1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciMethodData.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/barrierSet.hpp"
28 #include "gc/shared/c2/barrierSetC2.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/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(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 // If the loop has the shape of a counted loop but with a long
764 // induction variable, transform the loop in a loop nest: an inner
765 // loop that iterates for at most max int iterations with an integer
766 // induction variable and an outer loop that iterates over the full
767 // range of long values from the initial loop in (at most) max int
768 // steps. That is:
769 //
770 // x: for (long phi = init; phi < limit; phi += stride) {
771 // // phi := Phi(L, init, incr)
772 // // incr := AddL(phi, longcon(stride))
773 // long incr = phi + stride;
774 // ... use phi and incr ...
775 // }
776 //
777 // OR:
778 //
779 // x: for (long phi = init; (phi += stride) < limit; ) {
780 // // phi := Phi(L, AddL(init, stride), incr)
781 // // incr := AddL(phi, longcon(stride))
782 // long incr = phi + stride;
783 // ... use phi and (phi + stride) ...
784 // }
785 //
786 // ==transform=>
787 //
788 // const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0
789 // assert(stride <= inner_iters_limit); // else abort transform
790 // assert((extralong)limit + stride <= LONG_MAX); // else deopt
791 // outer_head: for (long outer_phi = init;;) {
792 // // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi)))
793 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi));
794 // long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max);
795 // assert(inner_iters_actual == (int)inner_iters_actual);
796 // int inner_phi, inner_incr;
797 // x: for (inner_phi = 0;; inner_phi = inner_incr) {
798 // // inner_phi := Phi(x, intcon(0), inner_incr)
799 // // inner_incr := AddI(inner_phi, intcon(stride))
800 // inner_incr = inner_phi + stride;
801 // if (inner_incr < inner_iters_actual) {
802 // ... use phi=>(outer_phi+inner_phi) ...
803 // continue;
804 // }
805 // else break;
806 // }
807 // if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit
808 // continue;
809 // else break;
810 // }
811 //
812 // The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int
813 // loops with long range checks transformed to int range checks in the inner loop.
814 bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {
815 Node* x = loop->_head;
816 // Only for inner loops
817 if (loop->_child != nullptr || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) {
818 return false;
819 }
820
821 if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) {
822 return false;
823 }
824
825 BaseCountedLoopNode* head = x->as_BaseCountedLoop();
826 BasicType bt = x->as_BaseCountedLoop()->bt();
827
828 check_counted_loop_shape(loop, x, bt);
829
830 #ifndef PRODUCT
831 if (bt == T_LONG) {
832 AtomicAccess::inc(&_long_loop_candidates);
833 }
834 #endif
835
836 jlong stride_con_long = head->stride_con();
837 assert(stride_con_long != 0, "missed some peephole opt");
838 // We can't iterate for more than max int at a time.
839 if (stride_con_long != (jint)stride_con_long || stride_con_long == min_jint) {
840 assert(bt == T_LONG, "only for long loops");
841 return false;
842 }
843 jint stride_con = checked_cast<jint>(stride_con_long);
844 // The number of iterations for the integer count loop: guarantee no
845 // overflow: max_jint - stride_con max. -1 so there's no need for a
846 // loop limit check if the exit test is <= or >=.
847 int iters_limit = max_jint - ABS(stride_con) - 1;
848 #ifdef ASSERT
849 if (bt == T_LONG && StressLongCountedLoop > 0) {
850 iters_limit = iters_limit / StressLongCountedLoop;
851 }
852 #endif
853 // At least 2 iterations so counted loop construction doesn't fail
854 if (iters_limit/ABS(stride_con) < 2) {
855 return false;
856 }
857
858 assert(iters_limit > 0, "can't be negative");
859
860 PhiNode* phi = head->phi()->as_Phi();
861
862 Node* back_control = head->in(LoopNode::LoopBackControl);
863
864 // data nodes on back branch not supported
865 if (back_control->outcnt() > 1) {
866 return false;
867 }
868
869 Node* limit = head->limit();
870 // We'll need to use the loop limit before the inner loop is entered
871 if (!is_dominator(get_ctrl(limit), x)) {
872 return false;
873 }
874
875 IfNode* exit_test = head->loopexit();
876
877 assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
878
879 Node_List range_checks;
880 iters_limit = extract_long_range_checks(loop, stride_con, iters_limit, phi, range_checks);
881
882 if (bt == T_INT) {
883 // The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further.
884 if (range_checks.size() == 0) {
885 return false;
886 }
887 }
888
889 // Take what we know about the number of iterations of the long counted loop into account when computing the limit of
890 // the inner loop.
891 Node* init = head->init_trip();
892 const TypeInteger* lo = _igvn.type(init)->is_integer(bt);
893 const TypeInteger* hi = _igvn.type(limit)->is_integer(bt);
894 if (stride_con < 0) {
895 swap(lo, hi);
896 }
897 if (hi->hi_as_long() <= lo->lo_as_long()) {
898 // not a loop after all
899 return false;
900 }
901
902 if (range_checks.size() > 0) {
903 // This transformation requires peeling one iteration. Also, if it has range checks and they are eliminated by Loop
904 // Predication, then 2 Hoisted Check Predicates are added for one range check. Finally, transforming a long range
905 // check requires extra logic to be executed before the loop is entered and for the outer loop. As a result, the
906 // transformations can't pay off for a small number of iterations: roughly, if the loop runs for 3 iterations, it's
907 // going to execute as many range checks once transformed with range checks eliminated (1 peeled iteration with
908 // range checks + 2 predicates per range checks) as it would have not transformed. It also has to pay for the extra
909 // logic on loop entry and for the outer loop.
910 loop->compute_trip_count(this, bt);
911 if (head->is_CountedLoop() && head->as_CountedLoop()->has_exact_trip_count()) {
912 if (head->as_CountedLoop()->trip_count() <= 3) {
913 return false;
914 }
915 } else {
916 loop->compute_profile_trip_cnt(this);
917 if (!head->is_profile_trip_failed() && head->profile_trip_cnt() <= 3) {
918 return false;
919 }
920 }
921 }
922
923 if (try_make_short_running_loop(loop, stride_con, range_checks, iters_limit)) {
924 C->set_major_progress();
925 return true;
926 }
927
928 julong orig_iters = (julong)hi->hi_as_long() - lo->lo_as_long();
929 iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
930
931 // We need a safepoint to insert Parse Predicates for the inner loop.
932 SafePointNode* safepoint;
933 if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) {
934 // Loop is strip mined: use the safepoint of the outer strip mined loop
935 OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
936 assert(outer_loop != nullptr, "no outer loop");
937 safepoint = outer_loop->outer_safepoint();
938 outer_loop->transform_to_counted_loop(&_igvn, this);
939 exit_test = head->loopexit();
940 } else {
941 safepoint = find_safepoint(back_control, x, loop);
942 }
943
944 IfFalseNode* exit_branch = exit_test->false_proj();
945 Node* entry_control = head->in(LoopNode::EntryControl);
946
947 // Clone the control flow of the loop to build an outer loop
948 Node* outer_back_branch = back_control->clone();
949 Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
950 Node* inner_exit_branch = exit_branch->clone();
951
952 LoopNode* outer_head = new LoopNode(entry_control, outer_back_branch);
953 IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_head, outer_back_branch);
954
955 const bool body_populated = true;
956 register_control(outer_head, outer_ilt, entry_control, body_populated);
957
958 _igvn.register_new_node_with_optimizer(inner_exit_branch);
959 set_loop(inner_exit_branch, outer_ilt);
960 set_idom(inner_exit_branch, exit_test, dom_depth(exit_branch));
961
962 outer_exit_test->set_req(0, inner_exit_branch);
963 register_control(outer_exit_test, outer_ilt, inner_exit_branch, body_populated);
964
965 _igvn.replace_input_of(exit_branch, 0, outer_exit_test);
966 set_idom(exit_branch, outer_exit_test, dom_depth(exit_branch));
967
968 outer_back_branch->set_req(0, outer_exit_test);
969 register_control(outer_back_branch, outer_ilt, outer_exit_test, body_populated);
970
971 _igvn.replace_input_of(x, LoopNode::EntryControl, outer_head);
972 set_idom(x, outer_head, dom_depth(x));
973
974 // add an iv phi to the outer loop and use it to compute the inner
975 // loop iteration limit
976 Node* outer_phi = phi->clone();
977 outer_phi->set_req(0, outer_head);
978 register_new_node(outer_phi, outer_head);
979
980 Node* inner_iters_max = nullptr;
981 if (stride_con > 0) {
982 inner_iters_max = MaxNode::max_diff_with_zero(limit, outer_phi, TypeInteger::bottom(bt), _igvn);
983 } else {
984 inner_iters_max = MaxNode::max_diff_with_zero(outer_phi, limit, TypeInteger::bottom(bt), _igvn);
985 }
986
987 Node* inner_iters_limit = _igvn.integercon(iters_limit, bt);
988 // inner_iters_max may not fit in a signed integer (iterating from
989 // Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned
990 // min.
991 const TypeInteger* inner_iters_actual_range = TypeInteger::make(0, iters_limit, Type::WidenMin, bt);
992 Node* inner_iters_actual = MaxNode::unsigned_min(inner_iters_max, inner_iters_limit, inner_iters_actual_range, _igvn);
993
994 Node* inner_iters_actual_int;
995 if (bt == T_LONG) {
996 inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
997 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
998 // When the inner loop is transformed to a counted loop, a loop limit check is not expected to be needed because
999 // the loop limit is less or equal to max_jint - stride - 1 (if stride is positive but a similar argument exists for
1000 // a negative stride). We add a CastII here to guarantee that, when the counted loop is created in a subsequent loop
1001 // opts pass, an accurate range of values for the limits is found.
1002 const TypeInt* inner_iters_actual_int_range = TypeInt::make(0, iters_limit, Type::WidenMin);
1003 inner_iters_actual_int = new CastIINode(outer_head, inner_iters_actual_int, inner_iters_actual_int_range, ConstraintCastNode::DependencyType::NonFloatingNonNarrowing);
1004 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1005 } else {
1006 inner_iters_actual_int = inner_iters_actual;
1007 }
1008
1009 Node* int_zero = intcon(0);
1010 if (stride_con < 0) {
1011 inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
1012 _igvn.register_new_node_with_optimizer(inner_iters_actual_int);
1013 }
1014
1015 // Clone the iv data nodes as an integer iv
1016 Node* int_stride = intcon(stride_con);
1017 Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
1018 Node* inner_incr = new AddINode(inner_phi, int_stride);
1019 Node* inner_cmp = nullptr;
1020 inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
1021 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
1022 inner_phi->set_req(LoopNode::EntryControl, int_zero);
1023 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
1024 register_new_node(inner_phi, x);
1025 register_new_node(inner_incr, x);
1026 register_new_node(inner_cmp, x);
1027 register_new_node(inner_bol, x);
1028
1029 _igvn.replace_input_of(exit_test, 1, inner_bol);
1030
1031 // Clone inner loop phis to outer loop
1032 for (uint i = 0; i < head->outcnt(); i++) {
1033 Node* u = head->raw_out(i);
1034 if (u->is_Phi() && u != inner_phi && u != phi) {
1035 assert(u->in(0) == head, "inconsistent");
1036 Node* clone = u->clone();
1037 clone->set_req(0, outer_head);
1038 register_new_node(clone, outer_head);
1039 _igvn.replace_input_of(u, LoopNode::EntryControl, clone);
1040 }
1041 }
1042
1043 // Replace inner loop long iv phi as inner loop int iv phi + outer
1044 // loop iv phi
1045 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
1046
1047 set_subtree_ctrl(inner_iters_actual_int, body_populated);
1048
1049 LoopNode* inner_head = create_inner_head(loop, head, exit_test);
1050
1051 // Summary of steps from initial loop to loop nest:
1052 //
1053 // == old IR nodes =>
1054 //
1055 // entry_control: {...}
1056 // x:
1057 // for (long phi = init;;) {
1058 // // phi := Phi(x, init, incr)
1059 // // incr := AddL(phi, longcon(stride))
1060 // exit_test:
1061 // if (phi < limit)
1062 // back_control: fallthrough;
1063 // else
1064 // exit_branch: break;
1065 // long incr = phi + stride;
1066 // ... use phi and incr ...
1067 // phi = incr;
1068 // }
1069 //
1070 // == new IR nodes (just before final peel) =>
1071 //
1072 // entry_control: {...}
1073 // long adjusted_limit = limit + stride; //because phi_incr != nullptr
1074 // assert(!limit_check_required || (extralong)limit + stride == adjusted_limit); // else deopt
1075 // ulong inner_iters_limit = max_jint - ABS(stride) - 1; //near 0x7FFFFFF0
1076 // outer_head:
1077 // for (long outer_phi = init;;) {
1078 // // outer_phi := phi->clone(), in(0):=outer_head, => Phi(outer_head, init, incr)
1079 // // REPLACE phi => AddL(outer_phi, I2L(inner_phi))
1080 // // REPLACE incr => AddL(outer_phi, I2L(inner_incr))
1081 // // SO THAT outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_incr)))
1082 // ulong inner_iters_max = (ulong) MAX(0, ((extralong)adjusted_limit - outer_phi) * SGN(stride));
1083 // int inner_iters_actual_int = (int) MIN(inner_iters_limit, inner_iters_max) * SGN(stride);
1084 // inner_head: x: //in(1) := outer_head
1085 // int inner_phi;
1086 // for (inner_phi = 0;;) {
1087 // // inner_phi := Phi(x, intcon(0), inner_phi + stride)
1088 // int inner_incr = inner_phi + stride;
1089 // bool inner_bol = (inner_incr < inner_iters_actual_int);
1090 // exit_test: //exit_test->in(1) := inner_bol;
1091 // if (inner_bol) // WAS (phi < limit)
1092 // back_control: fallthrough;
1093 // else
1094 // inner_exit_branch: break; //exit_branch->clone()
1095 // ... use phi=>(outer_phi+inner_phi) ...
1096 // inner_phi = inner_phi + stride; // inner_incr
1097 // }
1098 // outer_exit_test: //exit_test->clone(), in(0):=inner_exit_branch
1099 // if ((outer_phi+inner_phi) < limit) // WAS (phi < limit)
1100 // outer_back_branch: fallthrough; //back_control->clone(), in(0):=outer_exit_test
1101 // else
1102 // exit_branch: break; //in(0) := outer_exit_test
1103 // }
1104
1105 if (bt == T_INT) {
1106 outer_phi = new ConvI2LNode(outer_phi);
1107 register_new_node(outer_phi, outer_head);
1108 }
1109
1110 transform_long_range_checks(stride_con, range_checks, outer_phi, inner_iters_actual_int,
1111 inner_phi, iv_add, inner_head);
1112 // Peel one iteration of the loop and use the safepoint at the end
1113 // of the peeled iteration to insert Parse Predicates. If no well
1114 // positioned safepoint peel to guarantee a safepoint in the outer
1115 // loop.
1116 if (safepoint != nullptr || !loop->_has_call) {
1117 old_new.clear();
1118 do_peeling(loop, old_new);
1119 } else {
1120 C->set_major_progress();
1121 }
1122
1123 if (safepoint != nullptr) {
1124 SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
1125
1126 if (ShortRunningLongLoop) {
1127 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, inner_head, outer_ilt, cloned_sfpt);
1128 }
1129 if (UseLoopPredicate) {
1130 add_parse_predicate(Deoptimization::Reason_predicate, inner_head, outer_ilt, cloned_sfpt);
1131 if (UseProfiledLoopPredicate) {
1132 add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
1133 }
1134 }
1135
1136 if (UseAutoVectorizationPredicate) {
1137 // We only want to use the auto-vectorization check as a trap once per bci. And
1138 // PhaseIdealLoop::add_parse_predicate only checks trap limits per method, so
1139 // we do a custom check here.
1140 if (!C->too_many_traps(cloned_sfpt->jvms()->method(), cloned_sfpt->jvms()->bci(), Deoptimization::Reason_auto_vectorization_check)) {
1141 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, inner_head, outer_ilt, cloned_sfpt);
1142 }
1143 }
1144
1145 add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
1146 }
1147
1148 #ifndef PRODUCT
1149 if (bt == T_LONG) {
1150 AtomicAccess::inc(&_long_loop_nests);
1151 }
1152 #endif
1153
1154 inner_head->mark_loop_nest_inner_loop();
1155 outer_head->mark_loop_nest_outer_loop();
1156
1157 return true;
1158 }
1159
1160 // Make a copy of Parse/Template Assertion predicates below existing predicates at the loop passed as argument
1161 class CloneShortLoopPredicateVisitor : public PredicateVisitor {
1162 ClonePredicateToTargetLoop _clone_predicate_to_loop;
1163 PhaseIdealLoop* const _phase;
1164 Node* const _new_init;
1165
1166 public:
1167 CloneShortLoopPredicateVisitor(LoopNode* target_loop_head,
1168 Node* new_init,
1169 const NodeInSingleLoopBody &node_in_loop_body,
1170 PhaseIdealLoop* phase)
1171 : _clone_predicate_to_loop(target_loop_head, node_in_loop_body, phase),
1172 _phase(phase),
1173 _new_init(new_init) {
1174 }
1175 NONCOPYABLE(CloneShortLoopPredicateVisitor);
1176
1177 using PredicateVisitor::visit;
1178
1179 void visit(const ParsePredicate& parse_predicate) override {
1180 _clone_predicate_to_loop.clone_parse_predicate(parse_predicate, true);
1181 parse_predicate.kill(_phase->igvn());
1182 }
1183
1184 void visit(const TemplateAssertionPredicate& template_assertion_predicate) override {
1185 _clone_predicate_to_loop.clone_template_assertion_predicate_and_replace_init(template_assertion_predicate, _new_init);
1186 template_assertion_predicate.kill(_phase->igvn());
1187 }
1188 };
1189
1190 // For an int counted loop, try_make_short_running_loop() transforms the loop from:
1191 // for (int = start; i < stop; i+= stride) { ... }
1192 // to
1193 // for (int = 0; i < stop - start; i+= stride) { ... }
1194 // Template Assertion Predicates added so far were with an init value of start. They need to be updated with the new
1195 // init value of 0 (otherwise when a template assertion predicate is turned into an initialized assertion predicate, it
1196 // performs an incorrect check):
1197 // zero
1198 // init |
1199 // | ===> OpaqueLoopInit init
1200 // OpaqueLoopInit \ /
1201 // AddI
1202 //
1203 Node* PhaseIdealLoop::new_assertion_predicate_opaque_init(Node* entry_control, Node* init, Node* int_zero) {
1204 OpaqueLoopInitNode* new_opaque_init = new OpaqueLoopInitNode(C, int_zero);
1205 register_new_node(new_opaque_init, entry_control);
1206 Node* new_init = new AddINode(new_opaque_init, init);
1207 register_new_node(new_init, entry_control);
1208 return new_init;
1209 }
1210
1211 // If the loop is either statically known to run for a small enough number of iterations or if profile data indicates
1212 // that, we don't want an outer loop because the overhead of having an outer loop whose backedge is never taken, has a
1213 // measurable cost. Furthermore, creating the loop nest usually causes one iteration of the loop to be peeled so
1214 // predicates can be set up. If the loop is short running, then it's an extra iteration that's run with range checks
1215 // (compared to an int counted loop with int range checks).
1216 //
1217 // In the short running case, turn the loop into a regular loop again and transform the long range checks:
1218 // - LongCountedLoop: Create LoopNode but keep the loop limit type with a CastLL node to avoid that we later try to
1219 // create a Loop Limit Check when turning the LoopNode into a CountedLoopNode.
1220 // - CountedLoop: Can be reused.
1221 bool PhaseIdealLoop::try_make_short_running_loop(IdealLoopTree* loop, jint stride_con, const Node_List &range_checks,
1222 const uint iters_limit) {
1223 if (!ShortRunningLongLoop) {
1224 return false;
1225 }
1226 BaseCountedLoopNode* head = loop->_head->as_BaseCountedLoop();
1227 BasicType bt = head->bt();
1228 Node* entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1229
1230 loop->compute_trip_count(this, bt);
1231 // Loop must run for no more than iter_limits as it guarantees no overflow of scale * iv in long range checks (see
1232 // comment above PhaseIdealLoop::transform_long_range_checks()).
1233 // iters_limit / ABS(stride_con) is the largest trip count for which we know it's correct to not create a loop nest:
1234 // it's always beneficial to have a single loop rather than a loop nest, so we try to apply this transformation as
1235 // often as possible.
1236 bool known_short_running_loop = head->trip_count() <= iters_limit / ABS(stride_con);
1237 bool profile_short_running_loop = false;
1238 if (!known_short_running_loop) {
1239 loop->compute_profile_trip_cnt(this);
1240 if (StressShortRunningLongLoop) {
1241 profile_short_running_loop = true;
1242 } else {
1243 profile_short_running_loop = !head->is_profile_trip_failed() && head->profile_trip_cnt() <= iters_limit / ABS(stride_con);
1244 }
1245 }
1246
1247 if (!known_short_running_loop && !profile_short_running_loop) {
1248 return false;
1249 }
1250
1251 Node* limit = head->limit();
1252 Node* init = head->init_trip();
1253
1254 Node* new_limit;
1255 if (stride_con > 0) {
1256 new_limit = SubNode::make(limit, init, bt);
1257 } else {
1258 new_limit = SubNode::make(init, limit, bt);
1259 }
1260 register_new_node(new_limit, entry_control);
1261
1262 Node* int_zero = intcon(0);
1263 PhiNode* phi = head->phi()->as_Phi();
1264 if (profile_short_running_loop) {
1265 // Add a Short Running Long Loop Predicate. It's the first predicate in the predicate chain before entering a loop
1266 // because a cast that's control dependent on the Short Running Long Loop Predicate is added to narrow the limit and
1267 // future predicates may be dependent on the new limit (so have to be between the loop and Short Running Long Loop
1268 // Predicate). The current limit could, itself, be dependent on an existing predicate. Clone parse and template
1269 // assertion predicates below existing predicates to get proper ordering of predicates when walking from the loop
1270 // up: future predicates, Short Running Long Loop Predicate, existing predicates.
1271 //
1272 // Existing Hoisted
1273 // Check Predicates
1274 // |
1275 // New Short Running Long
1276 // Loop Predicate
1277 // |
1278 // Cloned Parse Predicates and
1279 // Template Assertion Predicates
1280 // (future predicates added here)
1281 // |
1282 // Loop
1283 const Predicates predicates_before_cloning(entry_control);
1284 const PredicateBlock* short_running_long_loop_predicate_block = predicates_before_cloning.short_running_long_loop_predicate_block();
1285 if (!short_running_long_loop_predicate_block->has_parse_predicate()) { // already trapped
1286 return false;
1287 }
1288 Node* new_init = new_assertion_predicate_opaque_init(entry_control, init, int_zero);
1289
1290 PredicateIterator predicate_iterator(entry_control);
1291 NodeInSingleLoopBody node_in_short_loop_body(this, loop);
1292 CloneShortLoopPredicateVisitor clone_short_loop_predicates_visitor(head, new_init, node_in_short_loop_body, this);
1293 predicate_iterator.for_each(clone_short_loop_predicates_visitor);
1294
1295 entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1296
1297 const Predicates predicates_after_cloning(entry_control);
1298
1299 ParsePredicateSuccessProj* short_running_loop_predicate_proj = predicates_after_cloning.
1300 short_running_long_loop_predicate_block()->
1301 parse_predicate_success_proj();
1302 assert(short_running_loop_predicate_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
1303
1304 const jlong iters_limit_long = iters_limit;
1305 Node* cmp_limit = CmpNode::make(new_limit, _igvn.integercon(iters_limit_long, bt), bt);
1306 Node* bol = new BoolNode(cmp_limit, BoolTest::le);
1307 Node* new_predicate_proj = create_new_if_for_predicate(short_running_loop_predicate_proj,
1308 nullptr,
1309 Deoptimization::Reason_short_running_long_loop,
1310 Op_If);
1311 Node* iff = new_predicate_proj->in(0);
1312 _igvn.replace_input_of(iff, 1, bol);
1313 register_new_node(cmp_limit, iff->in(0));
1314 register_new_node(bol, iff->in(0));
1315 new_limit = ConstraintCastNode::make_cast_for_basic_type(new_predicate_proj, new_limit,
1316 TypeInteger::make(1, iters_limit_long, Type::WidenMin, bt),
1317 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, bt);
1318 register_new_node(new_limit, new_predicate_proj);
1319
1320 #ifndef PRODUCT
1321 if (TraceLoopLimitCheck) {
1322 tty->print_cr("Short Long Loop Check Predicate generated:");
1323 DEBUG_ONLY(bol->dump(2);)
1324 }
1325 #endif
1326 entry_control = head->skip_strip_mined()->in(LoopNode::EntryControl);
1327 } else if (bt == T_LONG) {
1328 // We're turning a long counted loop into a regular loop that will be converted into an int counted loop. That loop
1329 // won't need loop limit check predicates (iters_limit guarantees that). Add a cast to make sure that, whatever
1330 // transformation happens by the time the counted loop is created (in a subsequent pass of loop opts), C2 knows
1331 // enough about the loop's limit that it doesn't try to add loop limit check predicates.
1332 const Predicates predicates(entry_control);
1333 const TypeLong* new_limit_t = new_limit->Value(&_igvn)->is_long();
1334 new_limit = ConstraintCastNode::make_cast_for_basic_type(predicates.entry(), new_limit,
1335 TypeLong::make(0, new_limit_t->_hi, new_limit_t->_widen),
1336 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, bt);
1337 register_new_node(new_limit, predicates.entry());
1338 } else {
1339 assert(bt == T_INT && known_short_running_loop, "only CountedLoop statically known to be short running");
1340 PredicateIterator predicate_iterator(entry_control);
1341 Node* new_init = new_assertion_predicate_opaque_init(entry_control, init, int_zero);
1342 UpdateInitForTemplateAssertionPredicates update_init_for_template_assertion_predicates(new_init, this);
1343 predicate_iterator.for_each(update_init_for_template_assertion_predicates);
1344 }
1345 IfNode* exit_test = head->loopexit();
1346
1347 if (bt == T_LONG) {
1348 // The loop is short running so new_limit fits into an int: either we determined that statically or added a guard
1349 new_limit = new ConvL2INode(new_limit);
1350 register_new_node(new_limit, entry_control);
1351 }
1352
1353 if (stride_con < 0) {
1354 new_limit = new SubINode(int_zero, new_limit);
1355 register_new_node(new_limit, entry_control);
1356 }
1357
1358 // Clone the iv data nodes as an integer iv
1359 Node* int_stride = intcon(stride_con);
1360 Node* inner_phi = new PhiNode(head, TypeInt::INT);
1361 Node* inner_incr = new AddINode(inner_phi, int_stride);
1362 Node* inner_cmp = new CmpINode(inner_incr, new_limit);
1363 Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
1364 inner_phi->set_req(LoopNode::EntryControl, int_zero);
1365 inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
1366 register_new_node(inner_phi, head);
1367 register_new_node(inner_incr, head);
1368 register_new_node(inner_cmp, head);
1369 register_new_node(inner_bol, head);
1370
1371 _igvn.replace_input_of(exit_test, 1, inner_bol);
1372
1373 // Replace inner loop long iv phi as inner loop int iv phi + outer
1374 // loop iv phi
1375 Node* iv_add = loop_nest_replace_iv(phi, inner_phi, init, head, bt);
1376
1377 LoopNode* inner_head = head;
1378 if (bt == T_LONG) {
1379 // Turn the loop back to a counted loop
1380 inner_head = create_inner_head(loop, head, exit_test);
1381 } else {
1382 // Use existing counted loop
1383 revert_to_normal_loop(head);
1384 }
1385
1386 if (bt == T_INT) {
1387 init = new ConvI2LNode(init);
1388 register_new_node(init, entry_control);
1389 }
1390
1391 transform_long_range_checks(stride_con, range_checks, init, new_limit,
1392 inner_phi, iv_add, inner_head);
1393
1394 inner_head->mark_loop_nest_inner_loop();
1395
1396 return true;
1397 }
1398
1399 int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1400 Node_List& range_checks) {
1401 const jlong min_iters = 2;
1402 jlong reduced_iters_limit = iters_limit;
1403 jlong original_iters_limit = iters_limit;
1404 for (uint i = 0; i < loop->_body.size(); i++) {
1405 Node* c = loop->_body.at(i);
1406 if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {
1407 IfProjNode* if_proj = c->as_IfProj();
1408 CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1409 if (call != nullptr) {
1410 Node* range = nullptr;
1411 Node* offset = nullptr;
1412 jlong scale = 0;
1413 if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1414 loop->is_invariant(range) && loop->is_invariant(offset) &&
1415 scale != min_jlong &&
1416 original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {
1417 assert(scale == (jint)scale, "scale should be an int");
1418 reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1419 range_checks.push(c);
1420 }
1421 }
1422 }
1423 }
1424
1425 return checked_cast<int>(reduced_iters_limit);
1426 }
1427
1428 // One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1429 // Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1430 // 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
1431 // i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1432
1433 // 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,
1434 // 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)
1435 // 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].
1436
1437 // 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
1438 // is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1439 // Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1440
1441 // 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
1442 // 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
1443 // chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1444 // logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1445
1446 // 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
1447 // ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1448
1449 // 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].
1450 // 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
1451 // 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
1452 // 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
1453 // down (from C and 0 again).
1454
1455 // 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).
1456 // (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1457 // 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1458 // 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1459 // PhaseIdealLoop::add_constraint.)
1460
1461 // 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
1462 // 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
1463 // 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.
1464 // Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1465 // values for the 32-bit counter j within which the range checks can be eliminated.
1466
1467 // 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
1468 // 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
1469 // allowed as an input to the R.C., as long as the R.C. as a whole fails.
1470
1471 // If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1472
1473 // 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
1474 // 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),
1475 // 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)
1476
1477 // 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).
1478 // 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.
1479 // If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1480
1481 // Case A: Some Negatives (but no overflow).
1482 // Number line:
1483 // |s64_min . . . 0 . . . s64_max|
1484 // | . Q_min..Q_max . 0 . . . . | s64 negative
1485 // | . . . . R=0 R< R< R< R< | (against R values)
1486 // | . . . Q_min..0..Q_max . . . | small mixed
1487 // | . . . . R R R< R< R< | (against R values)
1488 //
1489 // 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<.
1490 //
1491 // So, if Q_min <s64 0, then use this test:
1492 // j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1493 // j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1494 // Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1495 // j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1496
1497 // If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1498
1499 // Case B: No Negatives.
1500 // Number line:
1501 // |s64_min . . . 0 . . . s64_max|
1502 // | . . . . 0 Q_min..Q_max . . | small positive
1503 // | . . . . R> R R R< R< | (against R values)
1504 // | . . . . 0 . Q_min..Q_max . | s64 positive
1505 // | . . . . R> R> R R R< | (against R values)
1506 //
1507 // 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.
1508 // Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1509 //
1510 // So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1511 // j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1512 // More generally:
1513 // j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1514
1515 // Case C: Overflow in the 64-bit domain
1516 // Number line:
1517 // |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1518 // | . . . . R> R> R> R> R | (against R values)
1519 //
1520 // In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1521 // The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1522 // (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1523 //
1524 // So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1525 // j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1526 // More generally:
1527 // j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1528 //
1529 // Dropping the bad bound means only Q_min is used to reduce the range of R:
1530 // j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1531 //
1532 // Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1533 // clamp(X, L, H) := max(L, min(X, H))
1534 // When degenerately L > H, it returns L not H.
1535 //
1536 // All of the formulas above can be merged into a single one:
1537 // L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1538 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1539 // = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1540 // Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1541 // R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1542 // replacement R.C.:
1543 // j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1544 // or equivalently:
1545 // j*K + L_2 <u32 R_2
1546 // where
1547 // L_2 = Q_first - L_clamp
1548 // R_2 = R_clamp - L_clamp
1549 //
1550 // Note on why R is never negative:
1551 //
1552 // Various details of this transformation would break badly if R could be negative, so this transformation only
1553 // operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1554 // know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1555 // inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1556 // non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1557 // the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1558 // R.C., but R values from a.length and inside checkIndex are good to go.
1559 //
1560 void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1561 Node* inner_iters_actual_int, Node* inner_phi,
1562 Node* iv_add, LoopNode* inner_head) {
1563 Node* long_zero = longcon(0);
1564 Node* int_zero = intcon(0);
1565 Node* long_one = longcon(1);
1566 Node* int_stride = intcon(checked_cast<int>(stride_con));
1567
1568 for (uint i = 0; i < range_checks.size(); i++) {
1569 ProjNode* proj = range_checks.at(i)->as_Proj();
1570 RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1571 jlong scale = 0;
1572 Node* offset = nullptr;
1573 Node* rc_bol = rc->in(1);
1574 Node* rc_cmp = rc_bol->in(1);
1575 if (rc_cmp->Opcode() == Op_CmpU) {
1576 // could be shared and have already been taken care of
1577 continue;
1578 }
1579 bool short_scale = false;
1580 bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1581 assert(ok, "inconsistent: was tested before");
1582 Node* range = rc_cmp->in(2);
1583 Node* c = rc->in(0);
1584 Node* entry_control = inner_head->in(LoopNode::EntryControl);
1585
1586 Node* R = range;
1587 Node* K = longcon(scale);
1588
1589 Node* L = offset;
1590
1591 if (short_scale) {
1592 // This converts:
1593 // (int)i*K + L <u64 R
1594 // with K an int into:
1595 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1596 // to protect against an overflow of (int)i*K
1597 //
1598 // Because if (int)i*K overflows, there are K,L where:
1599 // (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1600 // But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1601 // i*(long)K + L <u64 R is true.
1602 //
1603 // As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1604 //
1605 // It's always true that:
1606 // (int)i*K <u64 (long)max_jint + 1
1607 // which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1608 // As a consequence:
1609 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1610 // is always false in case of overflow of i*K
1611 //
1612 // Note, there are also K,L where i*K overflows and
1613 // i*K + L <u64 R is true, but
1614 // i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1615 // So this transformation could cause spurious deoptimizations and failed range check elimination
1616 // (but not incorrect execution) for unlikely corner cases with overflow.
1617 // If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1618 Node* max_jint_plus_one_long = longcon((jlong)max_jint + 1);
1619 Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1620 register_new_node(max_range, entry_control);
1621 R = MaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1622 set_subtree_ctrl(R, true);
1623 }
1624
1625 Node* C = outer_phi;
1626
1627 // Start with 64-bit values:
1628 // i*K + L <u64 R
1629 // (C+j)*K + L <u64 R
1630 // j*K + Q <u64 R where Q = Q_first = C*K+L
1631 Node* Q_first = new MulLNode(C, K);
1632 register_new_node(Q_first, entry_control);
1633 Q_first = new AddLNode(Q_first, L);
1634 register_new_node(Q_first, entry_control);
1635
1636 // Compute endpoints of the range of values j*K + Q.
1637 // Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1638 Node* Q_min = Q_first;
1639
1640 // Compute the exact ending value B_2 (which is really A_2 if S < 0)
1641 Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1642 register_new_node(B_2, entry_control);
1643 B_2 = new SubINode(B_2, int_stride);
1644 register_new_node(B_2, entry_control);
1645 B_2 = new ConvI2LNode(B_2);
1646 register_new_node(B_2, entry_control);
1647
1648 Node* Q_max = new MulLNode(B_2, K);
1649 register_new_node(Q_max, entry_control);
1650 Q_max = new AddLNode(Q_max, Q_first);
1651 register_new_node(Q_max, entry_control);
1652
1653 if (scale * stride_con < 0) {
1654 swap(Q_min, Q_max);
1655 }
1656 // Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1657
1658 // L_clamp = Q_min < 0 ? 0 : Q_min
1659 Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1660 register_new_node(Q_min_cmp, entry_control);
1661 Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1662 register_new_node(Q_min_bool, entry_control);
1663 Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1664 register_new_node(L_clamp, entry_control);
1665 // (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1666
1667 Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1668 register_new_node(Q_max_plus_one, entry_control);
1669
1670 // H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1671 // (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1672 Node* max_jlong_long = longcon(max_jlong);
1673 Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1674 register_new_node(Q_max_cmp, entry_control);
1675 Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1676 register_new_node(Q_max_bool, entry_control);
1677 Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1678 register_new_node(H_clamp, entry_control);
1679 // (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1680
1681 // R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1682 // that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1683 // or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1684 // and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1685 // or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1686 Node* R_2 = clamp(R, L_clamp, H_clamp);
1687 R_2 = new SubLNode(R_2, L_clamp);
1688 register_new_node(R_2, entry_control);
1689 R_2 = new ConvL2INode(R_2, TypeInt::POS);
1690 register_new_node(R_2, entry_control);
1691
1692 // L_2 = Q_first - L_clamp
1693 // We are subtracting L_clamp from both sides of the <u32 comparison.
1694 // If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1695 // If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1696 Node* L_2 = new SubLNode(Q_first, L_clamp);
1697 register_new_node(L_2, entry_control);
1698 L_2 = new ConvL2INode(L_2, TypeInt::INT);
1699 register_new_node(L_2, entry_control);
1700
1701 // Transform the range check using the computed values L_2/R_2
1702 // from: i*K + L <u64 R
1703 // to: j*K + L_2 <u32 R_2
1704 // that is:
1705 // (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1706 K = intcon(checked_cast<int>(scale));
1707 Node* scaled_iv = new MulINode(inner_phi, K);
1708 register_new_node(scaled_iv, c);
1709 Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1710 register_new_node(scaled_iv_plus_offset, c);
1711
1712 Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1713 register_new_node(new_rc_cmp, c);
1714
1715 _igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1716 }
1717 }
1718
1719 Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {
1720 Node* min = MaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1721 set_subtree_ctrl(min, true);
1722 Node* max = MaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1723 set_subtree_ctrl(max, true);
1724 return max;
1725 }
1726
1727 LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1728 IfNode* exit_test) {
1729 LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1730 IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1731 _igvn.register_new_node_with_optimizer(new_inner_head);
1732 _igvn.register_new_node_with_optimizer(new_inner_exit);
1733 loop->_body.push(new_inner_head);
1734 loop->_body.push(new_inner_exit);
1735 loop->_body.yank(head);
1736 loop->_body.yank(exit_test);
1737 set_loop(new_inner_head, loop);
1738 set_loop(new_inner_exit, loop);
1739 set_idom(new_inner_head, idom(head), dom_depth(head));
1740 set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1741 replace_node_and_forward_ctrl(head, new_inner_head);
1742 replace_node_and_forward_ctrl(exit_test, new_inner_exit);
1743 loop->_head = new_inner_head;
1744 return new_inner_head;
1745 }
1746
1747 #ifdef ASSERT
1748 void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {
1749 Node* back_control = loop_exit_control(x, loop);
1750 assert(back_control != nullptr, "no back control");
1751
1752 BoolTest::mask mask = BoolTest::illegal;
1753 float cl_prob = 0;
1754 Node* incr = nullptr;
1755 Node* limit = nullptr;
1756
1757 Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1758 assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1759
1760 Node* phi_incr = nullptr;
1761 incr = loop_iv_incr(incr, x, loop, phi_incr);
1762 assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1763
1764 Node* xphi = nullptr;
1765 Node* stride = loop_iv_stride(incr, xphi);
1766
1767 assert(stride != nullptr, "no stride");
1768
1769 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x);
1770
1771 assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1772
1773 jlong stride_con = stride->get_integer_as_long(bt);
1774
1775 assert(condition_stride_ok(mask, stride_con), "illegal condition");
1776
1777 assert(mask != BoolTest::ne, "unexpected condition");
1778 assert(phi_incr == nullptr, "bad loop shape");
1779 assert(cmp->in(1) == incr, "bad exit test shape");
1780
1781 // Safepoint on backedge not supported
1782 assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1783 }
1784 #endif
1785
1786 #ifdef ASSERT
1787 // convert an int counted loop to a long counted to stress handling of
1788 // long counted loops
1789 bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {
1790 Unique_Node_List iv_nodes;
1791 Node_List old_new;
1792 iv_nodes.push(cmp);
1793 bool failed = false;
1794
1795 for (uint i = 0; i < iv_nodes.size() && !failed; i++) {
1796 Node* n = iv_nodes.at(i);
1797 switch(n->Opcode()) {
1798 case Op_Phi: {
1799 Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1800 old_new.map(n->_idx, clone);
1801 break;
1802 }
1803 case Op_CmpI: {
1804 Node* clone = new CmpLNode(nullptr, nullptr);
1805 old_new.map(n->_idx, clone);
1806 break;
1807 }
1808 case Op_AddI: {
1809 Node* clone = new AddLNode(nullptr, nullptr);
1810 old_new.map(n->_idx, clone);
1811 break;
1812 }
1813 case Op_CastII: {
1814 failed = true;
1815 break;
1816 }
1817 default:
1818 DEBUG_ONLY(n->dump());
1819 fatal("unexpected");
1820 }
1821
1822 for (uint i = 1; i < n->req(); i++) {
1823 Node* in = n->in(i);
1824 if (in == nullptr) {
1825 continue;
1826 }
1827 if (ctrl_is_member(loop, in)) {
1828 iv_nodes.push(in);
1829 }
1830 }
1831 }
1832
1833 if (failed) {
1834 for (uint i = 0; i < iv_nodes.size(); i++) {
1835 Node* n = iv_nodes.at(i);
1836 Node* clone = old_new[n->_idx];
1837 if (clone != nullptr) {
1838 _igvn.remove_dead_node(clone);
1839 }
1840 }
1841 return false;
1842 }
1843
1844 for (uint i = 0; i < iv_nodes.size(); i++) {
1845 Node* n = iv_nodes.at(i);
1846 Node* clone = old_new[n->_idx];
1847 for (uint i = 1; i < n->req(); i++) {
1848 Node* in = n->in(i);
1849 if (in == nullptr) {
1850 continue;
1851 }
1852 Node* in_clone = old_new[in->_idx];
1853 if (in_clone == nullptr) {
1854 assert(_igvn.type(in)->isa_int(), "");
1855 in_clone = new ConvI2LNode(in);
1856 _igvn.register_new_node_with_optimizer(in_clone);
1857 set_subtree_ctrl(in_clone, false);
1858 }
1859 if (in_clone->in(0) == nullptr) {
1860 in_clone->set_req(0, C->top());
1861 clone->set_req(i, in_clone);
1862 in_clone->set_req(0, nullptr);
1863 } else {
1864 clone->set_req(i, in_clone);
1865 }
1866 }
1867 _igvn.register_new_node_with_optimizer(clone);
1868 }
1869 set_ctrl(old_new[phi->_idx], phi->in(0));
1870
1871 for (uint i = 0; i < iv_nodes.size(); i++) {
1872 Node* n = iv_nodes.at(i);
1873 Node* clone = old_new[n->_idx];
1874 set_subtree_ctrl(clone, false);
1875 Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1876 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1877 Node* u = n->fast_out(i);
1878 if (iv_nodes.member(u)) {
1879 continue;
1880 }
1881 if (m == nullptr) {
1882 m = new ConvL2INode(clone);
1883 _igvn.register_new_node_with_optimizer(m);
1884 set_subtree_ctrl(m, false);
1885 }
1886 _igvn.rehash_node_delayed(u);
1887 int nb = u->replace_edge(n, m, &_igvn);
1888 --i, imax -= nb;
1889 }
1890 }
1891 return true;
1892 }
1893 #endif
1894
1895 //------------------------------is_counted_loop--------------------------------
1896 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt) {
1897 PhaseGVN *gvn = &_igvn;
1898
1899 Node* back_control = loop_exit_control(x, loop);
1900 if (back_control == nullptr) {
1901 return false;
1902 }
1903
1904 BoolTest::mask bt = BoolTest::illegal;
1905 float cl_prob = 0;
1906 Node* incr = nullptr;
1907 Node* limit = nullptr;
1908 Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1909 if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {
1910 return false; // Avoid pointer & float & 64-bit compares
1911 }
1912
1913 // Trip-counter increment must be commutative & associative.
1914 if (incr->Opcode() == Op_Cast(iv_bt)) {
1915 incr = incr->in(1);
1916 }
1917
1918 Node* phi_incr = nullptr;
1919 incr = loop_iv_incr(incr, x, loop, phi_incr);
1920 if (incr == nullptr) {
1921 return false;
1922 }
1923
1924 Node* trunc1 = nullptr;
1925 Node* trunc2 = nullptr;
1926 const TypeInteger* iv_trunc_t = nullptr;
1927 Node* orig_incr = incr;
1928 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {
1929 return false; // Funny increment opcode
1930 }
1931 assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1932
1933 Node* xphi = nullptr;
1934 Node* stride = loop_iv_stride(incr, xphi);
1935
1936 if (stride == nullptr) {
1937 return false;
1938 }
1939
1940 // Iteratively uncast the loop induction variable
1941 // until no more CastII/CastLL nodes are found.
1942 while (xphi->Opcode() == Op_Cast(iv_bt)) {
1943 xphi = xphi->in(1);
1944 }
1945
1946 // Stride must be constant
1947 jlong stride_con = stride->get_integer_as_long(iv_bt);
1948 assert(stride_con != 0, "missed some peephole opt");
1949
1950 PhiNode* phi = loop_iv_phi(xphi, phi_incr, x);
1951
1952 if (phi == nullptr ||
1953 (trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1954 (trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {
1955 return false;
1956 }
1957
1958 Node* iftrue = back_control;
1959 uint iftrue_op = iftrue->Opcode();
1960 Node* iff = iftrue->in(0);
1961 BoolNode* test = iff->in(1)->as_Bool();
1962
1963 const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1964 if (trunc1 != nullptr) {
1965 // When there is a truncation, we must be sure that after the truncation
1966 // the trip counter will end up higher than the limit, otherwise we are looking
1967 // at an endless loop. Can happen with range checks.
1968
1969 // Example:
1970 // int i = 0;
1971 // while (true)
1972 // sum + = array[i];
1973 // i++;
1974 // i = i && 0x7fff;
1975 // }
1976 //
1977 // If the array is shorter than 0x8000 this exits through a AIOOB
1978 // - Counted loop transformation is ok
1979 // If the array is longer then this is an endless loop
1980 // - No transformation can be done.
1981
1982 const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1983 if (limit_t->hi_as_long() > incr_t->hi_as_long()) {
1984 // if the limit can have a higher value than the increment (before the phi)
1985 return false;
1986 }
1987 }
1988
1989 Node *init_trip = phi->in(LoopNode::EntryControl);
1990
1991 // If iv trunc type is smaller than int, check for possible wrap.
1992 if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {
1993 assert(trunc1 != nullptr, "must have found some truncation");
1994
1995 // Get a better type for the phi (filtered thru if's)
1996 const TypeInteger* phi_ft = filtered_type(phi);
1997
1998 // Can iv take on a value that will wrap?
1999 //
2000 // Ensure iv's limit is not within "stride" of the wrap value.
2001 //
2002 // Example for "short" type
2003 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
2004 // If the stride is +10, then the last value of the induction
2005 // variable before the increment (phi_ft->_hi) must be
2006 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
2007 // ensure no truncation occurs after the increment.
2008
2009 if (stride_con > 0) {
2010 if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
2011 iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {
2012 return false; // truncation may occur
2013 }
2014 } else if (stride_con < 0) {
2015 if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
2016 iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {
2017 return false; // truncation may occur
2018 }
2019 }
2020 // No possibility of wrap so truncation can be discarded
2021 // Promote iv type to Int
2022 } else {
2023 assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
2024 }
2025
2026 if (!condition_stride_ok(bt, stride_con)) {
2027 return false;
2028 }
2029
2030 const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
2031
2032 if (stride_con > 0) {
2033 if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {
2034 return false; // cyclic loop
2035 }
2036 } else {
2037 if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {
2038 return false; // cyclic loop
2039 }
2040 }
2041
2042 if (phi_incr != nullptr && bt != BoolTest::ne) {
2043 // check if there is a possibility of IV overflowing after the first increment
2044 if (stride_con > 0) {
2045 if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {
2046 return false;
2047 }
2048 } else {
2049 if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {
2050 return false;
2051 }
2052 }
2053 }
2054
2055 // =================================================
2056 // ---- SUCCESS! Found A Trip-Counted Loop! -----
2057 //
2058
2059 if (x->Opcode() == Op_Region) {
2060 // x has not yet been transformed to Loop or LongCountedLoop.
2061 // This should only happen if we are inside an infinite loop.
2062 // It happens like this:
2063 // build_loop_tree -> do not attach infinite loop and nested loops
2064 // beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
2065 // build_loop_tree -> find and attach infinite and nested loops
2066 // counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
2067 assert(x->as_Region()->is_in_infinite_subgraph(),
2068 "x can only be a Region and not Loop if inside infinite loop");
2069 // Come back later when Region is transformed to LoopNode
2070 return false;
2071 }
2072
2073 assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
2074 C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
2075
2076 // ===================================================
2077 // We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
2078 // This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
2079 // At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
2080 // But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
2081 // to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
2082 // we trap and resume execution before the loop without having executed any iteration of the loop, yet.
2083 //
2084 // These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
2085 // Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
2086 // we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
2087 //
2088 // In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
2089 // for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
2090 //
2091 //
2092 // The loop to be converted does not always need to have the often used shape:
2093 //
2094 // i = init
2095 // i = init loop:
2096 // do { ...
2097 // // ... equivalent i+=stride
2098 // i+=stride <==> if (i < limit)
2099 // } while (i < limit); goto loop
2100 // exit:
2101 // ...
2102 //
2103 // where the loop exit check uses the post-incremented iv phi and a '<'-operator.
2104 //
2105 // We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
2106 // in the loop exit check:
2107 //
2108 // i = init
2109 // loop:
2110 // ...
2111 // if (i <= limit)
2112 // i+=stride
2113 // goto loop
2114 // exit:
2115 // ...
2116 //
2117 // Let's define the following terms:
2118 // - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
2119 // - iv_post_i: The post-incremented iv phi after the i-th iteration.
2120 //
2121 // The iv_pre_i and iv_post_i have the following relation:
2122 // iv_pre_i + stride = iv_post_i
2123 //
2124 // When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
2125 // iv_post_i < adjusted_limit
2126 //
2127 // If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
2128 // (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
2129 // -> adjusted_limit = limit.
2130 // (LE2) iv_post_i <= limit:
2131 // iv_post_i < limit + 1
2132 // -> adjusted limit = limit + 1
2133 // (LE3) iv_pre_i < limit:
2134 // iv_pre_i + stride < limit + stride
2135 // iv_post_i < limit + stride
2136 // -> adjusted_limit = limit + stride
2137 // (LE4) iv_pre_i <= limit:
2138 // iv_pre_i < limit + 1
2139 // iv_pre_i + stride < limit + stride + 1
2140 // iv_post_i < limit + stride + 1
2141 // -> adjusted_limit = limit + stride + 1
2142 //
2143 // Note that:
2144 // (AL) limit <= adjusted_limit.
2145 //
2146 // The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
2147 // loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
2148 // (INV) For i = 1..n, min_int <= iv_post_i <= max_int
2149 //
2150 // To prove (INV), we require the following two conditions/assumptions:
2151 // (i): adjusted_limit - 1 + stride <= max_int
2152 // (ii): init < limit
2153 //
2154 // If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
2155 // induction by assuming (i) and (ii).
2156 //
2157 // Proof by Induction
2158 // ------------------
2159 // > Base case (i = 1): We show that (INV) holds after the first iteration:
2160 // min_int <= iv_post_1 = init + stride <= max_int
2161 // Proof:
2162 // First, we note that (ii) implies
2163 // (iii) init <= limit - 1
2164 // max_int >= adjusted_limit - 1 + stride [using (i)]
2165 // >= limit - 1 + stride [using (AL)]
2166 // >= init + stride [using (iii)]
2167 // >= min_int [using stride > 0, no underflow]
2168 // Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
2169 //
2170 // Note that to prove the base case we need (i) and (ii).
2171 //
2172 // > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
2173 // min_int <= iv_post_j <= max_int
2174 // > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
2175 // min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int
2176 // Proof:
2177 // If iv_post_j >= adjusted_limit:
2178 // We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
2179 // also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.
2180 // If iv_post_j < adjusted_limit:
2181 // First, we note that:
2182 // (iv) iv_post_j <= adjusted_limit - 1
2183 // max_int >= adjusted_limit - 1 + stride [using (i)]
2184 // >= iv_post_j + stride [using (iv)]
2185 // >= min_int [using stride > 0, no underflow]
2186 //
2187 // Note that to prove the step case we only need (i).
2188 //
2189 // Thus, by assuming (i) and (ii), we proved (INV).
2190 //
2191 //
2192 // It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
2193 //
2194 // (1) Loop Limit Check Predicate for (i):
2195 // Using (i): adjusted_limit - 1 + stride <= max_int
2196 //
2197 // This condition is now restated to use limit instead of adjusted_limit:
2198 //
2199 // To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
2200 // max_int - stride + 1 >= adjusted_limit
2201 // We can merge the two constants into
2202 // canonicalized_correction = stride - 1
2203 // which gives us
2204 // max_int - canonicalized_correction >= adjusted_limit
2205 //
2206 // To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
2207 // adjusted_limit = limit + limit_correction
2208 // Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
2209 // max_int - canonicalized_correction - limit_correction >= limit
2210 // Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
2211 // (v) final_correction = canonicalized_correction + limit_correction
2212 //
2213 // which gives us:
2214 //
2215 // Final predicate condition:
2216 // max_int - final_correction >= limit
2217 //
2218 // However, we need to be careful that (v) does not over- or underflow.
2219 // We know that:
2220 // canonicalized_correction = stride - 1
2221 // and
2222 // limit_correction <= stride + 1
2223 // and thus
2224 // canonicalized_correction + limit_correction <= 2 * stride
2225 // To prevent an over- or underflow of (v), we must ensure that
2226 // 2 * stride <= max_int
2227 // which can safely be checked without over- or underflow with
2228 // (vi) stride != min_int AND abs(stride) <= max_int / 2
2229 //
2230 // We could try to further optimize the cases where (vi) does not hold but given that such large strides are
2231 // very uncommon and the loop would only run for a very few iterations anyway, we simply bail out if (vi) fails.
2232 //
2233 // (2) Loop Limit Check Predicate for (ii):
2234 // Using (ii): init < limit
2235 //
2236 // This Loop Limit Check Predicate is not required if we can prove at compile time that either:
2237 // (2.1) type(init) < type(limit)
2238 // In this case, we know:
2239 // all possible values of init < all possible values of limit
2240 // and we can skip the predicate.
2241 //
2242 // (2.2) init < limit is already checked before (i.e. found as a dominating check)
2243 // In this case, we do not need to re-check the condition and can skip the predicate.
2244 // This is often found for while- and for-loops which have the following shape:
2245 //
2246 // if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.
2247 // i = init;
2248 // if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate
2249 // do {
2250 // i += stride;
2251 // } while (i < limit);
2252 // }
2253 //
2254 // (2.3) init + stride <= max_int
2255 // In this case, there is no overflow of the iv phi after the first loop iteration.
2256 // In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
2257 // init < limit
2258 // In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
2259 // compile time that init + stride <= max_int then we have trivially proven the base case and that
2260 // there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
2261 // again and can skip the predicate.
2262
2263 // Check (vi) and bail out if the stride is too big.
2264 if (stride_con == min_signed_integer(iv_bt) || (ABS(stride_con) > max_signed_integer(iv_bt) / 2)) {
2265 return false;
2266 }
2267
2268 // Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
2269 const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
2270
2271 // Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
2272 const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
2273 const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
2274
2275 const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
2276 const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
2277 const jlong final_correction = canonicalized_correction + limit_correction;
2278
2279 int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
2280 Node* init_control = x->in(LoopNode::EntryControl);
2281
2282 // If sov==0, limit's type always satisfies the condition, for
2283 // example, when it is an array length.
2284 if (sov != 0) {
2285 if (sov < 0) {
2286 return false; // Bailout: integer overflow is certain.
2287 }
2288 // (1) Loop Limit Check Predicate is required because we could not statically prove that
2289 // limit + final_correction = adjusted_limit - 1 + stride <= max_int
2290 assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
2291 const Predicates predicates(init_control);
2292 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2293 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2294 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2295 #ifdef ASSERT
2296 if (TraceLoopLimitCheck) {
2297 tty->print("Missing Loop Limit Check Parse Predicate:");
2298 loop->dump_head();
2299 x->dump(1);
2300 }
2301 #endif
2302 return false;
2303 }
2304
2305 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2306 if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {
2307 return false;
2308 }
2309
2310 Node* cmp_limit;
2311 Node* bol;
2312
2313 if (stride_con > 0) {
2314 cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2315 bol = new BoolNode(cmp_limit, BoolTest::le);
2316 } else {
2317 cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2318 bol = new BoolNode(cmp_limit, BoolTest::ge);
2319 }
2320
2321 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2322 }
2323
2324 // (2.3)
2325 const bool init_plus_stride_could_overflow =
2326 (stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2327 (stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2328 // (2.1)
2329 const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2330 (stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2331
2332 if (init_gte_limit && // (2.1)
2333 ((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2334 !has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)
2335 // (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2336 // We use the following condition:
2337 // - stride > 0: init < limit
2338 // - stride < 0: init > limit
2339 //
2340 // This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2341 // a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2342 // check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2343
2344 const Predicates predicates(init_control);
2345 const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2346 if (!loop_limit_check_predicate_block->has_parse_predicate()) {
2347 // The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2348 #ifdef ASSERT
2349 if (TraceLoopLimitCheck) {
2350 tty->print("Missing Loop Limit Check Parse Predicate:");
2351 loop->dump_head();
2352 x->dump(1);
2353 }
2354 #endif
2355 return false;
2356 }
2357
2358 ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2359 Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2360 if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2361 !is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {
2362 return false;
2363 }
2364
2365 Node* cmp_limit;
2366 Node* bol;
2367
2368 if (stride_con > 0) {
2369 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2370 bol = new BoolNode(cmp_limit, BoolTest::lt);
2371 } else {
2372 cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2373 bol = new BoolNode(cmp_limit, BoolTest::gt);
2374 }
2375
2376 insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2377 }
2378
2379 if (bt == BoolTest::ne) {
2380 // Now we need to canonicalize the loop condition if it is 'ne'.
2381 assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2382 if (stride_con > 0) {
2383 // 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2384 bt = BoolTest::lt;
2385 } else {
2386 assert(stride_con < 0, "must be");
2387 // 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2388 bt = BoolTest::gt;
2389 }
2390 }
2391
2392 Node* sfpt = nullptr;
2393 if (loop->_child == nullptr) {
2394 sfpt = find_safepoint(back_control, x, loop);
2395 } else {
2396 sfpt = iff->in(0);
2397 if (sfpt->Opcode() != Op_SafePoint) {
2398 sfpt = nullptr;
2399 }
2400 }
2401
2402 if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {
2403 Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2404 if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2405 iv_bt == T_LONG) &&
2406 sfpt == nullptr) {
2407 // Leaving the safepoint on the backedge and creating a
2408 // CountedLoop will confuse optimizations. We can't move the
2409 // safepoint around because its jvm state wouldn't match a new
2410 // location. Give up on that loop.
2411 return false;
2412 }
2413 if (is_deleteable_safept(backedge_sfpt)) {
2414 replace_node_and_forward_ctrl(backedge_sfpt, iftrue);
2415 if (loop->_safepts != nullptr) {
2416 loop->_safepts->yank(backedge_sfpt);
2417 }
2418 loop->_tail = iftrue;
2419 }
2420 }
2421
2422
2423 #ifdef ASSERT
2424 if (iv_bt == T_INT &&
2425 !x->as_Loop()->is_loop_nest_inner_loop() &&
2426 StressLongCountedLoop > 0 &&
2427 trunc1 == nullptr &&
2428 convert_to_long_loop(cmp, phi, loop)) {
2429 return false;
2430 }
2431 #endif
2432
2433 Node* adjusted_limit = limit;
2434 if (phi_incr != nullptr) {
2435 // If compare points directly to the phi we need to adjust
2436 // the compare so that it points to the incr. Limit have
2437 // to be adjusted to keep trip count the same and we
2438 // should avoid int overflow.
2439 //
2440 // i = init; do {} while(i++ < limit);
2441 // is converted to
2442 // i = init; do {} while(++i < limit+1);
2443 //
2444 adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2445 }
2446
2447 if (includes_limit) {
2448 // The limit check guaranties that 'limit <= (max_jint - stride)' so
2449 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2450 //
2451 Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2452 adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2453 if (bt == BoolTest::le)
2454 bt = BoolTest::lt;
2455 else if (bt == BoolTest::ge)
2456 bt = BoolTest::gt;
2457 else
2458 ShouldNotReachHere();
2459 }
2460 set_subtree_ctrl(adjusted_limit, false);
2461
2462 // Build a canonical trip test.
2463 // Clone code, as old values may be in use.
2464 incr = incr->clone();
2465 incr->set_req(1,phi);
2466 incr->set_req(2,stride);
2467 incr = _igvn.register_new_node_with_optimizer(incr);
2468 set_early_ctrl(incr, false);
2469 _igvn.rehash_node_delayed(phi);
2470 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2471
2472 // If phi type is more restrictive than Int, raise to
2473 // Int to prevent (almost) infinite recursion in igvn
2474 // which can only handle integer types for constants or minint..maxint.
2475 if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {
2476 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2477 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2478 nphi = _igvn.register_new_node_with_optimizer(nphi);
2479 set_ctrl(nphi, get_ctrl(phi));
2480 _igvn.replace_node(phi, nphi);
2481 phi = nphi->as_Phi();
2482 }
2483 cmp = cmp->clone();
2484 cmp->set_req(1,incr);
2485 cmp->set_req(2, adjusted_limit);
2486 cmp = _igvn.register_new_node_with_optimizer(cmp);
2487 set_ctrl(cmp, iff->in(0));
2488
2489 test = test->clone()->as_Bool();
2490 (*(BoolTest*)&test->_test)._test = bt;
2491 test->set_req(1,cmp);
2492 _igvn.register_new_node_with_optimizer(test);
2493 set_ctrl(test, iff->in(0));
2494
2495 // Replace the old IfNode with a new LoopEndNode
2496 Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2497 IfNode *le = lex->as_If();
2498 uint dd = dom_depth(iff);
2499 set_idom(le, le->in(0), dd); // Update dominance for loop exit
2500 set_loop(le, loop);
2501
2502 // Get the loop-exit control
2503 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2504
2505 // Need to swap loop-exit and loop-back control?
2506 if (iftrue_op == Op_IfFalse) {
2507 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2508 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2509
2510 loop->_tail = back_control = ift2;
2511 set_loop(ift2, loop);
2512 set_loop(iff2, get_loop(iffalse));
2513
2514 // Lazy update of 'get_ctrl' mechanism.
2515 replace_node_and_forward_ctrl(iffalse, iff2);
2516 replace_node_and_forward_ctrl(iftrue, ift2);
2517
2518 // Swap names
2519 iffalse = iff2;
2520 iftrue = ift2;
2521 } else {
2522 _igvn.rehash_node_delayed(iffalse);
2523 _igvn.rehash_node_delayed(iftrue);
2524 iffalse->set_req_X( 0, le, &_igvn );
2525 iftrue ->set_req_X( 0, le, &_igvn );
2526 }
2527
2528 set_idom(iftrue, le, dd+1);
2529 set_idom(iffalse, le, dd+1);
2530 assert(iff->outcnt() == 0, "should be dead now");
2531 replace_node_and_forward_ctrl(iff, le); // fix 'get_ctrl'
2532
2533 Node* entry_control = init_control;
2534 bool strip_mine_loop = iv_bt == T_INT &&
2535 loop->_child == nullptr &&
2536 sfpt != nullptr &&
2537 !loop->_has_call &&
2538 is_deleteable_safept(sfpt);
2539 IdealLoopTree* outer_ilt = nullptr;
2540 if (strip_mine_loop) {
2541 outer_ilt = create_outer_strip_mined_loop(init_control, loop, cl_prob, le->_fcnt,
2542 entry_control, iffalse);
2543 }
2544
2545 // Now setup a new CountedLoopNode to replace the existing LoopNode
2546 BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2547 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2548 // The following assert is approximately true, and defines the intention
2549 // of can_be_counted_loop. It fails, however, because phase->type
2550 // is not yet initialized for this loop and its parts.
2551 //assert(l->can_be_counted_loop(this), "sanity");
2552 _igvn.register_new_node_with_optimizer(l);
2553 set_loop(l, loop);
2554 loop->_head = l;
2555 // Fix all data nodes placed at the old loop head.
2556 // Uses the lazy-update mechanism of 'get_ctrl'.
2557 replace_node_and_forward_ctrl(x, l);
2558 set_idom(l, entry_control, dom_depth(entry_control) + 1);
2559
2560 if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {
2561 // Check for immediately preceding SafePoint and remove
2562 if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {
2563 if (strip_mine_loop) {
2564 Node* outer_le = outer_ilt->_tail->in(0);
2565 Node* sfpt_clone = sfpt->clone();
2566 sfpt_clone->set_req(0, iffalse);
2567 outer_le->set_req(0, sfpt_clone);
2568
2569 Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2570 if (polladdr != nullptr && polladdr->is_Load()) {
2571 // Polling load should be pinned outside inner loop.
2572 Node* new_polladdr = polladdr->clone();
2573 new_polladdr->set_req(0, iffalse);
2574 _igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2575 set_ctrl(new_polladdr, iffalse);
2576 sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2577 }
2578 // When this code runs, loop bodies have not yet been populated.
2579 const bool body_populated = false;
2580 register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2581 set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2582 }
2583 replace_node_and_forward_ctrl(sfpt, sfpt->in(TypeFunc::Control));
2584 if (loop->_safepts != nullptr) {
2585 loop->_safepts->yank(sfpt);
2586 }
2587 }
2588 }
2589
2590 #ifdef ASSERT
2591 assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2592 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2593 #endif
2594 #ifndef PRODUCT
2595 if (TraceLoopOpts) {
2596 tty->print("Counted ");
2597 loop->dump_head();
2598 }
2599 #endif
2600
2601 C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2602
2603 // Capture bounds of the loop in the induction variable Phi before
2604 // subsequent transformation (iteration splitting) obscures the
2605 // bounds
2606 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2607
2608 if (strip_mine_loop) {
2609 l->mark_strip_mined();
2610 l->verify_strip_mined(1);
2611 outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2612 loop = outer_ilt;
2613 }
2614
2615 #ifndef PRODUCT
2616 if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {
2617 AtomicAccess::inc(&_long_loop_counted_loops);
2618 }
2619 #endif
2620 if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {
2621 l->mark_loop_nest_outer_loop();
2622 }
2623
2624 return true;
2625 }
2626
2627 // Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2628 // If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2629 bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2630 const BasicType iv_bt, Node* loop_entry) {
2631 // Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2632 // successfully find a dominated test with the If node below.
2633 Node* cmp_limit;
2634 Node* bol;
2635 if (stride_con > 0) {
2636 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2637 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2638 } else {
2639 cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2640 bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2641 }
2642
2643 // Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2644 IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2645 // Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2646 IfFalseNode* if_false = new IfFalseNode(iff);
2647 IfTrueNode* if_true = new IfTrueNode(iff);
2648 Node* dominated_iff = _igvn.transform(iff);
2649 // ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2650 const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2651
2652 // Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2653 _igvn.replace_input_of(iff, 0, C->top());
2654 _igvn.replace_input_of(iff, 1, C->top());
2655 return found_dominating_test;
2656 }
2657
2658 //----------------------exact_limit-------------------------------------------
2659 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
2660 assert(loop->_head->is_CountedLoop(), "");
2661 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2662 assert(cl->is_valid_counted_loop(T_INT), "");
2663
2664 if (cl->stride_con() == 1 ||
2665 cl->stride_con() == -1 ||
2666 cl->limit()->Opcode() == Op_LoopLimit) {
2667 // Old code has exact limit (it could be incorrect in case of int overflow).
2668 // Loop limit is exact with stride == 1. And loop may already have exact limit.
2669 return cl->limit();
2670 }
2671 Node *limit = nullptr;
2672 #ifdef ASSERT
2673 BoolTest::mask bt = cl->loopexit()->test_trip();
2674 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2675 #endif
2676 if (cl->has_exact_trip_count()) {
2677 // Simple case: loop has constant boundaries.
2678 // Use jlongs to avoid integer overflow.
2679 int stride_con = cl->stride_con();
2680 jlong init_con = cl->init_trip()->get_int();
2681 jlong limit_con = cl->limit()->get_int();
2682 julong trip_cnt = cl->trip_count();
2683 jlong final_con = init_con + trip_cnt*stride_con;
2684 int final_int = (int)final_con;
2685 // The final value should be in integer range since the loop
2686 // is counted and the limit was checked for overflow.
2687 assert(final_con == (jlong)final_int, "final value should be integer");
2688 limit = _igvn.intcon(final_int);
2689 } else {
2690 // Create new LoopLimit node to get exact limit (final iv value).
2691 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2692 register_new_node(limit, cl->in(LoopNode::EntryControl));
2693 }
2694 assert(limit != nullptr, "sanity");
2695 return limit;
2696 }
2697
2698 //------------------------------Ideal------------------------------------------
2699 // Return a node which is more "ideal" than the current node.
2700 // Attempt to convert into a counted-loop.
2701 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2702 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
2703 phase->C->set_major_progress();
2704 }
2705 return RegionNode::Ideal(phase, can_reshape);
2706 }
2707
2708 #ifdef ASSERT
2709 void LoopNode::verify_strip_mined(int expect_skeleton) const {
2710 const OuterStripMinedLoopNode* outer = nullptr;
2711 const CountedLoopNode* inner = nullptr;
2712 if (is_strip_mined()) {
2713 if (!is_valid_counted_loop(T_INT)) {
2714 return; // Skip malformed counted loop
2715 }
2716 assert(is_CountedLoop(), "no Loop should be marked strip mined");
2717 inner = as_CountedLoop();
2718 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2719 } else if (is_OuterStripMinedLoop()) {
2720 outer = this->as_OuterStripMinedLoop();
2721 inner = outer->unique_ctrl_out()->as_CountedLoop();
2722 assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2723 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2724 }
2725 if (inner != nullptr || outer != nullptr) {
2726 assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2727 Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2728 Node* outer_le = outer_tail->in(0);
2729 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2730 Node* sfpt = outer_le->in(0);
2731 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2732 Node* inner_out = sfpt->in(0);
2733 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2734 assert(cle == inner->loopexit_or_null(), "mismatch");
2735 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2736 if (has_skeleton) {
2737 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2738 assert(outer->outcnt() == 2, "only control nodes");
2739 } else {
2740 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2741 uint phis = 0;
2742 uint be_loads = 0;
2743 Node* be = inner->in(LoopNode::LoopBackControl);
2744 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
2745 Node* u = inner->fast_out(i);
2746 if (u->is_Phi()) {
2747 phis++;
2748 for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {
2749 Node* n = be->fast_out(j);
2750 if (n->is_Load()) {
2751 assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2752 do {
2753 n = n->raw_out(0);
2754 } while (!n->is_Phi());
2755 if (n == u) {
2756 be_loads++;
2757 break;
2758 }
2759 }
2760 }
2761 }
2762 }
2763 assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2764 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
2765 Node* u = outer->fast_out(i);
2766 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2767 }
2768 uint stores = 0;
2769 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
2770 Node* u = inner_out->fast_out(i);
2771 if (u->is_Store()) {
2772 stores++;
2773 }
2774 }
2775 // Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2776 // not guaranteed to be optimized out.
2777 assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2778 }
2779 assert(sfpt->outcnt() == 1, "no data node");
2780 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2781 }
2782 }
2783 #endif
2784
2785 //=============================================================================
2786 //------------------------------Ideal------------------------------------------
2787 // Return a node which is more "ideal" than the current node.
2788 // Attempt to convert into a counted-loop.
2789 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2790 return RegionNode::Ideal(phase, can_reshape);
2791 }
2792
2793 //------------------------------dump_spec--------------------------------------
2794 // Dump special per-node info
2795 #ifndef PRODUCT
2796 void CountedLoopNode::dump_spec(outputStream *st) const {
2797 LoopNode::dump_spec(st);
2798 if (stride_is_con()) {
2799 st->print("stride: %d ",stride_con());
2800 }
2801 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
2802 if (is_main_loop()) st->print("main of N%d", _idx);
2803 if (is_post_loop()) st->print("post of N%d", _main_idx);
2804 if (is_strip_mined()) st->print(" strip mined");
2805 if (is_multiversion_fast_loop()) { st->print(" multiversion_fast"); }
2806 if (is_multiversion_slow_loop()) { st->print(" multiversion_slow"); }
2807 if (is_multiversion_delayed_slow_loop()) { st->print(" multiversion_delayed_slow"); }
2808 }
2809 #endif
2810
2811 //=============================================================================
2812 jlong BaseCountedLoopEndNode::stride_con() const {
2813 return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2814 }
2815
2816
2817 BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {
2818 if (bt == T_INT) {
2819 return new CountedLoopEndNode(control, test, prob, cnt);
2820 }
2821 assert(bt == T_LONG, "unsupported");
2822 return new LongCountedLoopEndNode(control, test, prob, cnt);
2823 }
2824
2825 //=============================================================================
2826 //------------------------------Value-----------------------------------------
2827 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
2828 const Type* init_t = phase->type(in(Init));
2829 const Type* limit_t = phase->type(in(Limit));
2830 const Type* stride_t = phase->type(in(Stride));
2831 // Either input is TOP ==> the result is TOP
2832 if (init_t == Type::TOP) return Type::TOP;
2833 if (limit_t == Type::TOP) return Type::TOP;
2834 if (stride_t == Type::TOP) return Type::TOP;
2835
2836 int stride_con = stride_t->is_int()->get_con();
2837 if (stride_con == 1)
2838 return bottom_type(); // Identity
2839
2840 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
2841 // Use jlongs to avoid integer overflow.
2842 jlong init_con = init_t->is_int()->get_con();
2843 jlong limit_con = limit_t->is_int()->get_con();
2844 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2845 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2846 jlong final_con = init_con + stride_con*trip_count;
2847 int final_int = (int)final_con;
2848 // The final value should be in integer range in almost all cases,
2849 // since the loop is counted and the limit was checked for overflow.
2850 // There some exceptions, for example:
2851 // - During CCP, there might be a temporary overflow from PhiNodes, see JDK-8309266.
2852 // - During PhaseIdealLoop::split_thru_phi, the LoopLimitNode floats possibly far above
2853 // the loop and its predicates, and we might get constants on one side of the phi that
2854 // would lead to overflows. Such a code path would never lead us to enter the loop
2855 // because of the loop limit overflow check that happens after the LoopLimitNode
2856 // computation with overflow, but before we enter the loop, see JDK-8335747.
2857 if (final_con == (jlong)final_int) {
2858 return TypeInt::make(final_int);
2859 } else {
2860 return bottom_type();
2861 }
2862 }
2863
2864 return bottom_type(); // TypeInt::INT
2865 }
2866
2867 //------------------------------Ideal------------------------------------------
2868 // Return a node which is more "ideal" than the current node.
2869 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2870 if (phase->type(in(Init)) == Type::TOP ||
2871 phase->type(in(Limit)) == Type::TOP ||
2872 phase->type(in(Stride)) == Type::TOP)
2873 return nullptr; // Dead
2874
2875 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2876 if (stride_con == 1)
2877 return nullptr; // Identity
2878
2879 // Delay following optimizations until all loop optimizations
2880 // done to keep Ideal graph simple.
2881 if (!can_reshape || !phase->C->post_loop_opts_phase()) {
2882 phase->C->record_for_post_loop_opts_igvn(this);
2883 return nullptr;
2884 }
2885
2886 const TypeInt* init_t = phase->type(in(Init) )->is_int();
2887 const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2888 jlong stride_p;
2889 jlong lim, ini;
2890 julong max;
2891 if (stride_con > 0) {
2892 stride_p = stride_con;
2893 lim = limit_t->_hi;
2894 ini = init_t->_lo;
2895 max = (julong)max_jint;
2896 } else {
2897 stride_p = -(jlong)stride_con;
2898 lim = init_t->_hi;
2899 ini = limit_t->_lo;
2900 max = (julong)(juint)min_jint; // double cast to get 0x0000000080000000, not 0xffffffff80000000
2901 }
2902 julong range = lim - ini + stride_p;
2903 if (range <= max) {
2904 // Convert to integer expression if it is not overflow.
2905 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2906 Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2907 Node *bias = phase->transform(new AddINode(range, stride_m));
2908 Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2909 Node *span = phase->transform(new MulINode(trip, in(Stride)));
2910 return new AddINode(span, in(Init)); // exact limit
2911 }
2912
2913 if (is_power_of_2(stride_p) || // divisor is 2^n
2914 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
2915 // Convert to long expression to avoid integer overflow
2916 // and let igvn optimizer convert this division.
2917 //
2918 Node* init = phase->transform( new ConvI2LNode(in(Init)));
2919 Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2920 Node* stride = phase->longcon(stride_con);
2921 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2922
2923 Node *range = phase->transform(new SubLNode(limit, init));
2924 Node *bias = phase->transform(new AddLNode(range, stride_m));
2925 Node *span;
2926 if (stride_con > 0 && is_power_of_2(stride_p)) {
2927 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2928 // and avoid generating rounding for division. Zero trip guard should
2929 // guarantee that init < limit but sometimes the guard is missing and
2930 // we can get situation when init > limit. Note, for the empty loop
2931 // optimization zero trip guard is generated explicitly which leaves
2932 // only RCE predicate where exact limit is used and the predicate
2933 // will simply fail forcing recompilation.
2934 Node* neg_stride = phase->longcon(-stride_con);
2935 span = phase->transform(new AndLNode(bias, neg_stride));
2936 } else {
2937 Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2938 span = phase->transform(new MulLNode(trip, stride));
2939 }
2940 // Convert back to int
2941 Node *span_int = phase->transform(new ConvL2INode(span));
2942 return new AddINode(span_int, in(Init)); // exact limit
2943 }
2944
2945 return nullptr; // No progress
2946 }
2947
2948 //------------------------------Identity---------------------------------------
2949 // If stride == 1 return limit node.
2950 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
2951 int stride_con = phase->type(in(Stride))->is_int()->get_con();
2952 if (stride_con == 1 || stride_con == -1)
2953 return in(Limit);
2954 return this;
2955 }
2956
2957 //=============================================================================
2958 //----------------------match_incr_with_optional_truncation--------------------
2959 // Match increment with optional truncation:
2960 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2961 // Return null for failure. Success returns the increment node.
2962 Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2963 const TypeInteger** trunc_type,
2964 BasicType bt) {
2965 // Quick cutouts:
2966 if (expr == nullptr || expr->req() != 3) return nullptr;
2967
2968 Node *t1 = nullptr;
2969 Node *t2 = nullptr;
2970 Node* n1 = expr;
2971 int n1op = n1->Opcode();
2972 const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2973
2974 if (bt == T_INT) {
2975 // Try to strip (n1 & M) or (n1 << N >> N) from n1.
2976 if (n1op == Op_AndI &&
2977 n1->in(2)->is_Con() &&
2978 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
2979 // %%% This check should match any mask of 2**K-1.
2980 t1 = n1;
2981 n1 = t1->in(1);
2982 n1op = n1->Opcode();
2983 trunc_t = TypeInt::CHAR;
2984 } else if (n1op == Op_RShiftI &&
2985 n1->in(1) != nullptr &&
2986 n1->in(1)->Opcode() == Op_LShiftI &&
2987 n1->in(2) == n1->in(1)->in(2) &&
2988 n1->in(2)->is_Con()) {
2989 jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2990 // %%% This check should match any shift in [1..31].
2991 if (shift == 16 || shift == 8) {
2992 t1 = n1;
2993 t2 = t1->in(1);
2994 n1 = t2->in(1);
2995 n1op = n1->Opcode();
2996 if (shift == 16) {
2997 trunc_t = TypeInt::SHORT;
2998 } else if (shift == 8) {
2999 trunc_t = TypeInt::BYTE;
3000 }
3001 }
3002 }
3003 }
3004
3005 // If (maybe after stripping) it is an AddI, we won:
3006 if (n1op == Op_Add(bt)) {
3007 *trunc1 = t1;
3008 *trunc2 = t2;
3009 *trunc_type = trunc_t;
3010 return n1;
3011 }
3012
3013 // failed
3014 return nullptr;
3015 }
3016
3017 IfNode* CountedLoopNode::find_multiversion_if_from_multiversion_fast_main_loop() {
3018 assert(is_main_loop() && is_multiversion_fast_loop(), "must be multiversion fast main loop");
3019 CountedLoopEndNode* pre_end = find_pre_loop_end();
3020 if (pre_end == nullptr) { return nullptr; }
3021 Node* pre_entry = pre_end->loopnode()->in(LoopNode::EntryControl);
3022 const Predicates predicates(pre_entry);
3023 IfTrueNode* before_predicates = predicates.entry()->isa_IfTrue();
3024 if (before_predicates != nullptr &&
3025 before_predicates->in(0)->in(1)->is_OpaqueMultiversioning()) {
3026 return before_predicates->in(0)->as_If();
3027 }
3028 return nullptr;
3029 }
3030
3031 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
3032 if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {
3033 verify_strip_mined(expect_skeleton);
3034 return in(EntryControl)->as_Loop();
3035 }
3036 return this;
3037 }
3038
3039 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
3040 assert(is_strip_mined(), "not a strip mined loop");
3041 Node* c = in(EntryControl);
3042 if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {
3043 return nullptr;
3044 }
3045 return c->as_OuterStripMinedLoop();
3046 }
3047
3048 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
3049 Node* c = in(LoopBackControl);
3050 if (c == nullptr || c->is_top()) {
3051 return nullptr;
3052 }
3053 return c->as_IfTrue();
3054 }
3055
3056 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
3057 LoopNode* l = outer_loop();
3058 if (l == nullptr) {
3059 return nullptr;
3060 }
3061 return l->outer_loop_tail();
3062 }
3063
3064 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
3065 IfTrueNode* proj = outer_loop_tail();
3066 if (proj == nullptr) {
3067 return nullptr;
3068 }
3069 Node* c = proj->in(0);
3070 if (c == nullptr || c->is_top() || c->outcnt() != 2) {
3071 return nullptr;
3072 }
3073 return c->as_OuterStripMinedLoopEnd();
3074 }
3075
3076 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
3077 LoopNode* l = outer_loop();
3078 if (l == nullptr) {
3079 return nullptr;
3080 }
3081 return l->outer_loop_end();
3082 }
3083
3084 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
3085 IfNode* le = outer_loop_end();
3086 if (le == nullptr) {
3087 return nullptr;
3088 }
3089 IfFalseNode* c = le->false_proj_or_null();
3090 if (c == nullptr) {
3091 return nullptr;
3092 }
3093 return c->as_IfFalse();
3094 }
3095
3096 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
3097 LoopNode* l = outer_loop();
3098 if (l == nullptr) {
3099 return nullptr;
3100 }
3101 return l->outer_loop_exit();
3102 }
3103
3104 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
3105 IfNode* le = outer_loop_end();
3106 if (le == nullptr) {
3107 return nullptr;
3108 }
3109 Node* c = le->in(0);
3110 if (c == nullptr || c->is_top()) {
3111 return nullptr;
3112 }
3113 assert(c->Opcode() == Op_SafePoint, "broken outer loop");
3114 return c->as_SafePoint();
3115 }
3116
3117 SafePointNode* CountedLoopNode::outer_safepoint() const {
3118 LoopNode* l = outer_loop();
3119 if (l == nullptr) {
3120 return nullptr;
3121 }
3122 return l->outer_safepoint();
3123 }
3124
3125 Node* CountedLoopNode::skip_assertion_predicates_with_halt() {
3126 Node* ctrl = in(LoopNode::EntryControl);
3127 if (ctrl == nullptr) {
3128 // Dying loop.
3129 return nullptr;
3130 }
3131 if (is_main_loop()) {
3132 ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
3133 }
3134 if (is_main_loop() || is_post_loop()) {
3135 AssertionPredicates assertion_predicates(ctrl);
3136 return assertion_predicates.entry();
3137 }
3138 return ctrl;
3139 }
3140
3141
3142 int CountedLoopNode::stride_con() const {
3143 CountedLoopEndNode* cle = loopexit_or_null();
3144 return cle != nullptr ? cle->stride_con() : 0;
3145 }
3146
3147 BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {
3148 if (bt == T_INT) {
3149 return new CountedLoopNode(entry, backedge);
3150 }
3151 assert(bt == T_LONG, "unsupported");
3152 return new LongCountedLoopNode(entry, backedge);
3153 }
3154
3155 void OuterStripMinedLoopNode::fix_sunk_stores_when_back_to_counted_loop(PhaseIterGVN* igvn,
3156 PhaseIdealLoop* iloop) const {
3157 CountedLoopNode* inner_cl = inner_counted_loop();
3158 IfFalseNode* cle_out = inner_loop_exit();
3159
3160 if (cle_out->outcnt() > 1) {
3161 // Look for chains of stores that were sunk
3162 // out of the inner loop and are in the outer loop
3163 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
3164 Node* u = cle_out->fast_out(i);
3165 if (u->is_Store()) {
3166 int alias_idx = igvn->C->get_alias_index(u->adr_type());
3167 Node* first = u;
3168 for (;;) {
3169 Node* next = first->in(MemNode::Memory);
3170 if (!next->is_Store() || next->in(0) != cle_out) {
3171 break;
3172 }
3173 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
3174 first = next;
3175 }
3176 Node* last = u;
3177 for (;;) {
3178 Node* next = nullptr;
3179 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
3180 Node* uu = last->fast_out(j);
3181 if (uu->is_Store() && uu->in(0) == cle_out) {
3182 assert(next == nullptr, "only one in the outer loop");
3183 next = uu;
3184 assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
3185 }
3186 }
3187 if (next == nullptr) {
3188 break;
3189 }
3190 last = next;
3191 }
3192 Node* phi = nullptr;
3193 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
3194 Node* uu = inner_cl->fast_out(j);
3195 if (uu->is_Phi()) {
3196 Node* be = uu->in(LoopNode::LoopBackControl);
3197 if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {
3198 assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
3199 }
3200 if (be == last || be == first->in(MemNode::Memory)) {
3201 assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
3202 assert(phi == nullptr, "only one phi");
3203 phi = uu;
3204 }
3205 }
3206 }
3207 #ifdef ASSERT
3208 for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {
3209 Node* uu = inner_cl->fast_out(j);
3210 if (uu->is_memory_phi()) {
3211 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
3212 assert(phi == uu, "what's that phi?");
3213 } else if (uu->adr_type() == TypePtr::BOTTOM) {
3214 Node* n = uu->in(LoopNode::LoopBackControl);
3215 uint limit = igvn->C->live_nodes();
3216 uint i = 0;
3217 while (n != uu) {
3218 i++;
3219 assert(i < limit, "infinite loop");
3220 if (n->is_Proj()) {
3221 n = n->in(0);
3222 } else if (n->is_SafePoint() || n->is_MemBar()) {
3223 n = n->in(TypeFunc::Memory);
3224 } else if (n->is_Phi()) {
3225 n = n->in(1);
3226 } else if (n->is_MergeMem()) {
3227 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
3228 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
3229 n = n->in(MemNode::Memory);
3230 } else {
3231 n->dump();
3232 ShouldNotReachHere();
3233 }
3234 }
3235 }
3236 }
3237 }
3238 #endif
3239 if (phi == nullptr) {
3240 // If an entire chains was sunk, the
3241 // inner loop has no phi for that memory
3242 // slice, create one for the outer loop
3243 phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
3244 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
3245 phi->set_req(LoopNode::LoopBackControl, last);
3246 phi = register_new_node(phi, inner_cl, igvn, iloop);
3247 igvn->replace_input_of(first, MemNode::Memory, phi);
3248 } else {
3249 // Or fix the outer loop fix to include
3250 // that chain of stores.
3251 Node* be = phi->in(LoopNode::LoopBackControl);
3252 assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
3253 if (be == first->in(MemNode::Memory)) {
3254 if (be == phi->in(LoopNode::LoopBackControl)) {
3255 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
3256 } else {
3257 igvn->replace_input_of(be, MemNode::Memory, last);
3258 }
3259 } else {
3260 #ifdef ASSERT
3261 if (be == phi->in(LoopNode::LoopBackControl)) {
3262 assert(phi->in(LoopNode::LoopBackControl) == last, "");
3263 } else {
3264 assert(be->in(MemNode::Memory) == last, "");
3265 }
3266 #endif
3267 }
3268 }
3269 }
3270 }
3271 }
3272 }
3273
3274 // The outer strip mined loop is initially only partially constructed. In particular Phis are omitted.
3275 // See comment above: PhaseIdealLoop::create_outer_strip_mined_loop()
3276 // We're now in the process of finishing the construction of the outer loop. For each Phi in the inner loop, a Phi in
3277 // the outer loop was just now created. However, Sunk Stores cause an extra challenge:
3278 // 1) If all Stores in the inner loop were sunk for a particular memory slice, there's no Phi left for that memory slice
3279 // 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
3280 // added for each chain of sunk Stores for a particular memory slice.
3281 // 2) If some Stores were sunk and some left in the inner loop, a Phi was already created in the outer loop but
3282 // its backedge input wasn't wired correctly to the last Store of the chain: the backedge input was set to the
3283 // backedge of the inner loop Phi instead, but it needs to be the last Store of the chain in the outer loop. We now
3284 // have to fix that too.
3285 void OuterStripMinedLoopNode::handle_sunk_stores_when_finishing_construction(PhaseIterGVN* igvn) {
3286 IfFalseNode* cle_exit_proj = inner_loop_exit();
3287
3288 // Find Sunk stores: Sunk stores are pinned on the loop exit projection of the inner loop. Indeed, because Sunk Stores
3289 // modify the memory state captured by the SafePoint in the outer strip mined loop, they must be above it. The
3290 // SafePoint's control input is the loop exit projection. It's also the only control out of the inner loop above the
3291 // SafePoint.
3292 #ifdef ASSERT
3293 int stores_in_outer_loop_cnt = 0;
3294 for (DUIterator_Fast imax, i = cle_exit_proj->fast_outs(imax); i < imax; i++) {
3295 Node* u = cle_exit_proj->fast_out(i);
3296 if (u->is_Store()) {
3297 stores_in_outer_loop_cnt++;
3298 }
3299 }
3300 #endif
3301
3302 // Sunk stores are reachable from the memory state of the outer loop safepoint
3303 SafePointNode* safepoint = outer_safepoint();
3304 MergeMemNode* mm = safepoint->in(TypeFunc::Memory)->isa_MergeMem();
3305 if (mm == nullptr) {
3306 // There is no MergeMem, which should only happen if there was no memory node
3307 // sunk out of the loop.
3308 assert(stores_in_outer_loop_cnt == 0, "inconsistent");
3309 return;
3310 }
3311 DEBUG_ONLY(int stores_in_outer_loop_cnt2 = 0);
3312 for (MergeMemStream mms(mm); mms.next_non_empty();) {
3313 Node* mem = mms.memory();
3314 // Traverse up the chain of stores to find the first store pinned
3315 // at the loop exit projection.
3316 Node* last = mem;
3317 Node* first = nullptr;
3318 while (mem->is_Store() && mem->in(0) == cle_exit_proj) {
3319 DEBUG_ONLY(stores_in_outer_loop_cnt2++);
3320 first = mem;
3321 mem = mem->in(MemNode::Memory);
3322 }
3323 if (first != nullptr) {
3324 // Found a chain of Stores that were sunk
3325 // Do we already have a memory Phi for that slice on the outer loop? If that is the case, that Phi was created
3326 // by cloning an inner loop Phi. The inner loop Phi should have mem, the memory state of the first Store out of
3327 // the inner loop, as input on the backedge. So does the outer loop Phi given it's a clone.
3328 Node* phi = nullptr;
3329 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3330 Node* u = mem->fast_out(i);
3331 if (u->is_Phi() && u->in(0) == this && u->in(LoopBackControl) == mem) {
3332 assert(phi == nullptr, "there should be only one");
3333 phi = u;
3334 PRODUCT_ONLY(break);
3335 }
3336 }
3337 if (phi == nullptr) {
3338 // No outer loop Phi? create one
3339 phi = PhiNode::make(this, last);
3340 phi->set_req(EntryControl, mem);
3341 phi = igvn->transform(phi);
3342 igvn->replace_input_of(first, MemNode::Memory, phi);
3343 } else {
3344 // Fix memory state along the backedge: it should be the last sunk Store of the chain
3345 igvn->replace_input_of(phi, LoopBackControl, last);
3346 }
3347 }
3348 }
3349 assert(stores_in_outer_loop_cnt == stores_in_outer_loop_cnt2, "inconsistent");
3350 }
3351
3352 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
3353 verify_strip_mined(1);
3354 // Look for the outer & inner strip mined loop, reduce number of
3355 // iterations of the inner loop, set exit condition of outer loop,
3356 // construct required phi nodes for outer loop.
3357 CountedLoopNode* inner_cl = inner_counted_loop();
3358 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
3359 if (LoopStripMiningIter == 0) {
3360 remove_outer_loop_and_safepoint(igvn);
3361 return;
3362 }
3363 if (LoopStripMiningIter == 1) {
3364 transform_to_counted_loop(igvn, nullptr);
3365 return;
3366 }
3367 Node* inner_iv_phi = inner_cl->phi();
3368 if (inner_iv_phi == nullptr) {
3369 IfNode* outer_le = outer_loop_end();
3370 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3371 igvn->replace_node(outer_le, iff);
3372 inner_cl->clear_strip_mined();
3373 return;
3374 }
3375 CountedLoopEndNode* inner_cle = inner_counted_loop_end();
3376
3377 int stride = inner_cl->stride_con();
3378 // For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
3379 // except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
3380 // stride, the method is guaranteed to return at the next check below.
3381 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
3382 int scaled_iters = (int)scaled_iters_long;
3383 if ((jlong)scaled_iters != scaled_iters_long) {
3384 // Remove outer loop and safepoint (too few iterations)
3385 remove_outer_loop_and_safepoint(igvn);
3386 return;
3387 }
3388 jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
3389 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
3390 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
3391 assert(iter_estimate > 0, "broken");
3392 if (iter_estimate <= short_scaled_iters) {
3393 // Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
3394 remove_outer_loop_and_safepoint(igvn);
3395 return;
3396 }
3397 if (iter_estimate <= scaled_iters_long) {
3398 // We would only go through one iteration of
3399 // the outer loop: drop the outer loop but
3400 // keep the safepoint so we don't run for
3401 // too long without a safepoint
3402 IfNode* outer_le = outer_loop_end();
3403 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3404 igvn->replace_node(outer_le, iff);
3405 inner_cl->clear_strip_mined();
3406 return;
3407 }
3408
3409 IfTrueNode* cle_tail = inner_cle->true_proj();
3410 ResourceMark rm;
3411 Node_List old_new;
3412 if (cle_tail->outcnt() > 1) {
3413 // Look for nodes on backedge of inner loop and clone them
3414 Unique_Node_List backedge_nodes;
3415 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
3416 Node* u = cle_tail->fast_out(i);
3417 if (u != inner_cl) {
3418 assert(!u->is_CFG(), "control flow on the backedge?");
3419 backedge_nodes.push(u);
3420 }
3421 }
3422 uint last = igvn->C->unique();
3423 for (uint next = 0; next < backedge_nodes.size(); next++) {
3424 Node* n = backedge_nodes.at(next);
3425 old_new.map(n->_idx, n->clone());
3426 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3427 Node* u = n->fast_out(i);
3428 assert(!u->is_CFG(), "broken");
3429 if (u->_idx >= last) {
3430 continue;
3431 }
3432 if (!u->is_Phi()) {
3433 backedge_nodes.push(u);
3434 } else {
3435 assert(u->in(0) == inner_cl, "strange phi on the backedge");
3436 }
3437 }
3438 }
3439 // Put the clones on the outer loop backedge
3440 Node* le_tail = outer_loop_tail();
3441 for (uint next = 0; next < backedge_nodes.size(); next++) {
3442 Node *n = old_new[backedge_nodes.at(next)->_idx];
3443 for (uint i = 1; i < n->req(); i++) {
3444 if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {
3445 n->set_req(i, old_new[n->in(i)->_idx]);
3446 }
3447 }
3448 if (n->in(0) != nullptr && n->in(0) == cle_tail) {
3449 n->set_req(0, le_tail);
3450 }
3451 igvn->register_new_node_with_optimizer(n);
3452 }
3453 }
3454
3455 Node* iv_phi = nullptr;
3456 // Make a clone of each phi in the inner loop for the outer loop
3457 // When Stores were Sunk, after this step, a Phi may still be missing or its backedge incorrectly wired. See
3458 // handle_sunk_stores_when_finishing_construction()
3459 for (uint i = 0; i < inner_cl->outcnt(); i++) {
3460 Node* u = inner_cl->raw_out(i);
3461 if (u->is_Phi()) {
3462 assert(u->in(0) == inner_cl, "inconsistent");
3463 Node* phi = u->clone();
3464 phi->set_req(0, this);
3465 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3466 if (be != nullptr) {
3467 phi->set_req(LoopNode::LoopBackControl, be);
3468 }
3469 phi = igvn->transform(phi);
3470 igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3471 if (u == inner_iv_phi) {
3472 iv_phi = phi;
3473 }
3474 }
3475 }
3476
3477 handle_sunk_stores_when_finishing_construction(igvn);
3478
3479 if (iv_phi != nullptr) {
3480 // Now adjust the inner loop's exit condition
3481 Node* limit = inner_cl->limit();
3482 // If limit < init for stride > 0 (or limit > init for stride < 0),
3483 // the loop body is run only once. Given limit - init (init - limit resp.)
3484 // would be negative, the unsigned comparison below would cause
3485 // the loop body to be run for LoopStripMiningIter.
3486 Node* max = nullptr;
3487 if (stride > 0) {
3488 max = MaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3489 } else {
3490 max = MaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3491 }
3492 // sub is positive and can be larger than the max signed int
3493 // value. Use an unsigned min.
3494 Node* const_iters = igvn->intcon(scaled_iters);
3495 Node* min = MaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3496 // min is the number of iterations for the next inner loop execution:
3497 // unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3498 // unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3499
3500 Node* new_limit = nullptr;
3501 if (stride > 0) {
3502 new_limit = igvn->transform(new AddINode(min, iv_phi));
3503 } else {
3504 new_limit = igvn->transform(new SubINode(iv_phi, min));
3505 }
3506 Node* inner_cmp = inner_cle->cmp_node();
3507 Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3508 Node* outer_bol = inner_bol;
3509 // cmp node for inner loop may be shared
3510 inner_cmp = inner_cmp->clone();
3511 inner_cmp->set_req(2, new_limit);
3512 inner_bol = inner_bol->clone();
3513 inner_bol->set_req(1, igvn->transform(inner_cmp));
3514 igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3515 // Set the outer loop's exit condition too
3516 igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3517 } else {
3518 assert(false, "should be able to adjust outer loop");
3519 IfNode* outer_le = outer_loop_end();
3520 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3521 igvn->replace_node(outer_le, iff);
3522 inner_cl->clear_strip_mined();
3523 }
3524 }
3525
3526 void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3527 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3528 CountedLoopEndNode* cle = inner_cl->loopexit();
3529 Node* inner_test = cle->in(1);
3530 IfNode* outer_le = outer_loop_end();
3531 CountedLoopEndNode* inner_cle = inner_cl->loopexit();
3532 Node* safepoint = outer_safepoint();
3533
3534 fix_sunk_stores_when_back_to_counted_loop(igvn, iloop);
3535
3536 // make counted loop exit test always fail
3537 ConINode* zero = igvn->intcon(0);
3538 if (iloop != nullptr) {
3539 iloop->set_root_as_ctrl(zero);
3540 }
3541 igvn->replace_input_of(cle, 1, zero);
3542 // replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3543 Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3544 register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3545 if (iloop == nullptr) {
3546 igvn->replace_node(outer_le, new_end);
3547 } else {
3548 iloop->replace_node_and_forward_ctrl(outer_le, new_end);
3549 }
3550 // the backedge of the inner loop must be rewired to the new loop end
3551 IfTrueNode* backedge = cle->true_proj();
3552 igvn->replace_input_of(backedge, 0, new_end);
3553 if (iloop != nullptr) {
3554 iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3555 }
3556 // make the outer loop go away
3557 igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3558 igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3559 inner_cl->clear_strip_mined();
3560 if (iloop != nullptr) {
3561 Unique_Node_List wq;
3562 wq.push(safepoint);
3563
3564 IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3565 IdealLoopTree* loop = iloop->get_loop(inner_cl);
3566
3567 for (uint i = 0; i < wq.size(); i++) {
3568 Node* n = wq.at(i);
3569 for (uint j = 0; j < n->req(); ++j) {
3570 Node* in = n->in(j);
3571 if (in == nullptr || in->is_CFG()) {
3572 continue;
3573 }
3574 if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {
3575 continue;
3576 }
3577 wq.push(in);
3578 }
3579 assert(!loop->_body.contains(n), "Shouldn't append node to body twice");
3580 loop->_body.push(n);
3581 }
3582 iloop->set_loop(safepoint, loop);
3583 loop->_body.push(safepoint);
3584 iloop->set_loop(safepoint->in(0), loop);
3585 loop->_body.push(safepoint->in(0));
3586 outer_loop_ilt->_tail = igvn->C->top();
3587 }
3588 }
3589
3590 void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {
3591 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3592 Node* outer_sfpt = outer_safepoint();
3593 Node* outer_out = outer_loop_exit();
3594 igvn->replace_node(outer_out, outer_sfpt->in(0));
3595 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3596 inner_cl->clear_strip_mined();
3597 }
3598
3599 Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {
3600 if (iloop == nullptr) {
3601 return igvn->transform(node);
3602 }
3603 iloop->register_new_node(node, ctrl);
3604 return node;
3605 }
3606
3607 Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3608 PhaseIdealLoop* iloop) {
3609 if (iloop == nullptr) {
3610 return igvn->transform(node);
3611 }
3612 iloop->register_control(node, iloop->get_loop(loop), idom);
3613 return node;
3614 }
3615
3616 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
3617 if (!in(0)) return Type::TOP;
3618 if (phase->type(in(0)) == Type::TOP)
3619 return Type::TOP;
3620
3621 // Until expansion, the loop end condition is not set so this should not constant fold.
3622 if (is_expanded(phase)) {
3623 return IfNode::Value(phase);
3624 }
3625
3626 return TypeTuple::IFBOTH;
3627 }
3628
3629 bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {
3630 // The outer strip mined loop head only has Phi uses after expansion
3631 if (phase->is_IterGVN()) {
3632 IfTrueNode* backedge = true_proj_or_null();
3633 if (backedge != nullptr) {
3634 Node* head = backedge->unique_ctrl_out_or_null();
3635 if (head != nullptr && head->is_OuterStripMinedLoop()) {
3636 if (head->find_out_with(Op_Phi) != nullptr) {
3637 return true;
3638 }
3639 }
3640 }
3641 }
3642 return false;
3643 }
3644
3645 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3646 if (remove_dead_region(phase, can_reshape)) return this;
3647
3648 return nullptr;
3649 }
3650
3651 //------------------------------filtered_type--------------------------------
3652 // Return a type based on condition control flow
3653 // A successful return will be a type that is restricted due
3654 // to a series of dominating if-tests, such as:
3655 // if (i < 10) {
3656 // if (i > 0) {
3657 // here: "i" type is [1..10)
3658 // }
3659 // }
3660 // or a control flow merge
3661 // if (i < 10) {
3662 // do {
3663 // phi( , ) -- at top of loop type is [min_int..10)
3664 // i = ?
3665 // } while ( i < 10)
3666 //
3667 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
3668 assert(n && n->bottom_type()->is_int(), "must be int");
3669 const TypeInt* filtered_t = nullptr;
3670 if (!n->is_Phi()) {
3671 assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3672 filtered_t = filtered_type_from_dominators(n, n_ctrl);
3673
3674 } else {
3675 Node* phi = n->as_Phi();
3676 Node* region = phi->in(0);
3677 assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3678 if (region && region != C->top()) {
3679 for (uint i = 1; i < phi->req(); i++) {
3680 Node* val = phi->in(i);
3681 Node* use_c = region->in(i);
3682 const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3683 if (val_t != nullptr) {
3684 if (filtered_t == nullptr) {
3685 filtered_t = val_t;
3686 } else {
3687 filtered_t = filtered_t->meet(val_t)->is_int();
3688 }
3689 }
3690 }
3691 }
3692 }
3693 const TypeInt* n_t = _igvn.type(n)->is_int();
3694 if (filtered_t != nullptr) {
3695 n_t = n_t->join(filtered_t)->is_int();
3696 }
3697 return n_t;
3698 }
3699
3700
3701 //------------------------------filtered_type_from_dominators--------------------------------
3702 // Return a possibly more restrictive type for val based on condition control flow of dominators
3703 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
3704 if (val->is_Con()) {
3705 return val->bottom_type()->is_int();
3706 }
3707 uint if_limit = 10; // Max number of dominating if's visited
3708 const TypeInt* rtn_t = nullptr;
3709
3710 if (use_ctrl && use_ctrl != C->top()) {
3711 Node* val_ctrl = get_ctrl(val);
3712 uint val_dom_depth = dom_depth(val_ctrl);
3713 Node* pred = use_ctrl;
3714 uint if_cnt = 0;
3715 while (if_cnt < if_limit) {
3716 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
3717 if_cnt++;
3718 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3719 if (if_t != nullptr) {
3720 if (rtn_t == nullptr) {
3721 rtn_t = if_t;
3722 } else {
3723 rtn_t = rtn_t->join(if_t)->is_int();
3724 }
3725 }
3726 }
3727 pred = idom(pred);
3728 if (pred == nullptr || pred == C->top()) {
3729 break;
3730 }
3731 // Stop if going beyond definition block of val
3732 if (dom_depth(pred) < val_dom_depth) {
3733 break;
3734 }
3735 }
3736 }
3737 return rtn_t;
3738 }
3739
3740
3741 //------------------------------dump_spec--------------------------------------
3742 // Dump special per-node info
3743 #ifndef PRODUCT
3744 void CountedLoopEndNode::dump_spec(outputStream *st) const {
3745 if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {
3746 BoolTest bt( test_trip()); // Added this for g++.
3747
3748 st->print("[");
3749 bt.dump_on(st);
3750 st->print("]");
3751 }
3752 st->print(" ");
3753 IfNode::dump_spec(st);
3754 }
3755 #endif
3756
3757 //=============================================================================
3758 //------------------------------is_member--------------------------------------
3759 // Is 'l' a member of 'this'?
3760 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
3761 while( l->_nest > _nest ) l = l->_parent;
3762 return l == this;
3763 }
3764
3765 //------------------------------set_nest---------------------------------------
3766 // Set loop tree nesting depth. Accumulate _has_call bits.
3767 int IdealLoopTree::set_nest( uint depth ) {
3768 assert(depth <= SHRT_MAX, "sanity");
3769 _nest = depth;
3770 int bits = _has_call;
3771 if( _child ) bits |= _child->set_nest(depth+1);
3772 if( bits ) _has_call = 1;
3773 if( _next ) bits |= _next ->set_nest(depth );
3774 return bits;
3775 }
3776
3777 //------------------------------split_fall_in----------------------------------
3778 // Split out multiple fall-in edges from the loop header. Move them to a
3779 // private RegionNode before the loop. This becomes the loop landing pad.
3780 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
3781 PhaseIterGVN &igvn = phase->_igvn;
3782 uint i;
3783
3784 // Make a new RegionNode to be the landing pad.
3785 RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3786 phase->set_loop(landing_pad,_parent);
3787 // If _head was irreducible loop entry, landing_pad may now be too
3788 landing_pad->set_loop_status(_head->as_Region()->loop_status());
3789 // Gather all the fall-in control paths into the landing pad
3790 uint icnt = fall_in_cnt;
3791 uint oreq = _head->req();
3792 for( i = oreq-1; i>0; i-- )
3793 if( !phase->is_member( this, _head->in(i) ) )
3794 landing_pad->set_req(icnt--,_head->in(i));
3795
3796 // Peel off PhiNode edges as well
3797 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3798 Node *oj = _head->fast_out(j);
3799 if( oj->is_Phi() ) {
3800 PhiNode* old_phi = oj->as_Phi();
3801 assert( old_phi->region() == _head, "" );
3802 igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3803 Node *p = PhiNode::make_blank(landing_pad, old_phi);
3804 uint icnt = fall_in_cnt;
3805 for( i = oreq-1; i>0; i-- ) {
3806 if( !phase->is_member( this, _head->in(i) ) ) {
3807 p->init_req(icnt--, old_phi->in(i));
3808 // Go ahead and clean out old edges from old phi
3809 old_phi->del_req(i);
3810 }
3811 }
3812 // Search for CSE's here, because ZKM.jar does a lot of
3813 // loop hackery and we need to be a little incremental
3814 // with the CSE to avoid O(N^2) node blow-up.
3815 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3816 if( p2 ) { // Found CSE
3817 p->destruct(&igvn); // Recover useless new node
3818 p = p2; // Use old node
3819 } else {
3820 igvn.register_new_node_with_optimizer(p, old_phi);
3821 }
3822 // Make old Phi refer to new Phi.
3823 old_phi->add_req(p);
3824 // Check for the special case of making the old phi useless and
3825 // disappear it. In JavaGrande I have a case where this useless
3826 // Phi is the loop limit and prevents recognizing a CountedLoop
3827 // which in turn prevents removing an empty loop.
3828 Node *id_old_phi = old_phi->Identity(&igvn);
3829 if( id_old_phi != old_phi ) { // Found a simple identity?
3830 // Note that I cannot call 'replace_node' here, because
3831 // that will yank the edge from old_phi to the Region and
3832 // I'm mid-iteration over the Region's uses.
3833 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
3834 Node* use = old_phi->last_out(i);
3835 igvn.rehash_node_delayed(use);
3836 uint uses_found = 0;
3837 for (uint j = 0; j < use->len(); j++) {
3838 if (use->in(j) == old_phi) {
3839 if (j < use->req()) use->set_req (j, id_old_phi);
3840 else use->set_prec(j, id_old_phi);
3841 uses_found++;
3842 }
3843 }
3844 i -= uses_found; // we deleted 1 or more copies of this edge
3845 }
3846 }
3847 igvn._worklist.push(old_phi);
3848 }
3849 }
3850 // Finally clean out the fall-in edges from the RegionNode
3851 for( i = oreq-1; i>0; i-- ) {
3852 if( !phase->is_member( this, _head->in(i) ) ) {
3853 _head->del_req(i);
3854 }
3855 }
3856 igvn.rehash_node_delayed(_head);
3857 // Transform landing pad
3858 igvn.register_new_node_with_optimizer(landing_pad, _head);
3859 // Insert landing pad into the header
3860 _head->add_req(landing_pad);
3861 }
3862
3863 //------------------------------split_outer_loop-------------------------------
3864 // Split out the outermost loop from this shared header.
3865 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
3866 PhaseIterGVN &igvn = phase->_igvn;
3867
3868 // Find index of outermost loop; it should also be my tail.
3869 uint outer_idx = 1;
3870 while( _head->in(outer_idx) != _tail ) outer_idx++;
3871
3872 // Make a LoopNode for the outermost loop.
3873 Node *ctl = _head->in(LoopNode::EntryControl);
3874 Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3875 outer = igvn.register_new_node_with_optimizer(outer, _head);
3876 phase->set_created_loop_node();
3877
3878 // Outermost loop falls into '_head' loop
3879 _head->set_req(LoopNode::EntryControl, outer);
3880 _head->del_req(outer_idx);
3881 // Split all the Phis up between '_head' loop and 'outer' loop.
3882 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
3883 Node *out = _head->fast_out(j);
3884 if( out->is_Phi() ) {
3885 PhiNode *old_phi = out->as_Phi();
3886 assert( old_phi->region() == _head, "" );
3887 Node *phi = PhiNode::make_blank(outer, old_phi);
3888 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3889 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3890 phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3891 // Make old Phi point to new Phi on the fall-in path
3892 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3893 old_phi->del_req(outer_idx);
3894 }
3895 }
3896
3897 // Use the new loop head instead of the old shared one
3898 _head = outer;
3899 phase->set_loop(_head, this);
3900 }
3901
3902 //------------------------------fix_parent-------------------------------------
3903 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
3904 loop->_parent = parent;
3905 if( loop->_child ) fix_parent( loop->_child, loop );
3906 if( loop->_next ) fix_parent( loop->_next , parent );
3907 }
3908
3909 //------------------------------estimate_path_freq-----------------------------
3910 static float estimate_path_freq( Node *n ) {
3911 // Try to extract some path frequency info
3912 IfNode *iff;
3913 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
3914 uint nop = n->Opcode();
3915 if( nop == Op_SafePoint ) { // Skip any safepoint
3916 n = n->in(0);
3917 continue;
3918 }
3919 if( nop == Op_CatchProj ) { // Get count from a prior call
3920 // Assume call does not always throw exceptions: means the call-site
3921 // count is also the frequency of the fall-through path.
3922 assert( n->is_CatchProj(), "" );
3923 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3924 return 0.0f; // Assume call exception path is rare
3925 Node *call = n->in(0)->in(0)->in(0);
3926 assert( call->is_Call(), "expect a call here" );
3927 const JVMState *jvms = ((CallNode*)call)->jvms();
3928 ciMethodData* methodData = jvms->method()->method_data();
3929 if (!methodData->is_mature()) return 0.0f; // No call-site data
3930 ciProfileData* data = methodData->bci_to_data(jvms->bci());
3931 if ((data == nullptr) || !data->is_CounterData()) {
3932 // no call profile available, try call's control input
3933 n = n->in(0);
3934 continue;
3935 }
3936 return data->as_CounterData()->count()/FreqCountInvocations;
3937 }
3938 // See if there's a gating IF test
3939 Node *n_c = n->in(0);
3940 if( !n_c->is_If() ) break; // No estimate available
3941 iff = n_c->as_If();
3942 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3943 // Compute how much count comes on this path
3944 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3945 // Have no count info. Skip dull uncommon-trap like branches.
3946 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3947 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3948 break;
3949 // Skip through never-taken branch; look for a real loop exit.
3950 n = iff->in(0);
3951 }
3952 return 0.0f; // No estimate available
3953 }
3954
3955 //------------------------------merge_many_backedges---------------------------
3956 // Merge all the backedges from the shared header into a private Region.
3957 // Feed that region as the one backedge to this loop.
3958 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
3959 uint i;
3960
3961 // Scan for the top 2 hottest backedges
3962 float hotcnt = 0.0f;
3963 float warmcnt = 0.0f;
3964 uint hot_idx = 0;
3965 // Loop starts at 2 because slot 1 is the fall-in path
3966 for( i = 2; i < _head->req(); i++ ) {
3967 float cnt = estimate_path_freq(_head->in(i));
3968 if( cnt > hotcnt ) { // Grab hottest path
3969 warmcnt = hotcnt;
3970 hotcnt = cnt;
3971 hot_idx = i;
3972 } else if( cnt > warmcnt ) { // And 2nd hottest path
3973 warmcnt = cnt;
3974 }
3975 }
3976
3977 // See if the hottest backedge is worthy of being an inner loop
3978 // by being much hotter than the next hottest backedge.
3979 if( hotcnt <= 0.0001 ||
3980 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3981
3982 // Peel out the backedges into a private merge point; peel
3983 // them all except optionally hot_idx.
3984 PhaseIterGVN &igvn = phase->_igvn;
3985
3986 Node *hot_tail = nullptr;
3987 // Make a Region for the merge point
3988 Node *r = new RegionNode(1);
3989 for( i = 2; i < _head->req(); i++ ) {
3990 if( i != hot_idx )
3991 r->add_req( _head->in(i) );
3992 else hot_tail = _head->in(i);
3993 }
3994 igvn.register_new_node_with_optimizer(r, _head);
3995 // Plug region into end of loop _head, followed by hot_tail
3996 while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
3997 igvn.replace_input_of(_head, 2, r);
3998 if( hot_idx ) _head->add_req(hot_tail);
3999
4000 // Split all the Phis up between '_head' loop and the Region 'r'
4001 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
4002 Node *out = _head->fast_out(j);
4003 if( out->is_Phi() ) {
4004 PhiNode* n = out->as_Phi();
4005 igvn.hash_delete(n); // Delete from hash before hacking edges
4006 Node *hot_phi = nullptr;
4007 Node *phi = new PhiNode(r, n->type(), n->adr_type());
4008 // Check all inputs for the ones to peel out
4009 uint j = 1;
4010 for( uint i = 2; i < n->req(); i++ ) {
4011 if( i != hot_idx )
4012 phi->set_req( j++, n->in(i) );
4013 else hot_phi = n->in(i);
4014 }
4015 // Register the phi but do not transform until whole place transforms
4016 igvn.register_new_node_with_optimizer(phi, n);
4017 // Add the merge phi to the old Phi
4018 while( n->req() > 3 ) n->del_req( n->req()-1 );
4019 igvn.replace_input_of(n, 2, phi);
4020 if( hot_idx ) n->add_req(hot_phi);
4021 }
4022 }
4023
4024
4025 // Insert a new IdealLoopTree inserted below me. Turn it into a clone
4026 // of self loop tree. Turn self into a loop headed by _head and with
4027 // tail being the new merge point.
4028 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
4029 phase->set_loop(_tail,ilt); // Adjust tail
4030 _tail = r; // Self's tail is new merge point
4031 phase->set_loop(r,this);
4032 ilt->_child = _child; // New guy has my children
4033 _child = ilt; // Self has new guy as only child
4034 ilt->_parent = this; // new guy has self for parent
4035 ilt->_nest = _nest; // Same nesting depth (for now)
4036
4037 // Starting with 'ilt', look for child loop trees using the same shared
4038 // header. Flatten these out; they will no longer be loops in the end.
4039 IdealLoopTree **pilt = &_child;
4040 while( ilt ) {
4041 if( ilt->_head == _head ) {
4042 uint i;
4043 for( i = 2; i < _head->req(); i++ )
4044 if( _head->in(i) == ilt->_tail )
4045 break; // Still a loop
4046 if( i == _head->req() ) { // No longer a loop
4047 // Flatten ilt. Hang ilt's "_next" list from the end of
4048 // ilt's '_child' list. Move the ilt's _child up to replace ilt.
4049 IdealLoopTree **cp = &ilt->_child;
4050 while( *cp ) cp = &(*cp)->_next; // Find end of child list
4051 *cp = ilt->_next; // Hang next list at end of child list
4052 *pilt = ilt->_child; // Move child up to replace ilt
4053 ilt->_head = nullptr; // Flag as a loop UNIONED into parent
4054 ilt = ilt->_child; // Repeat using new ilt
4055 continue; // do not advance over ilt->_child
4056 }
4057 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
4058 phase->set_loop(_head,ilt);
4059 }
4060 pilt = &ilt->_child; // Advance to next
4061 ilt = *pilt;
4062 }
4063
4064 if( _child ) fix_parent( _child, this );
4065 }
4066
4067 //------------------------------beautify_loops---------------------------------
4068 // Split shared headers and insert loop landing pads.
4069 // Insert a LoopNode to replace the RegionNode.
4070 // Return TRUE if loop tree is structurally changed.
4071 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
4072 bool result = false;
4073 // Cache parts in locals for easy
4074 PhaseIterGVN &igvn = phase->_igvn;
4075
4076 igvn.hash_delete(_head); // Yank from hash before hacking edges
4077
4078 // Check for multiple fall-in paths. Peel off a landing pad if need be.
4079 int fall_in_cnt = 0;
4080 for( uint i = 1; i < _head->req(); i++ )
4081 if( !phase->is_member( this, _head->in(i) ) )
4082 fall_in_cnt++;
4083 assert( fall_in_cnt, "at least 1 fall-in path" );
4084 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
4085 split_fall_in( phase, fall_in_cnt );
4086
4087 // Swap inputs to the _head and all Phis to move the fall-in edge to
4088 // the left.
4089 fall_in_cnt = 1;
4090 while( phase->is_member( this, _head->in(fall_in_cnt) ) )
4091 fall_in_cnt++;
4092 if( fall_in_cnt > 1 ) {
4093 // Since I am just swapping inputs I do not need to update def-use info
4094 Node *tmp = _head->in(1);
4095 igvn.rehash_node_delayed(_head);
4096 _head->set_req( 1, _head->in(fall_in_cnt) );
4097 _head->set_req( fall_in_cnt, tmp );
4098 // Swap also all Phis
4099 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
4100 Node* phi = _head->fast_out(i);
4101 if( phi->is_Phi() ) {
4102 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
4103 tmp = phi->in(1);
4104 phi->set_req( 1, phi->in(fall_in_cnt) );
4105 phi->set_req( fall_in_cnt, tmp );
4106 }
4107 }
4108 }
4109 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
4110 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
4111
4112 // If I am a shared header (multiple backedges), peel off the many
4113 // backedges into a private merge point and use the merge point as
4114 // the one true backedge.
4115 if (_head->req() > 3) {
4116 // Merge the many backedges into a single backedge but leave
4117 // the hottest backedge as separate edge for the following peel.
4118 if (!_irreducible) {
4119 merge_many_backedges( phase );
4120 }
4121
4122 // When recursively beautify my children, split_fall_in can change
4123 // loop tree structure when I am an irreducible loop. Then the head
4124 // of my children has a req() not bigger than 3. Here we need to set
4125 // result to true to catch that case in order to tell the caller to
4126 // rebuild loop tree. See issue JDK-8244407 for details.
4127 result = true;
4128 }
4129
4130 // If I have one hot backedge, peel off myself loop.
4131 // I better be the outermost loop.
4132 if (_head->req() > 3 && !_irreducible) {
4133 split_outer_loop( phase );
4134 result = true;
4135
4136 } else if (!_head->is_Loop() && !_irreducible) {
4137 // Make a new LoopNode to replace the old loop head
4138 Node *l = new LoopNode( _head->in(1), _head->in(2) );
4139 l = igvn.register_new_node_with_optimizer(l, _head);
4140 phase->set_created_loop_node();
4141 // Go ahead and replace _head
4142 phase->_igvn.replace_node( _head, l );
4143 _head = l;
4144 phase->set_loop(_head, this);
4145 }
4146
4147 // Now recursively beautify nested loops
4148 if( _child ) result |= _child->beautify_loops( phase );
4149 if( _next ) result |= _next ->beautify_loops( phase );
4150 return result;
4151 }
4152
4153 //------------------------------allpaths_check_safepts----------------------------
4154 // Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
4155 // safepoint encountered. Helper for check_safepts.
4156 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
4157 assert(stack.size() == 0, "empty stack");
4158 stack.push(_head);
4159 visited.clear();
4160 visited.set(_head->_idx);
4161 while (stack.size() > 0) {
4162 Node* n = stack.pop();
4163 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
4164 // Terminate this path
4165 } else if (n->Opcode() == Op_SafePoint) {
4166 if (_phase->get_loop(n) != this) {
4167 if (_required_safept == nullptr) _required_safept = new Node_List();
4168 // save the first we run into on that path: closest to the tail if the head has a single backedge
4169 _required_safept->push(n);
4170 }
4171 // Terminate this path
4172 } else {
4173 uint start = n->is_Region() ? 1 : 0;
4174 uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
4175 for (uint i = start; i < end; i++) {
4176 Node* in = n->in(i);
4177 assert(in->is_CFG(), "must be");
4178 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
4179 stack.push(in);
4180 }
4181 }
4182 }
4183 }
4184 }
4185
4186 //------------------------------check_safepts----------------------------
4187 // Given dominators, try to find loops with calls that must always be
4188 // executed (call dominates loop tail). These loops do not need non-call
4189 // safepoints (ncsfpt).
4190 //
4191 // A complication is that a safepoint in a inner loop may be needed
4192 // by an outer loop. In the following, the inner loop sees it has a
4193 // call (block 3) on every path from the head (block 2) to the
4194 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
4195 // in block 2, _but_ this leaves the outer loop without a safepoint.
4196 //
4197 // entry 0
4198 // |
4199 // v
4200 // outer 1,2,4 +-> 1
4201 // | \
4202 // | v
4203 // inner 2,3 | 2 <---+ ncsfpt in 2
4204 // | / \ |
4205 // | v v |
4206 // | 4 3 | call in 3
4207 // |_/ \ \_|
4208 // |
4209 // v
4210 // exit 5
4211 //
4212 // This method maintains a list (_required_safept) of ncsfpts that must
4213 // be protected for each loop. It only marks ncsfpts for prevervation,
4214 // and does not actually delete any of them.
4215 //
4216 // If some other method needs to delete a ncsfpt later, it will make sure
4217 // the ncsfpt is not in the list of all outer loops of the current loop.
4218 // See `PhaseIdealLoop::is_deleteable_safept`.
4219 //
4220 // The insights into the problem:
4221 // A) Counted loops are okay (i.e. do not need to preserve ncsfpts),
4222 // they will be handled in `IdealLoopTree::counted_loop`
4223 // B) Innermost loops are okay because there's no inner loops that can
4224 // delete their ncsfpts. Only outer loops need to mark safepoints for
4225 // protection, because only loops further in can accidentally delete
4226 // their ncsfpts
4227 // C) If an outer loop has a call that's guaranteed to execute (on the
4228 // idom-path), then that loop is okay. Because the call will always
4229 // perform a safepoint poll, regardless of what safepoints are deleted
4230 // from its inner loops
4231 // D) Similarly, if an outer loop has a ncsfpt on the idom-path that isn't
4232 // inside any nested loop, then that loop is okay
4233 // E) Otherwise, if an outer loop's ncsfpt on the idom-path is nested in
4234 // an inner loop, we need to prevent the inner loop from deleting it
4235 //
4236 // There are two analyses:
4237 // 1) The first, and cheaper one, scans the loop body from
4238 // tail to head following the idom (immediate dominator)
4239 // chain, looking for the cases (C,D,E) above.
4240 // Since inner loops are scanned before outer loops, there is summary
4241 // information about inner loops. Inner loops can be skipped over
4242 // when the tail of an inner loop is encountered.
4243 //
4244 // 2) The second, invoked if the first fails to find a call or ncsfpt on
4245 // the idom path (which is rare), scans all predecessor control paths
4246 // from the tail to the head, terminating a path when a call or sfpt
4247 // is encountered, to find the ncsfpt's that are closest to the tail.
4248 //
4249 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
4250 // Bottom up traversal
4251 IdealLoopTree* ch = _child;
4252 if (_child) _child->check_safepts(visited, stack);
4253 if (_next) _next ->check_safepts(visited, stack);
4254
4255 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {
4256 bool has_call = false; // call on dom-path
4257 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
4258 Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
4259 if (!_irreducible) {
4260 // Scan the dom-path nodes from tail to head
4261 for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
4262 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
4263 has_call = true;
4264 _has_sfpt = 1; // Then no need for a safept!
4265 break;
4266 } else if (n->Opcode() == Op_SafePoint) {
4267 if (_phase->get_loop(n) == this) {
4268 // We found a local ncsfpt.
4269 // Continue searching for a call that is guaranteed to be a safepoint.
4270 has_local_ncsfpt = true;
4271 } else if (nonlocal_ncsfpt == nullptr) {
4272 nonlocal_ncsfpt = n; // save the one closest to the tail
4273 }
4274 } else {
4275 IdealLoopTree* nlpt = _phase->get_loop(n);
4276 if (this != nlpt) {
4277 // If at an inner loop tail, see if the inner loop has already
4278 // recorded seeing a call on the dom-path (and stop.) If not,
4279 // jump to the head of the inner loop.
4280 assert(is_member(nlpt), "nested loop");
4281 Node* tail = nlpt->_tail;
4282 if (tail->in(0)->is_If()) tail = tail->in(0);
4283 if (n == tail) {
4284 // If inner loop has call on dom-path, so does outer loop
4285 if (nlpt->_has_sfpt) {
4286 has_call = true;
4287 _has_sfpt = 1;
4288 break;
4289 }
4290 // Skip to head of inner loop
4291 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
4292 n = nlpt->_head;
4293 if (_head == n) {
4294 // this and nlpt (inner loop) have the same loop head. This should not happen because
4295 // during beautify_loops we call merge_many_backedges. However, infinite loops may not
4296 // have been attached to the loop-tree during build_loop_tree before beautify_loops,
4297 // but then attached in the build_loop_tree afterwards, and so still have unmerged
4298 // backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
4299 // since we have reached the loop head of this.
4300 assert(_head->as_Region()->is_in_infinite_subgraph(),
4301 "only expect unmerged backedges in infinite loops");
4302 break;
4303 }
4304 }
4305 }
4306 }
4307 }
4308 }
4309 // Record safept's that this loop needs preserved when an
4310 // inner loop attempts to delete it's safepoints.
4311 if (_child != nullptr && !has_call && !has_local_ncsfpt) {
4312 if (nonlocal_ncsfpt != nullptr) {
4313 if (_required_safept == nullptr) _required_safept = new Node_List();
4314 _required_safept->push(nonlocal_ncsfpt);
4315 } else {
4316 // Failed to find a suitable safept on the dom-path. Now use
4317 // an all paths walk from tail to head, looking for safepoints to preserve.
4318 allpaths_check_safepts(visited, stack);
4319 }
4320 }
4321 }
4322 }
4323
4324 //---------------------------is_deleteable_safept----------------------------
4325 // Is safept not required by an outer loop?
4326 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
4327 assert(sfpt->Opcode() == Op_SafePoint, "");
4328 IdealLoopTree* lp = get_loop(sfpt)->_parent;
4329 while (lp != nullptr) {
4330 Node_List* sfpts = lp->_required_safept;
4331 if (sfpts != nullptr) {
4332 for (uint i = 0; i < sfpts->size(); i++) {
4333 if (sfpt == sfpts->at(i))
4334 return false;
4335 }
4336 }
4337 lp = lp->_parent;
4338 }
4339 return true;
4340 }
4341
4342 //---------------------------replace_parallel_iv-------------------------------
4343 // Replace parallel induction variable (parallel to trip counter)
4344 // This optimization looks for patterns similar to:
4345 //
4346 // int a = init2;
4347 // for (int iv = init; iv < limit; iv += stride_con) {
4348 // a += stride_con2;
4349 // }
4350 //
4351 // and transforms it to:
4352 //
4353 // int iv2 = init2
4354 // int iv = init
4355 // loop:
4356 // if (iv >= limit) goto exit
4357 // iv += stride_con
4358 // iv2 = init2 + (iv - init) * (stride_con2 / stride_con)
4359 // goto loop
4360 // exit:
4361 // ...
4362 //
4363 // Such transformation introduces more optimization opportunities. In this
4364 // particular example, the loop can be eliminated entirely given that
4365 // `stride_con2 / stride_con` is exact (i.e., no remainder). Checks are in
4366 // place to only perform this optimization if such a division is exact. This
4367 // example will be transformed into its semantic equivalence:
4368 //
4369 // int iv2 = (iv * stride_con2 / stride_con) + (init2 - (init * stride_con2 / stride_con))
4370 //
4371 // which corresponds to the structure of transformed subgraph.
4372 //
4373 // However, if there is a mismatch between types of the loop and the parallel
4374 // induction variable (e.g., a long-typed IV in an int-typed loop), type
4375 // conversions are required:
4376 //
4377 // long iv2 = ((long) iv * stride_con2 / stride_con) + (init2 - ((long) init * stride_con2 / stride_con))
4378 //
4379 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
4380 assert(loop->_head->is_CountedLoop(), "");
4381 CountedLoopNode *cl = loop->_head->as_CountedLoop();
4382 if (!cl->is_valid_counted_loop(T_INT)) {
4383 return; // skip malformed counted loop
4384 }
4385 Node *incr = cl->incr();
4386 if (incr == nullptr) {
4387 return; // Dead loop?
4388 }
4389 Node *init = cl->init_trip();
4390 Node *phi = cl->phi();
4391 jlong stride_con = cl->stride_con();
4392
4393 // Visit all children, looking for Phis
4394 for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
4395 Node *out = cl->out(i);
4396 // Look for other phis (secondary IVs). Skip dead ones
4397 if (!out->is_Phi() || out == phi || !has_node(out)) {
4398 continue;
4399 }
4400
4401 PhiNode* phi2 = out->as_Phi();
4402 Node* incr2 = phi2->in(LoopNode::LoopBackControl);
4403 // Look for induction variables of the form: X += constant
4404 if (phi2->region() != loop->_head ||
4405 incr2->req() != 3 ||
4406 incr2->in(1)->uncast() != phi2 ||
4407 incr2 == incr ||
4408 (incr2->Opcode() != Op_AddI && incr2->Opcode() != Op_AddL) ||
4409 !incr2->in(2)->is_Con()) {
4410 continue;
4411 }
4412
4413 if (incr2->in(1)->is_ConstraintCast() &&
4414 !(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {
4415 // Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
4416 continue;
4417 }
4418 // Check for parallel induction variable (parallel to trip counter)
4419 // via an affine function. In particular, count-down loops with
4420 // count-up array indices are common. We only RCE references off
4421 // the trip-counter, so we need to convert all these to trip-counter
4422 // expressions.
4423 Node* init2 = phi2->in(LoopNode::EntryControl);
4424
4425 // Determine the basic type of the stride constant (and the iv being incremented).
4426 BasicType stride_con2_bt = incr2->Opcode() == Op_AddI ? T_INT : T_LONG;
4427 jlong stride_con2 = incr2->in(2)->get_integer_as_long(stride_con2_bt);
4428
4429 // The ratio of the two strides cannot be represented as an int
4430 // if stride_con2 is min_jint (or min_jlong, respectively) and
4431 // stride_con is -1.
4432 if (stride_con2 == min_signed_integer(stride_con2_bt) && stride_con == -1) {
4433 continue;
4434 }
4435
4436 // The general case here gets a little tricky. We want to find the
4437 // GCD of all possible parallel IV's and make a new IV using this
4438 // GCD for the loop. Then all possible IVs are simple multiples of
4439 // the GCD. In practice, this will cover very few extra loops.
4440 // Instead we require 'stride_con2' to be a multiple of 'stride_con',
4441 // where +/-1 is the common case, but other integer multiples are
4442 // also easy to handle.
4443 jlong ratio_con = stride_con2 / stride_con;
4444
4445 if ((ratio_con * stride_con) != stride_con2) { // Check for exact (no remainder)
4446 continue;
4447 }
4448
4449 #ifndef PRODUCT
4450 if (TraceLoopOpts) {
4451 tty->print("Parallel IV: %d ", phi2->_idx);
4452 loop->dump_head();
4453 }
4454 #endif
4455
4456 // Convert to using the trip counter. The parallel induction
4457 // variable differs from the trip counter by a loop-invariant
4458 // amount, the difference between their respective initial values.
4459 // It is scaled by the 'ratio_con'.
4460 Node* ratio = integercon(ratio_con, stride_con2_bt);
4461
4462 Node* init_converted = insert_convert_node_if_needed(stride_con2_bt, init);
4463 Node* phi_converted = insert_convert_node_if_needed(stride_con2_bt, phi);
4464
4465 Node* ratio_init = MulNode::make(init_converted, ratio, stride_con2_bt);
4466 _igvn.register_new_node_with_optimizer(ratio_init, init_converted);
4467 set_early_ctrl(ratio_init, false);
4468
4469 Node* diff = SubNode::make(init2, ratio_init, stride_con2_bt);
4470 _igvn.register_new_node_with_optimizer(diff, init2);
4471 set_early_ctrl(diff, false);
4472
4473 Node* ratio_idx = MulNode::make(phi_converted, ratio, stride_con2_bt);
4474 _igvn.register_new_node_with_optimizer(ratio_idx, phi_converted);
4475 set_ctrl(ratio_idx, cl);
4476
4477 Node* add = AddNode::make(ratio_idx, diff, stride_con2_bt);
4478 _igvn.register_new_node_with_optimizer(add);
4479 set_ctrl(add, cl);
4480
4481 _igvn.replace_node( phi2, add );
4482 // Sometimes an induction variable is unused
4483 if (add->outcnt() == 0) {
4484 _igvn.remove_dead_node(add);
4485 }
4486 --i; // deleted this phi; rescan starting with next position
4487 }
4488 }
4489
4490 Node* PhaseIdealLoop::insert_convert_node_if_needed(BasicType target, Node* input) {
4491 BasicType source = _igvn.type(input)->basic_type();
4492 if (source == target) {
4493 return input;
4494 }
4495
4496 Node* converted = ConvertNode::create_convert(source, target, input);
4497 _igvn.register_new_node_with_optimizer(converted, input);
4498 set_early_ctrl(converted, false);
4499
4500 return converted;
4501 }
4502
4503 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
4504 Node* keep = nullptr;
4505 if (keep_one) {
4506 // Look for a safepoint on the idom-path.
4507 for (Node* i = tail(); i != _head; i = phase->idom(i)) {
4508 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
4509 keep = i;
4510 break; // Found one
4511 }
4512 }
4513 }
4514
4515 // Don't remove any safepoints if it is requested to keep a single safepoint and
4516 // no safepoint was found on idom-path. It is not safe to remove any safepoint
4517 // in this case since there's no safepoint dominating all paths in the loop body.
4518 bool prune = !keep_one || keep != nullptr;
4519
4520 // Delete other safepoints in this loop.
4521 Node_List* sfpts = _safepts;
4522 if (prune && sfpts != nullptr) {
4523 assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4524 for (uint i = 0; i < sfpts->size(); i++) {
4525 Node* n = sfpts->at(i);
4526 assert(phase->get_loop(n) == this, "");
4527 if (n != keep && phase->is_deleteable_safept(n)) {
4528 phase->replace_node_and_forward_ctrl(n, n->in(TypeFunc::Control));
4529 }
4530 }
4531 }
4532 }
4533
4534 //------------------------------counted_loop-----------------------------------
4535 // Convert to counted loops where possible
4536 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
4537
4538 // For grins, set the inner-loop flag here
4539 if (!_child) {
4540 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4541 }
4542
4543 IdealLoopTree* loop = this;
4544 if (_head->is_CountedLoop() ||
4545 phase->is_counted_loop(_head, loop, T_INT)) {
4546
4547 if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {
4548 // Indicate we do not need a safepoint here
4549 _has_sfpt = 1;
4550 }
4551
4552 // Remove safepoints
4553 bool keep_one_sfpt = !(_has_call || _has_sfpt);
4554 remove_safepoints(phase, keep_one_sfpt);
4555
4556 // Look for induction variables
4557 phase->replace_parallel_iv(this);
4558 } else if (_head->is_LongCountedLoop() ||
4559 phase->is_counted_loop(_head, loop, T_LONG)) {
4560 remove_safepoints(phase, true);
4561 } else {
4562 assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4563 if (_parent != nullptr && !_irreducible) {
4564 // Not a counted loop. Keep one safepoint.
4565 bool keep_one_sfpt = true;
4566 remove_safepoints(phase, keep_one_sfpt);
4567 }
4568 }
4569
4570 // Recursively
4571 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4572 assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4573 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4574 if (loop->_next) loop->_next ->counted_loop(phase);
4575 }
4576
4577
4578 // The Estimated Loop Clone Size:
4579 // CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4580 // where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4581 // respectively, used to ensure that the node usage estimates made are on the
4582 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
4583 // possible additional/excessive nodes generated due to data and control flow
4584 // merging, for edges reaching outside the loop.
4585 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
4586
4587 precond(0 < factor && factor < 16);
4588
4589 uint const bc = 13;
4590 uint const cc = 17;
4591 uint const sz = _body.size() + (_body.size() + 7) / 2;
4592 uint estimate = factor * (sz + bc) + cc;
4593
4594 assert((estimate - cc) / factor == sz + bc, "overflow");
4595
4596 return estimate + est_loop_flow_merge_sz();
4597 }
4598
4599 // The Estimated Loop (full-) Unroll Size:
4600 // UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4601 // where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4602 // node usage estimates made are on the safe side, for the most part. This is
4603 // a "light" version of the loop clone size calculation (above), based on the
4604 // assumption that most of the loop-construct overhead will be unraveled when
4605 // (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4606 // including an overflow check and returning UINT_MAX in case of an overflow.
4607 uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {
4608
4609 precond(factor > 0);
4610
4611 // Take into account that after unroll conjoined heads and tails will fold.
4612 uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4613 uint const cc = 7;
4614 uint const sz = b0 + (b0 + 15) / 16;
4615 uint estimate = factor * sz + cc;
4616
4617 if ((estimate - cc) / factor != sz) {
4618 return UINT_MAX;
4619 }
4620
4621 return estimate + est_loop_flow_merge_sz();
4622 }
4623
4624 // Estimate the growth effect (in nodes) of merging control and data flow when
4625 // cloning a loop body, based on the amount of control and data flow reaching
4626 // outside of the (current) loop body.
4627 uint IdealLoopTree::est_loop_flow_merge_sz() const {
4628
4629 uint ctrl_edge_out_cnt = 0;
4630 uint data_edge_out_cnt = 0;
4631
4632 for (uint i = 0; i < _body.size(); i++) {
4633 Node* node = _body.at(i);
4634 uint outcnt = node->outcnt();
4635
4636 for (uint k = 0; k < outcnt; k++) {
4637 Node* out = node->raw_out(k);
4638 if (out == nullptr) continue;
4639 if (out->is_CFG()) {
4640 if (!is_member(_phase->get_loop(out))) {
4641 ctrl_edge_out_cnt++;
4642 }
4643 } else if (_phase->has_ctrl(out)) {
4644 Node* ctrl = _phase->get_ctrl(out);
4645 assert(ctrl != nullptr, "must be");
4646 assert(ctrl->is_CFG(), "must be");
4647 if (!is_member(_phase->get_loop(ctrl))) {
4648 data_edge_out_cnt++;
4649 }
4650 }
4651 }
4652 }
4653 // Use data and control count (x2.0) in estimate iff both are > 0. This is
4654 // a rather pessimistic estimate for the most part, in particular for some
4655 // complex loops, but still not enough to capture all loops.
4656 if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
4657 return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4658 }
4659 return 0;
4660 }
4661
4662 #ifndef PRODUCT
4663 //------------------------------dump_head--------------------------------------
4664 // Dump 1 liner for loop header info
4665 void IdealLoopTree::dump_head() {
4666 tty->sp(2 * _nest);
4667 tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);
4668 if (_irreducible) tty->print(" IRREDUCIBLE");
4669 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4670 : _head->in(LoopNode::EntryControl);
4671 const Predicates predicates(entry);
4672 if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {
4673 tty->print(" limit_check");
4674 }
4675 if (predicates.short_running_long_loop_predicate_block()->is_non_empty()) {
4676 tty->print(" short_running");
4677 }
4678 if (UseLoopPredicate) {
4679 if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {
4680 tty->print(" profile_predicated");
4681 }
4682 if (predicates.loop_predicate_block()->is_non_empty()) {
4683 tty->print(" predicated");
4684 }
4685 }
4686 if (UseAutoVectorizationPredicate && predicates.auto_vectorization_check_block()->is_non_empty()) {
4687 tty->print(" auto_vectorization_check_predicate");
4688 }
4689 if (_head->is_CountedLoop()) {
4690 CountedLoopNode *cl = _head->as_CountedLoop();
4691 tty->print(" counted");
4692
4693 Node* init_n = cl->init_trip();
4694 if (init_n != nullptr && init_n->is_Con())
4695 tty->print(" [%d,", cl->init_trip()->get_int());
4696 else
4697 tty->print(" [int,");
4698 Node* limit_n = cl->limit();
4699 if (limit_n != nullptr && limit_n->is_Con())
4700 tty->print("%d),", cl->limit()->get_int());
4701 else
4702 tty->print("int),");
4703 int stride_con = cl->stride_con();
4704 if (stride_con > 0) tty->print("+");
4705 tty->print("%d", stride_con);
4706
4707 tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
4708
4709 if (cl->is_pre_loop ()) tty->print(" pre" );
4710 if (cl->is_main_loop()) tty->print(" main");
4711 if (cl->is_post_loop()) tty->print(" post");
4712 if (cl->is_vectorized_loop()) tty->print(" vector");
4713 if (range_checks_present()) tty->print(" rc ");
4714 if (cl->is_multiversion_fast_loop()) { tty->print(" multiversion_fast"); }
4715 if (cl->is_multiversion_slow_loop()) { tty->print(" multiversion_slow"); }
4716 if (cl->is_multiversion_delayed_slow_loop()) { tty->print(" multiversion_delayed_slow"); }
4717 }
4718 if (_has_call) tty->print(" has_call");
4719 if (_has_sfpt) tty->print(" has_sfpt");
4720 if (_rce_candidate) tty->print(" rce");
4721 if (_safepts != nullptr && _safepts->size() > 0) {
4722 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
4723 }
4724 if (_required_safept != nullptr && _required_safept->size() > 0) {
4725 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
4726 }
4727 if (Verbose) {
4728 tty->print(" body={"); _body.dump_simple(); tty->print(" }");
4729 }
4730 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
4731 tty->print(" strip_mined");
4732 }
4733 tty->cr();
4734 }
4735
4736 //------------------------------dump-------------------------------------------
4737 // Dump loops by loop tree
4738 void IdealLoopTree::dump() {
4739 dump_head();
4740 if (_child) _child->dump();
4741 if (_next) _next ->dump();
4742 }
4743
4744 #endif
4745
4746 static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
4747 if (loop == root) {
4748 if (loop->_child != nullptr) {
4749 log->begin_head("loop_tree");
4750 log->end_head();
4751 log_loop_tree_helper(root, loop->_child, log);
4752 log->tail("loop_tree");
4753 assert(loop->_next == nullptr, "what?");
4754 }
4755 } else if (loop != nullptr) {
4756 Node* head = loop->_head;
4757 log->begin_head("loop");
4758 log->print(" idx='%d' ", head->_idx);
4759 if (loop->_irreducible) log->print("irreducible='1' ");
4760 if (head->is_Loop()) {
4761 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
4762 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
4763 } else if (head->is_CountedLoop()) {
4764 CountedLoopNode* cl = head->as_CountedLoop();
4765 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());
4766 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
4767 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());
4768 }
4769 log->end_head();
4770 log_loop_tree_helper(root, loop->_child, log);
4771 log->tail("loop");
4772 log_loop_tree_helper(root, loop->_next, log);
4773 }
4774 }
4775
4776 void PhaseIdealLoop::log_loop_tree() {
4777 if (C->log() != nullptr) {
4778 log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4779 }
4780 }
4781
4782 // Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4783 // predicates will be removed during the next round of IGVN.
4784 void PhaseIdealLoop::eliminate_useless_predicates() const {
4785 if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {
4786 return; // No predicates left.
4787 }
4788
4789 EliminateUselessPredicates eliminate_useless_predicates(_igvn, _ltree_root);
4790 eliminate_useless_predicates.eliminate();
4791 }
4792
4793 // If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4794 void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {
4795 if (_zero_trip_guard_opaque_nodes.size() == 0) {
4796 return;
4797 }
4798 Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4799 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4800 IdealLoopTree* lpt = iter.current();
4801 if (lpt->_child == nullptr && lpt->is_counted()) {
4802 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4803 Node* opaque = head->is_canonical_loop_entry();
4804 if (opaque != nullptr) {
4805 useful_zero_trip_guard_opaques_nodes.push(opaque);
4806 #ifdef ASSERT
4807 // See PhaseIdealLoop::do_unroll
4808 // This property is required in do_unroll, but it may not hold after cloning a loop.
4809 // In such a case, we bail out from unrolling, and rely on IGVN to clean up the graph.
4810 // We are here before loop cloning (before iteration_split), so if this property
4811 // does not hold, it must come from the previous round of loop optimizations, meaning
4812 // that IGVN failed to clean it: we will catch that here.
4813 // On the other hand, if this assert passes, a bailout in do_unroll means that
4814 // this property was broken in the current round of loop optimization (between here
4815 // and do_unroll), so we give a chance to IGVN to make the property true again.
4816 if (head->is_main_loop()) {
4817 assert(opaque->outcnt() == 1, "opaque node should not be shared");
4818 assert(opaque->in(1) == head->limit(), "After IGVN cleanup, input of opaque node must be the limit.");
4819 }
4820 if (head->is_post_loop()) {
4821 assert(opaque->outcnt() == 1, "opaque node should not be shared");
4822 }
4823 #endif
4824 }
4825 }
4826 }
4827 for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {
4828 OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4829 DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4830 if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {
4831 IfNode* iff = opaque->if_node();
4832 IdealLoopTree* loop = get_loop(iff);
4833 while (loop != _ltree_root && loop != nullptr) {
4834 loop = loop->_parent;
4835 }
4836 if (loop == nullptr) {
4837 // unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4838 // We can't tell if the opaque node is useful or not
4839 assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4840 } else {
4841 assert(guarded_loop == nullptr, "");
4842 this->_igvn.replace_node(opaque, opaque->in(1));
4843 }
4844 } else {
4845 assert(guarded_loop != nullptr, "");
4846 }
4847 }
4848 }
4849
4850 void PhaseIdealLoop::eliminate_useless_multiversion_if() {
4851 if (_multiversion_opaque_nodes.size() == 0) {
4852 return;
4853 }
4854
4855 ResourceMark rm;
4856 Unique_Node_List useful_multiversioning_opaque_nodes;
4857
4858 // The OpaqueMultiversioning is only used from the fast main loop in AutoVectorization, to add
4859 // speculative runtime-checks to the multiversion_if. Thus, a OpaqueMultiversioning is only
4860 // useful if it can be found from a fast main loop. If it can not be found from a fast main loop,
4861 // then we cannot ever use that multiversion_if to add more speculative runtime-checks, and hence
4862 // it is useless. If it is still in delayed mode, i.e. has not yet had any runtime-checks added,
4863 // then we can let it constant fold towards the fast loop.
4864 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
4865 IdealLoopTree* lpt = iter.current();
4866 if (lpt->_child == nullptr && lpt->is_counted()) {
4867 CountedLoopNode* head = lpt->_head->as_CountedLoop();
4868 if (head->is_main_loop() && head->is_multiversion_fast_loop()) {
4869 // There are fast_loop pre/main/post loops, but the finding traversal starts at the main
4870 // loop, and traverses via the fast pre loop to the multiversion_if.
4871 IfNode* multiversion_if = head->find_multiversion_if_from_multiversion_fast_main_loop();
4872 if (multiversion_if != nullptr) {
4873 useful_multiversioning_opaque_nodes.push(multiversion_if->in(1)->as_OpaqueMultiversioning());
4874 } else {
4875 // We could not find the multiversion_if, and would never find it again. Remove the
4876 // multiversion marking for consistency.
4877 head->set_no_multiversion();
4878 }
4879 }
4880 }
4881 }
4882
4883 for (uint i = 0; i < _multiversion_opaque_nodes.size(); i++) {
4884 OpaqueMultiversioningNode* opaque = _multiversion_opaque_nodes.at(i)->as_OpaqueMultiversioning();
4885 if (!useful_multiversioning_opaque_nodes.member(opaque)) {
4886 if (opaque->is_delayed_slow_loop()) {
4887 // We cannot hack the node directly, otherwise the slow_loop will complain that it cannot
4888 // find the multiversioning opaque node. Instead, we mark the opaque node as useless, and
4889 // it can be constant folded during IGVN.
4890 opaque->mark_useless(_igvn);
4891 }
4892 }
4893 }
4894 }
4895
4896 //------------------------process_expensive_nodes-----------------------------
4897 // Expensive nodes have their control input set to prevent the GVN
4898 // from commoning them and as a result forcing the resulting node to
4899 // be in a more frequent path. Use CFG information here, to change the
4900 // control inputs so that some expensive nodes can be commoned while
4901 // not executed more frequently.
4902 bool PhaseIdealLoop::process_expensive_nodes() {
4903 assert(OptimizeExpensiveOps, "optimization off?");
4904
4905 // Sort nodes to bring similar nodes together
4906 C->sort_expensive_nodes();
4907
4908 bool progress = false;
4909
4910 for (int i = 0; i < C->expensive_count(); ) {
4911 Node* n = C->expensive_node(i);
4912 int start = i;
4913 // Find nodes similar to n
4914 i++;
4915 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4916 int end = i;
4917 // And compare them two by two
4918 for (int j = start; j < end; j++) {
4919 Node* n1 = C->expensive_node(j);
4920 if (is_node_unreachable(n1)) {
4921 continue;
4922 }
4923 for (int k = j+1; k < end; k++) {
4924 Node* n2 = C->expensive_node(k);
4925 if (is_node_unreachable(n2)) {
4926 continue;
4927 }
4928
4929 assert(n1 != n2, "should be pair of nodes");
4930
4931 Node* c1 = n1->in(0);
4932 Node* c2 = n2->in(0);
4933
4934 Node* parent_c1 = c1;
4935 Node* parent_c2 = c2;
4936
4937 // The call to get_early_ctrl_for_expensive() moves the
4938 // expensive nodes up but stops at loops that are in a if
4939 // branch. See whether we can exit the loop and move above the
4940 // If.
4941 if (c1->is_Loop()) {
4942 parent_c1 = c1->in(1);
4943 }
4944 if (c2->is_Loop()) {
4945 parent_c2 = c2->in(1);
4946 }
4947
4948 if (parent_c1 == parent_c2) {
4949 _igvn._worklist.push(n1);
4950 _igvn._worklist.push(n2);
4951 continue;
4952 }
4953
4954 // Look for identical expensive node up the dominator chain.
4955 if (is_dominator(c1, c2)) {
4956 c2 = c1;
4957 } else if (is_dominator(c2, c1)) {
4958 c1 = c2;
4959 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4960 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
4961 // Both branches have the same expensive node so move it up
4962 // before the if.
4963 c1 = c2 = idom(parent_c1->in(0));
4964 }
4965 // Do the actual moves
4966 if (n1->in(0) != c1) {
4967 _igvn.replace_input_of(n1, 0, c1);
4968 progress = true;
4969 }
4970 if (n2->in(0) != c2) {
4971 _igvn.replace_input_of(n2, 0, c2);
4972 progress = true;
4973 }
4974 }
4975 }
4976 }
4977
4978 return progress;
4979 }
4980
4981 //=============================================================================
4982 //----------------------------build_and_optimize-------------------------------
4983 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4984 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4985 void PhaseIdealLoop::build_and_optimize() {
4986 assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4987
4988 bool do_split_ifs = (_mode == LoopOptsDefault);
4989 bool skip_loop_opts = (_mode == LoopOptsNone);
4990 bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
4991
4992
4993 bool old_progress = C->major_progress();
4994 uint orig_worklist_size = _igvn._worklist.size();
4995
4996 // Reset major-progress flag for the driver's heuristics
4997 C->clear_major_progress();
4998
4999 #ifndef PRODUCT
5000 // Capture for later assert
5001 uint unique = C->unique();
5002 _loop_invokes++;
5003 _loop_work += unique;
5004 #endif
5005
5006 // True if the method has at least 1 irreducible loop
5007 _has_irreducible_loops = false;
5008
5009 _created_loop_node = false;
5010
5011 VectorSet visited;
5012 // Pre-grow the mapping from Nodes to IdealLoopTrees.
5013 _loop_or_ctrl.map(C->unique(), nullptr);
5014 memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
5015
5016 // Pre-build the top-level outermost loop tree entry
5017 _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
5018 // Do not need a safepoint at the top level
5019 _ltree_root->_has_sfpt = 1;
5020
5021 // Initialize Dominators.
5022 // Checked in clone_loop_predicate() during beautify_loops().
5023 _idom_size = 0;
5024 _idom = nullptr;
5025 _dom_depth = nullptr;
5026 _dom_stk = nullptr;
5027
5028 // Empty pre-order array
5029 allocate_preorders();
5030
5031 // Build a loop tree on the fly. Build a mapping from CFG nodes to
5032 // IdealLoopTree entries. Data nodes are NOT walked.
5033 build_loop_tree();
5034 // Check for bailout, and return
5035 if (C->failing()) {
5036 return;
5037 }
5038
5039 // Verify that the has_loops() flag set at parse time is consistent with the just built loop tree. When the back edge
5040 // is an exception edge, parsing doesn't set has_loops().
5041 assert(_ltree_root->_child == nullptr || C->has_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
5042 // No loops after all
5043 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
5044
5045 // There should always be an outer loop containing the Root and Return nodes.
5046 // If not, we have a degenerate empty program. Bail out in this case.
5047 if (!has_node(C->root())) {
5048 if (!_verify_only) {
5049 C->clear_major_progress();
5050 assert(false, "empty program detected during loop optimization");
5051 C->record_method_not_compilable("empty program detected during loop optimization");
5052 }
5053 return;
5054 }
5055
5056 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5057 // Nothing to do, so get out
5058 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
5059 !_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode);
5060 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
5061 bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode);
5062 if (stop_early && !do_expensive_nodes) {
5063 return;
5064 }
5065
5066 // Set loop nesting depth
5067 _ltree_root->set_nest( 0 );
5068
5069 // Split shared headers and insert loop landing pads.
5070 // Do not bother doing this on the Root loop of course.
5071 if( !_verify_me && !_verify_only && _ltree_root->_child ) {
5072 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
5073 if( _ltree_root->_child->beautify_loops( this ) ) {
5074 // Re-build loop tree!
5075 _ltree_root->_child = nullptr;
5076 _loop_or_ctrl.clear();
5077 reallocate_preorders();
5078 build_loop_tree();
5079 // Check for bailout, and return
5080 if (C->failing()) {
5081 return;
5082 }
5083 // Reset loop nesting depth
5084 _ltree_root->set_nest( 0 );
5085
5086 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
5087 }
5088 }
5089
5090 // Build Dominators for elision of null checks & loop finding.
5091 // Since nodes do not have a slot for immediate dominator, make
5092 // a persistent side array for that info indexed on node->_idx.
5093 _idom_size = C->unique();
5094 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
5095 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
5096 _dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
5097 memset( _dom_depth, 0, _idom_size * sizeof(uint) );
5098
5099 Dominators();
5100
5101 if (!_verify_only) {
5102 // As a side effect, Dominators removed any unreachable CFG paths
5103 // into RegionNodes. It doesn't do this test against Root, so
5104 // we do it here.
5105 for( uint i = 1; i < C->root()->req(); i++ ) {
5106 if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?
5107 _igvn.delete_input_of(C->root(), i);
5108 i--; // Rerun same iteration on compressed edges
5109 }
5110 }
5111
5112 // Given dominators, try to find inner loops with calls that must
5113 // always be executed (call dominates loop tail). These loops do
5114 // not need a separate safepoint.
5115 Node_List cisstack;
5116 _ltree_root->check_safepts(visited, cisstack);
5117 }
5118
5119 // Walk the DATA nodes and place into loops. Find earliest control
5120 // node. For CFG nodes, the _loop_or_ctrl array starts out and remains
5121 // holding the associated IdealLoopTree pointer. For DATA nodes, the
5122 // _loop_or_ctrl array holds the earliest legal controlling CFG node.
5123
5124 // Allocate stack with enough space to avoid frequent realloc
5125 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
5126 Node_Stack nstack(stack_size);
5127
5128 visited.clear();
5129 Node_List worklist;
5130 // Don't need C->root() on worklist since
5131 // it will be processed among C->top() inputs
5132 worklist.push(C->top());
5133 visited.set(C->top()->_idx); // Set C->top() as visited now
5134 build_loop_early( visited, worklist, nstack );
5135
5136 // Given early legal placement, try finding counted loops. This placement
5137 // is good enough to discover most loop invariants.
5138 if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
5139 _ltree_root->counted_loop( this );
5140 }
5141
5142 // Find latest loop placement. Find ideal loop placement.
5143 visited.clear();
5144 init_dom_lca_tags();
5145 // Need C->root() on worklist when processing outs
5146 worklist.push(C->root());
5147 NOT_PRODUCT( C->verify_graph_edges(); )
5148 worklist.push(C->top());
5149 build_loop_late( visited, worklist, nstack );
5150 if (C->failing()) { return; }
5151
5152 if (_verify_only) {
5153 C->restore_major_progress(old_progress);
5154 assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
5155 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
5156 return;
5157 }
5158
5159 // clear out the dead code after build_loop_late
5160 while (_deadlist.size()) {
5161 _igvn.remove_globally_dead_node(_deadlist.pop());
5162 }
5163
5164 eliminate_useless_zero_trip_guard();
5165 eliminate_useless_multiversion_if();
5166
5167 if (stop_early) {
5168 assert(do_expensive_nodes, "why are we here?");
5169 if (process_expensive_nodes()) {
5170 // If we made some progress when processing expensive nodes then
5171 // the IGVN may modify the graph in a way that will allow us to
5172 // make some more progress: we need to try processing expensive
5173 // nodes again.
5174 C->set_major_progress();
5175 }
5176 return;
5177 }
5178
5179 // Some parser-inserted loop predicates could never be used by loop
5180 // predication or they were moved away from loop during some optimizations.
5181 // For example, peeling. Eliminate them before next loop optimizations.
5182 eliminate_useless_predicates();
5183
5184 #ifndef PRODUCT
5185 C->verify_graph_edges();
5186 if (_verify_me) { // Nested verify pass?
5187 // Check to see if the verify _mode is broken
5188 assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
5189 return;
5190 }
5191 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
5192 if (TraceLoopOpts && C->has_loops()) {
5193 _ltree_root->dump();
5194 }
5195 #endif
5196
5197 if (skip_loop_opts) {
5198 C->restore_major_progress(old_progress);
5199 return;
5200 }
5201
5202 if (do_max_unroll) {
5203 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5204 IdealLoopTree* lpt = iter.current();
5205 if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {
5206 lpt->compute_trip_count(this, T_INT);
5207 if (!lpt->do_one_iteration_loop(this) &&
5208 !lpt->do_remove_empty_loop(this)) {
5209 AutoNodeBudget node_budget(this);
5210 if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
5211 lpt->policy_maximally_unroll(this)) {
5212 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
5213 do_maximally_unroll(lpt, worklist);
5214 }
5215 }
5216 }
5217 }
5218
5219 C->restore_major_progress(old_progress);
5220 return;
5221 }
5222
5223 if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) {
5224 return;
5225 }
5226
5227 if (ReassociateInvariants && !C->major_progress()) {
5228 // Reassociate invariants and prep for split_thru_phi
5229 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5230 IdealLoopTree* lpt = iter.current();
5231 if (!lpt->is_loop()) {
5232 continue;
5233 }
5234 Node* head = lpt->_head;
5235 if (!lpt->is_innermost()) continue;
5236
5237 // check for vectorized loops, any reassociation of invariants was already done
5238 if (head->is_CountedLoop() && head->as_CountedLoop()->is_unroll_only()) {
5239 continue;
5240 } else {
5241 AutoNodeBudget node_budget(this);
5242 lpt->reassociate_invariants(this);
5243 }
5244 // Because RCE opportunities can be masked by split_thru_phi,
5245 // look for RCE candidates and inhibit split_thru_phi
5246 // on just their loop-phi's for this pass of loop opts
5247 if (SplitIfBlocks && do_split_ifs &&
5248 head->is_BaseCountedLoop() &&
5249 head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
5250 (lpt->policy_range_check(this, true, T_LONG) ||
5251 (head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {
5252 lpt->_rce_candidate = 1; // = true
5253 }
5254 }
5255 }
5256
5257 // Check for aggressive application of split-if and other transforms
5258 // that require basic-block info (like cloning through Phi's)
5259 if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {
5260 visited.clear();
5261 split_if_with_blocks(visited, nstack);
5262 if (C->failing()) {
5263 return;
5264 }
5265 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
5266 }
5267
5268 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
5269 C->set_major_progress();
5270 }
5271
5272 // Perform loop predication before iteration splitting
5273 if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {
5274 _ltree_root->_child->loop_predication(this);
5275 }
5276
5277 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
5278 if (do_intrinsify_fill()) {
5279 C->set_major_progress();
5280 }
5281 }
5282
5283 // Perform iteration-splitting on inner loops. Split iterations to avoid
5284 // range checks or one-shot null checks.
5285
5286 // If split-if's didn't hack the graph too bad (no CFG changes)
5287 // then do loop opts.
5288 if (C->has_loops() && !C->major_progress()) {
5289 memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
5290 _ltree_root->_child->iteration_split( this, worklist );
5291 // No verify after peeling! GCM has hoisted code out of the loop.
5292 // After peeling, the hoisted code could sink inside the peeled area.
5293 // The peeling code does not try to recompute the best location for
5294 // all the code before the peeled area, so the verify pass will always
5295 // complain about it.
5296 }
5297
5298 // Check for bailout, and return
5299 if (C->failing()) {
5300 return;
5301 }
5302
5303 // Do verify graph edges in any case
5304 NOT_PRODUCT( C->verify_graph_edges(); );
5305
5306 if (!do_split_ifs) {
5307 // We saw major progress in Split-If to get here. We forced a
5308 // pass with unrolling and not split-if, however more split-if's
5309 // might make progress. If the unrolling didn't make progress
5310 // then the major-progress flag got cleared and we won't try
5311 // another round of Split-If. In particular the ever-common
5312 // instance-of/check-cast pattern requires at least 2 rounds of
5313 // Split-If to clear out.
5314 C->set_major_progress();
5315 }
5316
5317 // Repeat loop optimizations if new loops were seen
5318 if (created_loop_node()) {
5319 C->set_major_progress();
5320 }
5321
5322 // Auto-vectorize main-loop
5323 if (C->do_superword() && C->has_loops() && !C->major_progress()) {
5324 Compile::TracePhase tp(_t_autoVectorize);
5325
5326 // Shared data structures for all AutoVectorizations, to reduce allocations
5327 // of large arrays.
5328 VSharedData vshared;
5329 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
5330 IdealLoopTree* lpt = iter.current();
5331 AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
5332
5333 if (status == AutoVectorizeStatus::TriedAndFailed) {
5334 // We tried vectorization, but failed. From now on only unroll the loop.
5335 CountedLoopNode* cl = lpt->_head->as_CountedLoop();
5336 if (cl->has_passed_slp()) {
5337 C->set_major_progress();
5338 cl->set_notpassed_slp();
5339 cl->mark_do_unroll_only();
5340 }
5341 }
5342 }
5343 }
5344
5345 // Keep loop predicates and perform optimizations with them
5346 // until no more loop optimizations could be done.
5347 // After that switch predicates off and do more loop optimizations.
5348 if (!C->major_progress() && (C->parse_predicate_count() > 0)) {
5349 C->mark_parse_predicate_nodes_useless(_igvn);
5350 assert(C->parse_predicate_count() == 0, "should be zero now");
5351 if (TraceLoopOpts) {
5352 tty->print_cr("PredicatesOff");
5353 }
5354 C->set_major_progress();
5355 }
5356 }
5357
5358 #ifndef PRODUCT
5359 //------------------------------print_statistics-------------------------------
5360 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
5361 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
5362 volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
5363 volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
5364 volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
5365 void PhaseIdealLoop::print_statistics() {
5366 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);
5367 }
5368 #endif
5369
5370 #ifdef ASSERT
5371 // Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
5372 void PhaseIdealLoop::verify() const {
5373 ResourceMark rm;
5374 bool old_progress = C->major_progress();
5375 bool success = true;
5376
5377 PhaseIdealLoop phase_verify(_igvn, this);
5378 if (C->failing_internal()) {
5379 return;
5380 }
5381
5382 // Verify ctrl and idom of every node.
5383 success &= verify_idom_and_nodes(C->root(), &phase_verify);
5384
5385 // Verify loop-tree.
5386 success &= _ltree_root->verify_tree(phase_verify._ltree_root);
5387
5388 assert(success, "VerifyLoopOptimizations failed");
5389
5390 // Major progress was cleared by creating a verify version of PhaseIdealLoop.
5391 C->restore_major_progress(old_progress);
5392 }
5393
5394 // Perform a BFS starting at n, through all inputs.
5395 // Call verify_idom and verify_node on all nodes of BFS traversal.
5396 bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {
5397 Unique_Node_List worklist;
5398 worklist.push(root);
5399 bool success = true;
5400 for (uint i = 0; i < worklist.size(); i++) {
5401 Node* n = worklist.at(i);
5402 // process node
5403 success &= verify_idom(n, phase_verify);
5404 success &= verify_loop_ctrl(n, phase_verify);
5405 // visit inputs
5406 for (uint j = 0; j < n->req(); j++) {
5407 if (n->in(j) != nullptr) {
5408 worklist.push(n->in(j));
5409 }
5410 }
5411 }
5412 return success;
5413 }
5414
5415 // Verify dominator structure (IDOM).
5416 bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {
5417 // Verify IDOM for all CFG nodes (except root).
5418 if (!n->is_CFG() || n->is_Root()) {
5419 return true; // pass
5420 }
5421
5422 if (n->_idx >= _idom_size) {
5423 tty->print("CFG Node with no idom: ");
5424 n->dump();
5425 return false; // fail
5426 }
5427
5428 Node* id = idom_no_update(n);
5429 Node* id_verify = phase_verify->idom_no_update(n);
5430 if (id != id_verify) {
5431 tty->print("Mismatching idom for node: ");
5432 n->dump();
5433 tty->print(" We have idom: ");
5434 id->dump();
5435 tty->print(" Verify has idom: ");
5436 id_verify->dump();
5437 tty->cr();
5438 return false; // fail
5439 }
5440 return true; // pass
5441 }
5442
5443 // Verify "_loop_or_ctrl": control and loop membership.
5444 // (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5445 // (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5446 // (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5447 // (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5448 bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {
5449 const uint i = n->_idx;
5450 // The loop-tree was built from def to use (top-down).
5451 // The verification happens from use to def (bottom-up).
5452 // We may thus find nodes during verification that are not in the loop-tree.
5453 if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {
5454 if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {
5455 tty->print_cr("Was reachable in only one. this %d, verify %d.",
5456 _loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5457 n->dump();
5458 return false; // fail
5459 }
5460 // Not reachable for both.
5461 return true; // pass
5462 }
5463
5464 if (n->is_CFG() == has_ctrl(n)) {
5465 tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));
5466 n->dump();
5467 return false; // fail
5468 }
5469
5470 if (has_ctrl(n) != phase_verify->has_ctrl(n)) {
5471 tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));
5472 n->dump();
5473 return false; // fail
5474 } else if (has_ctrl(n)) {
5475 assert(phase_verify->has_ctrl(n), "sanity");
5476 // n is a data node.
5477 // Verify that its ctrl is the same.
5478
5479 // Broken part of VerifyLoopOptimizations (A)
5480 // Reason:
5481 // BUG, wrong control set for example in
5482 // PhaseIdealLoop::split_if_with_blocks
5483 // at "set_ctrl(x, new_ctrl);"
5484 /*
5485 if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5486 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
5487 tty->print("Mismatched control setting for: ");
5488 n->dump();
5489 if( fail++ > 10 ) return;
5490 Node *c = get_ctrl_no_update(n);
5491 tty->print("We have it as: ");
5492 if( c->in(0) ) c->dump();
5493 else tty->print_cr("N%d",c->_idx);
5494 tty->print("Verify thinks: ");
5495 if( loop_verify->has_ctrl(n) )
5496 loop_verify->get_ctrl_no_update(n)->dump();
5497 else
5498 loop_verify->get_loop_idx(n)->dump();
5499 tty->cr();
5500 }
5501 */
5502 return true; // pass
5503 } else {
5504 assert(!phase_verify->has_ctrl(n), "sanity");
5505 // n is a ctrl node.
5506 // Verify that not has_ctrl, and that get_loop_idx is the same.
5507
5508 // Broken part of VerifyLoopOptimizations (B)
5509 // Reason:
5510 // NeverBranch node for example is added to loop outside its scope.
5511 // Once we run build_loop_tree again, it is added to the correct loop.
5512 /*
5513 if (!C->major_progress()) {
5514 // Loop selection can be messed up if we did a major progress
5515 // operation, like split-if. Do not verify in that case.
5516 IdealLoopTree *us = get_loop_idx(n);
5517 IdealLoopTree *them = loop_verify->get_loop_idx(n);
5518 if( us->_head != them->_head || us->_tail != them->_tail ) {
5519 tty->print("Unequals loops for: ");
5520 n->dump();
5521 if( fail++ > 10 ) return;
5522 tty->print("We have it as: ");
5523 us->dump();
5524 tty->print("Verify thinks: ");
5525 them->dump();
5526 tty->cr();
5527 }
5528 }
5529 */
5530 return true; // pass
5531 }
5532 }
5533
5534 static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {
5535 assert(a != nullptr && b != nullptr, "must be");
5536 return a->_head->_idx - b->_head->_idx;
5537 }
5538
5539 GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {
5540 GrowableArray<IdealLoopTree*> children;
5541 IdealLoopTree* child = _child;
5542 while (child != nullptr) {
5543 assert(child->_parent == this, "all must be children of this");
5544 children.insert_sorted<compare_tree>(child);
5545 child = child->_next;
5546 }
5547 return children;
5548 }
5549
5550 // Verify that tree structures match. Because the CFG can change, siblings
5551 // within the loop tree can be reordered. We attempt to deal with that by
5552 // reordering the verify's loop tree if possible.
5553 bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {
5554 assert(_head == loop_verify->_head, "mismatched loop head");
5555 assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5556
5557 // Collect the children
5558 GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5559 GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5560
5561 bool success = true;
5562
5563 // Compare the two children lists
5564 for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {
5565 IdealLoopTree* child = nullptr;
5566 IdealLoopTree* child_verify = nullptr;
5567 // Read from both lists, if possible.
5568 if (i < children.length()) {
5569 child = children.at(i);
5570 }
5571 if (j < children_verify.length()) {
5572 child_verify = children_verify.at(j);
5573 }
5574 assert(child != nullptr || child_verify != nullptr, "must find at least one");
5575 if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {
5576 // We found two non-equal children. Select the smaller one.
5577 if (child->_head->_idx < child_verify->_head->_idx) {
5578 child_verify = nullptr;
5579 } else {
5580 child = nullptr;
5581 }
5582 }
5583 // Process the two children, or potentially log the failure if we only found one.
5584 if (child_verify == nullptr) {
5585 if (child->_irreducible && Compile::current()->major_progress()) {
5586 // Irreducible loops can pick a different header (one of its entries).
5587 } else {
5588 tty->print_cr("We have a loop that verify does not have");
5589 child->dump();
5590 success = false;
5591 }
5592 i++; // step for this
5593 } else if (child == nullptr) {
5594 if (child_verify->_irreducible && Compile::current()->major_progress()) {
5595 // Irreducible loops can pick a different header (one of its entries).
5596 } else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {
5597 // Infinite loops do not get attached to the loop-tree on their first visit.
5598 // "this" runs before "loop_verify". It is thus possible that we find the
5599 // infinite loop only for "child_verify". Only finding it with "child" would
5600 // mean that we lost it, which is not ok.
5601 } else {
5602 tty->print_cr("Verify has a loop that we do not have");
5603 child_verify->dump();
5604 success = false;
5605 }
5606 j++; // step for verify
5607 } else {
5608 assert(child->_head == child_verify->_head, "We have both and they are equal");
5609 success &= child->verify_tree(child_verify); // Recursion
5610 i++; // step for this
5611 j++; // step for verify
5612 }
5613 }
5614
5615 // Broken part of VerifyLoopOptimizations (D)
5616 // Reason:
5617 // split_if has to update the _tail, if it is modified. But that is done by
5618 // checking to what loop the iff belongs to. That info can be wrong, and then
5619 // we do not update the _tail correctly.
5620 /*
5621 Node *tail = _tail; // Inline a non-updating version of
5622 while( !tail->in(0) ) // the 'tail()' call.
5623 tail = tail->in(1);
5624 assert( tail == loop->_tail, "mismatched loop tail" );
5625 */
5626
5627 if (_head->is_CountedLoop()) {
5628 CountedLoopNode *cl = _head->as_CountedLoop();
5629
5630 Node* ctrl = cl->init_control();
5631 Node* back = cl->back_control();
5632 assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5633 assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5634 cl->loopexit(); // assert implied
5635 }
5636
5637 // Broken part of VerifyLoopOptimizations (E)
5638 // Reason:
5639 // PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5640 // to the loop body. Or maybe they are not added to the correct loop.
5641 // at "Node* x = n->clone();"
5642 /*
5643 // Innermost loops need to verify loop bodies,
5644 // but only if no 'major_progress'
5645 int fail = 0;
5646 if (!Compile::current()->major_progress() && _child == nullptr) {
5647 for( uint i = 0; i < _body.size(); i++ ) {
5648 Node *n = _body.at(i);
5649 if (n->outcnt() == 0) continue; // Ignore dead
5650 uint j;
5651 for( j = 0; j < loop->_body.size(); j++ )
5652 if( loop->_body.at(j) == n )
5653 break;
5654 if( j == loop->_body.size() ) { // Not found in loop body
5655 // Last ditch effort to avoid assertion: Its possible that we
5656 // have some users (so outcnt not zero) but are still dead.
5657 // Try to find from root.
5658 if (Compile::current()->root()->find(n->_idx)) {
5659 fail++;
5660 tty->print("We have that verify does not: ");
5661 n->dump();
5662 }
5663 }
5664 }
5665 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
5666 Node *n = loop->_body.at(i2);
5667 if (n->outcnt() == 0) continue; // Ignore dead
5668 uint j;
5669 for( j = 0; j < _body.size(); j++ )
5670 if( _body.at(j) == n )
5671 break;
5672 if( j == _body.size() ) { // Not found in loop body
5673 // Last ditch effort to avoid assertion: Its possible that we
5674 // have some users (so outcnt not zero) but are still dead.
5675 // Try to find from root.
5676 if (Compile::current()->root()->find(n->_idx)) {
5677 fail++;
5678 tty->print("Verify has that we do not: ");
5679 n->dump();
5680 }
5681 }
5682 }
5683 assert( !fail, "loop body mismatch" );
5684 }
5685 */
5686 return success;
5687 }
5688 #endif
5689
5690 //------------------------------set_idom---------------------------------------
5691 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
5692 _nesting.check(); // Check if a potential reallocation in the resource arena is safe
5693 uint idx = d->_idx;
5694 if (idx >= _idom_size) {
5695 uint newsize = next_power_of_2(idx);
5696 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
5697 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
5698 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5699 _idom_size = newsize;
5700 }
5701 _idom[idx] = n;
5702 _dom_depth[idx] = dom_depth;
5703 }
5704
5705 //------------------------------recompute_dom_depth---------------------------------------
5706 // The dominator tree is constructed with only parent pointers.
5707 // This recomputes the depth in the tree by first tagging all
5708 // nodes as "no depth yet" marker. The next pass then runs up
5709 // the dom tree from each node marked "no depth yet", and computes
5710 // the depth on the way back down.
5711 void PhaseIdealLoop::recompute_dom_depth() {
5712 uint no_depth_marker = C->unique();
5713 uint i;
5714 // Initialize depth to "no depth yet" and realize all lazy updates
5715 for (i = 0; i < _idom_size; i++) {
5716 // Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5717 if (_dom_depth[i] > 0 && _idom[i] != nullptr) {
5718 _dom_depth[i] = no_depth_marker;
5719
5720 // heal _idom if it has a fwd mapping in _loop_or_ctrl
5721 if (_idom[i]->in(0) == nullptr) {
5722 idom(i);
5723 }
5724 }
5725 }
5726 if (_dom_stk == nullptr) {
5727 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5728 if (init_size < 10) init_size = 10;
5729 _dom_stk = new GrowableArray<uint>(init_size);
5730 }
5731 // Compute new depth for each node.
5732 for (i = 0; i < _idom_size; i++) {
5733 uint j = i;
5734 // Run up the dom tree to find a node with a depth
5735 while (_dom_depth[j] == no_depth_marker) {
5736 _dom_stk->push(j);
5737 j = _idom[j]->_idx;
5738 }
5739 // Compute the depth on the way back down this tree branch
5740 uint dd = _dom_depth[j] + 1;
5741 while (_dom_stk->length() > 0) {
5742 uint j = _dom_stk->pop();
5743 _dom_depth[j] = dd;
5744 dd++;
5745 }
5746 }
5747 }
5748
5749 //------------------------------sort-------------------------------------------
5750 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5751 // loop tree, not the root.
5752 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
5753 if( !innermost ) return loop; // New innermost loop
5754
5755 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5756 assert( loop_preorder, "not yet post-walked loop" );
5757 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5758 IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5759
5760 // Insert at start of list
5761 while( l ) { // Insertion sort based on pre-order
5762 if( l == loop ) return innermost; // Already on list!
5763 int l_preorder = get_preorder(l->_head); // Cache pre-order number
5764 assert( l_preorder, "not yet post-walked l" );
5765 // Check header pre-order number to figure proper nesting
5766 if( loop_preorder > l_preorder )
5767 break; // End of insertion
5768 // If headers tie (e.g., shared headers) check tail pre-order numbers.
5769 // Since I split shared headers, you'd think this could not happen.
5770 // BUT: I must first do the preorder numbering before I can discover I
5771 // have shared headers, so the split headers all get the same preorder
5772 // number as the RegionNode they split from.
5773 if( loop_preorder == l_preorder &&
5774 get_preorder(loop->_tail) < get_preorder(l->_tail) )
5775 break; // Also check for shared headers (same pre#)
5776 pp = &l->_parent; // Chain up list
5777 l = *pp;
5778 }
5779 // Link into list
5780 // Point predecessor to me
5781 *pp = loop;
5782 // Point me to successor
5783 IdealLoopTree *p = loop->_parent;
5784 loop->_parent = l; // Point me to successor
5785 if( p ) sort( p, innermost ); // Insert my parents into list as well
5786 return innermost;
5787 }
5788
5789 //------------------------------build_loop_tree--------------------------------
5790 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5791 // bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5792 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5793 // tightest enclosing IdealLoopTree for post-walked.
5794 //
5795 // During my forward walk I do a short 1-layer lookahead to see if I can find
5796 // a loop backedge with that doesn't have any work on the backedge. This
5797 // helps me construct nested loops with shared headers better.
5798 //
5799 // Once I've done the forward recursion, I do the post-work. For each child
5800 // I check to see if there is a backedge. Backedges define a loop! I
5801 // insert an IdealLoopTree at the target of the backedge.
5802 //
5803 // During the post-work I also check to see if I have several children
5804 // belonging to different loops. If so, then this Node is a decision point
5805 // where control flow can choose to change loop nests. It is at this
5806 // decision point where I can figure out how loops are nested. At this
5807 // time I can properly order the different loop nests from my children.
5808 // Note that there may not be any backedges at the decision point!
5809 //
5810 // Since the decision point can be far removed from the backedges, I can't
5811 // order my loops at the time I discover them. Thus at the decision point
5812 // I need to inspect loop header pre-order numbers to properly nest my
5813 // loops. This means I need to sort my childrens' loops by pre-order.
5814 // The sort is of size number-of-control-children, which generally limits
5815 // it to size 2 (i.e., I just choose between my 2 target loops).
5816 void PhaseIdealLoop::build_loop_tree() {
5817 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5818 GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5819 Node *n = C->root();
5820 bltstack.push(n);
5821 int pre_order = 1;
5822 int stack_size;
5823
5824 while ( ( stack_size = bltstack.length() ) != 0 ) {
5825 n = bltstack.top(); // Leave node on stack
5826 if ( !is_visited(n) ) {
5827 // ---- Pre-pass Work ----
5828 // Pre-walked but not post-walked nodes need a pre_order number.
5829
5830 set_preorder_visited( n, pre_order ); // set as visited
5831
5832 // ---- Scan over children ----
5833 // Scan first over control projections that lead to loop headers.
5834 // This helps us find inner-to-outer loops with shared headers better.
5835
5836 // Scan children's children for loop headers.
5837 for ( int i = n->outcnt() - 1; i >= 0; --i ) {
5838 Node* m = n->raw_out(i); // Child
5839 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
5840 // Scan over children's children to find loop
5841 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5842 Node* l = m->fast_out(j);
5843 if( is_visited(l) && // Been visited?
5844 !is_postvisited(l) && // But not post-visited
5845 get_preorder(l) < pre_order ) { // And smaller pre-order
5846 // Found! Scan the DFS down this path before doing other paths
5847 bltstack.push(m);
5848 break;
5849 }
5850 }
5851 }
5852 }
5853 pre_order++;
5854 }
5855 else if ( !is_postvisited(n) ) {
5856 // Note: build_loop_tree_impl() adds out edges on rare occasions,
5857 // such as com.sun.rsasign.am::a.
5858 // For non-recursive version, first, process current children.
5859 // On next iteration, check if additional children were added.
5860 for ( int k = n->outcnt() - 1; k >= 0; --k ) {
5861 Node* u = n->raw_out(k);
5862 if ( u->is_CFG() && !is_visited(u) ) {
5863 bltstack.push(u);
5864 }
5865 }
5866 if ( bltstack.length() == stack_size ) {
5867 // There were no additional children, post visit node now
5868 (void)bltstack.pop(); // Remove node from stack
5869 pre_order = build_loop_tree_impl(n, pre_order);
5870 // Check for bailout
5871 if (C->failing()) {
5872 return;
5873 }
5874 // Check to grow _preorders[] array for the case when
5875 // build_loop_tree_impl() adds new nodes.
5876 check_grow_preorders();
5877 }
5878 }
5879 else {
5880 (void)bltstack.pop(); // Remove post-visited node from stack
5881 }
5882 }
5883 DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5884 }
5885
5886 //------------------------------build_loop_tree_impl---------------------------
5887 int PhaseIdealLoop::build_loop_tree_impl(Node* n, int pre_order) {
5888 // ---- Post-pass Work ----
5889 // Pre-walked but not post-walked nodes need a pre_order number.
5890
5891 // Tightest enclosing loop for this Node
5892 IdealLoopTree *innermost = nullptr;
5893
5894 // For all children, see if any edge is a backedge. If so, make a loop
5895 // for it. Then find the tightest enclosing loop for the self Node.
5896 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
5897 Node* m = n->fast_out(i); // Child
5898 if (n == m) continue; // Ignore control self-cycles
5899 if (!m->is_CFG()) continue;// Ignore non-CFG edges
5900
5901 IdealLoopTree *l; // Child's loop
5902 if (!is_postvisited(m)) { // Child visited but not post-visited?
5903 // Found a backedge
5904 assert(get_preorder(m) < pre_order, "should be backedge");
5905 // Check for the RootNode, which is already a LoopNode and is allowed
5906 // to have multiple "backedges".
5907 if (m == C->root()) { // Found the root?
5908 l = _ltree_root; // Root is the outermost LoopNode
5909 } else { // Else found a nested loop
5910 // Insert a LoopNode to mark this loop.
5911 l = new IdealLoopTree(this, m, n);
5912 } // End of Else found a nested loop
5913 if (!has_loop(m)) { // If 'm' does not already have a loop set
5914 set_loop(m, l); // Set loop header to loop now
5915 }
5916 } else { // Else not a nested loop
5917 if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5918 IdealLoopTree* m_loop = get_loop(m);
5919 l = m_loop; // Get previously determined loop
5920 // If successor is header of a loop (nest), move up-loop till it
5921 // is a member of some outer enclosing loop. Since there are no
5922 // shared headers (I've split them already) I only need to go up
5923 // at most 1 level.
5924 while (l && l->_head == m) { // Successor heads loop?
5925 l = l->_parent; // Move up 1 for me
5926 }
5927 // If this loop is not properly parented, then this loop
5928 // has no exit path out, i.e. its an infinite loop.
5929 if (!l) {
5930 // Make loop "reachable" from root so the CFG is reachable. Basically
5931 // insert a bogus loop exit that is never taken. 'm', the loop head,
5932 // points to 'n', one (of possibly many) fall-in paths. There may be
5933 // many backedges as well.
5934
5935 if (!_verify_only) {
5936 // Insert the NeverBranch between 'm' and it's control user.
5937 NeverBranchNode *iff = new NeverBranchNode( m );
5938 _igvn.register_new_node_with_optimizer(iff);
5939 set_loop(iff, m_loop);
5940 Node *if_t = new CProjNode( iff, 0 );
5941 _igvn.register_new_node_with_optimizer(if_t);
5942 set_loop(if_t, m_loop);
5943
5944 Node* cfg = nullptr; // Find the One True Control User of m
5945 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
5946 Node* x = m->fast_out(j);
5947 if (x->is_CFG() && x != m && x != iff)
5948 { cfg = x; break; }
5949 }
5950 assert(cfg != nullptr, "must find the control user of m");
5951 uint k = 0; // Probably cfg->in(0)
5952 while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5953 _igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5954
5955 // Now create the never-taken loop exit
5956 Node *if_f = new CProjNode( iff, 1 );
5957 _igvn.register_new_node_with_optimizer(if_f);
5958 set_loop(if_f, _ltree_root);
5959 // Find frame ptr for Halt. Relies on the optimizer
5960 // V-N'ing. Easier and quicker than searching through
5961 // the program structure.
5962 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5963 _igvn.register_new_node_with_optimizer(frame);
5964 // Halt & Catch Fire
5965 Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5966 _igvn.register_new_node_with_optimizer(halt);
5967 set_loop(halt, _ltree_root);
5968 _igvn.add_input_to(C->root(), halt);
5969 }
5970 set_loop(C->root(), _ltree_root);
5971 // move to outer most loop with same header
5972 l = m_loop;
5973 while (true) {
5974 IdealLoopTree* next = l->_parent;
5975 if (next == nullptr || next->_head != m) {
5976 break;
5977 }
5978 l = next;
5979 }
5980 // properly insert infinite loop in loop tree
5981 sort(_ltree_root, l);
5982 // fix child link from parent
5983 IdealLoopTree* p = l->_parent;
5984 l->_next = p->_child;
5985 p->_child = l;
5986 // code below needs enclosing loop
5987 l = l->_parent;
5988 }
5989 }
5990 if (is_postvisited(l->_head)) {
5991 // We are currently visiting l, but its head has already been post-visited.
5992 // l is irreducible: we just found a second entry m.
5993 _has_irreducible_loops = true;
5994 RegionNode* secondary_entry = m->as_Region();
5995
5996 if (!secondary_entry->can_be_irreducible_entry()) {
5997 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
5998 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
5999 return pre_order;
6000 }
6001
6002 // Walk up the loop-tree, mark all loops that are already post-visited as irreducible
6003 // Since m is a secondary entry to them all.
6004 while( is_postvisited(l->_head) ) {
6005 l->_irreducible = 1; // = true
6006 RegionNode* head = l->_head->as_Region();
6007 if (!head->can_be_irreducible_entry()) {
6008 assert(!VerifyNoNewIrreducibleLoops, "A new irreducible loop was created after parsing.");
6009 C->record_method_not_compilable("A new irreducible loop was created after parsing.");
6010 return pre_order;
6011 }
6012 l = l->_parent;
6013 // Check for bad CFG here to prevent crash, and bailout of compile
6014 if (l == nullptr) {
6015 #ifndef PRODUCT
6016 if (TraceLoopOpts) {
6017 tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");
6018 m->dump();
6019 }
6020 #endif
6021 // This is a rare case that we do not want to handle in C2.
6022 C->record_method_not_compilable("unhandled CFG detected during loop optimization");
6023 return pre_order;
6024 }
6025 }
6026 }
6027 if (!_verify_only) {
6028 C->set_has_irreducible_loop(_has_irreducible_loops);
6029 }
6030
6031 // This Node might be a decision point for loops. It is only if
6032 // it's children belong to several different loops. The sort call
6033 // does a trivial amount of work if there is only 1 child or all
6034 // children belong to the same loop. If however, the children
6035 // belong to different loops, the sort call will properly set the
6036 // _parent pointers to show how the loops nest.
6037 //
6038 // In any case, it returns the tightest enclosing loop.
6039 innermost = sort( l, innermost );
6040 }
6041
6042 // Def-use info will have some dead stuff; dead stuff will have no
6043 // loop decided on.
6044
6045 // Am I a loop header? If so fix up my parent's child and next ptrs.
6046 if( innermost && innermost->_head == n ) {
6047 assert( get_loop(n) == innermost, "" );
6048 IdealLoopTree *p = innermost->_parent;
6049 IdealLoopTree *l = innermost;
6050 while (p && l->_head == n) {
6051 l->_next = p->_child; // Put self on parents 'next child'
6052 p->_child = l; // Make self as first child of parent
6053 l = p; // Now walk up the parent chain
6054 p = l->_parent;
6055 }
6056 } else {
6057 // Note that it is possible for a LoopNode to reach here, if the
6058 // backedge has been made unreachable (hence the LoopNode no longer
6059 // denotes a Loop, and will eventually be removed).
6060
6061 // Record tightest enclosing loop for self. Mark as post-visited.
6062 set_loop(n, innermost);
6063 // Also record has_call flag early on
6064 if (innermost) {
6065 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
6066 // Do not count uncommon calls
6067 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
6068 Node *iff = n->in(0)->in(0);
6069 // No any calls for vectorized loops.
6070 if (C->do_superword() ||
6071 !iff->is_If() ||
6072 (n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
6073 iff->as_If()->_prob >= 0.01) {
6074 innermost->_has_call = 1;
6075 }
6076 }
6077 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
6078 // Disable loop optimizations if the loop has a scalar replaceable
6079 // allocation. This disabling may cause a potential performance lost
6080 // if the allocation is not eliminated for some reason.
6081 innermost->_allow_optimizations = false;
6082 innermost->_has_call = 1; // = true
6083 } else if (n->Opcode() == Op_SafePoint) {
6084 // Record all safepoints in this loop.
6085 if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
6086 innermost->_safepts->push(n);
6087 }
6088 }
6089 }
6090
6091 // Flag as post-visited now
6092 set_postvisited(n);
6093 return pre_order;
6094 }
6095
6096 #ifdef ASSERT
6097 //--------------------------verify_regions_in_irreducible_loops----------------
6098 // Iterate down from Root through CFG, verify for every region:
6099 // if it is in an irreducible loop it must be marked as such
6100 void PhaseIdealLoop::verify_regions_in_irreducible_loops() {
6101 ResourceMark rm;
6102 if (!_has_irreducible_loops) {
6103 // last build_loop_tree has not found any irreducible loops
6104 // hence no region has to be marked is_in_irreduible_loop
6105 return;
6106 }
6107
6108 RootNode* root = C->root();
6109 Unique_Node_List worklist; // visit all nodes once
6110 worklist.push(root);
6111 bool failure = false;
6112 for (uint i = 0; i < worklist.size(); i++) {
6113 Node* n = worklist.at(i);
6114 if (n->is_Region()) {
6115 RegionNode* region = n->as_Region();
6116 if (is_in_irreducible_loop(region) &&
6117 region->loop_status() == RegionNode::LoopStatus::Reducible) {
6118 failure = true;
6119 tty->print("irreducible! ");
6120 region->dump();
6121 }
6122 }
6123 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
6124 Node* use = n->fast_out(j);
6125 if (use->is_CFG()) {
6126 worklist.push(use); // push if was not pushed before
6127 }
6128 }
6129 }
6130 assert(!failure, "region in irreducible loop was marked as reducible");
6131 }
6132
6133 //---------------------------is_in_irreducible_loop-------------------------
6134 // Analogous to ciTypeFlow::Block::is_in_irreducible_loop
6135 bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {
6136 if (!_has_irreducible_loops) {
6137 return false; // no irreducible loop in graph
6138 }
6139 IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
6140 do {
6141 if (l->_irreducible) {
6142 return true; // found it
6143 }
6144 if (l == _ltree_root) {
6145 return false; // reached root, terimnate
6146 }
6147 l = l->_parent;
6148 } while (l != nullptr);
6149 assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
6150 // We have "l->_parent == nullptr", which happens only for infinite loops,
6151 // where no parent is attached to the loop. We did not find any irreducible
6152 // loop from this block out to lp. Thus lp only has one entry, and no exit
6153 // (it is infinite and reducible). We can always rewrite an infinite loop
6154 // that is nested inside other loops:
6155 // while(condition) { infinite_loop; }
6156 // with an equivalent program where the infinite loop is an outermost loop
6157 // that is not nested in any loop:
6158 // while(condition) { break; } infinite_loop;
6159 // Thus, we can understand lp as an outermost loop, and can terminate and
6160 // conclude: this block is in no irreducible loop.
6161 return false;
6162 }
6163 #endif
6164
6165 //------------------------------build_loop_early-------------------------------
6166 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6167 // First pass computes the earliest controlling node possible. This is the
6168 // controlling input with the deepest dominating depth.
6169 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6170 while (worklist.size() != 0) {
6171 // Use local variables nstack_top_n & nstack_top_i to cache values
6172 // on nstack's top.
6173 Node *nstack_top_n = worklist.pop();
6174 uint nstack_top_i = 0;
6175 //while_nstack_nonempty:
6176 while (true) {
6177 // Get parent node and next input's index from stack's top.
6178 Node *n = nstack_top_n;
6179 uint i = nstack_top_i;
6180 uint cnt = n->req(); // Count of inputs
6181 if (i == 0) { // Pre-process the node.
6182 if( has_node(n) && // Have either loop or control already?
6183 !has_ctrl(n) ) { // Have loop picked out already?
6184 // During "merge_many_backedges" we fold up several nested loops
6185 // into a single loop. This makes the members of the original
6186 // loop bodies pointing to dead loops; they need to move up
6187 // to the new UNION'd larger loop. I set the _head field of these
6188 // dead loops to null and the _parent field points to the owning
6189 // loop. Shades of UNION-FIND algorithm.
6190 IdealLoopTree *ilt;
6191 while( !(ilt = get_loop(n))->_head ) {
6192 // Normally I would use a set_loop here. But in this one special
6193 // case, it is legal (and expected) to change what loop a Node
6194 // belongs to.
6195 _loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
6196 }
6197 // Remove safepoints ONLY if I've already seen I don't need one.
6198 // (the old code here would yank a 2nd safepoint after seeing a
6199 // first one, even though the 1st did not dominate in the loop body
6200 // and thus could be avoided indefinitely)
6201 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
6202 is_deleteable_safept(n)) {
6203 Node *in = n->in(TypeFunc::Control);
6204 replace_node_and_forward_ctrl(n, in); // Pull safepoint now
6205 if (ilt->_safepts != nullptr) {
6206 ilt->_safepts->yank(n);
6207 }
6208 // Carry on with the recursion "as if" we are walking
6209 // only the control input
6210 if( !visited.test_set( in->_idx ) ) {
6211 worklist.push(in); // Visit this guy later, using worklist
6212 }
6213 // Get next node from nstack:
6214 // - skip n's inputs processing by setting i > cnt;
6215 // - we also will not call set_early_ctrl(n) since
6216 // has_node(n) == true (see the condition above).
6217 i = cnt + 1;
6218 }
6219 }
6220 } // if (i == 0)
6221
6222 // Visit all inputs
6223 bool done = true; // Assume all n's inputs will be processed
6224 while (i < cnt) {
6225 Node *in = n->in(i);
6226 ++i;
6227 if (in == nullptr) continue;
6228 if (in->pinned() && !in->is_CFG())
6229 set_ctrl(in, in->in(0));
6230 int is_visited = visited.test_set( in->_idx );
6231 if (!has_node(in)) { // No controlling input yet?
6232 assert( !in->is_CFG(), "CFG Node with no controlling input?" );
6233 assert( !is_visited, "visit only once" );
6234 nstack.push(n, i); // Save parent node and next input's index.
6235 nstack_top_n = in; // Process current input now.
6236 nstack_top_i = 0;
6237 done = false; // Not all n's inputs processed.
6238 break; // continue while_nstack_nonempty;
6239 } else if (!is_visited) {
6240 // This guy has a location picked out for him, but has not yet
6241 // been visited. Happens to all CFG nodes, for instance.
6242 // Visit him using the worklist instead of recursion, to break
6243 // cycles. Since he has a location already we do not need to
6244 // find his location before proceeding with the current Node.
6245 worklist.push(in); // Visit this guy later, using worklist
6246 }
6247 }
6248 if (done) {
6249 // All of n's inputs have been processed, complete post-processing.
6250
6251 // Compute earliest point this Node can go.
6252 // CFG, Phi, pinned nodes already know their controlling input.
6253 if (!has_node(n)) {
6254 // Record earliest legal location
6255 set_early_ctrl(n, false);
6256 }
6257 if (nstack.is_empty()) {
6258 // Finished all nodes on stack.
6259 // Process next node on the worklist.
6260 break;
6261 }
6262 // Get saved parent node and next input's index.
6263 nstack_top_n = nstack.node();
6264 nstack_top_i = nstack.index();
6265 nstack.pop();
6266 }
6267 } // while (true)
6268 }
6269 }
6270
6271 //------------------------------dom_lca_internal--------------------------------
6272 // Pair-wise LCA
6273 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
6274 if( !n1 ) return n2; // Handle null original LCA
6275 assert( n1->is_CFG(), "" );
6276 assert( n2->is_CFG(), "" );
6277 // find LCA of all uses
6278 uint d1 = dom_depth(n1);
6279 uint d2 = dom_depth(n2);
6280 while (n1 != n2) {
6281 if (d1 > d2) {
6282 n1 = idom(n1);
6283 d1 = dom_depth(n1);
6284 } else if (d1 < d2) {
6285 n2 = idom(n2);
6286 d2 = dom_depth(n2);
6287 } else {
6288 // Here d1 == d2. Due to edits of the dominator-tree, sections
6289 // of the tree might have the same depth. These sections have
6290 // to be searched more carefully.
6291
6292 // Scan up all the n1's with equal depth, looking for n2.
6293 Node *t1 = idom(n1);
6294 while (dom_depth(t1) == d1) {
6295 if (t1 == n2) return n2;
6296 t1 = idom(t1);
6297 }
6298 // Scan up all the n2's with equal depth, looking for n1.
6299 Node *t2 = idom(n2);
6300 while (dom_depth(t2) == d2) {
6301 if (t2 == n1) return n1;
6302 t2 = idom(t2);
6303 }
6304 // Move up to a new dominator-depth value as well as up the dom-tree.
6305 n1 = t1;
6306 n2 = t2;
6307 d1 = dom_depth(n1);
6308 d2 = dom_depth(n2);
6309 }
6310 }
6311 return n1;
6312 }
6313
6314 //------------------------------compute_idom-----------------------------------
6315 // Locally compute IDOM using dom_lca call. Correct only if the incoming
6316 // IDOMs are correct.
6317 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
6318 assert( region->is_Region(), "" );
6319 Node *LCA = nullptr;
6320 for( uint i = 1; i < region->req(); i++ ) {
6321 if( region->in(i) != C->top() )
6322 LCA = dom_lca( LCA, region->in(i) );
6323 }
6324 return LCA;
6325 }
6326
6327 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
6328 bool had_error = false;
6329 #ifdef ASSERT
6330 if (early != C->root()) {
6331 // Make sure that there's a dominance path from LCA to early
6332 Node* d = LCA;
6333 while (d != early) {
6334 if (d == C->root()) {
6335 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
6336 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
6337 had_error = true;
6338 break;
6339 }
6340 d = idom(d);
6341 }
6342 }
6343 #endif
6344 return had_error;
6345 }
6346
6347
6348 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
6349 // Compute LCA over list of uses
6350 bool had_error = false;
6351 Node *LCA = nullptr;
6352 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
6353 Node* c = n->fast_out(i);
6354 if (_loop_or_ctrl[c->_idx] == nullptr)
6355 continue; // Skip the occasional dead node
6356 if( c->is_Phi() ) { // For Phis, we must land above on the path
6357 for( uint j=1; j<c->req(); j++ ) {// For all inputs
6358 if( c->in(j) == n ) { // Found matching input?
6359 Node *use = c->in(0)->in(j);
6360 if (_verify_only && use->is_top()) continue;
6361 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6362 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6363 }
6364 }
6365 } else {
6366 // For CFG data-users, use is in the block just prior
6367 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
6368 LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
6369 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
6370 }
6371 }
6372 assert(!had_error, "bad dominance");
6373 return LCA;
6374 }
6375
6376 // Check the shape of the graph at the loop entry. In some cases,
6377 // the shape of the graph does not match the shape outlined below.
6378 // That is caused by the Opaque1 node "protecting" the shape of
6379 // the graph being removed by, for example, the IGVN performed
6380 // in PhaseIdealLoop::build_and_optimize().
6381 //
6382 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
6383 // loop unswitching, and IGVN, or a combination of them) can freely change
6384 // the graph's shape. As a result, the graph shape outlined below cannot
6385 // be guaranteed anymore.
6386 Node* CountedLoopNode::is_canonical_loop_entry() {
6387 if (!is_main_loop() && !is_post_loop()) {
6388 return nullptr;
6389 }
6390 Node* ctrl = skip_assertion_predicates_with_halt();
6391
6392 if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
6393 return nullptr;
6394 }
6395 Node* iffm = ctrl->in(0);
6396 if (iffm == nullptr || iffm->Opcode() != Op_If) {
6397 return nullptr;
6398 }
6399 Node* bolzm = iffm->in(1);
6400 if (bolzm == nullptr || !bolzm->is_Bool()) {
6401 return nullptr;
6402 }
6403 Node* cmpzm = bolzm->in(1);
6404 if (cmpzm == nullptr || !cmpzm->is_Cmp()) {
6405 return nullptr;
6406 }
6407
6408 uint input = is_main_loop() ? 2 : 1;
6409 if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {
6410 return nullptr;
6411 }
6412 bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
6413 #ifdef ASSERT
6414 bool found_opaque = false;
6415 for (uint i = 1; i < cmpzm->req(); i++) {
6416 Node* opnd = cmpzm->in(i);
6417 if (opnd && opnd->is_Opaque1()) {
6418 found_opaque = true;
6419 break;
6420 }
6421 }
6422 assert(found_opaque == res, "wrong pattern");
6423 #endif
6424 return res ? cmpzm->in(input) : nullptr;
6425 }
6426
6427 // Find pre loop end from main loop. Returns nullptr if none.
6428 CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {
6429 assert(is_main_loop(), "Can only find pre-loop from main-loop");
6430 // The loop cannot be optimized if the graph shape at the loop entry is
6431 // inappropriate.
6432 if (is_canonical_loop_entry() == nullptr) {
6433 return nullptr;
6434 }
6435
6436 Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
6437 if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {
6438 return nullptr;
6439 }
6440 CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
6441 CountedLoopNode* loop_node = pre_end->loopnode();
6442 if (loop_node == nullptr || !loop_node->is_pre_loop()) {
6443 return nullptr;
6444 }
6445 return pre_end;
6446 }
6447
6448 Node* CountedLoopNode::uncasted_init_trip(bool uncast) {
6449 Node* init = init_trip();
6450 if (uncast && init->is_CastII()) {
6451 // skip over the cast added by PhaseIdealLoop::cast_incr_before_loop() when pre/post/main loops are created because
6452 // it can get in the way of type propagation. For instance, the index tested by an Assertion Predicate, if the cast
6453 // is not skipped over, could be (1):
6454 // (AddI (CastII (AddI pre_loop_iv -2) int) 1)
6455 // while without the cast, it is (2):
6456 // (AddI (AddI pre_loop_iv -2) 1)
6457 // which is be transformed to (3):
6458 // (AddI pre_loop_iv -1)
6459 // The compiler may be able to constant fold the Assertion Predicate condition for (3) but not (1)
6460 assert(init->as_CastII()->carry_dependency() && skip_assertion_predicates_with_halt() == init->in(0), "casted iv phi from pre loop expected");
6461 init = init->in(1);
6462 }
6463 return init;
6464 }
6465
6466 //------------------------------get_late_ctrl----------------------------------
6467 // Compute latest legal control.
6468 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
6469 assert(early != nullptr, "early control should not be null");
6470
6471 Node* LCA = compute_lca_of_uses(n, early);
6472 #ifdef ASSERT
6473 if (LCA == C->root() && LCA != early) {
6474 // def doesn't dominate uses so print some useful debugging output
6475 compute_lca_of_uses(n, early, true);
6476 }
6477 #endif
6478
6479 if (n->is_Load() && LCA != early) {
6480 LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6481 }
6482
6483 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6484 return LCA;
6485 }
6486
6487 // if this is a load, check for anti-dependent stores
6488 // We use a conservative algorithm to identify potential interfering
6489 // instructions and for rescheduling the load. The users of the memory
6490 // input of this load are examined. Any use which is not a load and is
6491 // dominated by early is considered a potentially interfering store.
6492 // This can produce false positives.
6493 Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {
6494 int load_alias_idx = C->get_alias_index(n->adr_type());
6495 if (C->alias_type(load_alias_idx)->is_rewritable()) {
6496 Unique_Node_List worklist;
6497
6498 Node* mem = n->in(MemNode::Memory);
6499 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
6500 Node* s = mem->fast_out(i);
6501 worklist.push(s);
6502 }
6503 for (uint i = 0; i < worklist.size() && LCA != early; i++) {
6504 Node* s = worklist.at(i);
6505 if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6506 (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6507 s->is_Phi()) {
6508 continue;
6509 } else if (s->is_MergeMem()) {
6510 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6511 Node* s1 = s->fast_out(i);
6512 worklist.push(s1);
6513 }
6514 } else {
6515 Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6516 assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6517 if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {
6518 const TypePtr* adr_type = s->adr_type();
6519 if (s->is_ArrayCopy()) {
6520 // Copy to known instance needs destination type to test for aliasing
6521 const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6522 if (dest_type != TypeOopPtr::BOTTOM) {
6523 adr_type = dest_type;
6524 }
6525 }
6526 if (C->can_alias(adr_type, load_alias_idx)) {
6527 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6528 } else if (s->is_CFG() && s->is_Multi()) {
6529 // Look for the memory use of s (that is the use of its memory projection)
6530 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
6531 Node* s1 = s->fast_out(i);
6532 assert(s1->is_Proj(), "projection expected");
6533 if (_igvn.type(s1) == Type::MEMORY) {
6534 for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {
6535 Node* s2 = s1->fast_out(j);
6536 worklist.push(s2);
6537 }
6538 }
6539 }
6540 }
6541 }
6542 }
6543 }
6544 // For Phis only consider Region's inputs that were reached by following the memory edges
6545 if (LCA != early) {
6546 for (uint i = 0; i < worklist.size(); i++) {
6547 Node* s = worklist.at(i);
6548 if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {
6549 Node* r = s->in(0);
6550 for (uint j = 1; j < s->req(); j++) {
6551 Node* in = s->in(j);
6552 Node* r_in = r->in(j);
6553 // We can't reach any node from a Phi because we don't enqueue Phi's uses above
6554 if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {
6555 LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6556 }
6557 }
6558 }
6559 }
6560 }
6561 }
6562 return LCA;
6563 }
6564
6565 // Is CFG node 'dominator' dominating node 'n'?
6566 bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {
6567 if (dominator == n) {
6568 return true;
6569 }
6570 assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6571 uint dd = dom_depth(dominator);
6572 while (dom_depth(n) >= dd) {
6573 if (n == dominator) {
6574 return true;
6575 }
6576 n = idom(n);
6577 }
6578 return false;
6579 }
6580
6581 // Is CFG node 'dominator' strictly dominating node 'n'?
6582 bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {
6583 return dominator != n && is_dominator(dominator, n);
6584 }
6585
6586 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
6587 // Pair-wise LCA with tags.
6588 // Tag each index with the node 'tag' currently being processed
6589 // before advancing up the dominator chain using idom().
6590 // Later calls that find a match to 'tag' know that this path has already
6591 // been considered in the current LCA (which is input 'n1' by convention).
6592 // Since get_late_ctrl() is only called once for each node, the tag array
6593 // does not need to be cleared between calls to get_late_ctrl().
6594 // Algorithm trades a larger constant factor for better asymptotic behavior
6595 //
6596 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {
6597 uint d1 = dom_depth(n1);
6598 uint d2 = dom_depth(n2);
6599 jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6600
6601 do {
6602 if (d1 > d2) {
6603 // current lca is deeper than n2
6604 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6605 n1 = idom(n1);
6606 d1 = dom_depth(n1);
6607 } else if (d1 < d2) {
6608 // n2 is deeper than current lca
6609 jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6610 if (memo == tag) {
6611 return n1; // Return the current LCA
6612 }
6613 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6614 n2 = idom(n2);
6615 d2 = dom_depth(n2);
6616 } else {
6617 // Here d1 == d2. Due to edits of the dominator-tree, sections
6618 // of the tree might have the same depth. These sections have
6619 // to be searched more carefully.
6620
6621 // Scan up all the n1's with equal depth, looking for n2.
6622 _dom_lca_tags.at_put_grow(n1->_idx, tag);
6623 Node *t1 = idom(n1);
6624 while (dom_depth(t1) == d1) {
6625 if (t1 == n2) return n2;
6626 _dom_lca_tags.at_put_grow(t1->_idx, tag);
6627 t1 = idom(t1);
6628 }
6629 // Scan up all the n2's with equal depth, looking for n1.
6630 _dom_lca_tags.at_put_grow(n2->_idx, tag);
6631 Node *t2 = idom(n2);
6632 while (dom_depth(t2) == d2) {
6633 if (t2 == n1) return n1;
6634 _dom_lca_tags.at_put_grow(t2->_idx, tag);
6635 t2 = idom(t2);
6636 }
6637 // Move up to a new dominator-depth value as well as up the dom-tree.
6638 n1 = t1;
6639 n2 = t2;
6640 d1 = dom_depth(n1);
6641 d2 = dom_depth(n2);
6642 }
6643 } while (n1 != n2);
6644 return n1;
6645 }
6646
6647 //------------------------------init_dom_lca_tags------------------------------
6648 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
6649 // Intended use does not involve any growth for the array, so it could
6650 // be of fixed size.
6651 void PhaseIdealLoop::init_dom_lca_tags() {
6652 uint limit = C->unique() + 1;
6653 _dom_lca_tags.at_grow(limit, 0);
6654 _dom_lca_tags_round = 0;
6655 #ifdef ASSERT
6656 for (uint i = 0; i < limit; ++i) {
6657 assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6658 }
6659 #endif // ASSERT
6660 }
6661
6662 //------------------------------build_loop_late--------------------------------
6663 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6664 // Second pass finds latest legal placement, and ideal loop placement.
6665 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
6666 while (worklist.size() != 0) {
6667 Node *n = worklist.pop();
6668 // Only visit once
6669 if (visited.test_set(n->_idx)) continue;
6670 uint cnt = n->outcnt();
6671 uint i = 0;
6672 while (true) {
6673 assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6674 // Visit all children
6675 if (i < cnt) {
6676 Node* use = n->raw_out(i);
6677 ++i;
6678 // Check for dead uses. Aggressively prune such junk. It might be
6679 // dead in the global sense, but still have local uses so I cannot
6680 // easily call 'remove_dead_node'.
6681 if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?
6682 // Due to cycles, we might not hit the same fixed point in the verify
6683 // pass as we do in the regular pass. Instead, visit such phis as
6684 // simple uses of the loop head.
6685 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
6686 if( !visited.test(use->_idx) )
6687 worklist.push(use);
6688 } else if( !visited.test_set(use->_idx) ) {
6689 nstack.push(n, i); // Save parent and next use's index.
6690 n = use; // Process all children of current use.
6691 cnt = use->outcnt();
6692 i = 0;
6693 }
6694 } else {
6695 // Do not visit around the backedge of loops via data edges.
6696 // push dead code onto a worklist
6697 _deadlist.push(use);
6698 }
6699 } else {
6700 // All of n's children have been processed, complete post-processing.
6701 build_loop_late_post(n);
6702 if (C->failing()) { return; }
6703 if (nstack.is_empty()) {
6704 // Finished all nodes on stack.
6705 // Process next node on the worklist.
6706 break;
6707 }
6708 // Get saved parent node and next use's index. Visit the rest of uses.
6709 n = nstack.node();
6710 cnt = n->outcnt();
6711 i = nstack.index();
6712 nstack.pop();
6713 }
6714 }
6715 }
6716 }
6717
6718 // Verify that no data node is scheduled in the outer loop of a strip
6719 // mined loop.
6720 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
6721 #ifdef ASSERT
6722 if (get_loop(least)->_nest == 0) {
6723 return;
6724 }
6725 IdealLoopTree* loop = get_loop(least);
6726 Node* head = loop->_head;
6727 if (head->is_OuterStripMinedLoop() &&
6728 // Verification can't be applied to fully built strip mined loops
6729 head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
6730 Node* sfpt = head->as_Loop()->outer_safepoint();
6731 ResourceMark rm;
6732 Unique_Node_List wq;
6733 wq.push(sfpt);
6734 for (uint i = 0; i < wq.size(); i++) {
6735 Node *m = wq.at(i);
6736 for (uint i = 1; i < m->req(); i++) {
6737 Node* nn = m->in(i);
6738 if (nn == n) {
6739 return;
6740 }
6741 if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
6742 wq.push(nn);
6743 }
6744 }
6745 }
6746 ShouldNotReachHere();
6747 }
6748 #endif
6749 }
6750
6751
6752 //------------------------------build_loop_late_post---------------------------
6753 // Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6754 // Second pass finds latest legal placement, and ideal loop placement.
6755 void PhaseIdealLoop::build_loop_late_post(Node *n) {
6756 build_loop_late_post_work(n, true);
6757 }
6758
6759 // Class to visit all predicates in a predicate chain to find out which are dominated by a given node. Keeps track of
6760 // the entry to the earliest predicate that is still dominated by the given dominator. This class is used when trying to
6761 // legally skip all predicates when figuring out the latest placement such that a node does not interfere with Loop
6762 // Predication or creating a Loop Limit Check Predicate later.
6763 class DominatedPredicates : public UnifiedPredicateVisitor {
6764 Node* const _dominator;
6765 Node* _earliest_dominated_predicate_entry;
6766 bool _should_continue;
6767 PhaseIdealLoop* const _phase;
6768
6769 public:
6770 DominatedPredicates(Node* dominator, Node* start_node, PhaseIdealLoop* phase)
6771 : _dominator(dominator),
6772 _earliest_dominated_predicate_entry(start_node),
6773 _should_continue(true),
6774 _phase(phase) {}
6775 NONCOPYABLE(DominatedPredicates);
6776
6777 bool should_continue() const override {
6778 return _should_continue;
6779 }
6780
6781 // Returns the entry to the earliest predicate that is still dominated by the given dominator (all could be dominated).
6782 Node* earliest_dominated_predicate_entry() const {
6783 return _earliest_dominated_predicate_entry;
6784 }
6785
6786 void visit_predicate(const Predicate& predicate) override {
6787 Node* entry = predicate.entry();
6788 if (_phase->is_strict_dominator(entry, _dominator)) {
6789 _should_continue = false;
6790 } else {
6791 _earliest_dominated_predicate_entry = entry;
6792 }
6793 }
6794 };
6795
6796 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
6797
6798 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
6799 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6800 }
6801
6802 #ifdef ASSERT
6803 if (_verify_only && !n->is_CFG()) {
6804 // Check def-use domination.
6805 // We would like to expose this check in product but it appears to be expensive.
6806 compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
6807 }
6808 #endif
6809
6810 // CFG and pinned nodes already handled
6811 if( n->in(0) ) {
6812 if( n->in(0)->is_top() ) return; // Dead?
6813
6814 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
6815 // _must_ be pinned (they have to observe their control edge of course).
6816 // Unlike Stores (which modify an unallocable resource, the memory
6817 // state), Mods/Loads can float around. So free them up.
6818 switch( n->Opcode() ) {
6819 case Op_DivI:
6820 case Op_DivF:
6821 case Op_DivD:
6822 case Op_ModI:
6823 case Op_LoadB: // Same with Loads; they can sink
6824 case Op_LoadUB: // during loop optimizations.
6825 case Op_LoadUS:
6826 case Op_LoadD:
6827 case Op_LoadF:
6828 case Op_LoadI:
6829 case Op_LoadKlass:
6830 case Op_LoadNKlass:
6831 case Op_LoadL:
6832 case Op_LoadS:
6833 case Op_LoadP:
6834 case Op_LoadN:
6835 case Op_LoadRange:
6836 case Op_LoadD_unaligned:
6837 case Op_LoadL_unaligned:
6838 case Op_StrComp: // Does a bunch of load-like effects
6839 case Op_StrEquals:
6840 case Op_StrIndexOf:
6841 case Op_StrIndexOfChar:
6842 case Op_AryEq:
6843 case Op_VectorizedHashCode:
6844 case Op_CountPositives:
6845 pinned = false;
6846 }
6847 if (n->is_CMove() || n->is_ConstraintCast()) {
6848 pinned = false;
6849 }
6850 if( pinned ) {
6851 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6852 if( !chosen_loop->_child ) // Inner loop?
6853 chosen_loop->_body.push(n); // Collect inner loops
6854 return;
6855 }
6856 } else { // No slot zero
6857 if( n->is_CFG() ) { // CFG with no slot 0 is dead
6858 _loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6859 return;
6860 }
6861 assert(!n->is_CFG() || n->outcnt() == 0, "");
6862 }
6863
6864 // Do I have a "safe range" I can select over?
6865 Node *early = get_ctrl(n);// Early location already computed
6866
6867 // Compute latest point this Node can go
6868 Node *LCA = get_late_ctrl( n, early );
6869 // LCA is null due to uses being dead
6870 if( LCA == nullptr ) {
6871 #ifdef ASSERT
6872 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
6873 assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6874 }
6875 #endif
6876 _loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6877 _deadlist.push(n);
6878 return;
6879 }
6880 assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6881
6882 Node *legal = LCA; // Walk 'legal' up the IDOM chain
6883 Node *least = legal; // Best legal position so far
6884 while( early != legal ) { // While not at earliest legal
6885 if (legal->is_Start() && !early->is_Root()) {
6886 #ifdef ASSERT
6887 // Bad graph. Print idom path and fail.
6888 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
6889 assert(false, "Bad graph detected in build_loop_late");
6890 #endif
6891 C->record_method_not_compilable("Bad graph detected in build_loop_late");
6892 return;
6893 }
6894 // Find least loop nesting depth
6895 legal = idom(legal); // Bump up the IDOM tree
6896 // Check for lower nesting depth
6897 if( get_loop(legal)->_nest < get_loop(least)->_nest )
6898 least = legal;
6899 }
6900 assert(early == legal || legal != C->root(), "bad dominance of inputs");
6901
6902 if (least != early) {
6903 // Move the node above predicates as far up as possible so a
6904 // following pass of Loop Predication doesn't hoist a predicate
6905 // that depends on it above that node.
6906 const PredicateIterator predicate_iterator(least);
6907 DominatedPredicates dominated_predicates(early, least, this);
6908 predicate_iterator.for_each(dominated_predicates);
6909 least = dominated_predicates.earliest_dominated_predicate_entry();
6910 }
6911 // Try not to place code on a loop entry projection
6912 // which can inhibit range check elimination.
6913 if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {
6914 Node* ctrl_out = least->unique_ctrl_out_or_null();
6915 if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6916 least == ctrl_out->in(LoopNode::EntryControl) &&
6917 (ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {
6918 Node* least_dom = idom(least);
6919 if (get_loop(least_dom)->is_member(get_loop(least))) {
6920 least = least_dom;
6921 }
6922 }
6923 }
6924 // Don't extend live ranges of raw oops
6925 if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6926 !n->bottom_type()->isa_rawptr()) {
6927 least = early;
6928 }
6929
6930 #ifdef ASSERT
6931 // Broken part of VerifyLoopOptimizations (F)
6932 // Reason:
6933 // _verify_me->get_ctrl_no_update(n) seems to return wrong result
6934 /*
6935 // If verifying, verify that 'verify_me' has a legal location
6936 // and choose it as our location.
6937 if( _verify_me ) {
6938 Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6939 Node *legal = LCA;
6940 while( early != legal ) { // While not at earliest legal
6941 if( legal == v_ctrl ) break; // Check for prior good location
6942 legal = idom(legal) ;// Bump up the IDOM tree
6943 }
6944 // Check for prior good location
6945 if( legal == v_ctrl ) least = legal; // Keep prior if found
6946 }
6947 */
6948 #endif
6949
6950 // Assign discovered "here or above" point
6951 least = find_non_split_ctrl(least);
6952 verify_strip_mined_scheduling(n, least);
6953 set_ctrl(n, least);
6954
6955 // Collect inner loop bodies
6956 IdealLoopTree *chosen_loop = get_loop(least);
6957 if( !chosen_loop->_child ) // Inner loop?
6958 chosen_loop->_body.push(n);// Collect inner loops
6959
6960 if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {
6961 _zero_trip_guard_opaque_nodes.push(n);
6962 }
6963
6964 if (!_verify_only && n->Opcode() == Op_OpaqueMultiversioning) {
6965 _multiversion_opaque_nodes.push(n);
6966 }
6967 }
6968
6969 #ifdef ASSERT
6970 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
6971 tty->print_cr("%s", msg);
6972 tty->print("n: "); n->dump();
6973 tty->print("early(n): "); early->dump();
6974 if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6975 n->in(0) != early && !n->in(0)->is_Root()) {
6976 tty->print("n->in(0): "); n->in(0)->dump();
6977 }
6978 for (uint i = 1; i < n->req(); i++) {
6979 Node* in1 = n->in(i);
6980 if (in1 != nullptr && in1 != n && !in1->is_top()) {
6981 tty->print("n->in(%d): ", i); in1->dump();
6982 Node* in1_early = get_ctrl(in1);
6983 tty->print("early(n->in(%d)): ", i); in1_early->dump();
6984 if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6985 in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
6986 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
6987 }
6988 for (uint j = 1; j < in1->req(); j++) {
6989 Node* in2 = in1->in(j);
6990 if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {
6991 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
6992 Node* in2_early = get_ctrl(in2);
6993 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
6994 if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
6995 in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
6996 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
6997 }
6998 }
6999 }
7000 }
7001 }
7002 tty->cr();
7003 tty->print("LCA(n): "); LCA->dump();
7004 for (uint i = 0; i < n->outcnt(); i++) {
7005 Node* u1 = n->raw_out(i);
7006 if (u1 == n)
7007 continue;
7008 tty->print("n->out(%d): ", i); u1->dump();
7009 if (u1->is_CFG()) {
7010 for (uint j = 0; j < u1->outcnt(); j++) {
7011 Node* u2 = u1->raw_out(j);
7012 if (u2 != u1 && u2 != n && u2->is_CFG()) {
7013 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
7014 }
7015 }
7016 } else {
7017 Node* u1_later = get_ctrl(u1);
7018 tty->print("later(n->out(%d)): ", i); u1_later->dump();
7019 if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
7020 u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
7021 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
7022 }
7023 for (uint j = 0; j < u1->outcnt(); j++) {
7024 Node* u2 = u1->raw_out(j);
7025 if (u2 == n || u2 == u1)
7026 continue;
7027 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
7028 if (!u2->is_CFG()) {
7029 Node* u2_later = get_ctrl(u2);
7030 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
7031 if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
7032 u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
7033 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
7034 }
7035 }
7036 }
7037 }
7038 }
7039 dump_idoms(early, LCA);
7040 tty->cr();
7041 }
7042
7043 // Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
7044 class RealLCA {
7045 const PhaseIdealLoop* _phase;
7046 Node* _early;
7047 Node* _wrong_lca;
7048 uint _early_index;
7049 int _wrong_lca_index;
7050
7051 // Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
7052 // is different: Return the previously visited node which must be the real LCA.
7053 // The node lists also contain _early and _wrong_lca, respectively.
7054 Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {
7055 int early_index = early_with_idoms.size() - 1;
7056 int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
7057 bool found_difference = false;
7058 do {
7059 if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {
7060 // First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
7061 found_difference = true;
7062 break;
7063 }
7064 early_index--;
7065 wrong_lca_index--;
7066 } while (wrong_lca_index >= 0);
7067
7068 assert(early_index >= 0, "must always find an LCA - cannot be early");
7069 _early_index = early_index;
7070 _wrong_lca_index = wrong_lca_index;
7071 Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
7072 assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
7073 return real_lca;
7074 }
7075
7076 void dump(Node* real_lca) {
7077 tty->cr();
7078 tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());
7079 _phase->dump_idom(_early, _early_index + 1);
7080
7081 tty->cr();
7082 tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());
7083 _phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
7084
7085 tty->cr();
7086 tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",
7087 _early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
7088 if (_wrong_lca_index >= 0) {
7089 tty->print(" (idom[%d])", _wrong_lca_index);
7090 }
7091 tty->print_cr(":");
7092 real_lca->dump();
7093 }
7094
7095 public:
7096 RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
7097 : _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {
7098 assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
7099 }
7100
7101 void compute_and_dump() {
7102 ResourceMark rm;
7103 Unique_Node_List early_with_idoms;
7104 Unique_Node_List wrong_lca_with_idoms;
7105 early_with_idoms.push(_early);
7106 wrong_lca_with_idoms.push(_wrong_lca);
7107 _phase->get_idoms(_early, 10000, early_with_idoms);
7108 _phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
7109 Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
7110 dump(real_lca);
7111 }
7112 };
7113
7114 // Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
7115 void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {
7116 assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
7117 assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
7118
7119 RealLCA real_lca(this, early, wrong_lca);
7120 real_lca.compute_and_dump();
7121 }
7122 #endif // ASSERT
7123
7124 #ifndef PRODUCT
7125 //------------------------------dump-------------------------------------------
7126 void PhaseIdealLoop::dump() const {
7127 ResourceMark rm;
7128 Node_Stack stack(C->live_nodes() >> 2);
7129 Node_List rpo_list;
7130 VectorSet visited;
7131 visited.set(C->top()->_idx);
7132 rpo(C->root(), stack, visited, rpo_list);
7133 // Dump root loop indexed by last element in PO order
7134 dump(_ltree_root, rpo_list.size(), rpo_list);
7135 }
7136
7137 void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {
7138 loop->dump_head();
7139
7140 // Now scan for CFG nodes in the same loop
7141 for (uint j = idx; j > 0; j--) {
7142 Node* n = rpo_list[j-1];
7143 if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
7144 continue;
7145
7146 if (get_loop(n) != loop) { // Wrong loop nest
7147 if (get_loop(n)->_head == n && // Found nested loop?
7148 get_loop(n)->_parent == loop)
7149 dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
7150 continue;
7151 }
7152
7153 // Dump controlling node
7154 tty->sp(2 * loop->_nest);
7155 tty->print("C");
7156 if (n == C->root()) {
7157 n->dump();
7158 } else {
7159 Node* cached_idom = idom_no_update(n);
7160 Node* computed_idom = n->in(0);
7161 if (n->is_Region()) {
7162 computed_idom = compute_idom(n);
7163 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
7164 // any MultiBranch ctrl node), so apply a similar transform to
7165 // the cached idom returned from idom_no_update.
7166 cached_idom = find_non_split_ctrl(cached_idom);
7167 }
7168 tty->print(" ID:%d", computed_idom->_idx);
7169 n->dump();
7170 if (cached_idom != computed_idom) {
7171 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",
7172 computed_idom->_idx, cached_idom->_idx);
7173 }
7174 }
7175 // Dump nodes it controls
7176 for (uint k = 0; k < _loop_or_ctrl.max(); k++) {
7177 // (k < C->unique() && get_ctrl(find(k)) == n)
7178 if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {
7179 Node* m = C->root()->find(k);
7180 if (m && m->outcnt() > 0) {
7181 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
7182 tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",
7183 _loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
7184 }
7185 tty->sp(2 * loop->_nest + 1);
7186 m->dump();
7187 }
7188 }
7189 }
7190 }
7191 }
7192
7193 void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {
7194 if (has_ctrl(n)) {
7195 tty->print_cr("No idom for data nodes");
7196 } else {
7197 ResourceMark rm;
7198 Unique_Node_List idoms;
7199 get_idoms(n, count, idoms);
7200 dump_idoms_in_reverse(n, idoms);
7201 }
7202 }
7203
7204 void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {
7205 Node* next = n;
7206 for (uint i = 0; !next->is_Start() && i < count; i++) {
7207 next = idom(next);
7208 assert(!idoms.member(next), "duplicated idom is not possible");
7209 idoms.push(next);
7210 }
7211 }
7212
7213 void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {
7214 Node* next;
7215 uint padding = 3;
7216 uint node_index_padding_width = (C->unique() == 0 ? 0 : static_cast<int>(log10(static_cast<double>(C->unique())))) + 1;
7217 for (int i = idom_list.size() - 1; i >= 0; i--) {
7218 if (i == 9 || i == 99) {
7219 padding++;
7220 }
7221 next = idom_list[i];
7222 tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());
7223 }
7224 tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());
7225 }
7226 #endif // NOT PRODUCT
7227
7228 // Collect a R-P-O for the whole CFG.
7229 // Result list is in post-order (scan backwards for RPO)
7230 void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {
7231 stk.push(start, 0);
7232 visited.set(start->_idx);
7233
7234 while (stk.is_nonempty()) {
7235 Node* m = stk.node();
7236 uint idx = stk.index();
7237 if (idx < m->outcnt()) {
7238 stk.set_index(idx + 1);
7239 Node* n = m->raw_out(idx);
7240 if (n->is_CFG() && !visited.test_set(n->_idx)) {
7241 stk.push(n, 0);
7242 }
7243 } else {
7244 rpo_list.push(m);
7245 stk.pop();
7246 }
7247 }
7248 }
7249
7250 ConINode* PhaseIdealLoop::intcon(jint i) {
7251 ConINode* node = _igvn.intcon(i);
7252 set_root_as_ctrl(node);
7253 return node;
7254 }
7255
7256 ConLNode* PhaseIdealLoop::longcon(jlong i) {
7257 ConLNode* node = _igvn.longcon(i);
7258 set_root_as_ctrl(node);
7259 return node;
7260 }
7261
7262 ConNode* PhaseIdealLoop::makecon(const Type* t) {
7263 ConNode* node = _igvn.makecon(t);
7264 set_root_as_ctrl(node);
7265 return node;
7266 }
7267
7268 ConNode* PhaseIdealLoop::integercon(jlong l, BasicType bt) {
7269 ConNode* node = _igvn.integercon(l, bt);
7270 set_root_as_ctrl(node);
7271 return node;
7272 }
7273
7274 ConNode* PhaseIdealLoop::zerocon(BasicType bt) {
7275 ConNode* node = _igvn.zerocon(bt);
7276 set_root_as_ctrl(node);
7277 return node;
7278 }
7279
7280
7281 //=============================================================================
7282 //------------------------------LoopTreeIterator-------------------------------
7283
7284 // Advance to next loop tree using a preorder, left-to-right traversal.
7285 void LoopTreeIterator::next() {
7286 assert(!done(), "must not be done.");
7287 if (_curnt->_child != nullptr) {
7288 _curnt = _curnt->_child;
7289 } else if (_curnt->_next != nullptr) {
7290 _curnt = _curnt->_next;
7291 } else {
7292 while (_curnt != _root && _curnt->_next == nullptr) {
7293 _curnt = _curnt->_parent;
7294 }
7295 if (_curnt == _root) {
7296 _curnt = nullptr;
7297 assert(done(), "must be done.");
7298 } else {
7299 assert(_curnt->_next != nullptr, "must be more to do");
7300 _curnt = _curnt->_next;
7301 }
7302 }
7303 }