
4237 // protection, because only loops further in can accidentally delete
4238 // their ncsfpts
4239 // C) If an outer loop has a call that's guaranteed to execute (on the
4240 // idom-path), then that loop is okay. Because the call will always
4241 // perform a safepoint poll, regardless of what safepoints are deleted
4242 // from its inner loops
4243 // D) Similarly, if an outer loop has a ncsfpt on the idom-path that isn't
4244 // inside any nested loop, then that loop is okay
4245 // E) Otherwise, if an outer loop's ncsfpt on the idom-path is nested in
4246 // an inner loop, we need to prevent the inner loop from deleting it
4247 //
4248 // There are two analyses:
4249 // 1) The first, and cheaper one, scans the loop body from
4250 // tail to head following the idom (immediate dominator)
4251 // chain, looking for the cases (C,D,E) above.
4252 // Since inner loops are scanned before outer loops, there is summary
4253 // information about inner loops. Inner loops can be skipped over
4254 // when the tail of an inner loop is encountered.
4255 //
4256 // 2) The second, invoked if the first fails to find a call or ncsfpt on
4257 // the idom path (which is rare), scans all predecessor control paths
4258 // from the tail to the head, terminating a path when a call or sfpt
4259 // is encountered, to find the ncsfpt's that are closest to the tail.
4260 //
4261 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
4262 // Bottom up traversal
4263 IdealLoopTree* ch = _child;
4264 if (_child) _child->check_safepts(visited, stack);
4265 if (_next) _next ->check_safepts(visited, stack);
4266
4267 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {
4268 bool has_call = false; // call on dom-path
4269 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
4270 Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
4271 if (!_irreducible) {
4272 // Scan the dom-path nodes from tail to head
4273 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
4274 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
4275 has_call = true;
4276 _has_sfpt = 1; // Then no need for a safept!
4277 break;
4278 } else if (n->Opcode() == Op_SafePoint) {
4279 if (_phase->get_loop(n) == this) {
4280 // We found a local ncsfpt.
4281 // Continue searching for a call that is guaranteed to be a safepoint.
4282 has_local_ncsfpt = true;
4283 } else if (nonlocal_ncsfpt == nullptr) {
4284 nonlocal_ncsfpt = n; // save the one closest to the tail
4285 }
4286 } else {
4287 IdealLoopTree* nlpt = _phase->get_loop(n);
4288 if (this != nlpt) {
4289 // If at an inner loop tail, see if the inner loop has already
4290 // recorded seeing a call on the dom-path (and stop.) If not,
4291 // jump to the head of the inner loop.
4292 assert(is_member(nlpt), "nested loop");
4293 Node* tail = nlpt->_tail;
4294 if (tail->in(0)->is_If()) tail = tail->in(0);
4295 if (n == tail) {
4296 // If inner loop has call on dom-path, so does outer loop
4297 if (nlpt->_has_sfpt) {
4298 has_call = true;
4299 _has_sfpt = 1;
4300 break;
4301 }
4302 // Skip to head of inner loop
4303 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
4304 n = nlpt->_head;
4305 if (_head == n) {
4306 // this and nlpt (inner loop) have the same loop head. This should not happen because
4307 // during beautify_loops we call merge_many_backedges. However, infinite loops may not
4308 // have been attached to the loop-tree during build_loop_tree before beautify_loops,
4309 // but then attached in the build_loop_tree afterwards, and so still have unmerged
4310 // backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
4311 // since we have reached the loop head of this.
4312 assert(_head->as_Region()->is_in_infinite_subgraph(),
4313 "only expect unmerged backedges in infinite loops");
4314 break;
4315 }
4316 }
4317 }
4318 }
4319 }
4320 }
4321 // Record safept's that this loop needs preserved when an
4322 // inner loop attempts to delete it's safepoints.
4323 if (_child != nullptr && !has_call && !has_local_ncsfpt) {
4324 if (nonlocal_ncsfpt != nullptr) {
4325 if (_required_safept == nullptr) _required_safept = new Node_List();
4326 _required_safept->push(nonlocal_ncsfpt);
4327 } else {
4328 // Failed to find a suitable safept on the dom-path. Now use
4329 // an all paths walk from tail to head, looking for safepoints to preserve.
4330 allpaths_check_safepts(visited, stack);
4331 }
4332 }
4333 }
4334 }
4335
4336 //---------------------------is_deleteable_safept----------------------------
4337 // Is safept not required by an outer loop?
4338 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
4339 assert(sfpt->Opcode() == Op_SafePoint, "");
4340 IdealLoopTree* lp = get_loop(sfpt)->_parent;
4341 while (lp != nullptr) {
4342 Node_List* sfpts = lp->_required_safept;
4343 if (sfpts != nullptr) {
4344 for (uint i = 0; i < sfpts->size(); i++) {
4345 if (sfpt == sfpts->at(i))
4346 return false;
4347 }
4348 }
4349 lp = lp->_parent;
4350 }
4351 return true;
4352 }
4353
4354 //---------------------------replace_parallel_iv-------------------------------
4355 // Replace parallel induction variable (parallel to trip counter)
4356 // This optimization looks for patterns similar to:
4357 //
4358 // int a = init2;
4359 // for (int iv = init; iv < limit; iv += stride_con) {
4360 // a += stride_con2;
4361 // }
4362 //
4363 // and transforms it to:
4364 //
4365 // int iv2 = init2
4366 // int iv = init
4367 // loop:
4368 // if (iv >= limit) goto exit
4369 // iv += stride_con
4370 // iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
4371 // goto loop
4372 // exit:
4373 // ...
4374 //
4375 // Such transformation introduces more optimization opportunities. In this
4376 // particular example, the loop can be eliminated entirely given that
4377 // `stride_con2 / stride_con` is exact (i.e., no remainder). Checks are in
4378 // place to only perform this optimization if such a division is exact. This
4379 // example will be transformed into its semantic equivalence:
4380 //
4381 // int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
4382 //
4383 // which corresponds to the structure of transformed subgraph.
4384 //
4385 // However, if there is a mismatch between types of the loop and the parallel
4386 // induction variable (e.g., a long-typed IV in an int-typed loop), type
4387 // conversions are required:
4388 //
4389 // long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
4390 //
4391 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
4392 assert(loop->_head->is_CountedLoop(), "");
4393 CountedLoopNode *cl = loop->_head->as_CountedLoop();
4394 if (!cl->is_valid_counted_loop(T_INT)) {
4395 return; // skip malformed counted loop
4396 }
4397 Node *incr = cl->incr();
4398 if (incr == nullptr) {
4399 return; // Dead loop?
4400 }
4401 Node *init = cl->init_trip();
4402 Node *phi = cl->phi();
4403 jlong stride_con = cl->stride_con();
4404
4405 // Visit all children, looking for Phis
4406 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
4407 Node *out = cl->out(i);
4408 // Look for other phis (secondary IVs). Skip dead ones
4409 if (!out->is_Phi() || out == phi || !has_node(out)) {
4410 continue;
4411 }
4412
4413 PhiNode* phi2 = out->as_Phi();
4414 Node* incr2 = phi2->in(LoopNode::LoopBackControl);
4415 // Look for induction variables of the form: X += constant
4416 if (phi2->region() != loop->_head ||
4417 incr2->req() != 3 ||
4418 incr2->in(1)->uncast() != phi2 ||
4419 incr2 == incr ||
4420 (incr2->Opcode() != Op_AddI && incr2->Opcode() != Op_AddL) ||
4421 !incr2->in(2)->is_Con()) {
4422 continue;
4423 }
4424
4425 if (incr2->in(1)->is_ConstraintCast() &&
4426 !(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {
4427 // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
4428 continue;
4429 }
4430 // Check for parallel induction variable (parallel to trip counter)
4431 // via an affine function. In particular, count-down loops with
4432 // count-up array indices are common. We only RCE references off
4433 // the trip-counter, so we need to convert all these to trip-counter
4434 // expressions.
4435 Node* init2 = phi2->in(LoopNode::EntryControl);
4436
4437 // Determine the basic type of the stride constant (and the iv being incremented).
4438 BasicType stride_con2_bt = incr2->Opcode() == Op_AddI ? T_INT : T_LONG;
4439 jlong stride_con2 = incr2->in(2)->get_integer_as_long(stride_con2_bt);
4440
4441 // The ratio of the two strides cannot be represented as an int
4442 // if stride_con2 is min_jint (or min_jlong, respectively) and
4443 // stride_con is -1.
4444 if (stride_con2 == min_signed_integer(stride_con2_bt) && stride_con == -1) {
4445 continue;
4446 }
4447
4448 // The general case here gets a little tricky. We want to find the
4449 // GCD of all possible parallel IV's and make a new IV using this
4450 // GCD for the loop. Then all possible IVs are simple multiples of
4451 // the GCD. In practice, this will cover very few extra loops.
4452 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
4453 // where +/-1 is the common case, but other integer multiples are
4454 // also easy to handle.
4455 jlong ratio_con = stride_con2 / stride_con;
4456
4457 if ((ratio_con * stride_con) != stride_con2) { // Check for exact (no remainder)
4458 continue;
4459 }
4460
4461 #ifndef PRODUCT
4462 if (TraceLoopOpts) {
4463 tty->print("Parallel IV: %d ", phi2->_idx);
4464 loop->dump_head();
4465 }
4466 #endif
4467
4468 // Convert to using the trip counter. The parallel induction
4469 // variable differs from the trip counter by a loop-invariant
4470 // amount, the difference between their respective initial values.
4471 // It is scaled by the 'ratio_con'.
4472 Node* ratio = integercon(ratio_con, stride_con2_bt);
4473
4474 Node* init_converted = insert_convert_node_if_needed(stride_con2_bt, init);
4475 Node* phi_converted = insert_convert_node_if_needed(stride_con2_bt, phi);
4476
4477 Node* ratio_init = MulNode::make(init_converted, ratio, stride_con2_bt);
4478 _igvn.register_new_node_with_optimizer(ratio_init, init_converted);
4479 set_early_ctrl(ratio_init, false);
4480
4481 Node* diff = SubNode::make(init2, ratio_init, stride_con2_bt);
4482 _igvn.register_new_node_with_optimizer(diff, init2);
4483 set_early_ctrl(diff, false);
4484
4485 Node* ratio_idx = MulNode::make(phi_converted, ratio, stride_con2_bt);
4486 _igvn.register_new_node_with_optimizer(ratio_idx, phi_converted);
4487 set_ctrl(ratio_idx, cl);
4488
4489 Node* add = AddNode::make(ratio_idx, diff, stride_con2_bt);
4490 _igvn.register_new_node_with_optimizer(add);
4491 set_ctrl(add, cl);
4492
4493 _igvn.replace_node( phi2, add );
4494 // Sometimes an induction variable is unused
4495 if (add->outcnt() == 0) {
4496 _igvn.remove_dead_node(add);
4497 }
4498 --i; // deleted this phi; rescan starting with next position
4499 }
4500 }
4501
4502 Node* PhaseIdealLoop::insert_convert_node_if_needed(BasicType target, Node* input) {
4503 BasicType source = _igvn.type(input)->basic_type();
4504 if (source == target) {
4505 return input;
4506 }
4507
4508 Node* converted = ConvertNode::create_convert(source, target, input);
4509 _igvn.register_new_node_with_optimizer(converted, input);
4510 set_early_ctrl(converted, false);
4511
4512 return converted;
4513 }
4514
4515 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
4516 Node* keep = nullptr;
4517 if (keep_one) {
4518 // Look for a safepoint on the idom-path.
4519 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
4520 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
4521 keep = i;
4522 break; // Found one
4523 }
4524 }
4525 }
4526
4527 // Don't remove any safepoints if it is requested to keep a single safepoint and
4528 // no safepoint was found on idom-path. It is not safe to remove any safepoint
4529 // in this case since there's no safepoint dominating all paths in the loop body.
4530 bool prune = !keep_one || keep != nullptr;
4531
4532 // Delete other safepoints in this loop.
4533 Node_List* sfpts = _safepts;
4534 if (prune && sfpts != nullptr) {
4535 assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4536 for (uint i = 0; i < sfpts->size(); i++) {
4537 Node* n = sfpts->at(i);
4538 assert(phase->get_loop(n) == this, "");
4539 if (n != keep && phase->is_deleteable_safept(n)) {
4540 phase->replace_node_and_forward_ctrl(n, n->in(TypeFunc::Control));
4541 }
4542 }
4543 }
4544 }
4545
4546 //------------------------------counted_loop-----------------------------------
4547 // Convert to counted loops where possible
4548 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
4549
4550 // For grins, set the inner-loop flag here
4551 if (!_child) {
4552 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4553 }
4554
4555 IdealLoopTree* loop = this;
4556 if (_head->is_CountedLoop() ||
4557 phase->is_counted_loop(_head, loop, T_INT)) {
4558
4559 if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {
4560 // Indicate we do not need a safepoint here
4561 _has_sfpt = 1;
4562 }
4563
4564 // Remove safepoints
4565 bool keep_one_sfpt = !(_has_call || _has_sfpt);
4566 remove_safepoints(phase, keep_one_sfpt);
4567
4568 // Look for induction variables
4569 phase->replace_parallel_iv(this);
4570 } else if (_head->is_LongCountedLoop() ||
4571 phase->is_counted_loop(_head, loop, T_LONG)) {
4572 remove_safepoints(phase, true);
4573 } else {
4574 assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4575 if (_parent != nullptr && !_irreducible) {
4576 // Not a counted loop. Keep one safepoint.
4577 bool keep_one_sfpt = true;
4578 remove_safepoints(phase, keep_one_sfpt);
4579 }
4580 }
4581
4582 // Recursively
4583 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4584 assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4585 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4586 if (loop->_next) loop->_next ->counted_loop(phase);
4587 }
4588
4589
4590 // The Estimated Loop Clone Size:
4591 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4592 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4593 // respectively, used to ensure that the node usage estimates made are on the
4594 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
4595 // possible additional/excessive nodes generated due to data and control flow
4596 // merging, for edges reaching outside the loop.
4597 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
4598
4599 precond(0 < factor && factor < 16);
4600
4601 uint const bc = 13;
4602 uint const cc = 17;
4603 uint const sz = _body.size() + (_body.size() + 7) / 2;
4604 uint estimate = factor * (sz + bc) + cc;
4605
4606 assert((estimate - cc) / factor == sz + bc, "overflow");
4607
4608 return estimate + est_loop_flow_merge_sz();
4609 }
4610
4611 // The Estimated Loop (full-) Unroll Size:
4612 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4613 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4614 // node usage estimates made are on the safe side, for the most part. This is
4615 // a "light" version of the loop clone size calculation (above), based on the
4616 // assumption that most of the loop-construct overhead will be unraveled when
4617 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4618 // including an overflow check and returning UINT_MAX in case of an overflow.
4619 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
4620
4621 precond(factor > 0);
4622
4623 // Take into account that after unroll conjoined heads and tails will fold.
4624 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4625 uint const cc = 7;
4626 uint const sz = b0 + (b0 + 15) / 16;
4627 uint estimate = factor * sz + cc;
4628
4629 if ((estimate - cc) / factor != sz) {
4630 return UINT_MAX;
4631 }
4632
4633 return estimate + est_loop_flow_merge_sz();
4634 }
4635
4636 // Estimate the growth effect (in nodes) of merging control and data flow when
4637 // cloning a loop body, based on the amount of control and data flow reaching
4638 // outside of the (current) loop body.
4639 uint IdealLoopTree::est_loop_flow_merge_sz() const {
4640
4641 uint ctrl_edge_out_cnt = 0;
4642 uint data_edge_out_cnt = 0;
4643
4644 for (uint i = 0; i < _body.size(); i++) {
4645 Node* node = _body.at(i);
4646 uint outcnt = node->outcnt();
4647
4648 for (uint k = 0; k < outcnt; k++) {
4649 Node* out = node->raw_out(k);
4650 if (out == nullptr) continue;
4651 if (out->is_CFG()) {
4652 if (!is_member(_phase->get_loop(out))) {
4653 ctrl_edge_out_cnt++;
4654 }
4655 } else if (_phase->has_ctrl(out)) {
4656 Node* ctrl = _phase->get_ctrl(out);
4657 assert(ctrl != nullptr, "must be");
4658 assert(ctrl->is_CFG(), "must be");
4659 if (!is_member(_phase->get_loop(ctrl))) {
4660 data_edge_out_cnt++;
4661 }
4662 }
4663 }
4664 }
4665 // Use data and control count (x2.0) in estimate iff both are > 0. This is
4666 // a rather pessimistic estimate for the most part, in particular for some
4667 // complex loops, but still not enough to capture all loops.
4668 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
4669 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4670 }
4671 return 0;
4672 }
4673
4674 #ifndef PRODUCT
4675 //------------------------------dump_head--------------------------------------
4676 // Dump 1 liner for loop header info
4677 void IdealLoopTree::dump_head() {
4678 tty->sp(2 * _nest);
4679 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
4680 if (_irreducible) tty->print(" IRREDUCIBLE");
4681 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4682 : _head->in(LoopNode::EntryControl);
4683 const Predicates predicates(entry);
4684 if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {
4685 tty->print(" limit_check");
4686 }
4687 if (predicates.short_running_long_loop_predicate_block()->is_non_empty()) {
4688 tty->print(" short_running");
4689 }
4690 if (UseLoopPredicate) {
4691 if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {
4692 tty->print(" profile_predicated");
4693 }
4694 if (predicates.loop_predicate_block()->is_non_empty()) {
4695 tty->print(" predicated");
4696 }
4697 }
4698 if (UseAutoVectorizationPredicate && predicates.auto_vectorization_check_block()->is_non_empty()) {
4699 tty->print(" auto_vectorization_check_predicate");
4700 }
4701 if (_head->is_CountedLoop()) {
4702 CountedLoopNode *cl = _head->as_CountedLoop();
4703 tty->print(" counted");
4704
4705 Node* init_n = cl->init_trip();
4706 if (init_n != nullptr && init_n->is_Con())
4707 tty->print(" [%d,", cl->init_trip()->get_int());
4708 else
4709 tty->print(" [int,");
4710 Node* limit_n = cl->limit();
4711 if (limit_n != nullptr && limit_n->is_Con())
4712 tty->print("%d),", cl->limit()->get_int());
4713 else
4714 tty->print("int),");
4715 int stride_con = cl->stride_con();
4716 if (stride_con > 0) tty->print("+");
4717 tty->print("%d", stride_con);
4718
4719 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
4720
4721 if (cl->is_pre_loop ()) tty->print(" pre" );
4722 if (cl->is_main_loop()) tty->print(" main");
4723 if (cl->is_post_loop()) tty->print(" post");
4724 if (cl->is_vectorized_loop()) tty->print(" vector");
4725 if (range_checks_present()) tty->print(" rc ");
4726 if (cl->is_multiversion_fast_loop()) { tty->print(" multiversion_fast"); }
4727 if (cl->is_multiversion_slow_loop()) { tty->print(" multiversion_slow"); }
4728 if (cl->is_multiversion_delayed_slow_loop()) { tty->print(" multiversion_delayed_slow"); }
4729 }
4730 if (_has_call) tty->print(" has_call");
4731 if (_has_sfpt) tty->print(" has_sfpt");
4732 if (_rce_candidate) tty->print(" rce");
4733 if (_safepts != nullptr && _safepts->size() > 0) {
4734 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
4735 }
4736 if (_required_safept != nullptr && _required_safept->size() > 0) {
4737 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
4738 }
4739 if (Verbose) {
4740 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
4741 }
4742 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
4743 tty->print(" strip_mined");
4744 }
4745 tty->cr();
4746 }
4747
4748 //------------------------------dump-------------------------------------------
4749 // Dump loops by loop tree
4750 void IdealLoopTree::dump() {
4751 dump_head();
4752 if (_child) _child->dump();
4753 if (_next) _next ->dump();
4754 }
4755
4756 #endif
4757
4758 static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
4759 if (loop == root) {
4760 if (loop->_child != nullptr) {
4761 log->begin_head("loop_tree");
4762 log->end_head();
4763 log_loop_tree_helper(root, loop->_child, log);
4764 log->tail("loop_tree");
4765 assert(loop->_next == nullptr, "what?");
4766 }
4767 } else if (loop != nullptr) {
4768 Node* head = loop->_head;
4769 log->begin_head("loop");
4770 log->print(" idx='%d' ", head->_idx);
4771 if (loop->_irreducible) log->print("irreducible='1' ");
4772 if (head->is_Loop()) {
4773 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
4774 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
4775 } else if (head->is_CountedLoop()) {
4776 CountedLoopNode* cl = head->as_CountedLoop();
4777 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
4778 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
4779 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
4780 }
4781 log->end_head();
4782 log_loop_tree_helper(root, loop->_child, log);
4783 log->tail("loop");
4784 log_loop_tree_helper(root, loop->_next, log);
4785 }
4786 }
4787
4788 void PhaseIdealLoop::log_loop_tree() {
4789 if (C->log() != nullptr) {
4790 log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4791 }
4792 }
4793
4794 // Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4795 // predicates will be removed during the next round of IGVN.
4796 void PhaseIdealLoop::eliminate_useless_predicates() const {
4797 if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {
4798 return; // No predicates left.
4799 }
4800
4801 EliminateUselessPredicates eliminate_useless_predicates(_igvn, _ltree_root);
4802 eliminate_useless_predicates.eliminate();
4803 }
4804
4805 // If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4806 void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {
4807 if (_zero_trip_guard_opaque_nodes.size() == 0) {
4808 return;
4809 }
4810 Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4811 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4812 IdealLoopTree* lpt = iter.current();
4813 if (lpt->_child == nullptr && lpt->is_counted()) {
4814 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4815 Node* opaque = head->is_canonical_loop_entry();
4816 if (opaque != nullptr) {
4817 useful_zero_trip_guard_opaques_nodes.push(opaque);
4818 #ifdef ASSERT
4819 // See PhaseIdealLoop::do_unroll
4820 // This property is required in do_unroll, but it may not hold after cloning a loop.
4821 // In such a case, we bail out from unrolling, and rely on IGVN to clean up the graph.
4822 // We are here before loop cloning (before iteration_split), so if this property
4823 // does not hold, it must come from the previous round of loop optimizations, meaning
4824 // that IGVN failed to clean it: we will catch that here.
4825 // On the other hand, if this assert passes, a bailout in do_unroll means that
4826 // this property was broken in the current round of loop optimization (between here
4827 // and do_unroll), so we give a chance to IGVN to make the property true again.
4828 if (head->is_main_loop()) {
4829 assert(opaque->outcnt() == 1, "opaque node should not be shared");
4830 assert(opaque->in(1) == head->limit(), "After IGVN cleanup, input of opaque node must be the limit.");
4831 }
4832 if (head->is_post_loop()) {
4833 assert(opaque->outcnt() == 1, "opaque node should not be shared");
4834 }
4835 #endif
4836 }
4837 }
4838 }
4839 for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {
4840 OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4841 DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4842 if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {
4843 IfNode* iff = opaque->if_node();
4844 IdealLoopTree* loop = get_loop(iff);
4845 while (loop != _ltree_root && loop != nullptr) {
4846 loop = loop->_parent;
4847 }
4848 if (loop == nullptr) {
4849 // unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4850 // We can't tell if the opaque node is useful or not
4851 assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4852 } else {
4853 assert(guarded_loop == nullptr, "");
4854 this->_igvn.replace_node(opaque, opaque->in(1));
4855 }
4856 } else {
4857 assert(guarded_loop != nullptr, "");
4858 }
4859 }
4860 }
4861
4862 void PhaseIdealLoop::eliminate_useless_multiversion_if() {
4863 if (_multiversion_opaque_nodes.size() == 0) {
4864 return;
4865 }
4866
4867 ResourceMark rm;
4868 Unique_Node_List useful_multiversioning_opaque_nodes;
4869
4870 // The OpaqueMultiversioning is only used from the fast main loop in AutoVectorization, to add
4871 // speculative runtime-checks to the multiversion_if. Thus, a OpaqueMultiversioning is only
4872 // useful if it can be found from a fast main loop. If it can not be found from a fast main loop,
4873 // then we cannot ever use that multiversion_if to add more speculative runtime-checks, and hence
4874 // it is useless. If it is still in delayed mode, i.e. has not yet had any runtime-checks added,
4875 // then we can let it constant fold towards the fast loop.
4876 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4877 IdealLoopTree* lpt = iter.current();
4878 if (lpt->_child == nullptr && lpt->is_counted()) {
4879 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4880 if (head->is_main_loop() && head->is_multiversion_fast_loop()) {
4881 // There are fast_loop pre/main/post loops, but the finding traversal starts at the main
4882 // loop, and traverses via the fast pre loop to the multiversion_if.
4883 IfNode* multiversion_if = head->find_multiversion_if_from_multiversion_fast_main_loop();
4884 if (multiversion_if != nullptr) {
4885 useful_multiversioning_opaque_nodes.push(multiversion_if->in(1)->as_OpaqueMultiversioning());
4886 } else {
4887 // We could not find the multiversion_if, and would never find it again. Remove the
4888 // multiversion marking for consistency.
4889 head->set_no_multiversion();
4890 }
4891 }
4892 }
4893 }
4894
4895 for (uint i = 0; i < _multiversion_opaque_nodes.size(); i++) {
4896 OpaqueMultiversioningNode* opaque = _multiversion_opaque_nodes.at(i)->as_OpaqueMultiversioning();
4897 if (!useful_multiversioning_opaque_nodes.member(opaque)) {
4898 if (opaque->is_delayed_slow_loop()) {
4899 // We cannot hack the node directly, otherwise the slow_loop will complain that it cannot
4900 // find the multiversioning opaque node. Instead, we mark the opaque node as useless, and
4901 // it can be constant folded during IGVN.
4902 opaque->mark_useless(_igvn);
4903 }
4904 }
4905 }
4906 }
4907
4908 //------------------------process_expensive_nodes-----------------------------
4909 // Expensive nodes have their control input set to prevent the GVN
4910 // from commoning them and as a result forcing the resulting node to
4911 // be in a more frequent path. Use CFG information here, to change the
4912 // control inputs so that some expensive nodes can be commoned while
4913 // not executed more frequently.
4914 bool PhaseIdealLoop::process_expensive_nodes() {
4915 assert(OptimizeExpensiveOps, "optimization off?");
4916
4917 // Sort nodes to bring similar nodes together
4918 C->sort_expensive_nodes();
4919
4920 bool progress = false;
4921
4922 for (int i = 0; i < C->expensive_count(); ) {
4923 Node* n = C->expensive_node(i);
4924 int start = i;
4925 // Find nodes similar to n
4926 i++;
4927 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4928 int end = i;
4929 // And compare them two by two
4930 for (int j = start; j < end; j++) {
4931 Node* n1 = C->expensive_node(j);
4932 if (is_node_unreachable(n1)) {
4933 continue;
4934 }
4935 for (int k = j+1; k < end; k++) {
4936 Node* n2 = C->expensive_node(k);
4937 if (is_node_unreachable(n2)) {
4938 continue;
4939 }
4940
4941 assert(n1 != n2, "should be pair of nodes");
4942
4943 Node* c1 = n1->in(0);
4944 Node* c2 = n2->in(0);
4945
4946 Node* parent_c1 = c1;
4947 Node* parent_c2 = c2;
4948
4949 // The call to get_early_ctrl_for_expensive() moves the
4950 // expensive nodes up but stops at loops that are in a if
4951 // branch. See whether we can exit the loop and move above the
4952 // If.
4953 if (c1->is_Loop()) {
4954 parent_c1 = c1->in(1);
4955 }
4956 if (c2->is_Loop()) {
4957 parent_c2 = c2->in(1);
4958 }
4959
4960 if (parent_c1 == parent_c2) {
4961 _igvn._worklist.push(n1);
4962 _igvn._worklist.push(n2);
4963 continue;
4964 }
4965
4966 // Look for identical expensive node up the dominator chain.
4967 if (is_dominator(c1, c2)) {
4968 c2 = c1;
4969 } else if (is_dominator(c2, c1)) {
4970 c1 = c2;
4971 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4972 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
4973 // Both branches have the same expensive node so move it up
4974 // before the if.
4975 c1 = c2 = idom(parent_c1->in(0));
4976 }
4977 // Do the actual moves
4978 if (n1->in(0) != c1) {
4979 _igvn.replace_input_of(n1, 0, c1);
4980 progress = true;
4981 }
4982 if (n2->in(0) != c2) {
4983 _igvn.replace_input_of(n2, 0, c2);
4984 progress = true;
4985 }
4986 }
4987 }
4988 }
4989
4990 return progress;
4991 }
4992
4993 //=============================================================================
4994 //----------------------------build_and_optimize-------------------------------
4995 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4996 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4997 void PhaseIdealLoop::build_and_optimize() {
4998 assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4999
5000 bool do_split_ifs = (_mode == LoopOptsDefault);
5001 bool skip_loop_opts = (_mode == LoopOptsNone);
5002 bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
5003
5004
5005 bool old_progress = C->major_progress();
5006 uint orig_worklist_size = _igvn._worklist.size();
5007
5008 // Reset major-progress flag for the driver's heuristics
5009 C->clear_major_progress();
5010
5011 #ifndef PRODUCT
5012 // Capture for later assert
5013 uint unique = C->unique();
5014 _loop_invokes++;
5015 _loop_work += unique;
5016 #endif
5017
5018 // True if the method has at least 1 irreducible loop
5019 _has_irreducible_loops = false;
5020
5021 _created_loop_node = false;
5022
5023 VectorSet visited;
5024 // Pre-grow the mapping from Nodes to IdealLoopTrees.
5025 _loop_or_ctrl.map(C->unique(), nullptr);
5026 memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
5027
5028 // Pre-build the top-level outermost loop tree entry
5029 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
5030 // Do not need a safepoint at the top level
5031 _ltree_root->_has_sfpt = 1;
5032
5033 // Initialize Dominators.
5034 // Checked in clone_loop_predicate() during beautify_loops().
5035 _idom_size = 0;
5036 _idom = nullptr;
5037 _dom_depth = nullptr;
5038 _dom_stk = nullptr;
5039
5040 // Empty pre-order array
5041 allocate_preorders();
5042
5043 // Build a loop tree on the fly. Build a mapping from CFG nodes to
5044 // IdealLoopTree entries. Data nodes are NOT walked.
5045 build_loop_tree();
5046 // Check for bailout, and return
5047 if (C->failing()) {
5048 return;
5049 }
5050
5051 // Verify that the has_loops() flag set at parse time is consistent with the just built loop tree. When the back edge
5052 // is an exception edge, parsing doesn't set has_loops().
5053 assert(_ltree_root->_child == nullptr || C->has_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
5054 // No loops after all
5055 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
5056
5057 // There should always be an outer loop containing the Root and Return nodes.
5058 // If not, we have a degenerate empty program. Bail out in this case.
5059 if (!has_node(C->root())) {
5060 if (!_verify_only) {
5061 C->clear_major_progress();
5062 assert(false, "empty program detected during loop optimization");
5063 C->record_method_not_compilable("empty program detected during loop optimization");
5064 }
5065 return;
5066 }
5067
5068 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5069 // Nothing to do, so get out
5070 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
5071 !_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode);
5072 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
5073 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode);
5074 if (stop_early && !do_expensive_nodes) {
5075 return;
5076 }
5077
5078 // Set loop nesting depth
5079 _ltree_root->set_nest( 0 );
5080
5081 // Split shared headers and insert loop landing pads.
5082 // Do not bother doing this on the Root loop of course.
5083 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
5084 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
5085 if( _ltree_root->_child->beautify_loops( this ) ) {
5086 // Re-build loop tree!
5087 _ltree_root->_child = nullptr;
5088 _loop_or_ctrl.clear();
5089 reallocate_preorders();
5090 build_loop_tree();
5091 // Check for bailout, and return
5092 if (C->failing()) {
5093 return;
5094 }
5095 // Reset loop nesting depth
5096 _ltree_root->set_nest( 0 );
5097
5098 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
5099 }
5100 }
5101
5102 // Build Dominators for elision of null checks & loop finding.
5103 // Since nodes do not have a slot for immediate dominator, make
5104 // a persistent side array for that info indexed on node->_idx.
5105 _idom_size = C->unique();
5106 _idom = NEW_ARENA_ARRAY(&_arena, Node*, _idom_size);
5107 _dom_depth = NEW_ARENA_ARRAY(&_arena, uint, _idom_size);
5108 _dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
5109 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
5110
5111 Dominators();
5112
5113 if (!_verify_only) {
5114 // As a side effect, Dominators removed any unreachable CFG paths
5115 // into RegionNodes. It doesn't do this test against Root, so
5116 // we do it here.
5117 for( uint i = 1; i < C->root()->req(); i++ ) {
5118 if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?
5119 _igvn.delete_input_of(C->root(), i);
5120 i--; // Rerun same iteration on compressed edges
5121 }
5122 }
5123
5124 // Given dominators, try to find inner loops with calls that must
5125 // always be executed (call dominates loop tail). These loops do
5126 // not need a separate safepoint.
5127 Node_List cisstack;
5128 _ltree_root->check_safepts(visited, cisstack);
5129 }
5130
5131 // Walk the DATA nodes and place into loops. Find earliest control
5132 // node. For CFG nodes, the _loop_or_ctrl array starts out and remains
5133 // holding the associated IdealLoopTree pointer. For DATA nodes, the
5134 // _loop_or_ctrl array holds the earliest legal controlling CFG node.
5135
5136 // Allocate stack with enough space to avoid frequent realloc
5137 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
5138 Node_Stack nstack(stack_size);
5139
5140 visited.clear();
5141 Node_List worklist;
5142 // Don't need C->root() on worklist since
5143 // it will be processed among C->top() inputs
5144 worklist.push(C->top());
5145 visited.set(C->top()->_idx); // Set C->top() as visited now
5146 build_loop_early( visited, worklist, nstack );
5147
5148 // Given early legal placement, try finding counted loops. This placement
5149 // is good enough to discover most loop invariants.
5150 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
5151 _ltree_root->counted_loop( this );
5152 }
5153
5154 // Find latest loop placement. Find ideal loop placement.
5155 visited.clear();
5156 init_dom_lca_tags();
5157 // Need C->root() on worklist when processing outs
5158 worklist.push(C->root());
5159 NOT_PRODUCT( C->verify_graph_edges(); )
5160 worklist.push(C->top());
5161 build_loop_late( visited, worklist, nstack );
5162 if (C->failing()) { return; }
5163
5164 if (_verify_only) {
5165 C->restore_major_progress(old_progress);
5166 assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
5167 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
5168 return;
5169 }
5170
5171 // clear out the dead code after build_loop_late
5172 while (_deadlist.size()) {
5173 _igvn.remove_globally_dead_node(_deadlist.pop());
5174 }
5175
5176 eliminate_useless_zero_trip_guard();
5177 eliminate_useless_multiversion_if();
5178
5179 if (stop_early) {
5180 assert(do_expensive_nodes, "why are we here?");
5181 if (process_expensive_nodes()) {
5182 // If we made some progress when processing expensive nodes then
5183 // the IGVN may modify the graph in a way that will allow us to
5184 // make some more progress: we need to try processing expensive
5185 // nodes again.
5186 C->set_major_progress();
5187 }
5188 return;
5189 }
5190
5191 // Some parser-inserted loop predicates could never be used by loop
5192 // predication or they were moved away from loop during some optimizations.
5193 // For example, peeling. Eliminate them before next loop optimizations.
5194 eliminate_useless_predicates();
5195
5196 #ifndef PRODUCT
5197 C->verify_graph_edges();
5198 if (_verify_me) { // Nested verify pass?
5199 // Check to see if the verify _mode is broken
5200 assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
5201 return;
5202 }
5203 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
5204 if (TraceLoopOpts && C->has_loops()) {
5205 _ltree_root->dump();
5206 }
5207 #endif
5208
5209 if (skip_loop_opts) {
5210 C->restore_major_progress(old_progress);
5211 return;
5212 }
5213
5214 if (do_max_unroll) {
5215 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5216 IdealLoopTree* lpt = iter.current();
5217 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
5218 lpt->compute_trip_count(this, T_INT);
5219 if (!lpt->do_one_iteration_loop(this) &&
5220 !lpt->do_remove_empty_loop(this)) {
5221 AutoNodeBudget node_budget(this);
5222 if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
5223 lpt->policy_maximally_unroll(this)) {
5224 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
5225 do_maximally_unroll(lpt, worklist);
5226 }
5227 }
5228 }
5229 }
5230
5231 C->restore_major_progress(old_progress);
5232 return;
5233 }
5234
5235 if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) {
5236 return;
5237 }
5238
5239 if (ReassociateInvariants && !C->major_progress()) {
5240 // Reassociate invariants and prep for split_thru_phi
5241 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5242 IdealLoopTree* lpt = iter.current();
5243 if (!lpt->is_loop()) {
5244 continue;
5245 }
5246 Node* head = lpt->_head;
5247 if (!lpt->is_innermost()) continue;
5248
5249 // check for vectorized loops, any reassociation of invariants was already done
5250 if (head->is_CountedLoop() && head->as_CountedLoop()->is_unroll_only()) {
5251 continue;
5252 } else {
5253 AutoNodeBudget node_budget(this);
5254 lpt->reassociate_invariants(this);
5255 }
5256 // Because RCE opportunities can be masked by split_thru_phi,
5257 // look for RCE candidates and inhibit split_thru_phi
5258 // on just their loop-phi's for this pass of loop opts
5259 if (SplitIfBlocks && do_split_ifs &&
5260 head->is_BaseCountedLoop() &&
5261 head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
5262 (lpt->policy_range_check(this, true, T_LONG) ||
5263 (head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
5264 lpt->_rce_candidate = 1; // = true
5265 }
5266 }
5267 }
5268
5269 // Check for aggressive application of split-if and other transforms
5270 // that require basic-block info (like cloning through Phi's)
5271 if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
5272 visited.clear();
5273 split_if_with_blocks(visited, nstack);
5274 if (C->failing()) {
5275 return;
5276 }
5277 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
5278 }
5279
5280 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
5281 C->set_major_progress();
5282 }
5283
5284 // Perform loop predication before iteration splitting
5285 if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {
5286 _ltree_root->_child->loop_predication(this);
5287 }
5288
5289 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
5290 if (do_intrinsify_fill()) {
5291 C->set_major_progress();
5292 }
5293 }
5294
5295 // Perform iteration-splitting on inner loops. Split iterations to avoid
5296 // range checks or one-shot null checks.
5297
5298 // If split-if's didn't hack the graph too bad (no CFG changes)
5299 // then do loop opts.
5300 if (C->has_loops() && !C->major_progress()) {
5301 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
5302 _ltree_root->_child->iteration_split( this, worklist );
5303 // No verify after peeling! GCM has hoisted code out of the loop.
5304 // After peeling, the hoisted code could sink inside the peeled area.
5305 // The peeling code does not try to recompute the best location for
5306 // all the code before the peeled area, so the verify pass will always
5307 // complain about it.
5308 }
5309
5310 // Check for bailout, and return
5311 if (C->failing()) {
5312 return;
5313 }
5314
5315 // Do verify graph edges in any case
5316 NOT_PRODUCT( C->verify_graph_edges(); );
5317
5318 if (!do_split_ifs) {
5319 // We saw major progress in Split-If to get here. We forced a
5320 // pass with unrolling and not split-if, however more split-if's
5321 // might make progress. If the unrolling didn't make progress
5322 // then the major-progress flag got cleared and we won't try
5323 // another round of Split-If. In particular the ever-common
5324 // instance-of/check-cast pattern requires at least 2 rounds of
5325 // Split-If to clear out.
5326 C->set_major_progress();
5327 }
5328
5329 // Repeat loop optimizations if new loops were seen
5330 if (created_loop_node()) {
5331 C->set_major_progress();
5332 }
5333
5334 // Auto-vectorize main-loop
5335 if (C->do_superword() && C->has_loops() && !C->major_progress()) {
5336 Compile::TracePhase tp(_t_autoVectorize);
5337
5338 // Shared data structures for all AutoVectorizations, to reduce allocations
5339 // of large arrays.
5340 VSharedData vshared;
5341 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5342 IdealLoopTree* lpt = iter.current();
5343 AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
5344
5345 if (status == AutoVectorizeStatus::TriedAndFailed) {
5346 // We tried vectorization, but failed. From now on only unroll the loop.
5347 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
5348 if (cl->has_passed_slp()) {
5349 C->set_major_progress();
5350 cl->set_notpassed_slp();
5351 cl->mark_do_unroll_only();
5352 }
5353 }
5354 }
5355 }
5356
5357 // Keep loop predicates and perform optimizations with them
5358 // until no more loop optimizations could be done.
5359 // After that switch predicates off and do more loop optimizations.
5360 if (!C->major_progress() && (C->parse_predicate_count() > 0)) {
5361 C->mark_parse_predicate_nodes_useless(_igvn);
5362 assert(C->parse_predicate_count() == 0, "should be zero now");
5363 if (TraceLoopOpts) {
5364 tty->print_cr("PredicatesOff");
5365 }
5366 C->set_major_progress();
5367 }
5368 }
5369
5370 #ifndef PRODUCT
5371 //------------------------------print_statistics-------------------------------
5372 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
5373 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
5374 volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
5375 volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
5376 volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
5377 void PhaseIdealLoop::print_statistics() {
5378 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);
5379 }
5380 #endif
5381
5382 #ifdef ASSERT
5383 // Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
5384 void PhaseIdealLoop::verify() const {
5385 ResourceMark rm;
5386 bool old_progress = C->major_progress();
5387 bool success = true;
5388
5389 PhaseIdealLoop phase_verify(_igvn, this);
5390 if (C->failing_internal()) {
5391 return;
5392 }
5393
5394 // Verify ctrl and idom of every node.
5395 success &= verify_idom_and_nodes(C->root(), &phase_verify);
5396
5397 // Verify loop-tree.
5398 success &= _ltree_root->verify_tree(phase_verify._ltree_root);
5399
5400 assert(success, "VerifyLoopOptimizations failed");
5401
5402 // Major progress was cleared by creating a verify version of PhaseIdealLoop.
5403 C->restore_major_progress(old_progress);
5404 }
5405
5406 // Perform a BFS starting at n, through all inputs.
5407 // Call verify_idom and verify_node on all nodes of BFS traversal.
5408 bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {
5409 Unique_Node_List worklist;
5410 worklist.push(root);
5411 bool success = true;
5412 for (uint i = 0; i < worklist.size(); i++) {
5413 Node* n = worklist.at(i);
5414 // process node
5415 success &= verify_idom(n, phase_verify);
5416 success &= verify_loop_ctrl(n, phase_verify);
5417 // visit inputs
5418 for (uint j = 0; j < n->req(); j++) {
5419 if (n->in(j) != nullptr) {
5420 worklist.push(n->in(j));
5421 }
5422 }
5423 }
5424 return success;
5425 }
5426
5427 // Verify dominator structure (IDOM).
5428 bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {
5429 // Verify IDOM for all CFG nodes (except root).
5430 if (!n->is_CFG() || n->is_Root()) {
5431 return true; // pass
5432 }
5433
5434 if (n->_idx >= _idom_size) {
5435 tty->print("CFG Node with no idom: ");
5436 n->dump();
5437 return false; // fail
5438 }
5439
5440 Node* id = idom_no_update(n);
5441 Node* id_verify = phase_verify->idom_no_update(n);
5442 if (id != id_verify) {
5443 tty->print("Mismatching idom for node: ");
5444 n->dump();
5445 tty->print(" We have idom: ");
5446 id->dump();
5447 tty->print(" Verify has idom: ");
5448 id_verify->dump();
5449 tty->cr();
5450 return false; // fail
5451 }
5452 return true; // pass
5453 }
5454
5455 // Verify "_loop_or_ctrl": control and loop membership.
5456 // (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5457 // (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5458 // (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5459 // (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5460 bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {
5461 const uint i = n->_idx;
5462 // The loop-tree was built from def to use (top-down).
5463 // The verification happens from use to def (bottom-up).
5464 // We may thus find nodes during verification that are not in the loop-tree.
5465 if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {
5466 if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {
5467 tty->print_cr("Was reachable in only one. this %d, verify %d.",
5468 _loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5469 n->dump();
5470 return false; // fail
5471 }
5472 // Not reachable for both.
5473 return true; // pass
5474 }
5475
5476 if (n->is_CFG() == has_ctrl(n)) {
5477 tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));
5478 n->dump();
5479 return false; // fail
5480 }
5481
5482 if (has_ctrl(n) != phase_verify->has_ctrl(n)) {
5483 tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));
5484 n->dump();
5485 return false; // fail
5486 } else if (has_ctrl(n)) {
5487 assert(phase_verify->has_ctrl(n), "sanity");
5488 // n is a data node.
5489 // Verify that its ctrl is the same.
5490
5491 // Broken part of VerifyLoopOptimizations (A)
5492 // Reason:
5493 // BUG, wrong control set for example in
5494 // PhaseIdealLoop::split_if_with_blocks
5495 // at "set_ctrl(x, new_ctrl);"
5496 /*
5497 if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5498 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
5499 tty->print("Mismatched control setting for: ");
5500 n->dump();
5501 if( fail++ > 10 ) return;
5502 Node *c = get_ctrl_no_update(n);
5503 tty->print("We have it as: ");
5504 if( c->in(0) ) c->dump();
5505 else tty->print_cr("N%d",c->_idx);
5506 tty->print("Verify thinks: ");
5507 if( loop_verify->has_ctrl(n) )
5508 loop_verify->get_ctrl_no_update(n)->dump();
5509 else
5510 loop_verify->get_loop_idx(n)->dump();
5511 tty->cr();
5512 }
5513 */
5514 return true; // pass
5515 } else {
5516 assert(!phase_verify->has_ctrl(n), "sanity");
5517 // n is a ctrl node.
5518 // Verify that not has_ctrl, and that get_loop_idx is the same.
5519
5520 // Broken part of VerifyLoopOptimizations (B)
5521 // Reason:
5522 // NeverBranch node for example is added to loop outside its scope.
5523 // Once we run build_loop_tree again, it is added to the correct loop.
5524 /*
5525 if (!C->major_progress()) {
5526 // Loop selection can be messed up if we did a major progress
5527 // operation, like split-if. Do not verify in that case.
5528 IdealLoopTree *us = get_loop_idx(n);
5529 IdealLoopTree *them = loop_verify->get_loop_idx(n);
5530 if( us->_head != them->_head || us->_tail != them->_tail ) {
5531 tty->print("Unequals loops for: ");
5532 n->dump();
5533 if( fail++ > 10 ) return;
5534 tty->print("We have it as: ");
5535 us->dump();
5536 tty->print("Verify thinks: ");
5537 them->dump();
5538 tty->cr();
5539 }
5540 }
5541 */
5542 return true; // pass
5543 }
5544 }
5545
5546 static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {
5547 assert(a != nullptr && b != nullptr, "must be");
5548 return a->_head->_idx - b->_head->_idx;
5549 }
5550
5551 GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {
5552 GrowableArray<IdealLoopTree*> children;
5553 IdealLoopTree* child = _child;
5554 while (child != nullptr) {
5555 assert(child->_parent == this, "all must be children of this");
5556 children.insert_sorted<compare_tree>(child);
5557 child = child->_next;
5558 }
5559 return children;
5560 }
5561
5562 // Verify that tree structures match. Because the CFG can change, siblings
5563 // within the loop tree can be reordered. We attempt to deal with that by
5564 // reordering the verify's loop tree if possible.
5565 bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {
5566 assert(_head == loop_verify->_head, "mismatched loop head");
5567 assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5568
5569 // Collect the children
5570 GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5571 GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5572
5573 bool success = true;
5574
5575 // Compare the two children lists
5576 for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {
5577 IdealLoopTree* child = nullptr;
5578 IdealLoopTree* child_verify = nullptr;
5579 // Read from both lists, if possible.
5580 if (i < children.length()) {
5581 child = children.at(i);
5582 }
5583 if (j < children_verify.length()) {
5584 child_verify = children_verify.at(j);
5585 }
5586 assert(child != nullptr || child_verify != nullptr, "must find at least one");
5587 if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {
5588 // We found two non-equal children. Select the smaller one.
5589 if (child->_head->_idx < child_verify->_head->_idx) {
5590 child_verify = nullptr;
5591 } else {
5592 child = nullptr;
5593 }
5594 }
5595 // Process the two children, or potentially log the failure if we only found one.
5596 if (child_verify == nullptr) {
5597 if (child->_irreducible && Compile::current()->major_progress()) {
5598 // Irreducible loops can pick a different header (one of its entries).
5599 } else {
5600 tty->print_cr("We have a loop that verify does not have");
5601 child->dump();
5602 success = false;
5603 }
5604 i++; // step for this
5605 } else if (child == nullptr) {
5606 if (child_verify->_irreducible && Compile::current()->major_progress()) {
5607 // Irreducible loops can pick a different header (one of its entries).
5608 } else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {
5609 // Infinite loops do not get attached to the loop-tree on their first visit.
5610 // "this" runs before "loop_verify". It is thus possible that we find the
5611 // infinite loop only for "child_verify". Only finding it with "child" would
5612 // mean that we lost it, which is not ok.
5613 } else {
5614 tty->print_cr("Verify has a loop that we do not have");
5615 child_verify->dump();
5616 success = false;
5617 }
5618 j++; // step for verify
5619 } else {
5620 assert(child->_head == child_verify->_head, "We have both and they are equal");
5621 success &= child->verify_tree(child_verify); // Recursion
5622 i++; // step for this
5623 j++; // step for verify
5624 }
5625 }
5626
5627 // Broken part of VerifyLoopOptimizations (D)
5628 // Reason:
5629 // split_if has to update the _tail, if it is modified. But that is done by
5630 // checking to what loop the iff belongs to. That info can be wrong, and then
5631 // we do not update the _tail correctly.
5632 /*
5633 Node *tail = _tail; // Inline a non-updating version of
5634 while( !tail->in(0) ) // the 'tail()' call.
5635 tail = tail->in(1);
5636 assert( tail == loop->_tail, "mismatched loop tail" );
5637 */
5638
5639 if (_head->is_CountedLoop()) {
5640 CountedLoopNode *cl = _head->as_CountedLoop();
5641
5642 Node* ctrl = cl->init_control();
5643 Node* back = cl->back_control();
5644 assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5645 assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5646 cl->loopexit(); // assert implied
5647 }
5648
5649 // Broken part of VerifyLoopOptimizations (E)
5650 // Reason:
5651 // PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5652 // to the loop body. Or maybe they are not added to the correct loop.
5653 // at "Node* x = n->clone();"
5654 /*
5655 // Innermost loops need to verify loop bodies,
5656 // but only if no 'major_progress'
5657 int fail = 0;
5658 if (!Compile::current()->major_progress() && _child == nullptr) {
5659 for( uint i = 0; i < _body.size(); i++ ) {
5660 Node *n = _body.at(i);
5661 if (n->outcnt() == 0) continue; // Ignore dead
5662 uint j;
5663 for( j = 0; j < loop->_body.size(); j++ )
5664 if( loop->_body.at(j) == n )
5665 break;
5666 if( j == loop->_body.size() ) { // Not found in loop body
5667 // Last ditch effort to avoid assertion: Its possible that we
5668 // have some users (so outcnt not zero) but are still dead.
5669 // Try to find from root.
5670 if (Compile::current()->root()->find(n->_idx)) {
5671 fail++;
5672 tty->print("We have that verify does not: ");
5673 n->dump();
5674 }
5675 }
5676 }
5677 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
5678 Node *n = loop->_body.at(i2);
5679 if (n->outcnt() == 0) continue; // Ignore dead
5680 uint j;
5681 for( j = 0; j < _body.size(); j++ )
5682 if( _body.at(j) == n )
5683 break;
5684 if( j == _body.size() ) { // Not found in loop body
5685 // Last ditch effort to avoid assertion: Its possible that we
5686 // have some users (so outcnt not zero) but are still dead.
5687 // Try to find from root.
5688 if (Compile::current()->root()->find(n->_idx)) {
5689 fail++;
5690 tty->print("Verify has that we do not: ");
5691 n->dump();
5692 }
5693 }
5694 }
5695 assert( !fail, "loop body mismatch" );
5696 }
5697 */
5698 return success;
5699 }
5700 #endif
5701
5702 //------------------------------set_idom---------------------------------------
5703 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
5704 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
5705 uint idx = d->_idx;
5706 if (idx >= _idom_size) {
5707 uint newsize = next_power_of_2(idx);
5708 _idom = REALLOC_ARENA_ARRAY(&_arena, Node*, _idom,_idom_size,newsize);
5709 _dom_depth = REALLOC_ARENA_ARRAY(&_arena, uint, _dom_depth,_idom_size,newsize);
5710 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5711 _idom_size = newsize;
5712 }
5713 _idom[idx] = n;
5714 _dom_depth[idx] = dom_depth;
5715 }
5716
5717 //------------------------------recompute_dom_depth---------------------------------------
5718 // The dominator tree is constructed with only parent pointers.
5719 // This recomputes the depth in the tree by first tagging all
5720 // nodes as "no depth yet" marker. The next pass then runs up
5721 // the dom tree from each node marked "no depth yet", and computes
5722 // the depth on the way back down.
5723 void PhaseIdealLoop::recompute_dom_depth() {
5724 uint no_depth_marker = C->unique();
5725 uint i;
5726 // Initialize depth to "no depth yet" and realize all lazy updates
5727 for (i = 0; i < _idom_size; i++) {
5728 // Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5729 if (_dom_depth[i] > 0 && _idom[i] != nullptr) {
5730 _dom_depth[i] = no_depth_marker;
5731
5732 // heal _idom if it has a fwd mapping in _loop_or_ctrl
5733 if (_idom[i]->in(0) == nullptr) {
5734 idom(i);
5735 }
5736 }
5737 }
5738 if (_dom_stk == nullptr) {
5739 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5740 if (init_size < 10) init_size = 10;
5741 _dom_stk = new GrowableArray<uint>(init_size);
5742 }
5743 // Compute new depth for each node.
5744 for (i = 0; i < _idom_size; i++) {
5745 uint j = i;
5746 // Run up the dom tree to find a node with a depth
5747 while (_dom_depth[j] == no_depth_marker) {
5748 _dom_stk->push(j);
5749 j = _idom[j]->_idx;
5750 }
5751 // Compute the depth on the way back down this tree branch
5752 uint dd = _dom_depth[j] + 1;
5753 while (_dom_stk->length() > 0) {
5754 uint j = _dom_stk->pop();
5755 _dom_depth[j] = dd;
5756 dd++;
5757 }
5758 }
5759 }
5760
5761 //------------------------------sort-------------------------------------------
5762 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5763 // loop tree, not the root.
5764 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
5765 if( !innermost ) return loop; // New innermost loop
5766
5767 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5768 assert( loop_preorder, "not yet post-walked loop" );
5769 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5770 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5771
5772 // Insert at start of list
5773 while( l ) { // Insertion sort based on pre-order
5774 if( l == loop ) return innermost; // Already on list!
5775 int l_preorder = get_preorder(l->_head); // Cache pre-order number
5776 assert( l_preorder, "not yet post-walked l" );
5777 // Check header pre-order number to figure proper nesting
5778 if( loop_preorder > l_preorder )
5779 break; // End of insertion
5780 // If headers tie (e.g., shared headers) check tail pre-order numbers.
5781 // Since I split shared headers, you'd think this could not happen.
5782 // BUT: I must first do the preorder numbering before I can discover I
5783 // have shared headers, so the split headers all get the same preorder
5784 // number as the RegionNode they split from.
5785 if( loop_preorder == l_preorder &&
5786 get_preorder(loop->_tail) < get_preorder(l->_tail) )
5787 break; // Also check for shared headers (same pre#)
5788 pp = &l->_parent; // Chain up list
5789 l = *pp;
5790 }
5791 // Link into list
5792 // Point predecessor to me
5793 *pp = loop;
5794 // Point me to successor
5795 IdealLoopTree *p = loop->_parent;
5796 loop->_parent = l; // Point me to successor
5797 if( p ) sort( p, innermost ); // Insert my parents into list as well
5798 return innermost;
5799 }
5800
5801 //------------------------------build_loop_tree--------------------------------
5802 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5803 // bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5804 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5805 // tightest enclosing IdealLoopTree for post-walked.
5806 //
5807 // During my forward walk I do a short 1-layer lookahead to see if I can find
5808 // a loop backedge with that doesn't have any work on the backedge. This
5809 // helps me construct nested loops with shared headers better.
5810 //
5811 // Once I've done the forward recursion, I do the post-work. For each child
5812 // I check to see if there is a backedge. Backedges define a loop! I
5813 // insert an IdealLoopTree at the target of the backedge.
5814 //
5815 // During the post-work I also check to see if I have several children
5816 // belonging to different loops. If so, then this Node is a decision point
5817 // where control flow can choose to change loop nests. It is at this
5818 // decision point where I can figure out how loops are nested. At this
5819 // time I can properly order the different loop nests from my children.
5820 // Note that there may not be any backedges at the decision point!
5821 //
5822 // Since the decision point can be far removed from the backedges, I can't
5823 // order my loops at the time I discover them. Thus at the decision point
5824 // I need to inspect loop header pre-order numbers to properly nest my
5825 // loops. This means I need to sort my childrens' loops by pre-order.
5826 // The sort is of size number-of-control-children, which generally limits
5827 // it to size 2 (i.e., I just choose between my 2 target loops).
5828 void PhaseIdealLoop::build_loop_tree() {
5829 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5830 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5831 Node *n = C->root();
5832 bltstack.push(n);
5833 int pre_order = 1;
5834 int stack_size;
5835
5836 while ( ( stack_size = bltstack.length() ) != 0 ) {
5837 n = bltstack.top(); // Leave node on stack
5838 if ( !is_visited(n) ) {
5839 // ---- Pre-pass Work ----
5840 // Pre-walked but not post-walked nodes need a pre_order number.
5841
5842 set_preorder_visited( n, pre_order ); // set as visited
5843
5844 // ---- Scan over children ----
5845 // Scan first over control projections that lead to loop headers.
5846 // This helps us find inner-to-outer loops with shared headers better.
5847
5848 // Scan children's children for loop headers.
5849 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
5850 Node* m = n->raw_out(i); // Child
5851 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
5852 // Scan over children's children to find loop
5853 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5854 Node* l = m->fast_out(j);
5855 if( is_visited(l) && // Been visited?
5856 !is_postvisited(l) && // But not post-visited
5857 get_preorder(l) < pre_order ) { // And smaller pre-order
5858 // Found! Scan the DFS down this path before doing other paths
5859 bltstack.push(m);
5860 break;
5861 }
5862 }
5863 }
5864 }
5865 pre_order++;
5866 }
5867 else if ( !is_postvisited(n) ) {
5868 // Note: build_loop_tree_impl() adds out edges on rare occasions,
5869 // such as com.sun.rsasign.am::a.
5870 // For non-recursive version, first, process current children.
5871 // On next iteration, check if additional children were added.
5872 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
5873 Node* u = n->raw_out(k);
5874 if ( u->is_CFG() && !is_visited(u) ) {
5875 bltstack.push(u);
5876 }
5877 }
5878 if ( bltstack.length() == stack_size ) {
5879 // There were no additional children, post visit node now
5880 (void)bltstack.pop(); // Remove node from stack
5881 pre_order = build_loop_tree_impl(n, pre_order);
5882 // Check for bailout
5883 if (C->failing()) {
5884 return;
5885 }
5886 // Check to grow _preorders[] array for the case when
5887 // build_loop_tree_impl() adds new nodes.
5888 check_grow_preorders();
5889 }
5890 }
5891 else {
5892 (void)bltstack.pop(); // Remove post-visited node from stack
5893 }
5894 }
5895 DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5896 }
5897
5898 //------------------------------build_loop_tree_impl---------------------------
5899 int PhaseIdealLoop::build_loop_tree_impl(Node* n, int pre_order) {
5900 // ---- Post-pass Work ----
5901 // Pre-walked but not post-walked nodes need a pre_order number.
5902
5903 // Tightest enclosing loop for this Node
5904 IdealLoopTree *innermost = nullptr;
5905
5906 // For all children, see if any edge is a backedge. If so, make a loop
5907 // for it. Then find the tightest enclosing loop for the self Node.
5908 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5909 Node* m = n->fast_out(i); // Child
5910 if (n == m) continue; // Ignore control self-cycles
5911 if (!m->is_CFG()) continue;// Ignore non-CFG edges
5912
5913 IdealLoopTree *l; // Child's loop
5914 if (!is_postvisited(m)) { // Child visited but not post-visited?
5915 // Found a backedge
5916 assert(get_preorder(m) < pre_order, "should be backedge");
5917 // Check for the RootNode, which is already a LoopNode and is allowed
5918 // to have multiple "backedges".
5919 if (m == C->root()) { // Found the root?
5920 l = _ltree_root; // Root is the outermost LoopNode
5921 } else { // Else found a nested loop
5922 // Insert a LoopNode to mark this loop.
5923 l = new IdealLoopTree(this, m, n);
5924 } // End of Else found a nested loop
5925 if (!has_loop(m)) { // If 'm' does not already have a loop set
5926 set_loop(m, l); // Set loop header to loop now
5927 }
5928 } else { // Else not a nested loop
5929 if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5930 IdealLoopTree* m_loop = get_loop(m);
5931 l = m_loop; // Get previously determined loop
5932 // If successor is header of a loop (nest), move up-loop till it
5933 // is a member of some outer enclosing loop. Since there are no
5934 // shared headers (I've split them already) I only need to go up
5935 // at most 1 level.
5936 while (l && l->_head == m) { // Successor heads loop?
5937 l = l->_parent; // Move up 1 for me
5938 }
5939 // If this loop is not properly parented, then this loop
5940 // has no exit path out, i.e. its an infinite loop.
5941 if (!l) {
5942 // Make loop "reachable" from root so the CFG is reachable. Basically
5943 // insert a bogus loop exit that is never taken. 'm', the loop head,
5944 // points to 'n', one (of possibly many) fall-in paths. There may be
5945 // many backedges as well.
5946
5947 if (!_verify_only) {
5948 // Insert the NeverBranch between 'm' and it's control user.
5949 NeverBranchNode *iff = new NeverBranchNode( m );
5950 _igvn.register_new_node_with_optimizer(iff);
5951 set_loop(iff, m_loop);
5952 Node *if_t = new CProjNode( iff, 0 );
5953 _igvn.register_new_node_with_optimizer(if_t);
5954 set_loop(if_t, m_loop);
5955
5956 Node* cfg = nullptr; // Find the One True Control User of m
5957 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5958 Node* x = m->fast_out(j);
5959 if (x->is_CFG() && x != m && x != iff)
5960 { cfg = x; break; }
5961 }
5962 assert(cfg != nullptr, "must find the control user of m");
5963 uint k = 0; // Probably cfg->in(0)
5964 while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5965 _igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5966
5967 // Now create the never-taken loop exit
5968 Node *if_f = new CProjNode( iff, 1 );
5969 _igvn.register_new_node_with_optimizer(if_f);
5970 set_loop(if_f, _ltree_root);
5971 // Find frame ptr for Halt. Relies on the optimizer
5972 // V-N'ing. Easier and quicker than searching through
5973 // the program structure.
5974 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5975 _igvn.register_new_node_with_optimizer(frame);
5976 // Halt & Catch Fire
5977 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5978 _igvn.register_new_node_with_optimizer(halt);
5979 set_loop(halt, _ltree_root);
5980 _igvn.add_input_to(C->root(), halt);
5981 }
5982 set_loop(C->root(), _ltree_root);
5983 // move to outer most loop with same header
5984 l = m_loop;
5985 while (true) {
5986 IdealLoopTree* next = l->_parent;
5987 if (next == nullptr || next->_head != m) {
5988 break;
5989 }
5990 l = next;
5991 }
5992 // properly insert infinite loop in loop tree
5993 sort(_ltree_root, l);
5994 // fix child link from parent
5995 IdealLoopTree* p = l->_parent;
5996 l->_next = p->_child;
5997 p->_child = l;
5998 // code below needs enclosing loop
5999 l = l->_parent;
6000 }
6001 }
6002 if (is_postvisited(l->_head)) {
6003 // We are currently visiting l, but its head has already been post-visited.
6004 // l is irreducible: we just found a second entry m.
6005 _has_irreducible_loops = true;
6006 RegionNode* secondary_entry = m->as_Region();
6007
6008 if (!secondary_entry->can_be_irreducible_entry()) {
6009 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
6010 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
6011 return pre_order;
6012 }
6013
6014 // Walk up the loop-tree, mark all loops that are already post-visited as irreducible
6015 // Since m is a secondary entry to them all.
6016 while( is_postvisited(l->_head) ) {
6017 l->_irreducible = 1; // = true
6018 RegionNode* head = l->_head->as_Region();
6019 if (!head->can_be_irreducible_entry()) {
6020 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
6021 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
6022 return pre_order;
6023 }
6024 l = l->_parent;
6025 // Check for bad CFG here to prevent crash, and bailout of compile
6026 if (l == nullptr) {
6027 #ifndef PRODUCT
6028 if (TraceLoopOpts) {
6029 tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");
6030 m->dump();
6031 }
6032 #endif
6033 // This is a rare case that we do not want to handle in C2.
6034 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
6035 return pre_order;
6036 }
6037 }
6038 }
6039 if (!_verify_only) {
6040 C->set_has_irreducible_loop(_has_irreducible_loops);
6041 }
6042
6043 // This Node might be a decision point for loops. It is only if
6044 // it's children belong to several different loops. The sort call
6045 // does a trivial amount of work if there is only 1 child or all
6046 // children belong to the same loop. If however, the children
6047 // belong to different loops, the sort call will properly set the
6048 // _parent pointers to show how the loops nest.
6049 //
6050 // In any case, it returns the tightest enclosing loop.
6051 innermost = sort( l, innermost );
6052 }
6053
6054 // Def-use info will have some dead stuff; dead stuff will have no
6055 // loop decided on.
6056
6057 // Am I a loop header? If so fix up my parent's child and next ptrs.
6058 if( innermost && innermost->_head == n ) {
6059 assert( get_loop(n) == innermost, "" );
6060 IdealLoopTree *p = innermost->_parent;
6061 IdealLoopTree *l = innermost;
6062 while (p && l->_head == n) {
6063 l->_next = p->_child; // Put self on parents 'next child'
6064 p->_child = l; // Make self as first child of parent
6065 l = p; // Now walk up the parent chain
6066 p = l->_parent;
6067 }
6068 } else {
6069 // Note that it is possible for a LoopNode to reach here, if the
6070 // backedge has been made unreachable (hence the LoopNode no longer
6071 // denotes a Loop, and will eventually be removed).
6072
6073 // Record tightest enclosing loop for self. Mark as post-visited.
6074 set_loop(n, innermost);
6075 // Also record has_call flag early on
6076 if (innermost) {
6077 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
6078 // Do not count uncommon calls
6079 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
6080 Node *iff = n->in(0)->in(0);
6081 // No any calls for vectorized loops.
6082 if (C->do_superword() ||
6083 !iff->is_If() ||
6084 (n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
6085 iff->as_If()->_prob >= 0.01) {
6086 innermost->_has_call = 1;
6087 }
6088 }
6089 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
6090 // Disable loop optimizations if the loop has a scalar replaceable
6091 // allocation. This disabling may cause a potential performance lost
6092 // if the allocation is not eliminated for some reason.
6093 innermost->_allow_optimizations = false;
6094 innermost->_has_call = 1; // = true
6095 } else if (n->Opcode() == Op_SafePoint) {
6096 // Record all safepoints in this loop.
6097 if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
6098 innermost->_safepts->push(n);
6099 }
6100 }
6101 }
6102
6103 // Flag as post-visited now
6104 set_postvisited(n);
6105 return pre_order;
6106 }
6107
6108 #ifdef ASSERT
6109 //--------------------------verify_regions_in_irreducible_loops----------------
6110 // Iterate down from Root through CFG, verify for every region:
6111 // if it is in an irreducible loop it must be marked as such
6112 void PhaseIdealLoop::verify_regions_in_irreducible_loops() {
6113 ResourceMark rm;
6114 if (!_has_irreducible_loops) {
6115 // last build_loop_tree has not found any irreducible loops
6116 // hence no region has to be marked is_in_irreduible_loop
6117 return;
6118 }
6119
6120 RootNode* root = C->root();
6121 Unique_Node_List worklist; // visit all nodes once
6122 worklist.push(root);
6123 bool failure = false;
6124 for (uint i = 0; i < worklist.size(); i++) {
6125 Node* n = worklist.at(i);
6126 if (n->is_Region()) {
6127 RegionNode* region = n->as_Region();
6128 if (is_in_irreducible_loop(region) &&
6129 region->loop_status() == RegionNode::LoopStatus::Reducible) {
6130 failure = true;
6131 tty->print("irreducible! ");
6132 region->dump();
6133 }
6134 }
6135 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
6136 Node* use = n->fast_out(j);
6137 if (use->is_CFG()) {
6138 worklist.push(use); // push if was not pushed before
6139 }
6140 }
6141 }
6142 assert(!failure, "region in irreducible loop was marked as reducible");
6143 }
6144
6145 //---------------------------is_in_irreducible_loop-------------------------
6146 // Analogous to ciTypeFlow::Block::is_in_irreducible_loop
6147 bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {
6148 if (!_has_irreducible_loops) {
6149 return false; // no irreducible loop in graph
6150 }
6151 IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
6152 do {
6153 if (l->_irreducible) {
6154 return true; // found it
6155 }
6156 if (l == _ltree_root) {
6157 return false; // reached root, terimnate
6158 }
6159 l = l->_parent;
6160 } while (l != nullptr);
6161 assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
6162 // We have "l->_parent == nullptr", which happens only for infinite loops,
6163 // where no parent is attached to the loop. We did not find any irreducible
6164 // loop from this block out to lp. Thus lp only has one entry, and no exit
6165 // (it is infinite and reducible). We can always rewrite an infinite loop
6166 // that is nested inside other loops:
6167 // while(condition) { infinite_loop; }
6168 // with an equivalent program where the infinite loop is an outermost loop
6169 // that is not nested in any loop:
6170 // while(condition) { break; } infinite_loop;
6171 // Thus, we can understand lp as an outermost loop, and can terminate and
6172 // conclude: this block is in no irreducible loop.
6173 return false;
6174 }
6175 #endif
6176
6177 //------------------------------build_loop_early-------------------------------
6178 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6179 // First pass computes the earliest controlling node possible. This is the
6180 // controlling input with the deepest dominating depth.
6181 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6182 while (worklist.size() != 0) {
6183 // Use local variables nstack_top_n & nstack_top_i to cache values
6184 // on nstack's top.
6185 Node *nstack_top_n = worklist.pop();
6186 uint nstack_top_i = 0;
6187 //while_nstack_nonempty:
6188 while (true) {
6189 // Get parent node and next input's index from stack's top.
6190 Node *n = nstack_top_n;
6191 uint i = nstack_top_i;
6192 uint cnt = n->req(); // Count of inputs
6193 if (i == 0) { // Pre-process the node.
6194 if( has_node(n) && // Have either loop or control already?
6195 !has_ctrl(n) ) { // Have loop picked out already?
6196 // During "merge_many_backedges" we fold up several nested loops
6197 // into a single loop. This makes the members of the original
6198 // loop bodies pointing to dead loops; they need to move up
6199 // to the new UNION'd larger loop. I set the _head field of these
6200 // dead loops to null and the _parent field points to the owning
6201 // loop. Shades of UNION-FIND algorithm.
6202 IdealLoopTree *ilt;
6203 while( !(ilt = get_loop(n))->_head ) {
6204 // Normally I would use a set_loop here. But in this one special
6205 // case, it is legal (and expected) to change what loop a Node
6206 // belongs to.
6207 _loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
6208 }
6209 // Remove safepoints ONLY if I've already seen I don't need one.
6210 // (the old code here would yank a 2nd safepoint after seeing a
6211 // first one, even though the 1st did not dominate in the loop body
6212 // and thus could be avoided indefinitely)
6213 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
6214 is_deleteable_safept(n)) {
6215 Node *in = n->in(TypeFunc::Control);
6216 replace_node_and_forward_ctrl(n, in); // Pull safepoint now
6217 if (ilt->_safepts != nullptr) {
6218 ilt->_safepts->yank(n);
6219 }
6220 // Carry on with the recursion "as if" we are walking
6221 // only the control input
6222 if( !visited.test_set( in->_idx ) ) {
6223 worklist.push(in); // Visit this guy later, using worklist
6224 }
6225 // Get next node from nstack:
6226 // - skip n's inputs processing by setting i > cnt;
6227 // - we also will not call set_early_ctrl(n) since
6228 // has_node(n) == true (see the condition above).
6229 i = cnt + 1;
6230 }
6231 }
6232 } // if (i == 0)
6233
6234 // Visit all inputs
6235 bool done = true; // Assume all n's inputs will be processed
6236 while (i < cnt) {
6237 Node *in = n->in(i);
6238 ++i;
6239 if (in == nullptr) continue;
6240 if (in->pinned() && !in->is_CFG())
6241 set_ctrl(in, in->in(0));
6242 int is_visited = visited.test_set( in->_idx );
6243 if (!has_node(in)) { // No controlling input yet?
6244 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
6245 assert( !is_visited, "visit only once" );
6246 nstack.push(n, i); // Save parent node and next input's index.
6247 nstack_top_n = in; // Process current input now.
6248 nstack_top_i = 0;
6249 done = false; // Not all n's inputs processed.
6250 break; // continue while_nstack_nonempty;
6251 } else if (!is_visited) {
6252 // This guy has a location picked out for him, but has not yet
6253 // been visited. Happens to all CFG nodes, for instance.
6254 // Visit him using the worklist instead of recursion, to break
6255 // cycles. Since he has a location already we do not need to
6256 // find his location before proceeding with the current Node.
6257 worklist.push(in); // Visit this guy later, using worklist
6258 }
6259 }
6260 if (done) {
6261 // All of n's inputs have been processed, complete post-processing.
6262
6263 // Compute earliest point this Node can go.
6264 // CFG, Phi, pinned nodes already know their controlling input.
6265 if (!has_node(n)) {
6266 // Record earliest legal location
6267 set_early_ctrl(n, false);
6268 }
6269 if (nstack.is_empty()) {
6270 // Finished all nodes on stack.
6271 // Process next node on the worklist.
6272 break;
6273 }
6274 // Get saved parent node and next input's index.
6275 nstack_top_n = nstack.node();
6276 nstack_top_i = nstack.index();
6277 nstack.pop();
6278 }
6279 } // while (true)
6280 }
6281 }
6282
6283 //------------------------------dom_lca_internal--------------------------------
6284 // Pair-wise LCA
6285 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
6286 if( !n1 ) return n2; // Handle null original LCA
6287 assert( n1->is_CFG(), "" );
6288 assert( n2->is_CFG(), "" );
6289 // find LCA of all uses
6290 uint d1 = dom_depth(n1);
6291 uint d2 = dom_depth(n2);
6292 while (n1 != n2) {
6293 if (d1 > d2) {
6294 n1 = idom(n1);
6295 d1 = dom_depth(n1);
6296 } else if (d1 < d2) {
6297 n2 = idom(n2);
6298 d2 = dom_depth(n2);
6299 } else {
6300 // Here d1 == d2. Due to edits of the dominator-tree, sections
6301 // of the tree might have the same depth. These sections have
6302 // to be searched more carefully.
6303
6304 // Scan up all the n1's with equal depth, looking for n2.
6305 Node *t1 = idom(n1);
6306 while (dom_depth(t1) == d1) {
6307 if (t1 == n2) return n2;
6308 t1 = idom(t1);
6309 }
6310 // Scan up all the n2's with equal depth, looking for n1.
6311 Node *t2 = idom(n2);
6312 while (dom_depth(t2) == d2) {
6313 if (t2 == n1) return n1;
6314 t2 = idom(t2);
6315 }
6316 // Move up to a new dominator-depth value as well as up the dom-tree.
6317 n1 = t1;
6318 n2 = t2;
6319 d1 = dom_depth(n1);
6320 d2 = dom_depth(n2);
6321 }
6322 }
6323 return n1;
6324 }
6325
6326 //------------------------------compute_idom-----------------------------------
6327 // Locally compute IDOM using dom_lca call. Correct only if the incoming
6328 // IDOMs are correct.
6329 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
6330 assert( region->is_Region(), "" );
6331 Node *LCA = nullptr;
6332 for( uint i = 1; i < region->req(); i++ ) {
6333 if( region->in(i) != C->top() )
6334 LCA = dom_lca( LCA, region->in(i) );
6335 }
6336 return LCA;
6337 }
6338
6339 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
6340 bool had_error = false;
6341 #ifdef ASSERT
6342 if (early != C->root()) {
6343 // Make sure that there's a dominance path from LCA to early
6344 Node* d = LCA;
6345 while (d != early) {
6346 if (d == C->root()) {
6347 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
6348 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
6349 had_error = true;
6350 break;
6351 }
6352 d = idom(d);
6353 }
6354 }
6355 #endif
6356 return had_error;
6357 }
6358
6359
6360 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
6361 // Compute LCA over list of uses
6362 bool had_error = false;
6363 Node *LCA = nullptr;
6364 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
6365 Node* c = n->fast_out(i);
6366 if (_loop_or_ctrl[c->_idx] == nullptr)
6367 continue; // Skip the occasional dead node
6368 if( c->is_Phi() ) { // For Phis, we must land above on the path
6369 for( uint j=1; j<c->req(); j++ ) {// For all inputs
6370 if( c->in(j) == n ) { // Found matching input?
6371 Node *use = c->in(0)->in(j);
6372 if (_verify_only && use->is_top()) continue;
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 } else {
6378 // For CFG data-users, use is in the block just prior
6379 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
6380 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6381 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6382 }
6383 }
6384 assert(!had_error, "bad dominance");
6385 return LCA;
6386 }
6387
6388 // Check the shape of the graph at the loop entry. In some cases,
6389 // the shape of the graph does not match the shape outlined below.
6390 // That is caused by the Opaque1 node "protecting" the shape of
6391 // the graph being removed by, for example, the IGVN performed
6392 // in PhaseIdealLoop::build_and_optimize().
6393 //
6394 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
6395 // loop unswitching, and IGVN, or a combination of them) can freely change
6396 // the graph's shape. As a result, the graph shape outlined below cannot
6397 // be guaranteed anymore.
6398 Node* CountedLoopNode::is_canonical_loop_entry() {
6399 if (!is_main_loop() && !is_post_loop()) {
6400 return nullptr;
6401 }
6402 Node* ctrl = skip_assertion_predicates_with_halt();
6403
6404 if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
6405 return nullptr;
6406 }
6407 Node* iffm = ctrl->in(0);
6408 if (iffm == nullptr || iffm->Opcode() != Op_If) {
6409 return nullptr;
6410 }
6411 Node* bolzm = iffm->in(1);
6412 if (bolzm == nullptr || !bolzm->is_Bool()) {
6413 return nullptr;
6414 }
6415 Node* cmpzm = bolzm->in(1);
6416 if (cmpzm == nullptr || !cmpzm->is_Cmp()) {
6417 return nullptr;
6418 }
6419
6420 uint input = is_main_loop() ? 2 : 1;
6421 if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {
6422 return nullptr;
6423 }
6424 bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
6425 #ifdef ASSERT
6426 bool found_opaque = false;
6427 for (uint i = 1; i < cmpzm->req(); i++) {
6428 Node* opnd = cmpzm->in(i);
6429 if (opnd && opnd->is_Opaque1()) {
6430 found_opaque = true;
6431 break;
6432 }
6433 }
6434 assert(found_opaque == res, "wrong pattern");
6435 #endif
6436 return res ? cmpzm->in(input) : nullptr;
6437 }
6438
6439 // Find pre loop end from main loop. Returns nullptr if none.
6440 CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {
6441 assert(is_main_loop(), "Can only find pre-loop from main-loop");
6442 // The loop cannot be optimized if the graph shape at the loop entry is
6443 // inappropriate.
6444 if (is_canonical_loop_entry() == nullptr) {
6445 return nullptr;
6446 }
6447
6448 Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
6449 if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {
6450 return nullptr;
6451 }
6452 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
6453 CountedLoopNode* loop_node = pre_end->loopnode();
6454 if (loop_node == nullptr || !loop_node->is_pre_loop()) {
6455 return nullptr;
6456 }
6457 return pre_end;
6458 }
6459
6460 Node* CountedLoopNode::uncasted_init_trip(bool uncast) {
6461 Node* init = init_trip();
6462 if (uncast && init->is_CastII()) {
6463 // skip over the cast added by PhaseIdealLoop::cast_incr_before_loop() when pre/post/main loops are created because
6464 // it can get in the way of type propagation. For instance, the index tested by an Assertion Predicate, if the cast
6465 // is not skipped over, could be (1):
6466 // (AddI (CastII (AddI pre_loop_iv -2) int) 1)
6467 // while without the cast, it is (2):
6468 // (AddI (AddI pre_loop_iv -2) 1)
6469 // which is be transformed to (3):
6470 // (AddI pre_loop_iv -1)
6471 // The compiler may be able to constant fold the Assertion Predicate condition for (3) but not (1)
6472 assert(init->as_CastII()->carry_dependency() && skip_assertion_predicates_with_halt() == init->in(0), "casted iv phi from pre loop expected");
6473 init = init->in(1);
6474 }
6475 return init;
6476 }
6477
6478 //------------------------------get_late_ctrl----------------------------------
6479 // Compute latest legal control.
6480 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
6481 assert(early != nullptr, "early control should not be null");
6482
6483 Node* LCA = compute_lca_of_uses(n, early);
6484 #ifdef ASSERT
6485 if (LCA == C->root() && LCA != early) {
6486 // def doesn't dominate uses so print some useful debugging output
6487 compute_lca_of_uses(n, early, true);
6488 }
6489 #endif
6490
6491 if (n->is_Load() && LCA != early) {
6492 LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6493 }
6494
6495 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6496 return LCA;
6497 }
6498
6499 // if this is a load, check for anti-dependent stores
6500 // We use a conservative algorithm to identify potential interfering
6501 // instructions and for rescheduling the load. The users of the memory
6502 // input of this load are examined. Any use which is not a load and is
6503 // dominated by early is considered a potentially interfering store.
6504 // This can produce false positives.
6505 Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {
6506 int load_alias_idx = C->get_alias_index(n->adr_type());
6507 if (C->alias_type(load_alias_idx)->is_rewritable()) {
6508 Unique_Node_List worklist;
6509
6510 Node* mem = n->in(MemNode::Memory);
6511 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
6512 Node* s = mem->fast_out(i);
6513 worklist.push(s);
6514 }
6515 for (uint i = 0; i < worklist.size() && LCA != early; i++) {
6516 Node* s = worklist.at(i);
6517 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6518 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6519 s->is_Phi()) {
6520 continue;
6521 } else if (s->is_MergeMem()) {
6522 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6523 Node* s1 = s->fast_out(i);
6524 worklist.push(s1);
6525 }
6526 } else {
6527 Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6528 assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6529 if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {
6530 const TypePtr* adr_type = s->adr_type();
6531 if (s->is_ArrayCopy()) {
6532 // Copy to known instance needs destination type to test for aliasing
6533 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6534 if (dest_type != TypeOopPtr::BOTTOM) {
6535 adr_type = dest_type;
6536 }
6537 }
6538 if (C->can_alias(adr_type, load_alias_idx)) {
6539 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6540 } else if (s->is_CFG() && s->is_Multi()) {
6541 // Look for the memory use of s (that is the use of its memory projection)
6542 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6543 Node* s1 = s->fast_out(i);
6544 assert(s1->is_Proj(), "projection expected");
6545 if (_igvn.type(s1) == Type::MEMORY) {
6546 for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
6547 Node* s2 = s1->fast_out(j);
6548 worklist.push(s2);
6549 }
6550 }
6551 }
6552 }
6553 }
6554 }
6555 }
6556 // For Phis only consider Region's inputs that were reached by following the memory edges
6557 if (LCA != early) {
6558 for (uint i = 0; i < worklist.size(); i++) {
6559 Node* s = worklist.at(i);
6560 if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
6561 Node* r = s->in(0);
6562 for (uint j = 1; j < s->req(); j++) {
6563 Node* in = s->in(j);
6564 Node* r_in = r->in(j);
6565 // We can't reach any node from a Phi because we don't enqueue Phi's uses above
6566 if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
6567 LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6568 }
6569 }
6570 }
6571 }
6572 }
6573 }
6574 return LCA;
6575 }
6576
6577 // Is CFG node 'dominator' dominating node 'n'?
6578 bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {
6579 if (dominator == n) {
6580 return true;
6581 }
6582 assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6583 uint dd = dom_depth(dominator);
6584 while (dom_depth(n) >= dd) {
6585 if (n == dominator) {
6586 return true;
6587 }
6588 n = idom(n);
6589 }
6590 return false;
6591 }
6592
6593 // Is CFG node 'dominator' strictly dominating node 'n'?
6594 bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {
6595 return dominator != n && is_dominator(dominator, n);
6596 }
6597
6598 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
6599 // Pair-wise LCA with tags.
6600 // Tag each index with the node 'tag' currently being processed
6601 // before advancing up the dominator chain using idom().
6602 // Later calls that find a match to 'tag' know that this path has already
6603 // been considered in the current LCA (which is input 'n1' by convention).
6604 // Since get_late_ctrl() is only called once for each node, the tag array
6605 // does not need to be cleared between calls to get_late_ctrl().
6606 // Algorithm trades a larger constant factor for better asymptotic behavior
6607 //
6608 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
6609 uint d1 = dom_depth(n1);
6610 uint d2 = dom_depth(n2);
6611 jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6612
6613 do {
6614 if (d1 > d2) {
6615 // current lca is deeper than n2
6616 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6617 n1 = idom(n1);
6618 d1 = dom_depth(n1);
6619 } else if (d1 < d2) {
6620 // n2 is deeper than current lca
6621 jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6622 if (memo == tag) {
6623 return n1; // Return the current LCA
6624 }
6625 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6626 n2 = idom(n2);
6627 d2 = dom_depth(n2);
6628 } else {
6629 // Here d1 == d2. Due to edits of the dominator-tree, sections
6630 // of the tree might have the same depth. These sections have
6631 // to be searched more carefully.
6632
6633 // Scan up all the n1's with equal depth, looking for n2.
6634 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6635 Node *t1 = idom(n1);
6636 while (dom_depth(t1) == d1) {
6637 if (t1 == n2) return n2;
6638 _dom_lca_tags.at_put_grow(t1->_idx, tag);
6639 t1 = idom(t1);
6640 }
6641 // Scan up all the n2's with equal depth, looking for n1.
6642 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6643 Node *t2 = idom(n2);
6644 while (dom_depth(t2) == d2) {
6645 if (t2 == n1) return n1;
6646 _dom_lca_tags.at_put_grow(t2->_idx, tag);
6647 t2 = idom(t2);
6648 }
6649 // Move up to a new dominator-depth value as well as up the dom-tree.
6650 n1 = t1;
6651 n2 = t2;
6652 d1 = dom_depth(n1);
6653 d2 = dom_depth(n2);
6654 }
6655 } while (n1 != n2);
6656 return n1;
6657 }
6658
6659 //------------------------------init_dom_lca_tags------------------------------
6660 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
6661 // Intended use does not involve any growth for the array, so it could
6662 // be of fixed size.
6663 void PhaseIdealLoop::init_dom_lca_tags() {
6664 uint limit = C->unique() + 1;
6665 _dom_lca_tags.at_grow(limit, 0);
6666 _dom_lca_tags_round = 0;
6667 #ifdef ASSERT
6668 for (uint i = 0; i < limit; ++i) {
6669 assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6670 }
6671 #endif // ASSERT
6672 }
6673
6674 //------------------------------build_loop_late--------------------------------
6675 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6676 // Second pass finds latest legal placement, and ideal loop placement.
6677 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6678 while (worklist.size() != 0) {
6679 Node *n = worklist.pop();
6680 // Only visit once
6681 if (visited.test_set(n->_idx)) continue;
6682 uint cnt = n->outcnt();
6683 uint i = 0;
6684 while (true) {
6685 assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6686 // Visit all children
6687 if (i < cnt) {
6688 Node* use = n->raw_out(i);
6689 ++i;
6690 // Check for dead uses. Aggressively prune such junk. It might be
6691 // dead in the global sense, but still have local uses so I cannot
6692 // easily call 'remove_dead_node'.
6693 if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?
6694 // Due to cycles, we might not hit the same fixed point in the verify
6695 // pass as we do in the regular pass. Instead, visit such phis as
6696 // simple uses of the loop head.
6697 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
6698 if( !visited.test(use->_idx) )
6699 worklist.push(use);
6700 } else if( !visited.test_set(use->_idx) ) {
6701 nstack.push(n, i); // Save parent and next use's index.
6702 n = use; // Process all children of current use.
6703 cnt = use->outcnt();
6704 i = 0;
6705 }
6706 } else {
6707 // Do not visit around the backedge of loops via data edges.
6708 // push dead code onto a worklist
6709 _deadlist.push(use);
6710 }
6711 } else {
6712 // All of n's children have been processed, complete post-processing.
6713 build_loop_late_post(n);
6714 if (C->failing()) { return; }
6715 if (nstack.is_empty()) {
6716 // Finished all nodes on stack.
6717 // Process next node on the worklist.
6718 break;
6719 }
6720 // Get saved parent node and next use's index. Visit the rest of uses.
6721 n = nstack.node();
6722 cnt = n->outcnt();
6723 i = nstack.index();
6724 nstack.pop();
6725 }
6726 }
6727 }
6728 }
6729
6730 // Verify that no data node is scheduled in the outer loop of a strip
6731 // mined loop.
6732 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
6733 #ifdef ASSERT
6734 if (get_loop(least)->_nest == 0) {
6735 return;
6736 }
6737 IdealLoopTree* loop = get_loop(least);
6738 Node* head = loop->_head;
6739 if (head->is_OuterStripMinedLoop() &&
6740 // Verification can't be applied to fully built strip mined loops
6741 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
6742 Node* sfpt = head->as_Loop()->outer_safepoint();
6743 ResourceMark rm;
6744 Unique_Node_List wq;
6745 wq.push(sfpt);
6746 for (uint i = 0; i < wq.size(); i++) {
6747 Node *m = wq.at(i);
6748 for (uint i = 1; i < m->req(); i++) {
6749 Node* nn = m->in(i);
6750 if (nn == n) {
6751 return;
6752 }
6753 if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
6754 wq.push(nn);
6755 }
6756 }
6757 }
6758 ShouldNotReachHere();
6759 }
6760 #endif
6761 }
6762
6763
6764 //------------------------------build_loop_late_post---------------------------
6765 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6766 // Second pass finds latest legal placement, and ideal loop placement.
6767 void PhaseIdealLoop::build_loop_late_post(Node *n) {
6768 build_loop_late_post_work(n, true);
6769 }
6770
6771 // Class to visit all predicates in a predicate chain to find out which are dominated by a given node. Keeps track of
6772 // the entry to the earliest predicate that is still dominated by the given dominator. This class is used when trying to
6773 // legally skip all predicates when figuring out the latest placement such that a node does not interfere with Loop
6774 // Predication or creating a Loop Limit Check Predicate later.
6775 class DominatedPredicates : public UnifiedPredicateVisitor {
6776 Node* const _dominator;
6777 Node* _earliest_dominated_predicate_entry;
6778 bool _should_continue;
6779 PhaseIdealLoop* const _phase;
6780
6781 public:
6782 DominatedPredicates(Node* dominator, Node* start_node, PhaseIdealLoop* phase)
6783 : _dominator(dominator),
6784 _earliest_dominated_predicate_entry(start_node),
6785 _should_continue(true),
6786 _phase(phase) {}
6787 NONCOPYABLE(DominatedPredicates);
6788
6789 bool should_continue() const override {
6790 return _should_continue;
6791 }
6792
6793 // Returns the entry to the earliest predicate that is still dominated by the given dominator (all could be dominated).
6794 Node* earliest_dominated_predicate_entry() const {
6795 return _earliest_dominated_predicate_entry;
6796 }
6797
6798 void visit_predicate(const Predicate& predicate) override {
6799 Node* entry = predicate.entry();
6800 if (_phase->is_strict_dominator(entry, _dominator)) {
6801 _should_continue = false;
6802 } else {
6803 _earliest_dominated_predicate_entry = entry;
6804 }
6805 }
6806 };
6807
6808 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
6809
6810 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
6811 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6812 }
6813
6814 #ifdef ASSERT
6815 if (_verify_only && !n->is_CFG()) {
6816 // Check def-use domination.
6817 // We would like to expose this check in product but it appears to be expensive.
6818 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
6819 }
6820 #endif
6821
6822 // CFG and pinned nodes already handled
6823 if( n->in(0) ) {
6824 if( n->in(0)->is_top() ) return; // Dead?
6825
6826 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
6827 // _must_ be pinned (they have to observe their control edge of course).
6828 // Unlike Stores (which modify an unallocable resource, the memory
6829 // state), Mods/Loads can float around. So free them up.
6830 switch( n->Opcode() ) {
6831 case Op_DivI:
6832 case Op_DivF:
6833 case Op_DivD:
6834 case Op_ModI:
6835 case Op_LoadB: // Same with Loads; they can sink
6836 case Op_LoadUB: // during loop optimizations.
6837 case Op_LoadUS:
6838 case Op_LoadD:
6839 case Op_LoadF:
6840 case Op_LoadI:
6841 case Op_LoadKlass:
6842 case Op_LoadNKlass:
6843 case Op_LoadL:
6844 case Op_LoadS:
6845 case Op_LoadP:
6846 case Op_LoadN:
6847 case Op_LoadRange:
6848 case Op_LoadD_unaligned:
6849 case Op_LoadL_unaligned:
6850 case Op_StrComp: // Does a bunch of load-like effects
6851 case Op_StrEquals:
6852 case Op_StrIndexOf:
6853 case Op_StrIndexOfChar:
6854 case Op_AryEq:
6855 case Op_VectorizedHashCode:
6856 case Op_CountPositives:
6857 pinned = false;
6858 }
6859 if (n->is_CMove() || n->is_ConstraintCast()) {
6860 pinned = false;
6861 }
6862 if( pinned ) {
6863 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6864 if( !chosen_loop->_child ) // Inner loop?
6865 chosen_loop->_body.push(n); // Collect inner loops
6866 return;
6867 }
6868 } else { // No slot zero
6869 if( n->is_CFG() ) { // CFG with no slot 0 is dead
6870 _loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6871 return;
6872 }
6873 assert(!n->is_CFG() || n->outcnt() == 0, "");
6874 }
6875
6876 // Do I have a "safe range" I can select over?
6877 Node *early = get_ctrl(n);// Early location already computed
6878
6879 // Compute latest point this Node can go
6880 Node *LCA = get_late_ctrl( n, early );
6881 // LCA is null due to uses being dead
6882 if( LCA == nullptr ) {
6883 #ifdef ASSERT
6884 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
6885 assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6886 }
6887 #endif
6888 _loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6889 _deadlist.push(n);
6890 return;
6891 }
6892 assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6893
6894 Node *legal = LCA; // Walk 'legal' up the IDOM chain
6895 Node *least = legal; // Best legal position so far
6896 while( early != legal ) { // While not at earliest legal
6897 if (legal->is_Start() && !early->is_Root()) {
6898 #ifdef ASSERT
6899 // Bad graph. Print idom path and fail.
6900 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
6901 assert(false, "Bad graph detected in build_loop_late");
6902 #endif
6903 C->record_method_not_compilable("Bad graph detected in build_loop_late");
6904 return;
6905 }
6906 // Find least loop nesting depth
6907 legal = idom(legal); // Bump up the IDOM tree
6908 // Check for lower nesting depth
6909 if( get_loop(legal)->_nest < get_loop(least)->_nest )
6910 least = legal;
6911 }
6912 assert(early == legal || legal != C->root(), "bad dominance of inputs");
6913
6914 if (least != early) {
6915 // Move the node above predicates as far up as possible so a
6916 // following pass of Loop Predication doesn't hoist a predicate
6917 // that depends on it above that node.
6918 const PredicateIterator predicate_iterator(least);
6919 DominatedPredicates dominated_predicates(early, least, this);
6920 predicate_iterator.for_each(dominated_predicates);
6921 least = dominated_predicates.earliest_dominated_predicate_entry();
6922 }
6923 // Try not to place code on a loop entry projection
6924 // which can inhibit range check elimination.
6925 if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {
6926 Node* ctrl_out = least->unique_ctrl_out_or_null();
6927 if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6928 least == ctrl_out->in(LoopNode::EntryControl) &&
6929 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
6930 Node* least_dom = idom(least);
6931 if (get_loop(least_dom)->is_member(get_loop(least))) {
6932 least = least_dom;
6933 }
6934 }
6935 }
6936 // Don't extend live ranges of raw oops
6937 if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6938 !n->bottom_type()->isa_rawptr()) {
6939 least = early;
6940 }
6941
6942 #ifdef ASSERT
6943 // Broken part of VerifyLoopOptimizations (F)
6944 // Reason:
6945 // _verify_me->get_ctrl_no_update(n) seems to return wrong result
6946 /*
6947 // If verifying, verify that 'verify_me' has a legal location
6948 // and choose it as our location.
6949 if( _verify_me ) {
6950 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6951 Node *legal = LCA;
6952 while( early != legal ) { // While not at earliest legal
6953 if( legal == v_ctrl ) break; // Check for prior good location
6954 legal = idom(legal) ;// Bump up the IDOM tree
6955 }
6956 // Check for prior good location
6957 if( legal == v_ctrl ) least = legal; // Keep prior if found
6958 }
6959 */
6960 #endif
6961
6962 // Assign discovered "here or above" point
6963 least = find_non_split_ctrl(least);
6964 verify_strip_mined_scheduling(n, least);
6965 set_ctrl(n, least);
6966
6967 // Collect inner loop bodies
6968 IdealLoopTree *chosen_loop = get_loop(least);
6969 if( !chosen_loop->_child ) // Inner loop?
6970 chosen_loop->_body.push(n);// Collect inner loops
6971
6972 if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {
6973 _zero_trip_guard_opaque_nodes.push(n);
6974 }
6975
6976 if (!_verify_only && n->Opcode() == Op_OpaqueMultiversioning) {
6977 _multiversion_opaque_nodes.push(n);
6978 }
6979 }
6980
6981 #ifdef ASSERT
6982 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
6983 tty->print_cr("%s", msg);
6984 tty->print("n: "); n->dump();
6985 tty->print("early(n): "); early->dump();
6986 if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6987 n->in(0) != early && !n->in(0)->is_Root()) {
6988 tty->print("n->in(0): "); n->in(0)->dump();
6989 }
6990 for (uint i = 1; i < n->req(); i++) {
6991 Node* in1 = n->in(i);
6992 if (in1 != nullptr && in1 != n && !in1->is_top()) {
6993 tty->print("n->in(%d): ", i); in1->dump();
6994 Node* in1_early = get_ctrl(in1);
6995 tty->print("early(n->in(%d)): ", i); in1_early->dump();
6996 if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6997 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
6998 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
6999 }
7000 for (uint j = 1; j < in1->req(); j++) {
7001 Node* in2 = in1->in(j);
7002 if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {
7003 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
7004 Node* in2_early = get_ctrl(in2);
7005 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
7006 if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
7007 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
7008 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
7009 }
7010 }
7011 }
7012 }
7013 }
7014 tty->cr();
7015 tty->print("LCA(n): "); LCA->dump();
7016 for (uint i = 0; i < n->outcnt(); i++) {
7017 Node* u1 = n->raw_out(i);
7018 if (u1 == n)
7019 continue;
7020 tty->print("n->out(%d): ", i); u1->dump();
7021 if (u1->is_CFG()) {
7022 for (uint j = 0; j < u1->outcnt(); j++) {
7023 Node* u2 = u1->raw_out(j);
7024 if (u2 != u1 && u2 != n && u2->is_CFG()) {
7025 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
7026 }
7027 }
7028 } else {
7029 Node* u1_later = get_ctrl(u1);
7030 tty->print("later(n->out(%d)): ", i); u1_later->dump();
7031 if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
7032 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
7033 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
7034 }
7035 for (uint j = 0; j < u1->outcnt(); j++) {
7036 Node* u2 = u1->raw_out(j);
7037 if (u2 == n || u2 == u1)
7038 continue;
7039 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
7040 if (!u2->is_CFG()) {
7041 Node* u2_later = get_ctrl(u2);
7042 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
7043 if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
7044 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
7045 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
7046 }
7047 }
7048 }
7049 }
7050 }
7051 dump_idoms(early, LCA);
7052 tty->cr();
7053 }
7054
7055 // Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
7056 class RealLCA {
7057 const PhaseIdealLoop* _phase;
7058 Node* _early;
7059 Node* _wrong_lca;
7060 uint _early_index;
7061 int _wrong_lca_index;
7062
7063 // Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
7064 // is different: Return the previously visited node which must be the real LCA.
7065 // The node lists also contain _early and _wrong_lca, respectively.
7066 Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {
7067 int early_index = early_with_idoms.size() - 1;
7068 int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
7069 bool found_difference = false;
7070 do {
7071 if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {
7072 // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
7073 found_difference = true;
7074 break;
7075 }
7076 early_index--;
7077 wrong_lca_index--;
7078 } while (wrong_lca_index >= 0);
7079
7080 assert(early_index >= 0, "must always find an LCA - cannot be early");
7081 _early_index = early_index;
7082 _wrong_lca_index = wrong_lca_index;
7083 Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
7084 assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
7085 return real_lca;
7086 }
7087
7088 void dump(Node* real_lca) {
7089 tty->cr();
7090 tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
7091 _phase->dump_idom(_early, _early_index + 1);
7092
7093 tty->cr();
7094 tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());
7095 _phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
7096
7097 tty->cr();
7098 tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",
7099 _early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
7100 if (_wrong_lca_index >= 0) {
7101 tty->print(" (idom[%d])", _wrong_lca_index);
7102 }
7103 tty->print_cr(":");
7104 real_lca->dump();
7105 }
7106
7107 public:
7108 RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
7109 : _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {
7110 assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
7111 }
7112
7113 void compute_and_dump() {
7114 ResourceMark rm;
7115 Unique_Node_List early_with_idoms;
7116 Unique_Node_List wrong_lca_with_idoms;
7117 early_with_idoms.push(_early);
7118 wrong_lca_with_idoms.push(_wrong_lca);
7119 _phase->get_idoms(_early, 10000, early_with_idoms);
7120 _phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
7121 Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
7122 dump(real_lca);
7123 }
7124 };
7125
7126 // Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
7127 void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
7128 assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
7129 assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
7130
7131 RealLCA real_lca(this, early, wrong_lca);
7132 real_lca.compute_and_dump();
7133 }
7134 #endif // ASSERT
7135
7136 #ifndef PRODUCT
7137 //------------------------------dump-------------------------------------------
7138 void PhaseIdealLoop::dump() const {
7139 ResourceMark rm;
7140 Node_Stack stack(C->live_nodes() >> 2);
7141 Node_List rpo_list;
7142 VectorSet visited;
7143 visited.set(C->top()->_idx);
7144 rpo(C->root(), stack, visited, rpo_list);
7145 // Dump root loop indexed by last element in PO order
7146 dump(_ltree_root, rpo_list.size(), rpo_list);
7147 }
7148
7149 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
7150 loop->dump_head();
7151
7152 // Now scan for CFG nodes in the same loop
7153 for (uint j = idx; j > 0; j--) {
7154 Node* n = rpo_list[j-1];
7155 if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
7156 continue;
7157
7158 if (get_loop(n) != loop) { // Wrong loop nest
7159 if (get_loop(n)->_head == n && // Found nested loop?
7160 get_loop(n)->_parent == loop)
7161 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
7162 continue;
7163 }
7164
7165 // Dump controlling node
7166 tty->sp(2 * loop->_nest);
7167 tty->print("C");
7168 if (n == C->root()) {
7169 n->dump();
7170 } else {
7171 Node* cached_idom = idom_no_update(n);
7172 Node* computed_idom = n->in(0);
7173 if (n->is_Region()) {
7174 computed_idom = compute_idom(n);
7175 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
7176 // any MultiBranch ctrl node), so apply a similar transform to
7177 // the cached idom returned from idom_no_update.
7178 cached_idom = find_non_split_ctrl(cached_idom);
7179 }
7180 tty->print(" ID:%d", computed_idom->_idx);
7181 n->dump();
7182 if (cached_idom != computed_idom) {
7183 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
7184 computed_idom->_idx, cached_idom->_idx);
7185 }
7186 }
7187 // Dump nodes it controls
7188 for (uint k = 0; k < _loop_or_ctrl.max(); k++) {
7189 // (k < C->unique() && get_ctrl(find(k)) == n)
7190 if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {
7191 Node* m = C->root()->find(k);
7192 if (m && m->outcnt() > 0) {
7193 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
7194 tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",
7195 _loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
7196 }
7197 tty->sp(2 * loop->_nest + 1);
7198 m->dump();
7199 }
7200 }
7201 }
7202 }
7203 }
7204
7205 void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {
7206 if (has_ctrl(n)) {
7207 tty->print_cr("No idom for data nodes");
7208 } else {
7209 ResourceMark rm;
7210 Unique_Node_List idoms;
7211 get_idoms(n, count, idoms);
7212 dump_idoms_in_reverse(n, idoms);
7213 }
7214 }
7215
7216 void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
7217 Node* next = n;
7218 for (uint i = 0; !next->is_Start() && i < count; i++) {
7219 next = idom(next);
7220 assert(!idoms.member(next), "duplicated idom is not possible");
7221 idoms.push(next);
7222 }
7223 }
7224
7225 void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
7226 Node* next;
7227 uint padding = 3;
7228 uint node_index_padding_width = (C->unique() == 0 ? 0 : static_cast<int>(log10(static_cast<double>(C->unique())))) + 1;
7229 for (int i = idom_list.size() - 1; i >= 0; i--) {
7230 if (i == 9 || i == 99) {
7231 padding++;
7232 }
7233 next = idom_list[i];
7234 tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
7235 }
7236 tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
7237 }
7238 #endif // NOT PRODUCT
7239
7240 // Collect a R-P-O for the whole CFG.
7241 // Result list is in post-order (scan backwards for RPO)
7242 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
7243 stk.push(start, 0);
7244 visited.set(start->_idx);
7245
7246 while (stk.is_nonempty()) {
7247 Node* m = stk.node();
7248 uint idx = stk.index();
7249 if (idx < m->outcnt()) {
7250 stk.set_index(idx + 1);
7251 Node* n = m->raw_out(idx);
7252 if (n->is_CFG() && !visited.test_set(n->_idx)) {
7253 stk.push(n, 0);
7254 }
7255 } else {
7256 rpo_list.push(m);
7257 stk.pop();
7258 }
7259 }
7260 }
7261
7262 ConINode* PhaseIdealLoop::intcon(jint i) {
7263 ConINode* node = _igvn.intcon(i);
7264 set_root_as_ctrl(node);
7265 return node;
7266 }
7267
7268 ConLNode* PhaseIdealLoop::longcon(jlong i) {
7269 ConLNode* node = _igvn.longcon(i);
7270 set_root_as_ctrl(node);
7271 return node;
7272 }
7273
7274 ConNode* PhaseIdealLoop::makecon(const Type* t) {
7275 ConNode* node = _igvn.makecon(t);
7276 set_root_as_ctrl(node);
7277 return node;
7278 }
7279
7280 ConNode* PhaseIdealLoop::integercon(jlong l, BasicType bt) {
7281 ConNode* node = _igvn.integercon(l, bt);
7282 set_root_as_ctrl(node);
7283 return node;
7284 }
7285
7286 ConNode* PhaseIdealLoop::zerocon(BasicType bt) {
7287 ConNode* node = _igvn.zerocon(bt);
7288 set_root_as_ctrl(node);
7289 return node;
7290 }
7291
7292
7293 //=============================================================================
7294 //------------------------------LoopTreeIterator-------------------------------
7295
7296 // Advance to next loop tree using a preorder, left-to-right traversal.
7297 void LoopTreeIterator::next() {
7298 assert(!done(), "must not be done.");
7299 if (_curnt->_child != nullptr) {
7300 _curnt = _curnt->_child;
7301 } else if (_curnt->_next != nullptr) {
7302 _curnt = _curnt->_next;
7303 } else {
7304 while (_curnt != _root && _curnt->_next == nullptr) {
7305 _curnt = _curnt->_parent;
7306 }
7307 if (_curnt == _root) {
7308 _curnt = nullptr;
7309 assert(done(), "must be done.");
7310 } else {
7311 assert(_curnt->_next != nullptr, "must be more to do");
7312 _curnt = _curnt->_next;
7313 }
7314 }
7315 }