1 /*
2 * Copyright (c) 2000, 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 "compiler/compileLog.hpp"
26 #include "gc/shared/barrierSet.hpp"
27 #include "gc/shared/c2/barrierSetC2.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/convertnode.hpp"
34 #include "opto/divnode.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/mulnode.hpp"
38 #include "opto/opaquenode.hpp"
39 #include "opto/phase.hpp"
40 #include "opto/predicates.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/runtime.hpp"
43 #include "opto/subnode.hpp"
44 #include "opto/superword.hpp"
45 #include "opto/vectornode.hpp"
46 #include "runtime/globals_extension.hpp"
47 #include "runtime/stubRoutines.hpp"
48
49 //------------------------------is_loop_exit-----------------------------------
50 // Given an IfNode, return the loop-exiting projection or null if both
51 // arms remain in the loop.
52 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
53 if (iff->outcnt() != 2) return nullptr; // Ignore partially dead tests
54 PhaseIdealLoop *phase = _phase;
55 // Test is an IfNode, has 2 projections. If BOTH are in the loop
56 // we need loop unswitching instead of peeling.
57 if (!is_member(phase->get_loop(iff->raw_out(0))))
58 return iff->raw_out(0);
59 if (!is_member(phase->get_loop(iff->raw_out(1))))
60 return iff->raw_out(1);
61 return nullptr;
62 }
63
64
65 //=============================================================================
66
67
68 //------------------------------record_for_igvn----------------------------
69 // Put loop body on igvn work list
70 void IdealLoopTree::record_for_igvn() {
71 for (uint i = 0; i < _body.size(); i++) {
72 Node *n = _body.at(i);
73 _phase->_igvn._worklist.push(n);
74 }
75 // put body of outer strip mined loop on igvn work list as well
76 if (_head->is_CountedLoop() && _head->as_Loop()->is_strip_mined()) {
77 CountedLoopNode* l = _head->as_CountedLoop();
78 Node* outer_loop = l->outer_loop();
79 assert(outer_loop != nullptr, "missing piece of strip mined loop");
80 _phase->_igvn._worklist.push(outer_loop);
81 Node* outer_loop_tail = l->outer_loop_tail();
82 assert(outer_loop_tail != nullptr, "missing piece of strip mined loop");
83 _phase->_igvn._worklist.push(outer_loop_tail);
84 Node* outer_loop_end = l->outer_loop_end();
85 assert(outer_loop_end != nullptr, "missing piece of strip mined loop");
86 _phase->_igvn._worklist.push(outer_loop_end);
87 Node* outer_safepoint = l->outer_safepoint();
88 assert(outer_safepoint != nullptr, "missing piece of strip mined loop");
89 _phase->_igvn._worklist.push(outer_safepoint);
90 IfFalseNode* cle_out = _head->as_CountedLoop()->loopexit()->false_proj();
91 assert(cle_out != nullptr, "missing piece of strip mined loop");
92 _phase->_igvn._worklist.push(cle_out);
93 }
94 }
95
96 //------------------------------compute_exact_trip_count-----------------------
97 // Compute loop trip count if possible. Do not recalculate trip count for
98 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
99 void IdealLoopTree::compute_trip_count(PhaseIdealLoop* phase, BasicType loop_bt) {
100 if (!_head->as_Loop()->is_valid_counted_loop(loop_bt)) {
101 return;
102 }
103 BaseCountedLoopNode* cl = _head->as_BaseCountedLoop();
104 // Trip count may become nonexact for iteration split loops since
105 // RCE modifies limits. Note, _trip_count value is not reset since
106 // it is used to limit unrolling of main loop.
107 cl->set_nonexact_trip_count();
108
109 // Loop's test should be part of loop.
110 if (!phase->ctrl_is_member(this, cl->loopexit()->in(CountedLoopEndNode::TestValue)))
111 return; // Infinite loop
112
113 #ifdef ASSERT
114 BoolTest::mask bt = cl->loopexit()->test_trip();
115 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
116 bt == BoolTest::ne, "canonical test is expected");
117 #endif
118
119 Node* init_n = cl->init_trip();
120 Node* limit_n = cl->limit();
121 if (init_n != nullptr && limit_n != nullptr) {
122 jlong stride_con = cl->stride_con();
123 const TypeInteger* init_type = phase->_igvn.type(init_n)->is_integer(loop_bt);
124 const TypeInteger* limit_type = phase->_igvn.type(limit_n)->is_integer(loop_bt);
125
126 // compute trip count
127 // It used to be computed as:
128 // max(1, limit_con - init_con + stride_m) / stride_con
129 // with stride_m = stride_con - (stride_con > 0 ? 1 : -1)
130 // for int counted loops only and by promoting all values to long to avoid overflow
131 // This implements the computation for int and long counted loops in a way that promotion to the next larger integer
132 // type is not needed to protect against overflow.
133 //
134 // Use unsigned longs to avoid overflow: number of iteration is a positive number but can be really large for
135 // instance if init_con = min_jint, limit_con = max_jint
136 jlong init_con = (stride_con > 0) ? init_type->lo_as_long() : init_type->hi_as_long();
137 julong uinit_con = init_con;
138 jlong limit_con = (stride_con > 0) ? limit_type->hi_as_long() : limit_type->lo_as_long();
139 julong ulimit_con = limit_con;
140 // The loop body is always executed at least once even if init >= limit (for stride_con > 0) or
141 // init <= limit (for stride_con < 0).
142 julong udiff = 1;
143 if (stride_con > 0 && limit_con > init_con) {
144 udiff = ulimit_con - uinit_con;
145 } else if (stride_con < 0 && limit_con < init_con) {
146 udiff = uinit_con - ulimit_con;
147 }
148 // The loop runs for one more iteration if the limit is (stride > 0 in this example):
149 // init + k * stride + small_value, 0 < small_value < stride
150 julong utrip_count = udiff / ABS(stride_con);
151 if (utrip_count * ABS(stride_con) != udiff) {
152 // Guaranteed to not overflow because it can only happen for ABS(stride) > 1 in which case, utrip_count can't be
153 // max_juint/max_julong
154 utrip_count++;
155 }
156
157 #ifdef ASSERT
158 if (loop_bt == T_INT) {
159 // Use longs to avoid integer overflow.
160 jlong init_con = (stride_con > 0) ? init_type->is_int()->_lo : init_type->is_int()->_hi;
161 jlong limit_con = (stride_con > 0) ? limit_type->is_int()->_hi : limit_type->is_int()->_lo;
162 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
163 jlong trip_count = (limit_con - init_con + stride_m) / stride_con;
164 // The loop body is always executed at least once even if init >= limit (for stride_con > 0) or
165 // init <= limit (for stride_con < 0).
166 trip_count = MAX2(trip_count, (jlong)1);
167 assert(checked_cast<juint>(trip_count) == checked_cast<juint>(utrip_count), "incorrect trip count computation");
168 }
169 #endif
170
171 if (utrip_count < max_unsigned_integer(loop_bt)) {
172 if (init_n->is_Con() && limit_n->is_Con()) {
173 // Set exact trip count.
174 cl->set_exact_trip_count(utrip_count);
175 } else if (loop_bt == T_LONG || cl->as_CountedLoop()->unrolled_count() == 1) {
176 // Set maximum trip count before unrolling.
177 cl->set_trip_count(utrip_count);
178 }
179 }
180 }
181 }
182
183 //------------------------------compute_profile_trip_cnt----------------------------
184 // Compute loop trip count from profile data as
185 // (backedge_count + loop_exit_count) / loop_exit_count
186
187 float IdealLoopTree::compute_profile_trip_cnt_helper(Node* n) {
188 if (n->is_If()) {
189 IfNode *iff = n->as_If();
190 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
191 Node *exit = is_loop_exit(iff);
192 if (exit) {
193 float exit_prob = iff->_prob;
194 if (exit->Opcode() == Op_IfFalse) {
195 exit_prob = 1.0 - exit_prob;
196 }
197 if (exit_prob > PROB_MIN) {
198 float exit_cnt = iff->_fcnt * exit_prob;
199 return exit_cnt;
200 }
201 }
202 }
203 }
204 if (n->is_Jump()) {
205 JumpNode *jmp = n->as_Jump();
206 if (jmp->_fcnt != COUNT_UNKNOWN) {
207 float* probs = jmp->_probs;
208 float exit_prob = 0;
209 PhaseIdealLoop *phase = _phase;
210 for (DUIterator_Fast imax, i = jmp->fast_outs(imax); i < imax; i++) {
211 JumpProjNode* u = jmp->fast_out(i)->as_JumpProj();
212 if (!is_member(_phase->get_loop(u))) {
213 exit_prob += probs[u->_con];
214 }
215 }
216 return exit_prob * jmp->_fcnt;
217 }
218 }
219 return 0;
220 }
221
222 void IdealLoopTree::compute_profile_trip_cnt(PhaseIdealLoop *phase) {
223 if (!_head->is_Loop()) {
224 return;
225 }
226 LoopNode* head = _head->as_Loop();
227 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
228 return; // Already computed
229 }
230 float trip_cnt = (float)max_jint; // default is big
231
232 Node* back = head->in(LoopNode::LoopBackControl);
233 while (back != head) {
234 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
235 back->in(0) &&
236 back->in(0)->is_If() &&
237 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
238 back->in(0)->as_If()->_prob != PROB_UNKNOWN &&
239 (back->Opcode() == Op_IfTrue ? 1-back->in(0)->as_If()->_prob : back->in(0)->as_If()->_prob) > PROB_MIN) {
240 break;
241 }
242 back = phase->idom(back);
243 }
244 if (back != head) {
245 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
246 back->in(0), "if-projection exists");
247 IfNode* back_if = back->in(0)->as_If();
248 float loop_back_cnt = back_if->_fcnt * (back->Opcode() == Op_IfTrue ? back_if->_prob : (1 - back_if->_prob));
249
250 // Now compute a loop exit count
251 float loop_exit_cnt = 0.0f;
252 if (_child == nullptr) {
253 for (uint i = 0; i < _body.size(); i++) {
254 Node *n = _body[i];
255 loop_exit_cnt += compute_profile_trip_cnt_helper(n);
256 }
257 } else {
258 ResourceMark rm;
259 Unique_Node_List wq;
260 wq.push(back);
261 for (uint i = 0; i < wq.size(); i++) {
262 Node *n = wq.at(i);
263 assert(n->is_CFG(), "only control nodes");
264 if (n != head) {
265 if (n->is_Region()) {
266 for (uint j = 1; j < n->req(); j++) {
267 wq.push(n->in(j));
268 }
269 } else {
270 loop_exit_cnt += compute_profile_trip_cnt_helper(n);
271 wq.push(n->in(0));
272 }
273 }
274 }
275
276 }
277 if (loop_exit_cnt > 0.0f) {
278 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
279 } else {
280 // No exit count so use
281 trip_cnt = loop_back_cnt;
282 }
283 } else {
284 head->mark_profile_trip_failed();
285 }
286 #ifndef PRODUCT
287 if (TraceProfileTripCount) {
288 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
289 }
290 #endif
291 head->set_profile_trip_cnt(trip_cnt);
292 }
293
294 // Return nonzero index of invariant operand for an associative
295 // binary operation of (nonconstant) invariant and variant values.
296 // Helper for reassociate_invariants.
297 int IdealLoopTree::find_invariant(Node* n, PhaseIdealLoop* phase) {
298 bool in1_invar = this->is_invariant(n->in(1));
299 bool in2_invar = this->is_invariant(n->in(2));
300 if (in1_invar && !in2_invar) return 1;
301 if (!in1_invar && in2_invar) return 2;
302 return 0;
303 }
304
305 // Return TRUE if "n" is an associative cmp node. A cmp node is
306 // associative if it is only used for equals or not-equals
307 // comparisons of integers or longs. We cannot reassociate
308 // non-equality comparisons due to possibility of overflow.
309 bool IdealLoopTree::is_associative_cmp(Node* n) {
310 if (n->Opcode() != Op_CmpI && n->Opcode() != Op_CmpL) {
311 return false;
312 }
313 for (DUIterator i = n->outs(); n->has_out(i); i++) {
314 BoolNode* bool_out = n->out(i)->isa_Bool();
315 if (bool_out == nullptr || !(bool_out->_test._test == BoolTest::eq ||
316 bool_out->_test._test == BoolTest::ne)) {
317 return false;
318 }
319 }
320 return true;
321 }
322
323 // Return TRUE if "n" is an associative binary node. If "base" is
324 // not null, "n" must be re-associative with it.
325 bool IdealLoopTree::is_associative(Node* n, Node* base) {
326 int op = n->Opcode();
327 if (base != nullptr) {
328 assert(is_associative(base), "Base node should be associative");
329 int base_op = base->Opcode();
330 if (base_op == Op_AddI || base_op == Op_SubI || base_op == Op_CmpI) {
331 return op == Op_AddI || op == Op_SubI;
332 }
333 if (base_op == Op_AddL || base_op == Op_SubL || base_op == Op_CmpL) {
334 return op == Op_AddL || op == Op_SubL;
335 }
336 return op == base_op;
337 } else {
338 // Integer "add/sub/mul/and/or/xor" operations are associative. Integer
339 // "cmp" operations are associative if it is an equality comparison.
340 return op == Op_AddI || op == Op_AddL
341 || op == Op_SubI || op == Op_SubL
342 || op == Op_MulI || op == Op_MulL
343 || op == Op_AndI || op == Op_AndL
344 || op == Op_OrI || op == Op_OrL
345 || op == Op_XorI || op == Op_XorL
346 || is_associative_cmp(n);
347 }
348 }
349
350 // Reassociate invariant add and subtract expressions:
351 //
352 // inv1 + (x + inv2) => ( inv1 + inv2) + x
353 // (x + inv2) + inv1 => ( inv1 + inv2) + x
354 // inv1 + (x - inv2) => ( inv1 - inv2) + x
355 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
356 // (x + inv2) - inv1 => (-inv1 + inv2) + x
357 // (x - inv2) + inv1 => ( inv1 - inv2) + x
358 // (x - inv2) - inv1 => (-inv1 - inv2) + x
359 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
360 // inv1 - (x - inv2) => ( inv1 + inv2) - x
361 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
362 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
363 // inv1 - (x + inv2) => ( inv1 - inv2) - x
364 //
365 // Apply the same transformations to == and !=
366 // inv1 == (x + inv2) => ( inv1 - inv2 ) == x
367 // inv1 == (x - inv2) => ( inv1 + inv2 ) == x
368 // inv1 == (inv2 - x) => (-inv1 + inv2 ) == x
369 Node* IdealLoopTree::reassociate_add_sub_cmp(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop* phase) {
370 Node* n2 = n1->in(3 - inv1_idx);
371 bool n1_is_sub = n1->is_Sub() && !n1->is_Cmp();
372 bool n1_is_cmp = n1->is_Cmp();
373 bool n2_is_sub = n2->is_Sub();
374 assert(n1->is_Add() || n1_is_sub || n1_is_cmp, "Target node should be add, subtract, or compare");
375 assert(n2->is_Add() || (n2_is_sub && !n2->is_Cmp()), "Child node should be add or subtract");
376 Node* inv1 = n1->in(inv1_idx);
377 Node* inv2 = n2->in(inv2_idx);
378 Node* x = n2->in(3 - inv2_idx);
379
380 // Determine whether x, inv1, or inv2 should be negative in the transformed
381 // expression
382 bool neg_x = n2_is_sub && inv2_idx == 1;
383 bool neg_inv2 = (n2_is_sub && !n1_is_cmp && inv2_idx == 2) || (n1_is_cmp && !n2_is_sub);
384 bool neg_inv1 = (n1_is_sub && inv1_idx == 2) || (n1_is_cmp && inv2_idx == 1 && n2_is_sub);
385 if (n1_is_sub && inv1_idx == 1) {
386 neg_x = !neg_x;
387 neg_inv2 = !neg_inv2;
388 }
389
390 bool is_int = n2->bottom_type()->isa_int() != nullptr;
391 Node* inv1_c = phase->get_ctrl(inv1);
392 Node* n_inv1;
393 if (neg_inv1) {
394 if (is_int) {
395 n_inv1 = new SubINode(phase->intcon(0), inv1);
396 } else {
397 n_inv1 = new SubLNode(phase->longcon(0L), inv1);
398 }
399 phase->register_new_node(n_inv1, inv1_c);
400 } else {
401 n_inv1 = inv1;
402 }
403
404 Node* inv;
405 if (is_int) {
406 if (neg_inv2) {
407 inv = new SubINode(n_inv1, inv2);
408 } else {
409 inv = new AddINode(n_inv1, inv2);
410 }
411 phase->register_new_node(inv, phase->get_early_ctrl(inv));
412 if (n1_is_cmp) {
413 return new CmpINode(x, inv);
414 }
415 if (neg_x) {
416 return new SubINode(inv, x);
417 } else {
418 return new AddINode(x, inv);
419 }
420 } else {
421 if (neg_inv2) {
422 inv = new SubLNode(n_inv1, inv2);
423 } else {
424 inv = new AddLNode(n_inv1, inv2);
425 }
426 phase->register_new_node(inv, phase->get_early_ctrl(inv));
427 if (n1_is_cmp) {
428 return new CmpLNode(x, inv);
429 }
430 if (neg_x) {
431 return new SubLNode(inv, x);
432 } else {
433 return new AddLNode(x, inv);
434 }
435 }
436 }
437
438 // Reassociate invariant binary expressions with add/sub/mul/
439 // and/or/xor/cmp operators.
440 // For add/sub/cmp expressions: see "reassociate_add_sub_cmp"
441 //
442 // For mul/and/or/xor expressions:
443 //
444 // inv1 op (x op inv2) => (inv1 op inv2) op x
445 //
446 Node* IdealLoopTree::reassociate(Node* n1, PhaseIdealLoop *phase) {
447 if (!is_associative(n1) || n1->outcnt() == 0) return nullptr;
448 if (is_invariant(n1)) return nullptr;
449 // Don't mess with add of constant (igvn moves them to expression tree root.)
450 if (n1->is_Add() && n1->in(2)->is_Con()) return nullptr;
451
452 int inv1_idx = find_invariant(n1, phase);
453 if (!inv1_idx) return nullptr;
454 Node* n2 = n1->in(3 - inv1_idx);
455 if (!is_associative(n2, n1)) return nullptr;
456 int inv2_idx = find_invariant(n2, phase);
457 if (!inv2_idx) return nullptr;
458
459 if (!phase->may_require_nodes(10, 10)) return nullptr;
460
461 Node* result = nullptr;
462 switch (n1->Opcode()) {
463 case Op_AddI:
464 case Op_AddL:
465 case Op_SubI:
466 case Op_SubL:
467 case Op_CmpI:
468 case Op_CmpL:
469 result = reassociate_add_sub_cmp(n1, inv1_idx, inv2_idx, phase);
470 break;
471 case Op_MulI:
472 case Op_MulL:
473 case Op_AndI:
474 case Op_AndL:
475 case Op_OrI:
476 case Op_OrL:
477 case Op_XorI:
478 case Op_XorL: {
479 Node* inv1 = n1->in(inv1_idx);
480 Node* inv2 = n2->in(inv2_idx);
481 Node* x = n2->in(3 - inv2_idx);
482 Node* inv = n2->clone_with_data_edge(inv1, inv2);
483 phase->register_new_node(inv, phase->get_early_ctrl(inv));
484 result = n1->clone_with_data_edge(x, inv);
485 break;
486 }
487 default:
488 ShouldNotReachHere();
489 }
490
491 assert(result != nullptr, "");
492 phase->register_new_node_with_ctrl_of(result, n1);
493 phase->_igvn.replace_node(n1, result);
494 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
495 _body.yank(n1);
496 return result;
497 }
498
499 //---------------------reassociate_invariants-----------------------------
500 // Reassociate invariant expressions:
501 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
502 for (int i = _body.size() - 1; i >= 0; i--) {
503 Node *n = _body.at(i);
504 for (int j = 0; j < 5; j++) {
505 Node* nn = reassociate(n, phase);
506 if (nn == nullptr) break;
507 n = nn; // again
508 }
509 }
510 }
511
512 //------------------------------policy_peeling---------------------------------
513 // Return TRUE if the loop should be peeled, otherwise return FALSE. Peeling
514 // is applicable if we can make a loop-invariant test (usually a null-check)
515 // execute before we enter the loop. When TRUE, the estimated node budget is
516 // also requested.
517 bool IdealLoopTree::policy_peeling(PhaseIdealLoop *phase) {
518 uint estimate = estimate_peeling(phase);
519
520 return estimate == 0 ? false : phase->may_require_nodes(estimate);
521 }
522
523 // Perform actual policy and size estimate for the loop peeling transform, and
524 // return the estimated loop size if peeling is applicable, otherwise return
525 // zero. No node budget is allocated.
526 uint IdealLoopTree::estimate_peeling(PhaseIdealLoop *phase) {
527
528 // If nodes are depleted, some transform has miscalculated its needs.
529 assert(!phase->exceeding_node_budget(), "sanity");
530
531 // Peeling does loop cloning which can result in O(N^2) node construction.
532 if (_body.size() > 255 && !StressLoopPeeling) {
533 return 0; // Suppress too large body size.
534 }
535 // Optimistic estimate that approximates loop body complexity via data and
536 // control flow fan-out (instead of using the more pessimistic: BodySize^2).
537 uint estimate = est_loop_clone_sz(2);
538
539 if (phase->exceeding_node_budget(estimate)) {
540 return 0; // Too large to safely clone.
541 }
542
543 // Check for vectorized loops, any peeling done was already applied.
544 if (_head->is_CountedLoop()) {
545 CountedLoopNode* cl = _head->as_CountedLoop();
546 if (cl->is_unroll_only() || cl->trip_count() == 1) {
547 // Peeling is not legal here (cf. assert in do_peeling), we don't even stress peel!
548 return 0;
549 }
550 }
551
552 #ifndef PRODUCT
553 // It is now safe to peel or not.
554 if (StressLoopPeeling) {
555 LoopNode* loop_head = _head->as_Loop();
556 static constexpr uint max_peeling_opportunities = 5;
557 if (loop_head->_stress_peeling_attempts < max_peeling_opportunities) {
558 loop_head->_stress_peeling_attempts++;
559 // In case of stress, let's just pick randomly...
560 return ((phase->C->random() % 2) == 0) ? estimate : 0;
561 }
562 return 0;
563 }
564 // ...otherwise, let's apply our heuristic.
565 #endif
566
567 Node* test = tail();
568
569 while (test != _head) { // Scan till run off top of loop
570 if (test->is_If()) { // Test?
571 Node *ctrl = phase->get_ctrl(test->in(1));
572 if (ctrl->is_top()) {
573 return 0; // Found dead test on live IF? No peeling!
574 }
575 // Standard IF only has one input value to check for loop invariance.
576 assert(test->Opcode() == Op_If ||
577 test->Opcode() == Op_CountedLoopEnd ||
578 test->Opcode() == Op_LongCountedLoopEnd ||
579 test->Opcode() == Op_RangeCheck ||
580 test->Opcode() == Op_ParsePredicate,
581 "Check this code when new subtype is added");
582 // Condition is not a member of this loop?
583 if (!is_member(phase->get_loop(ctrl)) && is_loop_exit(test)) {
584 return estimate; // Found reason to peel!
585 }
586 }
587 // Walk up dominators to loop _head looking for test which is executed on
588 // every path through the loop.
589 test = phase->idom(test);
590 }
591 return 0;
592 }
593
594 //------------------------------peeled_dom_test_elim---------------------------
595 // If we got the effect of peeling, either by actually peeling or by making
596 // a pre-loop which must execute at least once, we can remove all
597 // loop-invariant dominated tests in the main body.
598 void PhaseIdealLoop::peeled_dom_test_elim(IdealLoopTree* loop, Node_List& old_new) {
599 bool progress = true;
600 while (progress) {
601 progress = false; // Reset for next iteration
602 Node* prev = loop->_head->in(LoopNode::LoopBackControl); // loop->tail();
603 Node* test = prev->in(0);
604 while (test != loop->_head) { // Scan till run off top of loop
605 int p_op = prev->Opcode();
606 assert(test != nullptr, "test cannot be null");
607 Node* test_cond = nullptr;
608 if ((p_op == Op_IfFalse || p_op == Op_IfTrue) && test->is_If()) {
609 test_cond = test->in(1);
610 }
611 if (test_cond != nullptr && // Test?
612 !test_cond->is_Con() && // And not already obvious?
613 // And condition is not a member of this loop?
614 !ctrl_is_member(loop, test_cond)) {
615 // Walk loop body looking for instances of this test
616 for (uint i = 0; i < loop->_body.size(); i++) {
617 Node* n = loop->_body.at(i);
618 // Check against cached test condition because dominated_by()
619 // replaces the test condition with a constant.
620 if (n->is_If() && n->in(1) == test_cond) {
621 // IfNode was dominated by version in peeled loop body
622 progress = true;
623 dominated_by(old_new[prev->_idx]->as_IfProj(), n->as_If());
624 }
625 }
626 }
627 prev = test;
628 test = idom(test);
629 } // End of scan tests in loop
630 } // End of while (progress)
631 }
632
633 //------------------------------do_peeling-------------------------------------
634 // Peel the first iteration of the given loop.
635 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
636 // The pre-loop illegally has 2 control users (old & new loops).
637 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
638 // Do this by making the old-loop fall-in edges act as if they came
639 // around the loopback from the prior iteration (follow the old-loop
640 // backedges) and then map to the new peeled iteration. This leaves
641 // the pre-loop with only 1 user (the new peeled iteration), but the
642 // peeled-loop backedge has 2 users.
643 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
644 // extra backedge user.
645 //
646 // orig
647 //
648 // stmt1
649 // |
650 // v
651 // predicates
652 // |
653 // v
654 // loop<----+
655 // | |
656 // stmt2 |
657 // | |
658 // v |
659 // if ^
660 // / \ |
661 // / \ |
662 // v v |
663 // false true |
664 // / \ |
665 // / ----+
666 // |
667 // v
668 // exit
669 //
670 //
671 // after clone loop
672 //
673 // stmt1
674 // |
675 // v
676 // predicates
677 // / \
678 // clone / \ orig
679 // / \
680 // / \
681 // v v
682 // +---->loop clone loop<----+
683 // | | | |
684 // | stmt2 clone stmt2 |
685 // | | | |
686 // | v v |
687 // ^ if clone If ^
688 // | / \ / \ |
689 // | / \ / \ |
690 // | v v v v |
691 // | true false false true |
692 // | / \ / \ |
693 // +---- \ / ----+
694 // \ /
695 // 1v v2
696 // region
697 // |
698 // v
699 // exit
700 //
701 //
702 // after peel and predicate move
703 //
704 // stmt1
705 // |
706 // v
707 // predicates
708 // /
709 // /
710 // clone / orig
711 // /
712 // / +----------+
713 // / | |
714 // / | |
715 // / | |
716 // v v |
717 // TOP-->loop clone loop<----+ |
718 // | | | |
719 // stmt2 clone stmt2 | |
720 // | | | ^
721 // v v | |
722 // if clone If ^ |
723 // / \ / \ | |
724 // / \ / \ | |
725 // v v v v | |
726 // true false false true | |
727 // | \ / \ | |
728 // | \ / ----+ ^
729 // | \ / |
730 // | 1v v2 |
731 // v region |
732 // | | |
733 // | v |
734 // | exit |
735 // | |
736 // +--------------->-----------------+
737 //
738 //
739 // final graph
740 //
741 // stmt1
742 // |
743 // v
744 // predicates
745 // |
746 // v
747 // stmt2 clone
748 // |
749 // v
750 // if clone
751 // / |
752 // / |
753 // v v
754 // false true
755 // | |
756 // | v
757 // | Initialized Assertion Predicates
758 // | |
759 // | v
760 // | loop<----+
761 // | | |
762 // | stmt2 |
763 // | | |
764 // | v |
765 // v if ^
766 // | / \ |
767 // | / \ |
768 // | v v |
769 // | false true |
770 // | | \ |
771 // v v --+
772 // region
773 // |
774 // v
775 // exit
776 //
777 void PhaseIdealLoop::do_peeling(IdealLoopTree *loop, Node_List &old_new) {
778
779 C->set_major_progress();
780 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
781 // 'pre' loop from the main and the 'pre' can no longer have its
782 // iterations adjusted. Therefore, we need to declare this loop as
783 // no longer a 'main' loop; it will need new pre and post loops before
784 // we can do further RCE.
785 #ifndef PRODUCT
786 if (TraceLoopOpts) {
787 tty->print("Peel ");
788 loop->dump_head();
789 }
790 #endif
791 LoopNode* head = loop->_head->as_Loop();
792
793 C->print_method(PHASE_BEFORE_LOOP_PEELING, 4, head);
794
795 bool counted_loop = head->is_CountedLoop();
796 if (counted_loop) {
797 CountedLoopNode *cl = head->as_CountedLoop();
798 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
799 cl->set_trip_count(cl->trip_count() - 1);
800 if (cl->is_main_loop()) {
801 cl->set_normal_loop();
802 if (cl->is_multiversion()) {
803 // Peeling also destroys the connection of the main loop
804 // to the multiversion_if.
805 cl->set_no_multiversion();
806 }
807 #ifndef PRODUCT
808 if (TraceLoopOpts) {
809 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
810 }
811 #endif
812 }
813 }
814
815 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
816 // The pre-loop illegally has 2 control users (old & new loops).
817 const uint first_node_index_in_post_loop_body = Compile::current()->unique();
818 LoopNode* outer_loop_head = head->skip_strip_mined();
819 clone_loop(loop, old_new, dom_depth(outer_loop_head), ControlAroundStripMined);
820
821 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
822 // Do this by making the old-loop fall-in edges act as if they came
823 // around the loopback from the prior iteration (follow the old-loop
824 // backedges) and then map to the new peeled iteration. This leaves
825 // the pre-loop with only 1 user (the new peeled iteration), but the
826 // peeled-loop backedge has 2 users.
827 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
828 _igvn.hash_delete(outer_loop_head);
829 outer_loop_head->set_req(LoopNode::EntryControl, new_entry);
830 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
831 Node* old = head->fast_out(j);
832 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
833 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
834 if (!new_exit_value) // Backedge value is ALSO loop invariant?
835 // Then loop body backedge value remains the same.
836 new_exit_value = old->in(LoopNode::LoopBackControl);
837 _igvn.hash_delete(old);
838 old->set_req(LoopNode::EntryControl, new_exit_value);
839 }
840 }
841
842
843 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
844 // extra backedge user.
845 Node* new_head = old_new[head->_idx];
846 _igvn.hash_delete(new_head);
847 new_head->set_req(LoopNode::LoopBackControl, C->top());
848 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
849 Node* use = new_head->fast_out(j2);
850 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
851 _igvn.hash_delete(use);
852 use->set_req(LoopNode::LoopBackControl, C->top());
853 }
854 }
855
856 // Step 4: Correct dom-depth info. Set to loop-head depth.
857
858 int dd_outer_loop_head = dom_depth(outer_loop_head);
859 set_idom(outer_loop_head, outer_loop_head->in(LoopNode::EntryControl), dd_outer_loop_head);
860 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
861 Node *old = loop->_body.at(j3);
862 Node *nnn = old_new[old->_idx];
863 if (!has_ctrl(nnn)) {
864 set_idom(nnn, idom(nnn), dd_outer_loop_head-1);
865 }
866 }
867
868 // Step 5: Assertion Predicates initialization
869 if (counted_loop) {
870 CountedLoopNode* cl = head->as_CountedLoop();
871 Node* init = cl->init_trip();
872 Node* init_ctrl = cl->skip_strip_mined()->in(LoopNode::EntryControl);
873 initialize_assertion_predicates_for_peeled_loop(new_head->as_CountedLoop(), cl,
874 first_node_index_in_post_loop_body, old_new);
875 cast_incr_before_loop(init, init_ctrl, cl);
876 }
877
878 // Now force out all loop-invariant dominating tests. The optimizer
879 // finds some, but we _know_ they are all useless.
880 peeled_dom_test_elim(loop,old_new);
881
882 loop->record_for_igvn();
883
884 C->print_method(PHASE_AFTER_LOOP_PEELING, 4, new_head);
885 }
886
887 //------------------------------policy_maximally_unroll------------------------
888 // Calculate the exact loop trip-count and return TRUE if loop can be fully,
889 // i.e. maximally, unrolled, otherwise return FALSE. When TRUE, the estimated
890 // node budget is also requested.
891 bool IdealLoopTree::policy_maximally_unroll(PhaseIdealLoop* phase) const {
892 CountedLoopNode* cl = _head->as_CountedLoop();
893 assert(cl->is_normal_loop(), "");
894 if (!cl->is_valid_counted_loop(T_INT)) {
895 return false; // Malformed counted loop.
896 }
897 if (!cl->has_exact_trip_count()) {
898 return false; // Trip count is not exact.
899 }
900
901 uint trip_count = cl->trip_count();
902 // Note, max_juint is used to indicate unknown trip count.
903 assert(trip_count > 1, "one-iteration loop should be optimized out already");
904 assert(trip_count < max_juint, "exact trip_count should be less than max_juint.");
905
906 // If nodes are depleted, some transform has miscalculated its needs.
907 assert(!phase->exceeding_node_budget(), "sanity");
908
909 // Allow the unrolled body to get larger than the standard loop size limit.
910 uint unroll_limit = (uint)LoopUnrollLimit * 4;
911 assert((intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
912 if (trip_count > unroll_limit || _body.size() > unroll_limit) {
913 return false;
914 }
915
916 uint new_body_size = est_loop_unroll_sz(trip_count);
917
918 if (new_body_size == UINT_MAX) { // Check for bad estimate (overflow).
919 return false;
920 }
921
922 // Fully unroll a loop with few iterations, regardless of other conditions,
923 // since the following (general) loop optimizations will split such loop in
924 // any case (into pre-main-post).
925 if (trip_count <= 3) {
926 return phase->may_require_nodes(new_body_size);
927 }
928
929 // Reject if unrolling will result in too much node construction.
930 if (new_body_size > unroll_limit || phase->exceeding_node_budget(new_body_size)) {
931 return false;
932 }
933
934 // Do not unroll a loop with String intrinsics code.
935 // String intrinsics are large and have loops.
936 for (uint k = 0; k < _body.size(); k++) {
937 Node* n = _body.at(k);
938 switch (n->Opcode()) {
939 case Op_StrComp:
940 case Op_StrEquals:
941 case Op_VectorizedHashCode:
942 case Op_StrIndexOf:
943 case Op_StrIndexOfChar:
944 case Op_EncodeISOArray:
945 case Op_AryEq:
946 case Op_CountPositives: {
947 return false;
948 }
949 } // switch
950 }
951
952 return phase->may_require_nodes(new_body_size);
953 }
954
955
956 //------------------------------policy_unroll----------------------------------
957 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll if
958 // the loop is a counted loop and the loop body is small enough. When TRUE,
959 // the estimated node budget is also requested.
960 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
961
962 CountedLoopNode *cl = _head->as_CountedLoop();
963 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
964
965 if (!cl->is_valid_counted_loop(T_INT)) {
966 return false; // Malformed counted loop
967 }
968
969 // If nodes are depleted, some transform has miscalculated its needs.
970 assert(!phase->exceeding_node_budget(), "sanity");
971
972 // Protect against over-unrolling.
973 // After split at least one iteration will be executed in pre-loop.
974 if (cl->trip_count() <= (cl->is_normal_loop() ? 2u : 1u)) {
975 return false;
976 }
977 _local_loop_unroll_limit = LoopUnrollLimit;
978 _local_loop_unroll_factor = 4;
979 int future_unroll_cnt = cl->unrolled_count() * 2;
980 if (!cl->is_vectorized_loop()) {
981 if (future_unroll_cnt > LoopMaxUnroll) return false;
982 } else {
983 // obey user constraints on vector mapped loops with additional unrolling applied
984 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1;
985 if ((future_unroll_cnt / unroll_constraint) > LoopMaxUnroll) return false;
986 }
987
988 const int stride_con = cl->stride_con();
989
990 // Check for initial stride being a small enough constant
991 const int initial_stride_sz = MAX2(1<<2, Matcher::max_vector_size(T_BYTE) / 2);
992 // Maximum stride size should protect against overflow, when doubling stride unroll_count times
993 const int max_stride_size = MIN2<int>(max_jint / 2 - 2, initial_stride_sz * future_unroll_cnt);
994 // No abs() use; abs(min_jint) = min_jint
995 if (stride_con < -max_stride_size || stride_con > max_stride_size) return false;
996
997 // Don't unroll if the next round of unrolling would push us
998 // over the expected trip count of the loop. One is subtracted
999 // from the expected trip count because the pre-loop normally
1000 // executes 1 iteration.
1001 if (UnrollLimitForProfileCheck > 0 &&
1002 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
1003 future_unroll_cnt > UnrollLimitForProfileCheck &&
1004 (float)future_unroll_cnt > cl->profile_trip_cnt() - 1.0) {
1005 return false;
1006 }
1007
1008 bool should_unroll = true;
1009
1010 // When unroll count is greater than LoopUnrollMin, don't unroll if:
1011 // the residual iterations are more than 10% of the trip count
1012 // and rounds of "unroll,optimize" are not making significant progress
1013 // Progress defined as current size less than 20% larger than previous size.
1014 if (phase->C->do_superword() &&
1015 cl->node_count_before_unroll() > 0 &&
1016 future_unroll_cnt > LoopUnrollMin &&
1017 is_residual_iters_large(future_unroll_cnt, cl) &&
1018 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
1019 if ((cl->slp_max_unroll() == 0) && !is_residual_iters_large(cl->unrolled_count(), cl)) {
1020 // cl->slp_max_unroll() = 0 means that the previous slp analysis never passed.
1021 // slp analysis may fail due to the loop IR is too complicated especially during the early stage
1022 // of loop unrolling analysis. But after several rounds of loop unrolling and other optimizations,
1023 // it's possible that the loop IR becomes simple enough to pass the slp analysis.
1024 // So we don't return immediately in hoping that the next slp analysis can succeed.
1025 should_unroll = false;
1026 future_unroll_cnt = cl->unrolled_count();
1027 } else {
1028 return false;
1029 }
1030 }
1031
1032 Node *init_n = cl->init_trip();
1033 Node *limit_n = cl->limit();
1034 if (limit_n == nullptr) return false; // We will dereference it below.
1035
1036 // Non-constant bounds.
1037 // Protect against over-unrolling when init or/and limit are not constant
1038 // (so that trip_count's init value is maxint) but iv range is known.
1039 if (init_n == nullptr || !init_n->is_Con() || !limit_n->is_Con()) {
1040 Node* phi = cl->phi();
1041 if (phi != nullptr) {
1042 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
1043 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
1044 int next_stride = stride_con * 2; // stride after this unroll
1045 if (next_stride > 0) {
1046 if (iv_type->_lo > max_jint - next_stride || // overflow
1047 iv_type->_lo + next_stride > iv_type->_hi) {
1048 return false; // over-unrolling
1049 }
1050 } else if (next_stride < 0) {
1051 if (iv_type->_hi < min_jint - next_stride || // overflow
1052 iv_type->_hi + next_stride < iv_type->_lo) {
1053 return false; // over-unrolling
1054 }
1055 }
1056 }
1057 }
1058
1059 // After unroll limit will be adjusted: new_limit = limit-stride.
1060 // Bailout if adjustment overflow.
1061 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
1062 if ((stride_con > 0 && ((min_jint + stride_con) > limit_type->_hi)) ||
1063 (stride_con < 0 && ((max_jint + stride_con) < limit_type->_lo)))
1064 return false; // overflow
1065
1066 // Rudimentary cost model to estimate loop unrolling
1067 // factor.
1068 // Adjust body_size to determine if we unroll or not
1069 uint body_size = _body.size();
1070 // Key test to unroll loop in CRC32 java code
1071 int xors_in_loop = 0;
1072 // Also count ModL, DivL, MulL, and other nodes that expand mightly
1073 for (uint k = 0; k < _body.size(); k++) {
1074 Node* n = _body.at(k);
1075 if (MemNode::barrier_data(n) != 0) {
1076 body_size += BarrierSet::barrier_set()->barrier_set_c2()->estimated_barrier_size(n);
1077 }
1078 switch (n->Opcode()) {
1079 case Op_XorI: xors_in_loop++; break; // CRC32 java code
1080 case Op_ModL: body_size += 30; break;
1081 case Op_DivL: body_size += 30; break;
1082 case Op_MulL: body_size += 10; break;
1083 case Op_RoundF:
1084 case Op_RoundD: {
1085 body_size += Matcher::scalar_op_pre_select_sz_estimate(n->Opcode(), n->bottom_type()->basic_type());
1086 } break;
1087 case Op_CountTrailingZerosV:
1088 case Op_CountLeadingZerosV:
1089 case Op_LoadVectorGather:
1090 case Op_LoadVectorGatherMasked:
1091 case Op_ReverseV:
1092 case Op_RoundVF:
1093 case Op_RoundVD:
1094 case Op_VectorCastD2X:
1095 case Op_VectorCastF2X:
1096 case Op_PopCountVI:
1097 case Op_PopCountVL: {
1098 const TypeVect* vt = n->bottom_type()->is_vect();
1099 body_size += Matcher::vector_op_pre_select_sz_estimate(n->Opcode(), vt->element_basic_type(), vt->length());
1100 } break;
1101 case Op_StrComp:
1102 case Op_StrEquals:
1103 case Op_StrIndexOf:
1104 case Op_StrIndexOfChar:
1105 case Op_EncodeISOArray:
1106 case Op_AryEq:
1107 case Op_VectorizedHashCode:
1108 case Op_CountPositives: {
1109 // Do not unroll a loop with String intrinsics code.
1110 // String intrinsics are large and have loops.
1111 return false;
1112 }
1113 } // switch
1114 }
1115
1116 if (phase->C->do_superword()) {
1117 // Only attempt slp analysis when user controls do not prohibit it
1118 if (!range_checks_present() && (LoopMaxUnroll > _local_loop_unroll_factor)) {
1119 // Once policy_slp_analysis succeeds, mark the loop with the
1120 // maximal unroll factor so that we minimize analysis passes
1121 if (future_unroll_cnt >= _local_loop_unroll_factor) {
1122 policy_unroll_slp_analysis(cl, phase, future_unroll_cnt);
1123 }
1124 }
1125 }
1126
1127 int slp_max_unroll_factor = cl->slp_max_unroll();
1128 if ((LoopMaxUnroll < slp_max_unroll_factor) && FLAG_IS_DEFAULT(LoopMaxUnroll) && UseSubwordForMaxVector) {
1129 LoopMaxUnroll = slp_max_unroll_factor;
1130 }
1131
1132 uint estimate = est_loop_clone_sz(2);
1133
1134 if (cl->has_passed_slp()) {
1135 if (slp_max_unroll_factor >= future_unroll_cnt) {
1136 return should_unroll && phase->may_require_nodes(estimate);
1137 }
1138 return false; // Loop too big.
1139 }
1140
1141 // Check for being too big
1142 if (body_size > (uint)_local_loop_unroll_limit) {
1143 if ((cl->is_subword_loop() || xors_in_loop >= 4) && body_size < 4u * LoopUnrollLimit) {
1144 return should_unroll && phase->may_require_nodes(estimate);
1145 }
1146 return false; // Loop too big.
1147 }
1148
1149 if (cl->is_unroll_only()) {
1150 if (TraceSuperWordLoopUnrollAnalysis) {
1151 tty->print_cr("policy_unroll passed vector loop(vlen=%d, factor=%d)\n",
1152 slp_max_unroll_factor, future_unroll_cnt);
1153 }
1154 }
1155
1156 // Unroll once! (Each trip will soon do double iterations)
1157 return should_unroll && phase->may_require_nodes(estimate);
1158 }
1159
1160 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_cnt) {
1161
1162 // If nodes are depleted, some transform has miscalculated its needs.
1163 assert(!phase->exceeding_node_budget(), "sanity");
1164
1165 // Enable this functionality target by target as needed
1166 if (SuperWordLoopUnrollAnalysis) {
1167 if (!cl->was_slp_analyzed()) {
1168 Compile::TracePhase tp(Phase::_t_autoVectorize);
1169
1170 VLoop vloop(this, true);
1171 if (vloop.check_preconditions()) {
1172 SuperWord::unrolling_analysis(vloop, _local_loop_unroll_factor);
1173 }
1174 }
1175
1176 if (cl->has_passed_slp()) {
1177 int slp_max_unroll_factor = cl->slp_max_unroll();
1178 if (slp_max_unroll_factor >= future_unroll_cnt) {
1179 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor;
1180 if (new_limit > LoopUnrollLimit) {
1181 if (TraceSuperWordLoopUnrollAnalysis) {
1182 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit);
1183 }
1184 _local_loop_unroll_limit = new_limit;
1185 }
1186 }
1187 }
1188 }
1189 }
1190
1191
1192 //------------------------------policy_range_check-----------------------------
1193 // Return TRUE or FALSE if the loop should be range-check-eliminated or not.
1194 // When TRUE, the estimated node budget is also requested.
1195 //
1196 // We will actually perform iteration-splitting, a more powerful form of RCE.
1197 bool IdealLoopTree::policy_range_check(PhaseIdealLoop* phase, bool provisional, BasicType bt) const {
1198 if (!provisional && !RangeCheckElimination) return false;
1199
1200 // If nodes are depleted, some transform has miscalculated its needs.
1201 assert(provisional || !phase->exceeding_node_budget(), "sanity");
1202
1203 if (_head->is_CountedLoop()) {
1204 CountedLoopNode *cl = _head->as_CountedLoop();
1205 // If we unrolled with no intention of doing RCE and we later changed our
1206 // minds, we got no pre-loop. Either we need to make a new pre-loop, or we
1207 // have to disallow RCE.
1208 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
1209
1210 // check for vectorized loops, some opts are no longer needed
1211 // RCE needs pre/main/post loops. Don't apply it on a single iteration loop.
1212 if (cl->is_unroll_only() || (cl->is_normal_loop() && cl->trip_count() == 1)) return false;
1213 } else {
1214 assert(provisional, "no long counted loop expected");
1215 }
1216
1217 BaseCountedLoopNode* cl = _head->as_BaseCountedLoop();
1218 Node *trip_counter = cl->phi();
1219 assert(!cl->is_LongCountedLoop() || bt == T_LONG, "only long range checks in long counted loops");
1220 assert(cl->is_valid_counted_loop(cl->bt()), "only for well formed loops");
1221
1222 // Check loop body for tests of trip-counter plus loop-invariant vs
1223 // loop-invariant.
1224 for (uint i = 0; i < _body.size(); i++) {
1225 Node *iff = _body[i];
1226 if (iff->Opcode() == Op_If ||
1227 iff->Opcode() == Op_RangeCheck) { // Test?
1228
1229 // Comparing trip+off vs limit
1230 Node* bol = iff->in(1);
1231 if (bol->req() != 2) {
1232 // Could be a dead constant test or another dead variant (e.g. a Phi with 2 inputs created with split_thru_phi).
1233 // Either way, skip this test.
1234 continue;
1235 }
1236 if (!bol->is_Bool()) {
1237 assert(bol->is_OpaqueNotNull() ||
1238 bol->is_OpaqueTemplateAssertionPredicate() ||
1239 bol->is_OpaqueInitializedAssertionPredicate() ||
1240 bol->is_OpaqueMultiversioning(),
1241 "Opaque node of a non-null-check or an Assertion Predicate or Multiversioning");
1242 continue;
1243 }
1244 if (bol->as_Bool()->_test._test == BoolTest::ne) {
1245 continue; // not RC
1246 }
1247 Node *cmp = bol->in(1);
1248
1249 if (provisional) {
1250 // Try to pattern match with either cmp inputs, do not check
1251 // whether one of the inputs is loop independent as it may not
1252 // have had a chance to be hoisted yet.
1253 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), trip_counter, bt, nullptr, nullptr) &&
1254 !phase->is_scaled_iv_plus_offset(cmp->in(2), trip_counter, bt, nullptr, nullptr)) {
1255 continue;
1256 }
1257 } else {
1258 Node *rc_exp = cmp->in(1);
1259 Node *limit = cmp->in(2);
1260 Node *limit_c = phase->get_ctrl(limit);
1261 if (limit_c == phase->C->top()) {
1262 return false; // Found dead test on live IF? No RCE!
1263 }
1264 if (is_member(phase->get_loop(limit_c))) {
1265 // Compare might have operands swapped; commute them
1266 rc_exp = cmp->in(2);
1267 limit = cmp->in(1);
1268 limit_c = phase->get_ctrl(limit);
1269 if (is_member(phase->get_loop(limit_c))) {
1270 continue; // Both inputs are loop varying; cannot RCE
1271 }
1272 }
1273
1274 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, bt, nullptr, nullptr)) {
1275 continue;
1276 }
1277 }
1278 // Found a test like 'trip+off vs limit'. Test is an IfNode, has two (2)
1279 // projections. If BOTH are in the loop we need loop unswitching instead
1280 // of iteration splitting.
1281 if (is_loop_exit(iff)) {
1282 // Found valid reason to split iterations (if there is room).
1283 // NOTE: Usually a gross overestimate.
1284 // Long range checks cause the loop to be transformed in a loop nest which only causes a fixed number of nodes
1285 // to be added
1286 return provisional || bt == T_LONG || phase->may_require_nodes(est_loop_clone_sz(2));
1287 }
1288 } // End of is IF
1289 }
1290
1291 return false;
1292 }
1293
1294 //------------------------------policy_peel_only-------------------------------
1295 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
1296 // for unrolling loops with NO array accesses.
1297 bool IdealLoopTree::policy_peel_only(PhaseIdealLoop *phase) const {
1298
1299 // If nodes are depleted, some transform has miscalculated its needs.
1300 assert(!phase->exceeding_node_budget(), "sanity");
1301
1302 // check for vectorized loops, any peeling done was already applied
1303 if (_head->is_CountedLoop() && _head->as_CountedLoop()->is_unroll_only()) {
1304 return false;
1305 }
1306
1307 for (uint i = 0; i < _body.size(); i++) {
1308 if (_body[i]->is_Mem()) {
1309 return false;
1310 }
1311 }
1312 // No memory accesses at all!
1313 return true;
1314 }
1315
1316 //------------------------------clone_up_backedge_goo--------------------------
1317 // If Node n lives in the back_ctrl block and cannot float, we clone a private
1318 // version of n in preheader_ctrl block and return that, otherwise return n.
1319 Node *PhaseIdealLoop::clone_up_backedge_goo(Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones) {
1320 if (get_ctrl(n) != back_ctrl) return n;
1321
1322 // Only visit once
1323 if (visited.test_set(n->_idx)) {
1324 Node *x = clones.find(n->_idx);
1325 return (x != nullptr) ? x : n;
1326 }
1327
1328 Node *x = nullptr; // If required, a clone of 'n'
1329 // Check for 'n' being pinned in the backedge.
1330 if (n->in(0) && n->in(0) == back_ctrl) {
1331 assert(clones.find(n->_idx) == nullptr, "dead loop");
1332 x = n->clone(); // Clone a copy of 'n' to preheader
1333 clones.push(x, n->_idx);
1334 x->set_req(0, preheader_ctrl); // Fix x's control input to preheader
1335 }
1336
1337 // Recursive fixup any other input edges into x.
1338 // If there are no changes we can just return 'n', otherwise
1339 // we need to clone a private copy and change it.
1340 for (uint i = 1; i < n->req(); i++) {
1341 Node *g = clone_up_backedge_goo(back_ctrl, preheader_ctrl, n->in(i), visited, clones);
1342 if (g != n->in(i)) {
1343 if (!x) {
1344 assert(clones.find(n->_idx) == nullptr, "dead loop");
1345 x = n->clone();
1346 clones.push(x, n->_idx);
1347 }
1348 x->set_req(i, g);
1349 }
1350 }
1351 if (x) { // x can legally float to pre-header location
1352 register_new_node(x, preheader_ctrl);
1353 return x;
1354 } else { // raise n to cover LCA of uses
1355 set_ctrl(n, find_non_split_ctrl(back_ctrl->in(0)));
1356 }
1357 return n;
1358 }
1359
1360 // When a counted loop is created, the loop phi type may be narrowed down. As a consequence, the control input of some
1361 // nodes may be cleared: in particular in the case of a division by the loop iv, the Div node would lose its control
1362 // dependency if the loop phi is never zero. After pre/main/post loops are created (and possibly unrolling), the
1363 // loop phi type is only correct if the loop is indeed reachable: there's an implicit dependency between the loop phi
1364 // type and the zero trip guard for the main or post loop and as a consequence a dependency between the Div node and the
1365 // zero trip guard. This makes the dependency explicit by adding a CastII for the loop entry input of the loop phi. If
1366 // the backedge of the main or post loop is removed, a Div node won't be able to float above the zero trip guard of the
1367 // loop and can't execute even if the loop is not reached.
1368 void PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, CountedLoopNode* loop) {
1369 Node* castii = new CastIINode(ctrl, incr, TypeInt::INT, ConstraintCastNode::DependencyType::NonFloatingNonNarrowing);
1370 register_new_node(castii, ctrl);
1371 Node* phi = loop->phi();
1372 assert(phi->in(LoopNode::EntryControl) == incr, "replacing wrong input?");
1373 _igvn.replace_input_of(phi, LoopNode::EntryControl, castii);
1374 }
1375
1376 #ifdef ASSERT
1377 void PhaseIdealLoop::ensure_zero_trip_guard_proj(Node* node, bool is_main_loop) {
1378 assert(node->is_IfProj(), "must be the zero trip guard If node");
1379 Node* zer_bol = node->in(0)->in(1);
1380 assert(zer_bol != nullptr && zer_bol->is_Bool(), "must be Bool");
1381 Node* zer_cmp = zer_bol->in(1);
1382 assert(zer_cmp != nullptr && zer_cmp->Opcode() == Op_CmpI, "must be CmpI");
1383 // For the main loop, the opaque node is the second input to zer_cmp, for the post loop it's the first input node
1384 Node* zer_opaq = zer_cmp->in(is_main_loop ? 2 : 1);
1385 assert(zer_opaq != nullptr && zer_opaq->Opcode() == Op_OpaqueZeroTripGuard, "must be OpaqueZeroTripGuard");
1386 }
1387 #endif
1388
1389 //------------------------------insert_pre_post_loops--------------------------
1390 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
1391 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
1392 // alignment. Useful to unroll loops that do no array accesses.
1393 void PhaseIdealLoop::insert_pre_post_loops(IdealLoopTree *loop, Node_List &old_new, bool peel_only) {
1394
1395 #ifndef PRODUCT
1396 if (TraceLoopOpts) {
1397 if (peel_only)
1398 tty->print("PeelMainPost ");
1399 else
1400 tty->print("PreMainPost ");
1401 loop->dump_head();
1402 }
1403 #endif
1404 C->set_major_progress();
1405
1406 // Find common pieces of the loop being guarded with pre & post loops
1407 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1408 assert(main_head->is_normal_loop(), "");
1409 CountedLoopEndNode *main_end = main_head->loopexit();
1410 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1411
1412 C->print_method(PHASE_BEFORE_PRE_MAIN_POST, 4, main_head);
1413
1414 Node *init = main_head->init_trip();
1415 Node *incr = main_end ->incr();
1416 Node *limit = main_end ->limit();
1417 Node *stride = main_end ->stride();
1418 Node *cmp = main_end ->cmp_node();
1419 BoolTest::mask b_test = main_end->test_trip();
1420
1421 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
1422 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
1423 if (bol->outcnt() != 1) {
1424 bol = bol->clone();
1425 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
1426 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
1427 }
1428 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
1429 if (cmp->outcnt() != 1) {
1430 cmp = cmp->clone();
1431 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
1432 _igvn.replace_input_of(bol, 1, cmp);
1433 }
1434
1435 // Add the post loop
1436 CountedLoopNode *post_head = nullptr;
1437 Node* post_incr = incr;
1438 Node* main_exit = insert_post_loop(loop, old_new, main_head, main_end, post_incr, limit, post_head);
1439 C->print_method(PHASE_AFTER_POST_LOOP, 4, post_head);
1440
1441 //------------------------------
1442 // Step B: Create Pre-Loop.
1443
1444 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1445 // loop pre-header illegally has 2 control users (old & new loops).
1446 LoopNode* outer_main_head = main_head;
1447 IdealLoopTree* outer_loop = loop;
1448 if (main_head->is_strip_mined()) {
1449 main_head->verify_strip_mined(1);
1450 outer_main_head = main_head->outer_loop();
1451 outer_loop = loop->_parent;
1452 assert(outer_loop->_head == outer_main_head, "broken loop tree");
1453 }
1454
1455 const uint first_node_index_in_pre_loop_body = Compile::current()->unique();
1456 uint dd_main_head = dom_depth(outer_main_head);
1457 clone_loop(loop, old_new, dd_main_head, ControlAroundStripMined);
1458 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1459 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1460 pre_head->set_pre_loop(main_head);
1461 Node *pre_incr = old_new[incr->_idx];
1462
1463 // Reduce the pre-loop trip count.
1464 pre_end->_prob = PROB_FAIR;
1465
1466 // Find the pre-loop normal exit.
1467 IfFalseNode* pre_exit = pre_end->false_proj();
1468 IfFalseNode* new_pre_exit = new IfFalseNode(pre_end);
1469 _igvn.register_new_node_with_optimizer(new_pre_exit);
1470 set_idom(new_pre_exit, pre_end, dd_main_head);
1471 set_loop(new_pre_exit, outer_loop->_parent);
1472
1473 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1474 // pre-loop, the main-loop may not execute at all. Later in life this
1475 // zero-trip guard will become the minimum-trip guard when we unroll
1476 // the main-loop.
1477 Node *min_opaq = new OpaqueZeroTripGuardNode(C, limit, b_test);
1478 Node *min_cmp = new CmpINode(pre_incr, min_opaq);
1479 Node *min_bol = new BoolNode(min_cmp, b_test);
1480 register_new_node(min_opaq, new_pre_exit);
1481 register_new_node(min_cmp , new_pre_exit);
1482 register_new_node(min_bol , new_pre_exit);
1483
1484 // Build the IfNode (assume the main-loop is executed always).
1485 IfNode *min_iff = new IfNode(new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN);
1486 _igvn.register_new_node_with_optimizer(min_iff);
1487 set_idom(min_iff, new_pre_exit, dd_main_head);
1488 set_loop(min_iff, outer_loop->_parent);
1489
1490 // Plug in the false-path, taken if we need to skip main-loop
1491 _igvn.hash_delete(pre_exit);
1492 pre_exit->set_req(0, min_iff);
1493 set_idom(pre_exit, min_iff, dd_main_head);
1494 set_idom(pre_exit->unique_ctrl_out(), min_iff, dd_main_head);
1495 // Make the true-path, must enter the main loop
1496 Node *min_taken = new IfTrueNode(min_iff);
1497 _igvn.register_new_node_with_optimizer(min_taken);
1498 set_idom(min_taken, min_iff, dd_main_head);
1499 set_loop(min_taken, outer_loop->_parent);
1500 // Plug in the true path
1501 _igvn.hash_delete(outer_main_head);
1502 outer_main_head->set_req(LoopNode::EntryControl, min_taken);
1503 set_idom(outer_main_head, min_taken, dd_main_head);
1504 assert(post_head->in(1)->is_IfProj(), "must be zero-trip guard If node projection of the post loop");
1505
1506 VectorSet visited;
1507 Node_Stack clones(main_head->back_control()->outcnt());
1508 // Step B3: Make the fall-in values to the main-loop come from the
1509 // fall-out values of the pre-loop.
1510 const uint last_node_index_in_pre_loop_body = Compile::current()->unique() - 1;
1511 for (DUIterator i2 = main_head->outs(); main_head->has_out(i2); i2++) {
1512 Node* main_phi = main_head->out(i2);
1513 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0) {
1514 Node* pre_phi = old_new[main_phi->_idx];
1515 Node* fallpre = clone_up_backedge_goo(pre_head->back_control(),
1516 main_head->skip_strip_mined()->in(LoopNode::EntryControl),
1517 pre_phi->in(LoopNode::LoopBackControl),
1518 visited, clones);
1519 _igvn.hash_delete(main_phi);
1520 main_phi->set_req(LoopNode::EntryControl, fallpre);
1521 }
1522 }
1523 DEBUG_ONLY(const uint last_node_index_from_backedge_goo = Compile::current()->unique() - 1);
1524
1525 DEBUG_ONLY(ensure_zero_trip_guard_proj(outer_main_head->in(LoopNode::EntryControl), true);)
1526 initialize_assertion_predicates_for_main_loop(pre_head, main_head, first_node_index_in_pre_loop_body,
1527 last_node_index_in_pre_loop_body,
1528 DEBUG_ONLY(last_node_index_from_backedge_goo COMMA) old_new);
1529 // CastII for the main loop:
1530 cast_incr_before_loop(pre_incr, min_taken, main_head);
1531
1532 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1533 // RCE and alignment may change this later.
1534 Node *cmp_end = pre_end->cmp_node();
1535 assert(cmp_end->in(2) == limit, "");
1536 Node *pre_limit = new AddINode(init, stride);
1537
1538 // Save the original loop limit in this Opaque1 node for
1539 // use by range check elimination.
1540 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit);
1541
1542 register_new_node(pre_limit, pre_head->in(LoopNode::EntryControl));
1543 register_new_node(pre_opaq , pre_head->in(LoopNode::EntryControl));
1544
1545 // Since no other users of pre-loop compare, I can hack limit directly
1546 assert(cmp_end->outcnt() == 1, "no other users");
1547 _igvn.hash_delete(cmp_end);
1548 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1549
1550 // Special case for not-equal loop bounds:
1551 // Change pre loop test, main loop test, and the
1552 // main loop guard test to use lt or gt depending on stride
1553 // direction:
1554 // positive stride use <
1555 // negative stride use >
1556 //
1557 // not-equal test is kept for post loop to handle case
1558 // when init > limit when stride > 0 (and reverse).
1559
1560 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1561
1562 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1563 // Modify pre loop end condition
1564 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1565 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1566 register_new_node(new_bol0, pre_head->in(0));
1567 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1568 // Modify main loop guard condition
1569 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1570 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1571 register_new_node(new_bol1, new_pre_exit);
1572 _igvn.hash_delete(min_iff);
1573 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1574 // Modify main loop end condition
1575 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1576 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1577 register_new_node(new_bol2, main_end->in(CountedLoopEndNode::TestControl));
1578 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1579 }
1580
1581 // Flag main loop
1582 main_head->set_main_loop();
1583 if (peel_only) {
1584 main_head->set_main_no_pre_loop();
1585 }
1586
1587 // Subtract a trip count for the pre-loop.
1588 main_head->set_trip_count(main_head->trip_count() - 1);
1589
1590 // It's difficult to be precise about the trip-counts
1591 // for the pre/post loops. They are usually very short,
1592 // so guess that 4 trips is a reasonable value.
1593 post_head->set_profile_trip_cnt(4.0);
1594 pre_head->set_profile_trip_cnt(4.0);
1595
1596 // Now force out all loop-invariant dominating tests. The optimizer
1597 // finds some, but we _know_ they are all useless.
1598 peeled_dom_test_elim(loop,old_new);
1599 loop->record_for_igvn();
1600
1601 C->print_method(PHASE_AFTER_PRE_MAIN_POST, 4, main_head);
1602 }
1603
1604 //------------------------------insert_vector_post_loop------------------------
1605 // Insert a copy of the atomic unrolled vectorized main loop as a post loop,
1606 // unroll_policy has already informed us that more unrolling is about to
1607 // happen to the main loop. The resultant post loop will serve as a
1608 // vectorized drain loop.
1609 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1610 if (!loop->_head->is_CountedLoop()) return;
1611
1612 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1613
1614 // only process vectorized main loops
1615 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return;
1616
1617 int slp_max_unroll_factor = cl->slp_max_unroll();
1618 int cur_unroll = cl->unrolled_count();
1619
1620 if (slp_max_unroll_factor == 0) return;
1621
1622 // only process atomic unroll vector loops (not super unrolled after vectorization)
1623 if (cur_unroll != slp_max_unroll_factor) return;
1624
1625 // we only ever process this one time
1626 if (cl->has_atomic_post_loop()) return;
1627
1628 if (!may_require_nodes(loop->est_loop_clone_sz(2))) {
1629 return;
1630 }
1631
1632 #ifndef PRODUCT
1633 if (TraceLoopOpts) {
1634 tty->print("PostVector ");
1635 loop->dump_head();
1636 }
1637 #endif
1638 C->set_major_progress();
1639
1640 // Find common pieces of the loop being guarded with pre & post loops
1641 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1642 CountedLoopEndNode *main_end = main_head->loopexit();
1643 // diagnostic to show loop end is not properly formed
1644 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1645
1646 // mark this loop as processed
1647 main_head->mark_has_atomic_post_loop();
1648
1649 Node *incr = main_end->incr();
1650 Node *limit = main_end->limit();
1651
1652 // In this case we throw away the result as we are not using it to connect anything else.
1653 C->print_method(PHASE_BEFORE_POST_LOOP, 4, main_head);
1654 CountedLoopNode *post_head = nullptr;
1655 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1656 C->print_method(PHASE_AFTER_POST_LOOP, 4, post_head);
1657
1658 // It's difficult to be precise about the trip-counts
1659 // for post loops. They are usually very short,
1660 // so guess that unit vector trips is a reasonable value.
1661 post_head->set_profile_trip_cnt(cur_unroll);
1662
1663 // Now force out all loop-invariant dominating tests. The optimizer
1664 // finds some, but we _know_ they are all useless.
1665 peeled_dom_test_elim(loop, old_new);
1666 loop->record_for_igvn();
1667 }
1668
1669 Node* PhaseIdealLoop::find_last_store_in_outer_loop(Node* store, const IdealLoopTree* outer_loop) {
1670 assert(store != nullptr && store->is_Store(), "starting point should be a store node");
1671 // Follow the memory uses until we get out of the loop.
1672 // Store nodes in the outer loop body were moved by PhaseIdealLoop::try_move_store_after_loop.
1673 // Because of the conditions in try_move_store_after_loop (no other usage in the loop body
1674 // except for the phi node associated with the loop head), we have the guarantee of a
1675 // linear memory subgraph within the outer loop body.
1676 Node* last = store;
1677 Node* unique_next = store;
1678 do {
1679 last = unique_next;
1680 for (DUIterator_Fast imax, l = last->fast_outs(imax); l < imax; l++) {
1681 Node* use = last->fast_out(l);
1682 if (use->is_Store() && use->in(MemNode::Memory) == last) {
1683 if (ctrl_is_member(outer_loop, use)) {
1684 assert(unique_next == last, "memory node should only have one usage in the loop body");
1685 unique_next = use;
1686 }
1687 }
1688 }
1689 } while (last != unique_next);
1690 return last;
1691 }
1692
1693 //------------------------------insert_post_loop-------------------------------
1694 // Insert post loops. Add a post loop to the given loop passed.
1695 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree* loop, Node_List& old_new,
1696 CountedLoopNode* main_head, CountedLoopEndNode* main_end,
1697 Node* incr, Node* limit, CountedLoopNode*& post_head) {
1698 IfNode* outer_main_end = main_end;
1699 IdealLoopTree* outer_loop = loop;
1700 if (main_head->is_strip_mined()) {
1701 main_head->verify_strip_mined(1);
1702 outer_main_end = main_head->outer_loop_end();
1703 outer_loop = loop->_parent;
1704 assert(outer_loop->_head == main_head->in(LoopNode::EntryControl), "broken loop tree");
1705 }
1706
1707 //------------------------------
1708 // Step A: Create a new post-Loop.
1709 IfFalseNode* main_exit = outer_main_end->false_proj();
1710 int dd_main_exit = dom_depth(main_exit);
1711
1712 // Step A1: Clone the loop body of main. The clone becomes the post-loop.
1713 // The main loop pre-header illegally has 2 control users (old & new loops).
1714 const uint first_node_index_in_cloned_loop_body = C->unique();
1715 clone_loop(loop, old_new, dd_main_exit, ControlAroundStripMined);
1716 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, "");
1717 post_head = old_new[main_head->_idx]->as_CountedLoop();
1718 post_head->set_normal_loop();
1719 post_head->set_post_loop(main_head);
1720
1721 // clone_loop() above changes the exit projection
1722 main_exit = outer_main_end->false_proj();
1723
1724 // Reduce the post-loop trip count.
1725 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd();
1726 post_end->_prob = PROB_FAIR;
1727
1728 // Build the main-loop normal exit.
1729 IfFalseNode *new_main_exit = new IfFalseNode(outer_main_end);
1730 _igvn.register_new_node_with_optimizer(new_main_exit);
1731 set_idom(new_main_exit, outer_main_end, dd_main_exit);
1732 set_loop(new_main_exit, outer_loop->_parent);
1733
1734 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
1735 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
1736 // (the previous loop trip-counter exit value) because we will be changing
1737 // the exit value (via additional unrolling) so we cannot constant-fold away the zero
1738 // trip guard until all unrolling is done.
1739 Node *zer_opaq = new OpaqueZeroTripGuardNode(C, incr, main_end->test_trip());
1740 Node *zer_cmp = new CmpINode(zer_opaq, limit);
1741 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip());
1742 register_new_node(zer_opaq, new_main_exit);
1743 register_new_node(zer_cmp, new_main_exit);
1744 register_new_node(zer_bol, new_main_exit);
1745
1746 // Build the IfNode
1747 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN);
1748 _igvn.register_new_node_with_optimizer(zer_iff);
1749 set_idom(zer_iff, new_main_exit, dd_main_exit);
1750 set_loop(zer_iff, outer_loop->_parent);
1751
1752 // Plug in the false-path, taken if we need to skip this post-loop
1753 _igvn.replace_input_of(main_exit, 0, zer_iff);
1754 set_idom(main_exit, zer_iff, dd_main_exit);
1755 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
1756 // Make the true-path, must enter this post loop
1757 Node *zer_taken = new IfTrueNode(zer_iff);
1758 _igvn.register_new_node_with_optimizer(zer_taken);
1759 set_idom(zer_taken, zer_iff, dd_main_exit);
1760 set_loop(zer_taken, outer_loop->_parent);
1761 // Plug in the true path
1762 _igvn.hash_delete(post_head);
1763 post_head->set_req(LoopNode::EntryControl, zer_taken);
1764 set_idom(post_head, zer_taken, dd_main_exit);
1765
1766 VectorSet visited;
1767 Node_Stack clones(main_head->back_control()->outcnt());
1768 // Step A3: Make the fall-in values to the post-loop come from the
1769 // fall-out values of the main-loop.
1770 for (DUIterator i = main_head->outs(); main_head->has_out(i); i++) {
1771 Node* main_phi = main_head->out(i);
1772 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0) {
1773 Node* cur_phi = old_new[main_phi->_idx];
1774 Node* fallnew = clone_up_backedge_goo(main_head->back_control(),
1775 post_head->init_control(),
1776 main_phi->in(LoopNode::LoopBackControl),
1777 visited, clones);
1778 _igvn.hash_delete(cur_phi);
1779 cur_phi->set_req(LoopNode::EntryControl, fallnew);
1780 }
1781 }
1782 // Store nodes that were moved to the outer loop by PhaseIdealLoop::try_move_store_after_loop
1783 // do not have an associated Phi node. Such nodes are attached to the false projection of the CountedLoopEnd node,
1784 // right after the execution of the inner CountedLoop.
1785 // We have to make sure that such stores in the post loop have the right memory inputs from the main loop
1786 // The moved store node is always attached right after the inner loop exit, and just before the safepoint
1787 const IfFalseNode* if_false = main_end->false_proj();
1788 for (DUIterator j = if_false->outs(); if_false->has_out(j); j++) {
1789 Node* store = if_false->out(j);
1790 if (store->is_Store()) {
1791 // We only make changes if the memory input of the store is outside the outer loop body,
1792 // as this is when we would normally expect a Phi as input. If the memory input
1793 // is in the loop body as well, then we can safely assume it is still correct as the entire
1794 // body was cloned as a unit
1795 if (!ctrl_is_member(outer_loop, store->in(MemNode::Memory))) {
1796 Node* mem_out = find_last_store_in_outer_loop(store, outer_loop);
1797 Node* store_new = old_new[store->_idx];
1798 store_new->set_req(MemNode::Memory, mem_out);
1799 }
1800 }
1801 }
1802
1803 DEBUG_ONLY(ensure_zero_trip_guard_proj(post_head->in(LoopNode::EntryControl), false);)
1804 initialize_assertion_predicates_for_post_loop(main_head, post_head, first_node_index_in_cloned_loop_body);
1805 cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1806 return new_main_exit;
1807 }
1808
1809 //------------------------------is_invariant-----------------------------
1810 // Return true if n is invariant
1811 bool IdealLoopTree::is_invariant(Node* n) const {
1812 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1813 if (n_c->is_top()) return false;
1814 return !is_member(_phase->get_loop(n_c));
1815 }
1816
1817 // Search the Assertion Predicates added by loop predication and/or range check elimination and update them according
1818 // to the new stride.
1819 void PhaseIdealLoop::update_main_loop_assertion_predicates(CountedLoopNode* new_main_loop_head,
1820 const int stride_con_before_unroll) {
1821 // Compute the value of the loop induction variable at the end of the
1822 // first iteration of the unrolled loop: init + new_stride_con - init_inc
1823 int unrolled_stride_con = stride_con_before_unroll * 2;
1824 Node* unrolled_stride = intcon(unrolled_stride_con);
1825
1826 Node* loop_entry = new_main_loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1827 PredicateIterator predicate_iterator(loop_entry);
1828 UpdateStrideForAssertionPredicates update_stride_for_assertion_predicates(unrolled_stride, new_main_loop_head, this);
1829 predicate_iterator.for_each(update_stride_for_assertion_predicates);
1830 }
1831
1832 // Source Loop: Cloned - peeled_loop_head
1833 // Target Loop: Original - remaining_loop_head
1834 void PhaseIdealLoop::initialize_assertion_predicates_for_peeled_loop(CountedLoopNode* peeled_loop_head,
1835 CountedLoopNode* remaining_loop_head,
1836 const uint first_node_index_in_cloned_loop_body,
1837 const Node_List& old_new) {
1838 const NodeInOriginalLoopBody node_in_original_loop_body(first_node_index_in_cloned_loop_body, old_new);
1839 create_assertion_predicates_at_loop(peeled_loop_head, remaining_loop_head, node_in_original_loop_body, true);
1840 }
1841
1842 // Source Loop: Cloned - pre_loop_head
1843 // Target Loop: Original - main_loop_head
1844 void PhaseIdealLoop::initialize_assertion_predicates_for_main_loop(CountedLoopNode* pre_loop_head,
1845 CountedLoopNode* main_loop_head,
1846 const uint first_node_index_in_pre_loop_body,
1847 const uint last_node_index_in_pre_loop_body,
1848 DEBUG_ONLY(const uint last_node_index_from_backedge_goo COMMA)
1849 const Node_List& old_new) {
1850 assert(first_node_index_in_pre_loop_body < last_node_index_in_pre_loop_body, "cloned some nodes");
1851 const NodeInMainLoopBody node_in_main_loop_body(first_node_index_in_pre_loop_body,
1852 last_node_index_in_pre_loop_body,
1853 DEBUG_ONLY(last_node_index_from_backedge_goo COMMA) old_new);
1854 create_assertion_predicates_at_main_or_post_loop(pre_loop_head, main_loop_head, node_in_main_loop_body, true);
1855 }
1856
1857 // Source Loop: Original - main_loop_head
1858 // Target Loop: Cloned - post_loop_head
1859 //
1860 // The post loop is cloned before the pre loop. Do not kill the old Template Assertion Predicates, yet. We need to clone
1861 // from them when creating the pre loop. Only then we can kill them.
1862 void PhaseIdealLoop::initialize_assertion_predicates_for_post_loop(CountedLoopNode* main_loop_head,
1863 CountedLoopNode* post_loop_head,
1864 const uint first_node_index_in_cloned_loop_body) {
1865 const NodeInClonedLoopBody node_in_cloned_loop_body(first_node_index_in_cloned_loop_body);
1866 create_assertion_predicates_at_main_or_post_loop(main_loop_head, post_loop_head, node_in_cloned_loop_body, false);
1867 }
1868
1869 void PhaseIdealLoop::create_assertion_predicates_at_loop(CountedLoopNode* source_loop_head,
1870 CountedLoopNode* target_loop_head,
1871 const NodeInLoopBody& _node_in_loop_body,
1872 const bool kill_old_template) {
1873 CreateAssertionPredicatesVisitor create_assertion_predicates_visitor(target_loop_head, this, _node_in_loop_body,
1874 kill_old_template);
1875 Node* source_loop_entry = source_loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1876 PredicateIterator predicate_iterator(source_loop_entry);
1877 predicate_iterator.for_each(create_assertion_predicates_visitor);
1878 }
1879
1880 void PhaseIdealLoop::create_assertion_predicates_at_main_or_post_loop(CountedLoopNode* source_loop_head,
1881 CountedLoopNode* target_loop_head,
1882 const NodeInLoopBody& _node_in_loop_body,
1883 const bool kill_old_template) {
1884 Node* old_target_loop_head_entry = target_loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1885 const uint node_index_before_new_assertion_predicate_nodes = C->unique();
1886 const bool need_to_rewire_old_target_loop_entry_dependencies = old_target_loop_head_entry->outcnt() > 1;
1887 create_assertion_predicates_at_loop(source_loop_head, target_loop_head, _node_in_loop_body, kill_old_template);
1888 if (need_to_rewire_old_target_loop_entry_dependencies) {
1889 rewire_old_target_loop_entry_dependency_to_new_entry(target_loop_head, old_target_loop_head_entry,
1890 node_index_before_new_assertion_predicate_nodes);
1891 }
1892 }
1893
1894 // Rewire any control dependent nodes on the old target loop entry before adding Assertion Predicate related nodes.
1895 // These have been added by PhaseIdealLoop::clone_up_backedge_goo() and assume to be ending up at the target loop entry
1896 // which is no longer the case when adding additional Assertion Predicates. Fix this by rewiring these nodes to the new
1897 // target loop entry which corresponds to the tail of the last Assertion Predicate before the target loop. This is safe
1898 // to do because these control dependent nodes on the old target loop entry created by clone_up_backedge_goo() were
1899 // pinned on the loop backedge before. The Assertion Predicates are not control dependent on these nodes in any way.
1900 void PhaseIdealLoop::rewire_old_target_loop_entry_dependency_to_new_entry(
1901 CountedLoopNode* target_loop_head, const Node* old_target_loop_entry,
1902 const uint node_index_before_new_assertion_predicate_nodes) {
1903 Node* new_main_loop_entry = target_loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1904 if (new_main_loop_entry == old_target_loop_entry) {
1905 // No Assertion Predicates added.
1906 return;
1907 }
1908
1909 for (DUIterator_Fast imax, i = old_target_loop_entry->fast_outs(imax); i < imax; i++) {
1910 Node* out = old_target_loop_entry->fast_out(i);
1911 if (!out->is_CFG() && out->_idx < node_index_before_new_assertion_predicate_nodes) {
1912 assert(out != target_loop_head->init_trip(), "CastII on loop entry?");
1913 _igvn.replace_input_of(out, 0, new_main_loop_entry);
1914 set_ctrl(out, new_main_loop_entry);
1915 --i;
1916 --imax;
1917 }
1918 }
1919 }
1920
1921 //------------------------------do_unroll--------------------------------------
1922 // Unroll the loop body one step - make each trip do 2 iterations.
1923 void PhaseIdealLoop::do_unroll(IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip) {
1924 assert(LoopUnrollLimit, "");
1925 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1926 CountedLoopEndNode *loop_end = loop_head->loopexit();
1927
1928 C->print_method(PHASE_BEFORE_LOOP_UNROLLING, 4, loop_head);
1929
1930 #ifndef PRODUCT
1931 if (TraceLoopOpts) {
1932 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1933 tty->print("Unroll %d(" JULONG_FORMAT_W(2) ") ", loop_head->unrolled_count()*2, loop_head->trip_count());
1934 } else {
1935 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1936 }
1937 loop->dump_head();
1938 }
1939
1940 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1941 Node_Stack stack(C->live_nodes() >> 2);
1942 Node_List rpo_list;
1943 VectorSet visited;
1944 visited.set(loop_head->_idx);
1945 rpo(loop_head, stack, visited, rpo_list);
1946 dump(loop, rpo_list.size(), rpo_list);
1947 }
1948 #endif
1949
1950 // Remember loop node count before unrolling to detect
1951 // if rounds of unroll,optimize are making progress
1952 loop_head->set_node_count_before_unroll(loop->_body.size());
1953
1954 Node *ctrl = loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1955 Node *limit = loop_head->limit();
1956 Node *init = loop_head->init_trip();
1957 Node *stride = loop_head->stride();
1958
1959 Node *opaq = nullptr;
1960 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1961 // Search for zero-trip guard.
1962
1963 // Check the shape of the graph at the loop entry. If an inappropriate
1964 // graph shape is encountered, the compiler bails out loop unrolling;
1965 // compilation of the method will still succeed.
1966 opaq = loop_head->is_canonical_loop_entry();
1967 if (opaq == nullptr) {
1968 return;
1969 }
1970 // Zero-trip test uses an 'opaque' node which is not shared, otherwise bail out.
1971 if (opaq->outcnt() != 1 || opaq->in(1) != limit) {
1972 #ifdef ASSERT
1973 // In rare cases, loop cloning (as for peeling, for instance) can break this by replacing
1974 // limit and the input of opaq by equivalent but distinct phis.
1975 // Next IGVN should clean it up. Let's try to detect we are in such a case.
1976 Unique_Node_List& worklist = loop->_phase->_igvn._worklist;
1977 assert(C->major_progress(), "The operation that replaced limit and opaq->in(1) (e.g. peeling) should have set major_progress");
1978 assert(opaq->in(1)->is_Phi() && limit->is_Phi(), "Nodes limit and opaq->in(1) should have been replaced by PhiNodes by fix_data_uses from clone_loop.");
1979 assert(worklist.member(opaq->in(1)) && worklist.member(limit), "Nodes limit and opaq->in(1) differ and should have been recorded for IGVN.");
1980 #endif
1981 return;
1982 }
1983 }
1984
1985 C->set_major_progress();
1986
1987 Node* new_limit = nullptr;
1988 const int stride_con = stride->get_int();
1989 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1990 uint old_trip_count = loop_head->trip_count();
1991 // Verify that unroll policy result is still valid.
1992 assert(old_trip_count > 1 && (!adjust_min_trip || stride_p <=
1993 MIN2<int>(max_jint / 2 - 2, MAX2(1<<3, Matcher::max_vector_size(T_BYTE)) * loop_head->unrolled_count())), "sanity");
1994
1995 // Adjust loop limit to keep valid iterations number after unroll.
1996 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1997 // which may overflow.
1998 if (!adjust_min_trip) {
1999 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
2000 "odd trip count for maximally unroll");
2001 // Don't need to adjust limit for maximally unroll since trip count is even.
2002 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
2003 // The trip count being exact means it has been set (using CountedLoopNode::set_exact_trip_count in compute_trip_count)
2004 assert(old_trip_count < max_juint, "sanity");
2005 // Loop's limit is constant. Loop's init could be constant when pre-loop
2006 // become peeled iteration.
2007 jlong init_con = init->get_int();
2008 // We can keep old loop limit if iterations count stays the same:
2009 // old_trip_count == new_trip_count * 2
2010 // Note: since old_trip_count >= 2 then new_trip_count >= 1
2011 // so we also don't need to adjust zero trip test.
2012 jlong limit_con = limit->get_int();
2013 // (stride_con*2) not overflow since stride_con <= 8.
2014 int new_stride_con = stride_con * 2;
2015 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
2016 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
2017 // New trip count should satisfy next conditions.
2018 assert(trip_count > 0 && (julong)trip_count <= (julong)max_juint/2, "sanity");
2019 uint new_trip_count = (uint)trip_count;
2020 // Since old_trip_count has been set to < max_juint (that is at most 2^32-2),
2021 // new_trip_count is lower than or equal to 2^31-1 and the multiplication cannot overflow.
2022 adjust_min_trip = (old_trip_count != new_trip_count*2);
2023 }
2024
2025 if (adjust_min_trip) {
2026 // Step 2: Adjust the trip limit if it is called for.
2027 // The adjustment amount is -stride. Need to make sure if the
2028 // adjustment underflows or overflows, then the main loop is skipped.
2029 Node* cmp = loop_end->cmp_node();
2030 assert(cmp->in(2) == limit, "sanity");
2031 assert(opaq != nullptr && opaq->in(1) == limit, "sanity");
2032
2033 // Verify that policy_unroll result is still valid.
2034 const TypeInt* limit_type = _igvn.type(limit)->is_int();
2035 assert((stride_con > 0 && ((min_jint + stride_con) <= limit_type->_hi)) ||
2036 (stride_con < 0 && ((max_jint + stride_con) >= limit_type->_lo)),
2037 "sanity");
2038
2039 if (limit->is_Con()) {
2040 // The check in policy_unroll and the assert above guarantee
2041 // no underflow if limit is constant.
2042 new_limit = intcon(limit->get_int() - stride_con);
2043 } else {
2044 // Limit is not constant. Int subtraction could lead to underflow.
2045 // (1) Convert to long.
2046 Node* limit_l = new ConvI2LNode(limit);
2047 register_new_node_with_ctrl_of(limit_l, limit);
2048 Node* stride_l = longcon(stride_con);
2049
2050 // (2) Subtract: compute in long, to prevent underflow.
2051 Node* new_limit_l = new SubLNode(limit_l, stride_l);
2052 register_new_node(new_limit_l, ctrl);
2053
2054 // (3) Clamp to int range, in case we had subtraction underflow.
2055 Node* underflow_clamp_l = longcon((stride_con > 0) ? min_jint : max_jint);
2056 Node* new_limit_no_underflow_l = nullptr;
2057 if (stride_con > 0) {
2058 // limit = MaxL(limit - stride, min_jint)
2059 new_limit_no_underflow_l = new MaxLNode(C, new_limit_l, underflow_clamp_l);
2060 } else {
2061 // limit = MinL(limit - stride, max_jint)
2062 new_limit_no_underflow_l = new MinLNode(C, new_limit_l, underflow_clamp_l);
2063 }
2064 register_new_node(new_limit_no_underflow_l, ctrl);
2065
2066 // (4) Convert back to int.
2067 new_limit = new ConvL2INode(new_limit_no_underflow_l);
2068 register_new_node(new_limit, ctrl);
2069 }
2070
2071 assert(new_limit != nullptr, "");
2072 // Replace in loop test.
2073 assert(loop_end->in(1)->in(1) == cmp, "sanity");
2074 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
2075 // Don't need to create new test since only one user.
2076 _igvn.hash_delete(cmp);
2077 cmp->set_req(2, new_limit);
2078 } else {
2079 // Create new test since it is shared.
2080 Node* ctrl2 = loop_end->in(0);
2081 Node* cmp2 = cmp->clone();
2082 cmp2->set_req(2, new_limit);
2083 register_new_node(cmp2, ctrl2);
2084 Node* bol2 = loop_end->in(1)->clone();
2085 bol2->set_req(1, cmp2);
2086 register_new_node(bol2, ctrl2);
2087 _igvn.replace_input_of(loop_end, 1, bol2);
2088 }
2089 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
2090 // Make it a 1-trip test (means at least 2 trips).
2091
2092 // Guard test uses an 'opaque' node which is not shared. Hence I
2093 // can edit it's inputs directly. Hammer in the new limit for the
2094 // minimum-trip guard.
2095 assert(opaq->outcnt() == 1, "");
2096 // Notify limit -> opaq -> CmpI, it may constant fold.
2097 _igvn.add_users_to_worklist(opaq->in(1));
2098 _igvn.replace_input_of(opaq, 1, new_limit);
2099 }
2100
2101 // Adjust max trip count. The trip count is intentionally rounded
2102 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
2103 // the main, unrolled, part of the loop will never execute as it is protected
2104 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
2105 // and later determined that part of the unrolled loop was dead.
2106 loop_head->set_trip_count(old_trip_count / 2);
2107
2108 // Double the count of original iterations in the unrolled loop body.
2109 loop_head->double_unrolled_count();
2110
2111 // ---------
2112 // Step 4: Clone the loop body. Move it inside the loop. This loop body
2113 // represents the odd iterations; since the loop trips an even number of
2114 // times its backedge is never taken. Kill the backedge.
2115 uint dd = dom_depth(loop_head);
2116 clone_loop(loop, old_new, dd, IgnoreStripMined);
2117
2118 // Make backedges of the clone equal to backedges of the original.
2119 // Make the fall-in from the original come from the fall-out of the clone.
2120 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
2121 Node* phi = loop_head->fast_out(j);
2122 if (phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0) {
2123 Node *newphi = old_new[phi->_idx];
2124 _igvn.hash_delete(phi);
2125 _igvn.hash_delete(newphi);
2126
2127 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
2128 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
2129 phi ->set_req(LoopNode::LoopBackControl, C->top());
2130 }
2131 }
2132 CountedLoopNode* clone_head = old_new[loop_head->_idx]->as_CountedLoop();
2133 _igvn.hash_delete(clone_head);
2134 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
2135 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
2136 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
2137 loop->_head = clone_head; // New loop header
2138
2139 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
2140 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
2141
2142 // Kill the clone's backedge
2143 Node *newcle = old_new[loop_end->_idx];
2144 _igvn.hash_delete(newcle);
2145 Node* one = intcon(1);
2146 newcle->set_req(1, one);
2147 // Force clone into same loop body
2148 uint max = loop->_body.size();
2149 for (uint k = 0; k < max; k++) {
2150 Node *old = loop->_body.at(k);
2151 Node *nnn = old_new[old->_idx];
2152 loop->_body.push(nnn);
2153 if (!has_ctrl(old)) {
2154 set_loop(nnn, loop);
2155 }
2156 }
2157
2158 loop->record_for_igvn();
2159 loop_head->clear_strip_mined();
2160
2161 update_main_loop_assertion_predicates(clone_head, stride_con);
2162
2163 #ifndef PRODUCT
2164 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
2165 tty->print("\nnew loop after unroll\n"); loop->dump_head();
2166 for (uint i = 0; i < loop->_body.size(); i++) {
2167 loop->_body.at(i)->dump();
2168 }
2169 if (C->clone_map().is_debug()) {
2170 tty->print("\nCloneMap\n");
2171 Dict* dict = C->clone_map().dict();
2172 DictI i(dict);
2173 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size());
2174 for (int ii = 0; i.test(); ++i, ++ii) {
2175 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key));
2176 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen());
2177 if (ii % 10 == 9) {
2178 tty->print_cr(" ");
2179 }
2180 }
2181 tty->print_cr(" ");
2182 }
2183 }
2184 #endif
2185
2186 C->print_method(PHASE_AFTER_LOOP_UNROLLING, 4, clone_head);
2187 }
2188
2189 //------------------------------do_maximally_unroll----------------------------
2190
2191 void PhaseIdealLoop::do_maximally_unroll(IdealLoopTree *loop, Node_List &old_new) {
2192 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2193 assert(cl->has_exact_trip_count(), "trip count is not exact");
2194 assert(cl->trip_count() > 0, "");
2195 #ifndef PRODUCT
2196 if (TraceLoopOpts) {
2197 tty->print("MaxUnroll " JULONG_FORMAT " ", cl->trip_count());
2198 loop->dump_head();
2199 }
2200 #endif
2201
2202 // If loop is tripping an odd number of times, peel odd iteration
2203 if ((cl->trip_count() & 1) == 1) {
2204 do_peeling(loop, old_new);
2205 }
2206
2207 // Now its tripping an even number of times remaining. Double loop body.
2208 // Do not adjust pre-guards; they are not needed and do not exist.
2209 if (cl->trip_count() > 0) {
2210 assert((cl->trip_count() & 1) == 0, "missed peeling");
2211 do_unroll(loop, old_new, false);
2212 }
2213 }
2214
2215 //------------------------------adjust_limit-----------------------------------
2216 // Helper function that computes new loop limit as (rc_limit-offset)/scale
2217 Node* PhaseIdealLoop::adjust_limit(bool is_positive_stride, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round) {
2218 Node* old_limit_long = new ConvI2LNode(old_limit);
2219 register_new_node(old_limit_long, pre_ctrl);
2220
2221 Node* sub = new SubLNode(rc_limit, offset);
2222 register_new_node(sub, pre_ctrl);
2223 Node* limit = new DivLNode(nullptr, sub, scale);
2224 register_new_node(limit, pre_ctrl);
2225
2226 // When the absolute value of scale is greater than one, the division
2227 // may round limit down/up, so add/sub one to/from the limit.
2228 if (round) {
2229 limit = new AddLNode(limit, _igvn.longcon(is_positive_stride ? -1 : 1));
2230 register_new_node(limit, pre_ctrl);
2231 }
2232
2233 // Clamp the limit to handle integer under-/overflows by using long values.
2234 // We only convert the limit back to int when we handled under-/overflows.
2235 // Note that all values are longs in the following computations.
2236 // When reducing the limit, clamp to [min_jint, old_limit]:
2237 // INT(MINL(old_limit, MAXL(limit, min_jint)))
2238 // - integer underflow of limit: MAXL chooses min_jint.
2239 // - integer overflow of limit: MINL chooses old_limit (<= MAX_INT < limit)
2240 // When increasing the limit, clamp to [old_limit, max_jint]:
2241 // INT(MAXL(old_limit, MINL(limit, max_jint)))
2242 // - integer overflow of limit: MINL chooses max_jint.
2243 // - integer underflow of limit: MAXL chooses old_limit (>= MIN_INT > limit)
2244 // INT() is finally converting the limit back to an integer value.
2245
2246 Node* inner_result_long = nullptr;
2247 Node* outer_result_long = nullptr;
2248 if (is_positive_stride) {
2249 inner_result_long = new MaxLNode(C, limit, _igvn.longcon(min_jint));
2250 outer_result_long = new MinLNode(C, inner_result_long, old_limit_long);
2251 } else {
2252 inner_result_long = new MinLNode(C, limit, _igvn.longcon(max_jint));
2253 outer_result_long = new MaxLNode(C, inner_result_long, old_limit_long);
2254 }
2255 register_new_node(inner_result_long, pre_ctrl);
2256 register_new_node(outer_result_long, pre_ctrl);
2257
2258 limit = new ConvL2INode(outer_result_long);
2259 register_new_node(limit, pre_ctrl);
2260 return limit;
2261 }
2262
2263 //------------------------------add_constraint---------------------------------
2264 // Constrain the main loop iterations so the conditions:
2265 // low_limit <= scale_con*I + offset < upper_limit
2266 // always hold true. That is, either increase the number of iterations in the
2267 // pre-loop or reduce the number of iterations in the main-loop until the condition
2268 // holds true in the main-loop. Stride, scale, offset and limit are all loop
2269 // invariant. Further, stride and scale are constants (offset and limit often are).
2270 void PhaseIdealLoop::add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit) {
2271 assert(_igvn.type(offset)->isa_long() != nullptr && _igvn.type(low_limit)->isa_long() != nullptr &&
2272 _igvn.type(upper_limit)->isa_long() != nullptr, "arguments should be long values");
2273
2274 // For a positive stride, we need to reduce the main-loop limit and
2275 // increase the pre-loop limit. This is reversed for a negative stride.
2276 bool is_positive_stride = (stride_con > 0);
2277
2278 // If the absolute scale value is greater one, division in 'adjust_limit' may require
2279 // rounding. Make sure the ABS method correctly handles min_jint.
2280 // Only do this for the pre-loop, one less iteration of the main loop doesn't hurt.
2281 bool round = ABS(scale_con) > 1;
2282
2283 Node* scale = longcon(scale_con);
2284
2285 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
2286 // Positive stride*scale: the affine function is increasing,
2287 // the pre-loop checks for underflow and the post-loop for overflow.
2288
2289 // The overflow limit: scale*I+offset < upper_limit
2290 // For the main-loop limit compute:
2291 // ( if (scale > 0) /* and stride > 0 */
2292 // I < (upper_limit-offset)/scale
2293 // else /* scale < 0 and stride < 0 */
2294 // I > (upper_limit-offset)/scale
2295 // )
2296 *main_limit = adjust_limit(is_positive_stride, scale, offset, upper_limit, *main_limit, pre_ctrl, false);
2297
2298 // The underflow limit: low_limit <= scale*I+offset
2299 // For the pre-loop limit compute:
2300 // NOT(scale*I+offset >= low_limit)
2301 // scale*I+offset < low_limit
2302 // ( if (scale > 0) /* and stride > 0 */
2303 // I < (low_limit-offset)/scale
2304 // else /* scale < 0 and stride < 0 */
2305 // I > (low_limit-offset)/scale
2306 // )
2307 *pre_limit = adjust_limit(!is_positive_stride, scale, offset, low_limit, *pre_limit, pre_ctrl, round);
2308 } else {
2309 // Negative stride*scale: the affine function is decreasing,
2310 // the pre-loop checks for overflow and the post-loop for underflow.
2311
2312 // The overflow limit: scale*I+offset < upper_limit
2313 // For the pre-loop limit compute:
2314 // NOT(scale*I+offset < upper_limit)
2315 // scale*I+offset >= upper_limit
2316 // scale*I+offset+1 > upper_limit
2317 // ( if (scale < 0) /* and stride > 0 */
2318 // I < (upper_limit-(offset+1))/scale
2319 // else /* scale > 0 and stride < 0 */
2320 // I > (upper_limit-(offset+1))/scale
2321 // )
2322 Node* one = longcon(1);
2323 Node* plus_one = new AddLNode(offset, one);
2324 register_new_node(plus_one, pre_ctrl);
2325 *pre_limit = adjust_limit(!is_positive_stride, scale, plus_one, upper_limit, *pre_limit, pre_ctrl, round);
2326
2327 // The underflow limit: low_limit <= scale*I+offset
2328 // For the main-loop limit compute:
2329 // scale*I+offset+1 > low_limit
2330 // ( if (scale < 0) /* and stride > 0 */
2331 // I < (low_limit-(offset+1))/scale
2332 // else /* scale > 0 and stride < 0 */
2333 // I > (low_limit-(offset+1))/scale
2334 // )
2335 *main_limit = adjust_limit(is_positive_stride, scale, plus_one, low_limit, *main_limit, pre_ctrl, false);
2336 }
2337 }
2338
2339 //----------------------------------is_iv------------------------------------
2340 // Return true if exp is the value (of type bt) of the given induction var.
2341 // This grammar of cases is recognized, where X is I|L according to bt:
2342 // VIV[iv] = iv | (CastXX VIV[iv]) | (ConvI2X VIV[iv])
2343 bool PhaseIdealLoop::is_iv(Node* exp, Node* iv, BasicType bt) {
2344 exp = exp->uncast();
2345 if (exp == iv && iv->bottom_type()->isa_integer(bt)) {
2346 return true;
2347 }
2348
2349 if (bt == T_LONG && iv->bottom_type()->isa_int() && exp->Opcode() == Op_ConvI2L && exp->in(1)->uncast() == iv) {
2350 return true;
2351 }
2352 return false;
2353 }
2354
2355 //------------------------------is_scaled_iv---------------------------------
2356 // Return true if exp is a constant times the given induction var (of type bt).
2357 // The multiplication is either done in full precision (exactly of type bt),
2358 // or else bt is T_LONG but iv is scaled using 32-bit arithmetic followed by a ConvI2L.
2359 // This grammar of cases is recognized, where X is I|L according to bt:
2360 // SIV[iv] = VIV[iv] | (CastXX SIV[iv])
2361 // | (MulX VIV[iv] ConX) | (MulX ConX VIV[iv])
2362 // | (LShiftX VIV[iv] ConI)
2363 // | (ConvI2L SIV[iv]) -- a "short-scale" can occur here; note recursion
2364 // | (SubX 0 SIV[iv]) -- same as MulX(iv, -scale); note recursion
2365 // | (AddX SIV[iv] SIV[iv]) -- sum of two scaled iv; note recursion
2366 // | (SubX SIV[iv] SIV[iv]) -- difference of two scaled iv; note recursion
2367 // VIV[iv] = [either iv or its value converted; see is_iv() above]
2368 // On success, the constant scale value is stored back to *p_scale.
2369 // The value (*p_short_scale) reports if such a ConvI2L conversion was present.
2370 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, BasicType bt, jlong* p_scale, bool* p_short_scale, int depth) {
2371 BasicType exp_bt = bt;
2372 exp = exp->uncast(); //strip casts
2373 assert(exp_bt == T_INT || exp_bt == T_LONG, "unexpected int type");
2374 if (is_iv(exp, iv, exp_bt)) {
2375 if (p_scale != nullptr) {
2376 *p_scale = 1;
2377 }
2378 if (p_short_scale != nullptr) {
2379 *p_short_scale = false;
2380 }
2381 return true;
2382 }
2383 if (exp_bt == T_LONG && iv->bottom_type()->isa_int() && exp->Opcode() == Op_ConvI2L) {
2384 exp = exp->in(1);
2385 exp_bt = T_INT;
2386 }
2387 int opc = exp->Opcode();
2388 int which = 0; // this is which subexpression we find the iv in
2389 // Can't use is_Mul() here as it's true for AndI and AndL
2390 if (opc == Op_Mul(exp_bt)) {
2391 if ((is_iv(exp->in(which = 1), iv, exp_bt) && exp->in(2)->is_Con()) ||
2392 (is_iv(exp->in(which = 2), iv, exp_bt) && exp->in(1)->is_Con())) {
2393 Node* factor = exp->in(which == 1 ? 2 : 1); // the other argument
2394 jlong scale = factor->find_integer_as_long(exp_bt, 0);
2395 if (scale == 0) {
2396 return false; // might be top
2397 }
2398 if (p_scale != nullptr) {
2399 *p_scale = scale;
2400 }
2401 if (p_short_scale != nullptr) {
2402 // (ConvI2L (MulI iv K)) can be 64-bit linear if iv is kept small enough...
2403 *p_short_scale = (exp_bt != bt && scale != 1);
2404 }
2405 return true;
2406 }
2407 } else if (opc == Op_LShift(exp_bt)) {
2408 if (is_iv(exp->in(1), iv, exp_bt) && exp->in(2)->is_Con()) {
2409 jint shift_amount = exp->in(2)->find_int_con(min_jint);
2410 if (shift_amount == min_jint) {
2411 return false; // might be top
2412 }
2413 jlong scale;
2414 if (exp_bt == T_INT) {
2415 scale = java_shift_left((jint)1, (juint)shift_amount);
2416 } else if (exp_bt == T_LONG) {
2417 scale = java_shift_left((jlong)1, (julong)shift_amount);
2418 }
2419 if (p_scale != nullptr) {
2420 *p_scale = scale;
2421 }
2422 if (p_short_scale != nullptr) {
2423 // (ConvI2L (MulI iv K)) can be 64-bit linear if iv is kept small enough...
2424 *p_short_scale = (exp_bt != bt && scale != 1);
2425 }
2426 return true;
2427 }
2428 } else if (opc == Op_Add(exp_bt)) {
2429 jlong scale_l = 0;
2430 jlong scale_r = 0;
2431 bool short_scale_l = false;
2432 bool short_scale_r = false;
2433 if (depth == 0 &&
2434 is_scaled_iv(exp->in(1), iv, exp_bt, &scale_l, &short_scale_l, depth + 1) &&
2435 is_scaled_iv(exp->in(2), iv, exp_bt, &scale_r, &short_scale_r, depth + 1)) {
2436 // AddX(iv*K1, iv*K2) => iv*(K1+K2)
2437 jlong scale_sum = java_add(scale_l, scale_r);
2438 if (scale_sum > max_signed_integer(exp_bt) || scale_sum <= min_signed_integer(exp_bt)) {
2439 // This logic is shared by int and long. For int, the result may overflow
2440 // as we use jlong to compute so do the check here. Long result may also
2441 // overflow but that's fine because result wraps.
2442 return false;
2443 }
2444 if (p_scale != nullptr) {
2445 *p_scale = scale_sum;
2446 }
2447 if (p_short_scale != nullptr) {
2448 *p_short_scale = short_scale_l && short_scale_r;
2449 }
2450 return true;
2451 }
2452 } else if (opc == Op_Sub(exp_bt)) {
2453 if (exp->in(1)->find_integer_as_long(exp_bt, -1) == 0) {
2454 jlong scale = 0;
2455 if (depth == 0 && is_scaled_iv(exp->in(2), iv, exp_bt, &scale, p_short_scale, depth + 1)) {
2456 // SubX(0, iv*K) => iv*(-K)
2457 if (scale == min_signed_integer(exp_bt)) {
2458 // This should work even if -K overflows, but let's not.
2459 return false;
2460 }
2461 scale = java_multiply(scale, (jlong)-1);
2462 if (p_scale != nullptr) {
2463 *p_scale = scale;
2464 }
2465 if (p_short_scale != nullptr) {
2466 // (ConvI2L (MulI iv K)) can be 64-bit linear if iv is kept small enough...
2467 *p_short_scale = *p_short_scale || (exp_bt != bt && scale != 1);
2468 }
2469 return true;
2470 }
2471 } else {
2472 jlong scale_l = 0;
2473 jlong scale_r = 0;
2474 bool short_scale_l = false;
2475 bool short_scale_r = false;
2476 if (depth == 0 &&
2477 is_scaled_iv(exp->in(1), iv, exp_bt, &scale_l, &short_scale_l, depth + 1) &&
2478 is_scaled_iv(exp->in(2), iv, exp_bt, &scale_r, &short_scale_r, depth + 1)) {
2479 // SubX(iv*K1, iv*K2) => iv*(K1-K2)
2480 jlong scale_diff = java_subtract(scale_l, scale_r);
2481 if (scale_diff > max_signed_integer(exp_bt) || scale_diff <= min_signed_integer(exp_bt)) {
2482 // This logic is shared by int and long. For int, the result may
2483 // overflow as we use jlong to compute so do the check here. Long
2484 // result may also overflow but that's fine because result wraps.
2485 return false;
2486 }
2487 if (p_scale != nullptr) {
2488 *p_scale = scale_diff;
2489 }
2490 if (p_short_scale != nullptr) {
2491 *p_short_scale = short_scale_l && short_scale_r;
2492 }
2493 return true;
2494 }
2495 }
2496 }
2497 // We could also recognize (iv*K1)*K2, even with overflow, but let's not.
2498 return false;
2499 }
2500
2501 //-------------------------is_scaled_iv_plus_offset--------------------------
2502 // Return true if exp is a simple linear transform of the given induction var.
2503 // The scale must be constant and the addition tree (if any) must be simple.
2504 // This grammar of cases is recognized, where X is I|L according to bt:
2505 //
2506 // OIV[iv] = SIV[iv] | (CastXX OIV[iv])
2507 // | (AddX SIV[iv] E) | (AddX E SIV[iv])
2508 // | (SubX SIV[iv] E) | (SubX E SIV[iv])
2509 // SSIV[iv] = (ConvI2X SIV[iv]) -- a "short scale" might occur here
2510 // SIV[iv] = [a possibly scaled value of iv; see is_scaled_iv() above]
2511 //
2512 // On success, the constant scale value is stored back to *p_scale unless null.
2513 // Likewise, the addend (perhaps a synthetic AddX node) is stored to *p_offset.
2514 // Also, (*p_short_scale) reports if a ConvI2L conversion was seen after a MulI,
2515 // meaning bt is T_LONG but iv was scaled using 32-bit arithmetic.
2516 // To avoid looping, the match is depth-limited, and so may fail to match the grammar to complex expressions.
2517 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, BasicType bt, jlong* p_scale, Node** p_offset, bool* p_short_scale, int depth) {
2518 assert(bt == T_INT || bt == T_LONG, "unexpected int type");
2519 jlong scale = 0; // to catch result from is_scaled_iv()
2520 BasicType exp_bt = bt;
2521 exp = exp->uncast();
2522 if (is_scaled_iv(exp, iv, exp_bt, &scale, p_short_scale)) {
2523 if (p_scale != nullptr) {
2524 *p_scale = scale;
2525 }
2526 if (p_offset != nullptr) {
2527 Node* zero = zerocon(bt);
2528 *p_offset = zero;
2529 }
2530 return true;
2531 }
2532 if (exp_bt != bt) {
2533 // We would now be matching inputs like (ConvI2L exp:(AddI (MulI iv S) E)).
2534 // It's hard to make 32-bit arithmetic linear if it overflows. Although we do
2535 // cope with overflowing multiplication by S, it would be even more work to
2536 // handle overflowing addition of E. So we bail out here on ConvI2L input.
2537 return false;
2538 }
2539 int opc = exp->Opcode();
2540 int which = 0; // this is which subexpression we find the iv in
2541 Node* offset = nullptr;
2542 if (opc == Op_Add(exp_bt)) {
2543 // Check for a scaled IV in (AddX (MulX iv S) E) or (AddX E (MulX iv S)).
2544 if (is_scaled_iv(exp->in(which = 1), iv, bt, &scale, p_short_scale) ||
2545 is_scaled_iv(exp->in(which = 2), iv, bt, &scale, p_short_scale)) {
2546 offset = exp->in(which == 1 ? 2 : 1); // the other argument
2547 if (p_scale != nullptr) {
2548 *p_scale = scale;
2549 }
2550 if (p_offset != nullptr) {
2551 *p_offset = offset;
2552 }
2553 return true;
2554 }
2555 // Check for more addends, like (AddX (AddX (MulX iv S) E1) E2), etc.
2556 if (is_scaled_iv_plus_extra_offset(exp->in(1), exp->in(2), iv, bt, p_scale, p_offset, p_short_scale, depth) ||
2557 is_scaled_iv_plus_extra_offset(exp->in(2), exp->in(1), iv, bt, p_scale, p_offset, p_short_scale, depth)) {
2558 return true;
2559 }
2560 } else if (opc == Op_Sub(exp_bt)) {
2561 if (is_scaled_iv(exp->in(which = 1), iv, bt, &scale, p_short_scale) ||
2562 is_scaled_iv(exp->in(which = 2), iv, bt, &scale, p_short_scale)) {
2563 // Match (SubX SIV[iv] E) as if (AddX SIV[iv] (SubX 0 E)), and
2564 // match (SubX E SIV[iv]) as if (AddX E (SubX 0 SIV[iv])).
2565 offset = exp->in(which == 1 ? 2 : 1); // the other argument
2566 if (which == 2) {
2567 // We can't handle a scale of min_jint (or min_jlong) here as -1 * min_jint = min_jint
2568 if (scale == min_signed_integer(bt)) {
2569 return false; // cannot negate the scale of the iv
2570 }
2571 scale = java_multiply(scale, (jlong)-1);
2572 }
2573 if (p_scale != nullptr) {
2574 *p_scale = scale;
2575 }
2576 if (p_offset != nullptr) {
2577 if (which == 1) { // must negate the extracted offset
2578 Node* zero = integercon(0, exp_bt);
2579 Node *ctrl_off = get_ctrl(offset);
2580 offset = SubNode::make(zero, offset, exp_bt);
2581 register_new_node(offset, ctrl_off);
2582 }
2583 *p_offset = offset;
2584 }
2585 return true;
2586 }
2587 }
2588 return false;
2589 }
2590
2591 // Helper for is_scaled_iv_plus_offset(), not called separately.
2592 // The caller encountered (AddX exp1 offset3) or (AddX offset3 exp1).
2593 // Here, exp1 is inspected to see if it is a simple linear transform of iv.
2594 // If so, the offset3 is combined with any other offset2 from inside exp1.
2595 bool PhaseIdealLoop::is_scaled_iv_plus_extra_offset(Node* exp1, Node* offset3, Node* iv,
2596 BasicType bt,
2597 jlong* p_scale, Node** p_offset,
2598 bool* p_short_scale, int depth) {
2599 // By the time we reach here, it is unlikely that exp1 is a simple iv*K.
2600 // If is a linear iv transform, it is probably an add or subtract.
2601 // Let's collect the internal offset2 from it.
2602 Node* offset2 = nullptr;
2603 if (offset3->is_Con() &&
2604 depth < 2 &&
2605 is_scaled_iv_plus_offset(exp1, iv, bt, p_scale,
2606 &offset2, p_short_scale, depth+1)) {
2607 if (p_offset != nullptr) {
2608 Node* ctrl_off2 = get_ctrl(offset2);
2609 Node* offset = AddNode::make(offset2, offset3, bt);
2610 register_new_node(offset, ctrl_off2);
2611 *p_offset = offset;
2612 }
2613 return true;
2614 }
2615 return false;
2616 }
2617
2618 //------------------------------do_range_check---------------------------------
2619 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
2620 void PhaseIdealLoop::do_range_check(IdealLoopTree* loop) {
2621 #ifndef PRODUCT
2622 if (TraceLoopOpts) {
2623 tty->print("RangeCheck ");
2624 loop->dump_head();
2625 }
2626 #endif
2627
2628 assert(RangeCheckElimination, "");
2629 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2630
2631 // protect against stride not being a constant
2632 if (!cl->stride_is_con()) {
2633 return;
2634 }
2635 // Find the trip counter; we are iteration splitting based on it
2636 Node *trip_counter = cl->phi();
2637 // Find the main loop limit; we will trim it's iterations
2638 // to not ever trip end tests
2639 Node *main_limit = cl->limit();
2640 Node* main_limit_ctrl = get_ctrl(main_limit);
2641
2642 // Check graph shape. Cannot optimize a loop if zero-trip
2643 // Opaque1 node is optimized away and then another round
2644 // of loop opts attempted.
2645 if (cl->is_canonical_loop_entry() == nullptr) {
2646 return;
2647 }
2648
2649 // Need to find the main-loop zero-trip guard
2650 Node *ctrl = cl->skip_assertion_predicates_with_halt();
2651 Node *iffm = ctrl->in(0);
2652 Node *opqzm = iffm->in(1)->in(1)->in(2);
2653 assert(opqzm->in(1) == main_limit, "do not understand situation");
2654
2655 // Find the pre-loop limit; we will expand its iterations to
2656 // not ever trip low tests.
2657 Node *p_f = iffm->in(0);
2658 // pre loop may have been optimized out
2659 if (p_f->Opcode() != Op_IfFalse) {
2660 return;
2661 }
2662 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2663 assert(pre_end->loopnode()->is_pre_loop(), "");
2664 Node *pre_opaq1 = pre_end->limit();
2665 // Occasionally it's possible for a pre-loop Opaque1 node to be
2666 // optimized away and then another round of loop opts attempted.
2667 // We can not optimize this particular loop in that case.
2668 if (pre_opaq1->Opcode() != Op_Opaque1) {
2669 return;
2670 }
2671 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2672 Node *pre_limit = pre_opaq->in(1);
2673 Node* pre_limit_ctrl = get_ctrl(pre_limit);
2674
2675 // Where do we put new limit calculations
2676 Node* pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2677 // Range check elimination optimizes out conditions whose parameters are loop invariant in the main loop. They usually
2678 // have control above the pre loop, but there's no guarantee that they do. There's no guarantee either that the pre
2679 // loop limit has control that's out of loop (a previous round of range check elimination could have set a limit that's
2680 // not loop invariant). new_limit_ctrl is used for both the pre and main loops. Early control for the main limit may be
2681 // below the pre loop entry and the pre limit and must be taken into account when initializing new_limit_ctrl.
2682 Node* new_limit_ctrl = dominated_node(pre_ctrl, pre_limit_ctrl, compute_early_ctrl(main_limit, main_limit_ctrl));
2683
2684 // Ensure the original loop limit is available from the
2685 // pre-loop Opaque1 node.
2686 Node *orig_limit = pre_opaq->original_loop_limit();
2687 if (orig_limit == nullptr || _igvn.type(orig_limit) == Type::TOP) {
2688 return;
2689 }
2690 // Must know if its a count-up or count-down loop
2691
2692 int stride_con = cl->stride_con();
2693 bool abs_stride_is_one = stride_con == 1 || stride_con == -1;
2694 Node* zero = longcon(0);
2695 Node* one = longcon(1);
2696 // Use symmetrical int range [-max_jint,max_jint]
2697 Node* mini = longcon(-max_jint);
2698
2699 Node* loop_entry = cl->skip_strip_mined()->in(LoopNode::EntryControl);
2700 assert(loop_entry->is_Proj() && loop_entry->in(0)->is_If(), "if projection only");
2701
2702 // if abs(stride) == 1, an Assertion Predicate for the final iv value is added. We don't know the final iv value until
2703 // we're done with range check elimination so use a place holder.
2704 Node* final_iv_placeholder = nullptr;
2705 if (abs_stride_is_one) {
2706 final_iv_placeholder = new Node(1);
2707 _igvn.set_type(final_iv_placeholder, TypeInt::INT);
2708 final_iv_placeholder->init_req(0, loop_entry);
2709 }
2710
2711 // Check loop body for tests of trip-counter plus loop-invariant vs loop-variant.
2712 for (uint i = 0; i < loop->_body.size(); i++) {
2713 Node *iff = loop->_body[i];
2714 if (iff->Opcode() == Op_If ||
2715 iff->Opcode() == Op_RangeCheck) { // Test?
2716 // Test is an IfNode, has 2 projections. If BOTH are in the loop
2717 // we need loop unswitching instead of iteration splitting.
2718 Node *exit = loop->is_loop_exit(iff);
2719 if (!exit) continue;
2720 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
2721
2722 // Get boolean condition to test
2723 Node *i1 = iff->in(1);
2724 if (!i1->is_Bool()) continue;
2725 BoolNode *bol = i1->as_Bool();
2726 BoolTest b_test = bol->_test;
2727 // Flip sense of test if exit condition is flipped
2728 if (flip) {
2729 b_test = b_test.negate();
2730 }
2731 // Get compare
2732 Node *cmp = bol->in(1);
2733
2734 // Look for trip_counter + offset vs limit
2735 Node *rc_exp = cmp->in(1);
2736 Node *limit = cmp->in(2);
2737 int scale_con= 1; // Assume trip counter not scaled
2738
2739 Node* limit_ctrl = get_ctrl(limit);
2740 if (loop->is_member(get_loop(limit_ctrl))) {
2741 // Compare might have operands swapped; commute them
2742 b_test = b_test.commute();
2743 rc_exp = cmp->in(2);
2744 limit = cmp->in(1);
2745 limit_ctrl = get_ctrl(limit);
2746 if (loop->is_member(get_loop(limit_ctrl))) {
2747 continue; // Both inputs are loop varying; cannot RCE
2748 }
2749 }
2750 // Here we know 'limit' is loop invariant
2751
2752 // 'limit' maybe pinned below the zero trip test (probably from a
2753 // previous round of rce), in which case, it can't be used in the
2754 // zero trip test expression which must occur before the zero test's if.
2755 if (is_dominator(ctrl, limit_ctrl)) {
2756 continue; // Don't rce this check but continue looking for other candidates.
2757 }
2758
2759 assert(is_dominator(compute_early_ctrl(limit, limit_ctrl), pre_end), "node pinned on loop exit test?");
2760
2761 // Check for scaled induction variable plus an offset
2762 Node *offset = nullptr;
2763
2764 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
2765 continue;
2766 }
2767
2768 Node* offset_ctrl = get_ctrl(offset);
2769 if (loop->is_member(get_loop(offset_ctrl))) {
2770 continue; // Offset is not really loop invariant
2771 }
2772 // Here we know 'offset' is loop invariant.
2773
2774 // As above for the 'limit', the 'offset' maybe pinned below the
2775 // zero trip test.
2776 if (is_dominator(ctrl, offset_ctrl)) {
2777 continue; // Don't rce this check but continue looking for other candidates.
2778 }
2779
2780 // offset and limit can have control set below the pre loop when they are not loop invariant in the pre loop.
2781 // Update their control (and the control of inputs as needed) to be above pre_end
2782 offset_ctrl = ensure_node_and_inputs_are_above_pre_end(pre_end, offset);
2783 limit_ctrl = ensure_node_and_inputs_are_above_pre_end(pre_end, limit);
2784
2785 // offset and limit could have control below new_limit_ctrl if they are not loop invariant in the pre loop.
2786 Node* next_limit_ctrl = dominated_node(new_limit_ctrl, offset_ctrl, limit_ctrl);
2787
2788 #ifdef ASSERT
2789 if (TraceRangeLimitCheck) {
2790 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
2791 bol->dump(2);
2792 }
2793 #endif
2794 // At this point we have the expression as:
2795 // scale_con * trip_counter + offset :: limit
2796 // where scale_con, offset and limit are loop invariant. Trip_counter
2797 // monotonically increases by stride_con, a constant. Both (or either)
2798 // stride_con and scale_con can be negative which will flip about the
2799 // sense of the test.
2800
2801 C->print_method(PHASE_BEFORE_RANGE_CHECK_ELIMINATION, 4, iff);
2802
2803 // Perform the limit computations in jlong to avoid overflow
2804 jlong lscale_con = scale_con;
2805 Node* int_offset = offset;
2806 offset = new ConvI2LNode(offset);
2807 register_new_node(offset, next_limit_ctrl);
2808 Node* int_limit = limit;
2809 limit = new ConvI2LNode(limit);
2810 register_new_node(limit, next_limit_ctrl);
2811
2812 // Adjust pre and main loop limits to guard the correct iteration set
2813 if (cmp->Opcode() == Op_CmpU) { // Unsigned compare is really 2 tests
2814 if (b_test._test == BoolTest::lt) { // Range checks always use lt
2815 // The underflow and overflow limits: 0 <= scale*I+offset < limit
2816 add_constraint(stride_con, lscale_con, offset, zero, limit, next_limit_ctrl, &pre_limit, &main_limit);
2817 Node* init = cl->uncasted_init_trip(true);
2818
2819 Node* opaque_init = new OpaqueLoopInitNode(C, init);
2820 register_new_node(opaque_init, loop_entry);
2821
2822 InitializedAssertionPredicateCreator initialized_assertion_predicate_creator(this);
2823 if (abs_stride_is_one) {
2824 // If the main loop becomes empty and the array access for this range check is sunk out of the loop, the index
2825 // for the array access will be set to the index value of the final iteration which could be out of loop.
2826 // Add an Initialized Assertion Predicate for that corner case. The final iv is computed from LoopLimit which
2827 // is the LoopNode::limit() only if abs(stride) == 1 otherwise the computation depends on LoopNode::init_trip()
2828 // as well. When LoopLimit only depends on LoopNode::limit(), there are cases where the zero trip guard for
2829 // the main loop doesn't constant fold after range check elimination but, the array access for the final
2830 // iteration of the main loop is out of bound and the index for that access is out of range for the range
2831 // check CastII.
2832 // Note that we do not need to emit a Template Assertion Predicate to update this predicate. When further
2833 // splitting this loop, the final IV will still be the same. When unrolling the loop, we will remove a
2834 // previously added Initialized Assertion Predicate here. But then abs(stride) is greater than 1, and we
2835 // cannot remove an empty loop with a constant limit when init is not a constant as well. We will use
2836 // a LoopLimitCheck node that can only be folded if the zero grip guard is also foldable.
2837 loop_entry = initialized_assertion_predicate_creator.create(final_iv_placeholder, loop_entry, stride_con,
2838 scale_con, int_offset, int_limit,
2839 AssertionPredicateType::FinalIv);
2840 }
2841
2842 // Add two Template Assertion Predicates to create new Initialized Assertion Predicates from when either
2843 // unrolling or splitting this main-loop further.
2844 TemplateAssertionPredicateCreator template_assertion_predicate_creator(cl, scale_con , int_offset, int_limit,
2845 this);
2846 loop_entry = template_assertion_predicate_creator.create(loop_entry);
2847
2848 // Initialized Assertion Predicate for the value of the initial main-loop.
2849 loop_entry = initialized_assertion_predicate_creator.create(init, loop_entry, stride_con, scale_con,
2850 int_offset, int_limit,
2851 AssertionPredicateType::InitValue);
2852
2853 } else {
2854 if (PrintOpto) {
2855 tty->print_cr("missed RCE opportunity");
2856 }
2857 continue; // In release mode, ignore it
2858 }
2859 } else { // Otherwise work on normal compares
2860 switch(b_test._test) {
2861 case BoolTest::gt:
2862 // Fall into GE case
2863 case BoolTest::ge:
2864 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
2865 lscale_con = -lscale_con;
2866 offset = new SubLNode(zero, offset);
2867 register_new_node(offset, next_limit_ctrl);
2868 limit = new SubLNode(zero, limit);
2869 register_new_node(limit, next_limit_ctrl);
2870 // Fall into LE case
2871 case BoolTest::le:
2872 if (b_test._test != BoolTest::gt) {
2873 // Convert X <= Y to X < Y+1
2874 limit = new AddLNode(limit, one);
2875 register_new_node(limit, next_limit_ctrl);
2876 }
2877 // Fall into LT case
2878 case BoolTest::lt:
2879 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
2880 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
2881 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2882 add_constraint(stride_con, lscale_con, offset, mini, limit, next_limit_ctrl, &pre_limit, &main_limit);
2883 break;
2884 default:
2885 if (PrintOpto) {
2886 tty->print_cr("missed RCE opportunity");
2887 }
2888 continue; // Unhandled case
2889 }
2890 }
2891 // Only update variable tracking control for new nodes if it's indeed a range check that can be eliminated (and
2892 // limits are updated)
2893 new_limit_ctrl = next_limit_ctrl;
2894
2895 // Kill the eliminated test
2896 C->set_major_progress();
2897 Node* kill_con = intcon(1-flip);
2898 _igvn.replace_input_of(iff, 1, kill_con);
2899 // Find surviving projection
2900 assert(iff->is_If(), "");
2901 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2902 // Find loads off the surviving projection; remove their control edge
2903 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2904 Node* cd = dp->fast_out(i); // Control-dependent node
2905 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
2906 // Allow the load to float around in the loop, or before it
2907 // but NOT before the pre-loop.
2908 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not null
2909 --i;
2910 --imax;
2911 }
2912 }
2913 } // End of is IF
2914 }
2915 if (loop_entry != cl->skip_strip_mined()->in(LoopNode::EntryControl)) {
2916 _igvn.replace_input_of(cl->skip_strip_mined(), LoopNode::EntryControl, loop_entry);
2917 set_idom(cl->skip_strip_mined(), loop_entry, dom_depth(cl->skip_strip_mined()));
2918 }
2919
2920 // Update loop limits
2921 if (pre_limit != orig_limit) {
2922 // Computed pre-loop limit can be outside of loop iterations range.
2923 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2924 : (Node*)new MaxINode(pre_limit, orig_limit);
2925 register_new_node(pre_limit, new_limit_ctrl);
2926 }
2927 // new pre_limit can push Bool/Cmp/Opaque nodes down (when one of the eliminated condition has parameters that are not
2928 // loop invariant in the pre loop.
2929 set_ctrl(pre_opaq, new_limit_ctrl);
2930 // Can't use new_limit_ctrl for Bool/Cmp because it can be out of loop while they are loop variant. Conservatively set
2931 // control to latest possible one.
2932 set_ctrl(pre_end->cmp_node(), pre_end->in(0));
2933 set_ctrl(pre_end->in(1), pre_end->in(0));
2934
2935 _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2936
2937 // Note:: we are making the main loop limit no longer precise;
2938 // need to round up based on stride.
2939 cl->set_nonexact_trip_count();
2940 Node *main_cle = cl->loopexit();
2941 Node *main_bol = main_cle->in(1);
2942 // Hacking loop bounds; need private copies of exit test
2943 if (main_bol->outcnt() > 1) { // BoolNode shared?
2944 main_bol = main_bol->clone(); // Clone a private BoolNode
2945 register_new_node(main_bol, main_cle->in(0));
2946 _igvn.replace_input_of(main_cle, 1, main_bol);
2947 }
2948 Node *main_cmp = main_bol->in(1);
2949 if (main_cmp->outcnt() > 1) { // CmpNode shared?
2950 main_cmp = main_cmp->clone(); // Clone a private CmpNode
2951 register_new_node(main_cmp, main_cle->in(0));
2952 _igvn.replace_input_of(main_bol, 1, main_cmp);
2953 }
2954 assert(main_limit == cl->limit() || get_ctrl(main_limit) == new_limit_ctrl, "wrong control for added limit");
2955 const TypeInt* orig_limit_t = _igvn.type(orig_limit)->is_int();
2956 bool upward = cl->stride_con() > 0;
2957 // The new loop limit is <= (for an upward loop) >= (for a downward loop) than the orig limit.
2958 // The expression that computes the new limit may be too complicated and the computed type of the new limit
2959 // may be too pessimistic. A CastII here guarantees it's not lost.
2960 main_limit = new CastIINode(pre_ctrl, main_limit, TypeInt::make(upward ? min_jint : orig_limit_t->_lo,
2961 upward ? orig_limit_t->_hi : max_jint, Type::WidenMax));
2962 register_new_node(main_limit, new_limit_ctrl);
2963 // Hack the now-private loop bounds
2964 _igvn.replace_input_of(main_cmp, 2, main_limit);
2965 if (abs_stride_is_one) {
2966 Node* final_iv = new SubINode(main_limit, cl->stride());
2967 register_new_node(final_iv, loop_entry);
2968 _igvn.replace_node(final_iv_placeholder, final_iv);
2969 }
2970 // The OpaqueNode is unshared by design
2971 assert(opqzm->outcnt() == 1, "cannot hack shared node");
2972 _igvn.replace_input_of(opqzm, 1, main_limit);
2973 // new main_limit can push opaque node for zero trip guard down (when one of the eliminated condition has parameters
2974 // that are not loop invariant in the pre loop).
2975 set_ctrl(opqzm, new_limit_ctrl);
2976 // Bool/Cmp nodes for zero trip guard should have been assigned control between the main and pre loop (because zero
2977 // trip guard depends on induction variable value out of pre loop) so shouldn't need to be adjusted
2978 assert(is_dominator(new_limit_ctrl, get_ctrl(iffm->in(1)->in(1))), "control of cmp should be below control of updated input");
2979
2980 C->print_method(PHASE_AFTER_RANGE_CHECK_ELIMINATION, 4, cl);
2981 }
2982
2983 // Adjust control for node and its inputs (and inputs of its inputs) to be above the pre end
2984 Node* PhaseIdealLoop::ensure_node_and_inputs_are_above_pre_end(CountedLoopEndNode* pre_end, Node* node) {
2985 Node* control = get_ctrl(node);
2986 assert(is_dominator(compute_early_ctrl(node, control), pre_end), "node pinned on loop exit test?");
2987
2988 if (is_dominator(control, pre_end)) {
2989 return control;
2990 }
2991 control = pre_end->in(0);
2992 ResourceMark rm;
2993 Unique_Node_List wq;
2994 wq.push(node);
2995 for (uint i = 0; i < wq.size(); i++) {
2996 Node* n = wq.at(i);
2997 assert(is_dominator(compute_early_ctrl(n, get_ctrl(n)), pre_end), "node pinned on loop exit test?");
2998 set_ctrl(n, control);
2999 for (uint j = 0; j < n->req(); j++) {
3000 Node* in = n->in(j);
3001 if (in != nullptr && has_ctrl(in) && !is_dominator(get_ctrl(in), pre_end)) {
3002 wq.push(in);
3003 }
3004 }
3005 }
3006 return control;
3007 }
3008
3009 bool IdealLoopTree::compute_has_range_checks() const {
3010 assert(_head->is_CountedLoop(), "");
3011 for (uint i = 0; i < _body.size(); i++) {
3012 Node *iff = _body[i];
3013 int iff_opc = iff->Opcode();
3014 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
3015 return true;
3016 }
3017 }
3018 return false;
3019 }
3020
3021 //------------------------------DCE_loop_body----------------------------------
3022 // Remove simplistic dead code from loop body
3023 void IdealLoopTree::DCE_loop_body() {
3024 for (uint i = 0; i < _body.size(); i++) {
3025 if (_body.at(i)->outcnt() == 0) {
3026 _body.map(i, _body.pop());
3027 i--; // Ensure we revisit the updated index.
3028 }
3029 }
3030 }
3031
3032
3033 //------------------------------adjust_loop_exit_prob--------------------------
3034 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
3035 // Replace with a 1-in-10 exit guess.
3036 void IdealLoopTree::adjust_loop_exit_prob(PhaseIdealLoop *phase) {
3037 Node *test = tail();
3038 while (test != _head) {
3039 uint top = test->Opcode();
3040 if (top == Op_IfTrue || top == Op_IfFalse) {
3041 int test_con = ((ProjNode*)test)->_con;
3042 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
3043 IfNode *iff = test->in(0)->as_If();
3044 if (iff->outcnt() == 2) { // Ignore dead tests
3045 Node *bol = iff->in(1);
3046 if (bol && bol->req() > 1 && bol->in(1) &&
3047 ((bol->in(1)->Opcode() == Op_CompareAndExchangeB) ||
3048 (bol->in(1)->Opcode() == Op_CompareAndExchangeS) ||
3049 (bol->in(1)->Opcode() == Op_CompareAndExchangeI) ||
3050 (bol->in(1)->Opcode() == Op_CompareAndExchangeL) ||
3051 (bol->in(1)->Opcode() == Op_CompareAndExchangeP) ||
3052 (bol->in(1)->Opcode() == Op_CompareAndExchangeN) ||
3053 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapB) ||
3054 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapS) ||
3055 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI) ||
3056 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL) ||
3057 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP) ||
3058 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN) ||
3059 (bol->in(1)->Opcode() == Op_CompareAndSwapB) ||
3060 (bol->in(1)->Opcode() == Op_CompareAndSwapS) ||
3061 (bol->in(1)->Opcode() == Op_CompareAndSwapI) ||
3062 (bol->in(1)->Opcode() == Op_CompareAndSwapL) ||
3063 (bol->in(1)->Opcode() == Op_CompareAndSwapP) ||
3064 (bol->in(1)->Opcode() == Op_CompareAndSwapN) ||
3065 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndExchangeP) ||
3066 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndExchangeN) ||
3067 (bol->in(1)->Opcode() == Op_ShenandoahWeakCompareAndSwapP) ||
3068 (bol->in(1)->Opcode() == Op_ShenandoahWeakCompareAndSwapN) ||
3069 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndSwapP) ||
3070 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndSwapN)))
3071 return; // Allocation loops RARELY take backedge
3072 // Find the OTHER exit path from the IF
3073 Node* ex = iff->proj_out(1-test_con);
3074 float p = iff->_prob;
3075 if (!phase->is_member(this, ex) && iff->_fcnt == COUNT_UNKNOWN) {
3076 if (top == Op_IfTrue) {
3077 if (p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
3078 iff->_prob = PROB_STATIC_FREQUENT;
3079 }
3080 } else {
3081 if (p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
3082 iff->_prob = PROB_STATIC_INFREQUENT;
3083 }
3084 }
3085 }
3086 }
3087 }
3088 test = phase->idom(test);
3089 }
3090 }
3091
3092 static CountedLoopNode* locate_pre_from_main(CountedLoopNode* main_loop) {
3093 assert(!main_loop->is_main_no_pre_loop(), "Does not have a pre loop");
3094 Node* ctrl = main_loop->skip_assertion_predicates_with_halt();
3095 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
3096 Node* iffm = ctrl->in(0);
3097 assert(iffm->Opcode() == Op_If, "");
3098 Node* p_f = iffm->in(0);
3099 assert(p_f->Opcode() == Op_IfFalse, "");
3100 CountedLoopNode* pre_loop = p_f->in(0)->as_CountedLoopEnd()->loopnode();
3101 assert(pre_loop->is_pre_loop(), "No pre loop found");
3102 return pre_loop;
3103 }
3104
3105 // Remove the main and post loops and make the pre loop execute all
3106 // iterations. Useful when the pre loop is found empty.
3107 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) {
3108 CountedLoopEndNode* pre_end = cl->loopexit();
3109 Node* pre_cmp = pre_end->cmp_node();
3110 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) {
3111 // Only safe to remove the main loop if the compiler optimized it
3112 // out based on an unknown number of iterations
3113 return;
3114 }
3115
3116 // Can we find the main loop?
3117 if (_next == nullptr) {
3118 return;
3119 }
3120
3121 Node* next_head = _next->_head;
3122 if (!next_head->is_CountedLoop()) {
3123 return;
3124 }
3125
3126 CountedLoopNode* main_head = next_head->as_CountedLoop();
3127 if (!main_head->is_main_loop() || main_head->is_main_no_pre_loop()) {
3128 return;
3129 }
3130
3131 // We found a main-loop after this pre-loop, but they might not belong together.
3132 if (locate_pre_from_main(main_head) != cl) {
3133 return;
3134 }
3135
3136 Node* main_iff = main_head->skip_assertion_predicates_with_halt()->in(0);
3137
3138 // Remove the Opaque1Node of the pre loop and make it execute all iterations
3139 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2));
3140 // Remove the OpaqueZeroTripGuardNode of the main loop so it can be optimized out
3141 Node* main_cmp = main_iff->in(1)->in(1);
3142 assert(main_cmp->in(2)->Opcode() == Op_OpaqueZeroTripGuard, "main loop has no opaque node?");
3143 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1));
3144 }
3145
3146 //------------------------------do_remove_empty_loop---------------------------
3147 // We always attempt remove empty loops. The approach is to replace the trip
3148 // counter with the value it will have on the last iteration. This will break
3149 // the loop.
3150 bool IdealLoopTree::do_remove_empty_loop(PhaseIdealLoop *phase) {
3151 if (!_head->is_CountedLoop()) {
3152 return false; // Dead loop
3153 }
3154 if (!empty_loop_candidate(phase)) {
3155 return false;
3156 }
3157 CountedLoopNode *cl = _head->as_CountedLoop();
3158 #ifdef ASSERT
3159 // Call collect_loop_core_nodes to exercise the assert that checks that it finds the right number of nodes
3160 if (empty_loop_with_extra_nodes_candidate(phase)) {
3161 Unique_Node_List wq;
3162 collect_loop_core_nodes(phase, wq);
3163 }
3164 #endif
3165 // Minimum size must be empty loop
3166 if (_body.size() > EMPTY_LOOP_SIZE) {
3167 // This loop has more nodes than an empty loop but, maybe they are only kept alive by the outer strip mined loop's
3168 // safepoint. If they go away once the safepoint is removed, that loop is empty.
3169 if (!empty_loop_with_data_nodes(phase)) {
3170 return false;
3171 }
3172 }
3173 phase->C->print_method(PHASE_BEFORE_REMOVE_EMPTY_LOOP, 4, cl);
3174 if (cl->is_pre_loop()) {
3175 // If the loop we are removing is a pre-loop then the main and post loop
3176 // can be removed as well.
3177 remove_main_post_loops(cl, phase);
3178 }
3179
3180 #ifdef ASSERT
3181 // Ensure at most one used phi exists, which is the iv.
3182 Node* iv = nullptr;
3183 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
3184 Node* n = cl->fast_out(i);
3185 if ((n->Opcode() == Op_Phi) && (n->outcnt() > 0)) {
3186 assert(iv == nullptr, "Too many phis");
3187 iv = n;
3188 }
3189 }
3190 assert(iv == cl->phi(), "Wrong phi");
3191 #endif
3192
3193 // main and post loops have explicitly created zero trip guard
3194 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
3195 if (needs_guard) {
3196 // Skip guard if values not overlap.
3197 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
3198 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
3199 int stride_con = cl->stride_con();
3200 if (stride_con > 0) {
3201 needs_guard = (init_t->_hi >= limit_t->_lo);
3202 } else {
3203 needs_guard = (init_t->_lo <= limit_t->_hi);
3204 }
3205 }
3206 if (needs_guard) {
3207 // Check for an obvious zero trip guard.
3208 Predicates predicates(cl->skip_strip_mined()->in(LoopNode::EntryControl));
3209 Node* in_ctrl = predicates.entry();
3210 if (in_ctrl->Opcode() == Op_IfTrue || in_ctrl->Opcode() == Op_IfFalse) {
3211 bool maybe_swapped = (in_ctrl->Opcode() == Op_IfFalse);
3212 // The test should look like just the backedge of a CountedLoop
3213 Node* iff = in_ctrl->in(0);
3214 if (iff->is_If()) {
3215 Node* bol = iff->in(1);
3216 if (bol->is_Bool()) {
3217 BoolTest test = bol->as_Bool()->_test;
3218 if (maybe_swapped) {
3219 test._test = test.commute();
3220 test._test = test.negate();
3221 }
3222 if (test._test == cl->loopexit()->test_trip()) {
3223 Node* cmp = bol->in(1);
3224 int init_idx = maybe_swapped ? 2 : 1;
3225 int limit_idx = maybe_swapped ? 1 : 2;
3226 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) {
3227 needs_guard = false;
3228 }
3229 }
3230 }
3231 }
3232 }
3233 }
3234
3235 #ifndef PRODUCT
3236 if (PrintOpto) {
3237 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
3238 this->dump_head();
3239 } else if (TraceLoopOpts) {
3240 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
3241 this->dump_head();
3242 }
3243 #endif
3244
3245 if (needs_guard) {
3246 // Peel the loop to ensure there's a zero trip guard
3247 Node_List old_new;
3248 phase->do_peeling(this, old_new);
3249 }
3250
3251 // Replace the phi at loop head with the final value of the last
3252 // iteration (exact_limit - stride), to make sure the loop exit value
3253 // is correct, for any users after the loop.
3254 // Note: the final value after increment should not overflow since
3255 // counted loop has limit check predicate.
3256 Node* phi = cl->phi();
3257 Node* exact_limit = phase->exact_limit(this);
3258
3259 // We need to pin the exact limit to prevent it from floating above the zero trip guard.
3260 Node* cast_ii = ConstraintCastNode::make_cast_for_basic_type(
3261 cl->in(LoopNode::EntryControl), exact_limit,
3262 phase->_igvn.type(exact_limit),
3263 ConstraintCastNode::DependencyType::NonFloatingNonNarrowing, T_INT);
3264 phase->register_new_node(cast_ii, cl->in(LoopNode::EntryControl));
3265
3266 Node* final_iv = new SubINode(cast_ii, cl->stride());
3267 phase->register_new_node(final_iv, cl->in(LoopNode::EntryControl));
3268 phase->_igvn.replace_node(phi, final_iv);
3269
3270 // Set loop-exit condition to false. Then the CountedLoopEnd will collapse,
3271 // because the back edge is never taken.
3272 Node* zero = phase->_igvn.intcon(0);
3273 phase->_igvn.replace_input_of(cl->loopexit(), CountedLoopEndNode::TestValue, zero);
3274
3275 phase->C->set_major_progress();
3276 phase->C->print_method(PHASE_AFTER_REMOVE_EMPTY_LOOP, 4, final_iv);
3277 return true;
3278 }
3279
3280 bool IdealLoopTree::empty_loop_candidate(PhaseIdealLoop* phase) const {
3281 CountedLoopNode *cl = _head->as_CountedLoop();
3282 if (!cl->is_valid_counted_loop(T_INT)) {
3283 return false; // Malformed loop
3284 }
3285 if (!phase->ctrl_is_member(this, cl->loopexit()->in(CountedLoopEndNode::TestValue))) {
3286 return false; // Infinite loop
3287 }
3288 return true;
3289 }
3290
3291 bool IdealLoopTree::empty_loop_with_data_nodes(PhaseIdealLoop* phase) const {
3292 CountedLoopNode* cl = _head->as_CountedLoop();
3293 if (!cl->is_strip_mined() || !empty_loop_with_extra_nodes_candidate(phase)) {
3294 return false;
3295 }
3296 Unique_Node_List empty_loop_nodes;
3297 Unique_Node_List wq;
3298
3299 // Start from all data nodes in the loop body that are not one of the EMPTY_LOOP_SIZE nodes expected in an empty body
3300 enqueue_data_nodes(phase, empty_loop_nodes, wq);
3301 // and now follow uses
3302 for (uint i = 0; i < wq.size(); ++i) {
3303 Node* n = wq.at(i);
3304 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3305 Node* u = n->fast_out(j);
3306 if (u->Opcode() == Op_SafePoint) {
3307 // found a safepoint. Maybe this loop's safepoint or another loop safepoint.
3308 if (!process_safepoint(phase, empty_loop_nodes, wq, u)) {
3309 return false;
3310 }
3311 } else {
3312 const Type* u_t = phase->_igvn.type(u);
3313 if (u_t == Type::CONTROL || u_t == Type::MEMORY || u_t == Type::ABIO) {
3314 // found a side effect
3315 return false;
3316 }
3317 wq.push(u);
3318 }
3319 }
3320 }
3321 // Nodes (ignoring the EMPTY_LOOP_SIZE nodes of the "core" of the loop) are kept alive by otherwise empty loops'
3322 // safepoints: kill them.
3323 for (uint i = 0; i < wq.size(); ++i) {
3324 Node* n = wq.at(i);
3325 phase->_igvn.replace_node(n, phase->C->top());
3326 }
3327
3328 #ifdef ASSERT
3329 for (uint i = 0; i < _body.size(); ++i) {
3330 Node* n = _body.at(i);
3331 assert(wq.member(n) || empty_loop_nodes.member(n), "missed a node in the body?");
3332 }
3333 #endif
3334
3335 return true;
3336 }
3337
3338 bool IdealLoopTree::process_safepoint(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes, Unique_Node_List& wq,
3339 Node* sfpt) const {
3340 CountedLoopNode* cl = _head->as_CountedLoop();
3341 if (cl->outer_safepoint() == sfpt) {
3342 // the current loop's safepoint
3343 return true;
3344 }
3345
3346 // Some other loop's safepoint. Maybe that loop is empty too.
3347 IdealLoopTree* sfpt_loop = phase->get_loop(sfpt);
3348 if (!sfpt_loop->_head->is_OuterStripMinedLoop()) {
3349 return false;
3350 }
3351 IdealLoopTree* sfpt_inner_loop = sfpt_loop->_child;
3352 CountedLoopNode* sfpt_cl = sfpt_inner_loop->_head->as_CountedLoop();
3353 assert(sfpt_cl->is_strip_mined(), "inconsistent");
3354
3355 if (empty_loop_nodes.member(sfpt_cl)) {
3356 // already taken care of
3357 return true;
3358 }
3359
3360 if (!sfpt_inner_loop->empty_loop_candidate(phase) || !sfpt_inner_loop->empty_loop_with_extra_nodes_candidate(phase)) {
3361 return false;
3362 }
3363
3364 // Enqueue the nodes of that loop for processing too
3365 sfpt_inner_loop->enqueue_data_nodes(phase, empty_loop_nodes, wq);
3366 return true;
3367 }
3368
3369 bool IdealLoopTree::empty_loop_with_extra_nodes_candidate(PhaseIdealLoop* phase) const {
3370 CountedLoopNode *cl = _head->as_CountedLoop();
3371 // No other control flow node in the loop body
3372 if (cl->loopexit()->in(0) != cl) {
3373 return false;
3374 }
3375
3376 if (phase->ctrl_is_member(this, cl->limit())) {
3377 return false;
3378 }
3379 return true;
3380 }
3381
3382 void IdealLoopTree::enqueue_data_nodes(PhaseIdealLoop* phase, Unique_Node_List& empty_loop_nodes,
3383 Unique_Node_List& wq) const {
3384 collect_loop_core_nodes(phase, empty_loop_nodes);
3385 for (uint i = 0; i < _body.size(); ++i) {
3386 Node* n = _body.at(i);
3387 if (!empty_loop_nodes.member(n)) {
3388 wq.push(n);
3389 }
3390 }
3391 }
3392
3393 // This collects the node that would be left if this body was empty
3394 void IdealLoopTree::collect_loop_core_nodes(PhaseIdealLoop* phase, Unique_Node_List& wq) const {
3395 uint before = wq.size();
3396 wq.push(_head->in(LoopNode::LoopBackControl));
3397 for (uint i = before; i < wq.size(); ++i) {
3398 Node* n = wq.at(i);
3399 for (uint j = 0; j < n->req(); ++j) {
3400 Node* in = n->in(j);
3401 if (in != nullptr) {
3402 if (phase->get_loop(phase->ctrl_or_self(in)) == this) {
3403 wq.push(in);
3404 }
3405 }
3406 }
3407 }
3408 assert(wq.size() - before == EMPTY_LOOP_SIZE, "expect the EMPTY_LOOP_SIZE nodes of this body if empty");
3409 }
3410
3411 //------------------------------do_one_iteration_loop--------------------------
3412 // Convert one-iteration loop into normal code.
3413 bool IdealLoopTree::do_one_iteration_loop(PhaseIdealLoop *phase) {
3414 if (!_head->as_Loop()->is_valid_counted_loop(T_INT)) {
3415 return false; // Only for counted loop
3416 }
3417 CountedLoopNode *cl = _head->as_CountedLoop();
3418 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
3419 return false;
3420 }
3421
3422 #ifndef PRODUCT
3423 if (TraceLoopOpts) {
3424 tty->print("OneIteration ");
3425 this->dump_head();
3426 }
3427 #endif
3428
3429 phase->C->print_method(PHASE_BEFORE_ONE_ITERATION_LOOP, 4, cl);
3430 Node *init_n = cl->init_trip();
3431 // Loop boundaries should be constant since trip count is exact.
3432 assert((cl->stride_con() > 0 && init_n->get_int() + cl->stride_con() >= cl->limit()->get_int()) ||
3433 (cl->stride_con() < 0 && init_n->get_int() + cl->stride_con() <= cl->limit()->get_int()), "should be one iteration");
3434 // Replace the phi at loop head with the value of the init_trip.
3435 // Then the CountedLoopEnd will collapse (backedge will not be taken)
3436 // and all loop-invariant uses of the exit values will be correct.
3437 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
3438 phase->C->set_major_progress();
3439 phase->C->print_method(PHASE_AFTER_ONE_ITERATION_LOOP, 4, init_n);
3440 return true;
3441 }
3442
3443 //=============================================================================
3444 //------------------------------iteration_split_impl---------------------------
3445 bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_new) {
3446 if (!_head->is_Loop()) {
3447 // Head could be a region with a NeverBranch that was added in beautify loops but the region was not
3448 // yet transformed into a LoopNode. Bail out and wait until beautify loops turns it into a Loop node.
3449 return false;
3450 }
3451 // Compute loop trip count if possible.
3452 compute_trip_count(phase, T_INT);
3453
3454 // Convert one-iteration loop into normal code.
3455 if (do_one_iteration_loop(phase)) {
3456 return true;
3457 }
3458 // Check and remove empty loops (spam micro-benchmarks)
3459 if (do_remove_empty_loop(phase)) {
3460 return true; // Here we removed an empty loop
3461 }
3462
3463 AutoNodeBudget node_budget(phase);
3464
3465 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
3466 // This removes loop-invariant tests (usually null checks).
3467 if (!_head->is_CountedLoop()) { // Non-counted loop
3468 if (PartialPeelLoop) {
3469 bool rc = phase->partial_peel(this, old_new);
3470 if (Compile::current()->failing()) { return false; }
3471 if (rc) {
3472 // Partial peel succeeded so terminate this round of loop opts
3473 return false;
3474 }
3475 }
3476 if (policy_peeling(phase)) { // Should we peel?
3477 if (PrintOpto) { tty->print_cr("should_peel"); }
3478 phase->do_peeling(this, old_new);
3479 } else if (policy_unswitching(phase)) {
3480 phase->do_unswitching(this, old_new);
3481 return false; // need to recalculate idom data
3482 } else if (phase->duplicate_loop_backedge(this, old_new)) {
3483 return false;
3484 } else if (_head->is_LongCountedLoop()) {
3485 phase->create_loop_nest(this, old_new);
3486 }
3487 return true;
3488 }
3489 CountedLoopNode *cl = _head->as_CountedLoop();
3490
3491 if (!cl->is_valid_counted_loop(T_INT)) return true; // Ignore various kinds of broken loops
3492
3493 // Do nothing special to pre- and post- loops
3494 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
3495
3496 // With multiversioning, we create a fast_loop and a slow_loop, and a multiversion_if that
3497 // decides which loop is taken at runtime. At first, the multiversion_if always takes the
3498 // fast_loop, and we only optimize the fast_loop. Since we are not sure if we will ever use
3499 // the slow_loop, we delay optimizations for it, so we do not waste compile time and code
3500 // size. If we never change the condition of the multiversion_if, the slow_loop is eventually
3501 // folded away after loop-opts. While optimizing the fast_loop, we may want to perform some
3502 // speculative optimization, for which we need a runtime-check. We add this runtime-check
3503 // condition to the multiversion_if. Now, it becomes possible to execute the slow_loop at
3504 // runtime, and we resume optimizations for slow_loop ("un-delay" it).
3505 // TLDR: If the slow_loop is still in "delay" mode, check if the multiversion_if was changed
3506 // and we should now resume optimizations for it.
3507 if (cl->is_multiversion_delayed_slow_loop() &&
3508 !phase->try_resume_optimizations_for_delayed_slow_loop(this)) {
3509 // We are still delayed, so wait with further loop-opts.
3510 return true;
3511 }
3512
3513 // Compute loop trip count from profile data
3514 compute_profile_trip_cnt(phase);
3515
3516 // Before attempting fancy unrolling, RCE or alignment, see if we want
3517 // to completely unroll this loop or do loop unswitching.
3518 if (cl->is_normal_loop()) {
3519 if (policy_unswitching(phase)) {
3520 phase->do_unswitching(this, old_new);
3521 return false; // need to recalculate idom data
3522 }
3523 if (policy_maximally_unroll(phase)) {
3524 // Here we did some unrolling and peeling. Eventually we will
3525 // completely unroll this loop and it will no longer be a loop.
3526 phase->do_maximally_unroll(this, old_new);
3527 return true;
3528 }
3529 if (StressDuplicateBackedge && phase->duplicate_loop_backedge(this, old_new)) {
3530 return false;
3531 }
3532 }
3533
3534 uint est_peeling = estimate_peeling(phase);
3535 bool should_peel = 0 < est_peeling;
3536
3537 // Counted loops may be peeled, or may need some iterations run up
3538 // front for RCE. Thus we clone a full loop up front whose trip count is
3539 // at least 1 (if peeling), but may be several more.
3540
3541 // The main loop will start cache-line aligned with at least 1
3542 // iteration of the unrolled body (zero-trip test required) and
3543 // will have some range checks removed.
3544
3545 // A post-loop will finish any odd iterations (leftover after
3546 // unrolling), plus any needed for RCE purposes.
3547
3548 bool should_unroll = policy_unroll(phase);
3549 bool should_rce = policy_range_check(phase, false, T_INT);
3550 bool should_rce_long = policy_range_check(phase, false, T_LONG);
3551
3552 // If not RCE'ing (iteration splitting), then we do not need a pre-loop.
3553 // We may still need to peel an initial iteration but we will not
3554 // be needing an unknown number of pre-iterations.
3555 //
3556 // Basically, if peel_only reports TRUE first time through, we will not
3557 // be able to later do RCE on this loop.
3558 bool peel_only = policy_peel_only(phase) && !should_rce;
3559
3560 // If we have any of these conditions (RCE, unrolling) met, then
3561 // we switch to the pre-/main-/post-loop model. This model also covers
3562 // peeling.
3563 if (should_rce || should_unroll) {
3564 if (cl->is_normal_loop()) { // Convert to 'pre/main/post' loops
3565 if (should_rce_long && phase->create_loop_nest(this, old_new)) {
3566 return true;
3567 }
3568 uint estimate = est_loop_clone_sz(3);
3569 if (!phase->may_require_nodes(estimate)) {
3570 return false;
3571 }
3572
3573 if (!peel_only) {
3574 // We are going to add pre-loop and post-loop (PreMainPost).
3575 // But should we also multiversion for auto-vectorization speculative
3576 // checks, i.e. fast and slow-paths?
3577 // Note: Just PeelMainPost is not sufficient, as we could never find the
3578 // multiversion_if again from the main loop: we need a nicely structured
3579 // pre-loop, a peeled iteration cannot easily be parsed through.
3580 phase->maybe_multiversion_for_auto_vectorization_runtime_checks(this, old_new);
3581 }
3582
3583 phase->insert_pre_post_loops(this, old_new, peel_only);
3584 }
3585 // Adjust the pre- and main-loop limits to let the pre and post loops run
3586 // with full checks, but the main-loop with no checks. Remove said checks
3587 // from the main body.
3588 if (should_rce) {
3589 phase->do_range_check(this);
3590 }
3591
3592 // Double loop body for unrolling. Adjust the minimum-trip test (will do
3593 // twice as many iterations as before) and the main body limit (only do
3594 // an even number of trips). If we are peeling, we might enable some RCE
3595 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
3596 // peeling.
3597 if (should_unroll && !should_peel) {
3598 if (SuperWordLoopUnrollAnalysis) {
3599 phase->insert_vector_post_loop(this, old_new);
3600 }
3601 phase->do_unroll(this, old_new, true);
3602 }
3603 } else { // Else we have an unchanged counted loop
3604 if (should_peel) { // Might want to peel but do nothing else
3605 if (phase->may_require_nodes(est_peeling)) {
3606 phase->do_peeling(this, old_new);
3607 }
3608 }
3609 if (should_rce_long) {
3610 phase->create_loop_nest(this, old_new);
3611 }
3612 }
3613 return true;
3614 }
3615
3616
3617 //=============================================================================
3618 //------------------------------iteration_split--------------------------------
3619 bool IdealLoopTree::iteration_split(PhaseIdealLoop* phase, Node_List &old_new) {
3620 // Recursively iteration split nested loops
3621 if (_child && !_child->iteration_split(phase, old_new)) {
3622 return false;
3623 }
3624
3625 // Clean out prior deadwood
3626 DCE_loop_body();
3627
3628 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
3629 // Replace with a 1-in-10 exit guess.
3630 if (!is_root() && is_loop()) {
3631 adjust_loop_exit_prob(phase);
3632 }
3633
3634 // Unrolling, RCE and peeling efforts, iff innermost loop.
3635 if (_allow_optimizations && is_innermost()) {
3636 if (!_has_call) {
3637 if (!iteration_split_impl(phase, old_new)) {
3638 return false;
3639 }
3640 } else {
3641 AutoNodeBudget node_budget(phase);
3642 if (policy_unswitching(phase)) {
3643 phase->do_unswitching(this, old_new);
3644 return false; // need to recalculate idom data
3645 }
3646 }
3647 }
3648
3649 if (_next && !_next->iteration_split(phase, old_new)) {
3650 return false;
3651 }
3652 return true;
3653 }
3654
3655
3656 //=============================================================================
3657 // Process all the loops in the loop tree and replace any fill
3658 // patterns with an intrinsic version.
3659 bool PhaseIdealLoop::do_intrinsify_fill() {
3660 bool changed = false;
3661 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3662 IdealLoopTree* lpt = iter.current();
3663 changed |= intrinsify_fill(lpt);
3664 }
3665 return changed;
3666 }
3667
3668
3669 // Examine an inner loop looking for a single store of an invariant
3670 // value in a unit stride loop,
3671 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
3672 Node*& shift, Node*& con) {
3673 const char* msg = nullptr;
3674 Node* msg_node = nullptr;
3675
3676 store_value = nullptr;
3677 con = nullptr;
3678 shift = nullptr;
3679
3680 // Process the loop looking for stores. If there are multiple
3681 // stores or extra control flow give at this point.
3682 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3683 for (uint i = 0; msg == nullptr && i < lpt->_body.size(); i++) {
3684 Node* n = lpt->_body.at(i);
3685 if (n->outcnt() == 0) continue; // Ignore dead
3686 if (n->is_Store()) {
3687 if (store != nullptr) {
3688 msg = "multiple stores";
3689 break;
3690 }
3691 int opc = n->Opcode();
3692 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass) {
3693 msg = "oop fills not handled";
3694 break;
3695 }
3696 Node* value = n->in(MemNode::ValueIn);
3697 if (!lpt->is_invariant(value)) {
3698 msg = "variant store value";
3699 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
3700 msg = "not array address";
3701 }
3702 store = n;
3703 store_value = value;
3704 } else if (n->is_If() && n != head->loopexit_or_null()) {
3705 msg = "extra control flow";
3706 msg_node = n;
3707 }
3708 }
3709
3710 if (store == nullptr) {
3711 // No store in loop
3712 return false;
3713 }
3714
3715 if (msg == nullptr && store->as_Mem()->is_mismatched_access()) {
3716 // This optimization does not currently support mismatched stores, where the
3717 // type of the value to be stored differs from the element type of the
3718 // destination array. Such patterns arise for example from memory segment
3719 // initialization. This limitation could be overcome by extending this
3720 // function's address matching logic and ensuring that the fill intrinsic
3721 // implementations support mismatched array filling.
3722 msg = "mismatched store";
3723 }
3724
3725 if (msg == nullptr && head->stride_con() != 1) {
3726 // could handle negative strides too
3727 if (head->stride_con() < 0) {
3728 msg = "negative stride";
3729 } else {
3730 msg = "non-unit stride";
3731 }
3732 }
3733
3734 if (msg == nullptr && !store->in(MemNode::Address)->is_AddP()) {
3735 msg = "can't handle store address";
3736 msg_node = store->in(MemNode::Address);
3737 }
3738
3739 if (msg == nullptr &&
3740 (!store->in(MemNode::Memory)->is_Phi() ||
3741 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
3742 msg = "store memory isn't proper phi";
3743 msg_node = store->in(MemNode::Memory);
3744 }
3745
3746 // Make sure there is an appropriate fill routine
3747 BasicType t = msg == nullptr ?
3748 store->adr_type()->isa_aryptr()->elem()->array_element_basic_type() : T_VOID;
3749 const char* fill_name;
3750 if (msg == nullptr &&
3751 StubRoutines::select_fill_function(t, false, fill_name) == nullptr) {
3752 msg = "unsupported store";
3753 msg_node = store;
3754 }
3755
3756 if (msg != nullptr) {
3757 #ifndef PRODUCT
3758 if (TraceOptimizeFill) {
3759 tty->print_cr("not fill intrinsic candidate: %s", msg);
3760 if (msg_node != nullptr) msg_node->dump();
3761 }
3762 #endif
3763 return false;
3764 }
3765
3766 // Make sure the address expression can be handled. It should be
3767 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()).
3768 Node* elements[4];
3769 Node* cast = nullptr;
3770 Node* conv = nullptr;
3771 bool found_index = false;
3772 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
3773 for (int e = 0; e < count; e++) {
3774 Node* n = elements[e];
3775 if (n->is_Con() && con == nullptr) {
3776 con = n;
3777 } else if (n->Opcode() == Op_LShiftX && shift == nullptr) {
3778 Node* value = n->in(1);
3779 #ifdef _LP64
3780 if (value->Opcode() == Op_ConvI2L) {
3781 conv = value;
3782 value = value->in(1);
3783 }
3784 if (value->Opcode() == Op_CastII &&
3785 value->as_CastII()->has_range_check()) {
3786 // Skip range check dependent CastII nodes
3787 cast = value;
3788 value = value->in(1);
3789 }
3790 #endif
3791 if (value != head->phi()) {
3792 msg = "unhandled shift in address";
3793 } else {
3794 if (type2aelembytes(t, true) != (1 << n->in(2)->get_int())) {
3795 msg = "scale doesn't match";
3796 } else {
3797 found_index = true;
3798 shift = n;
3799 }
3800 }
3801 } else if (n->Opcode() == Op_ConvI2L && conv == nullptr) {
3802 conv = n;
3803 n = n->in(1);
3804 if (n->Opcode() == Op_CastII &&
3805 n->as_CastII()->has_range_check()) {
3806 // Skip range check dependent CastII nodes
3807 cast = n;
3808 n = n->in(1);
3809 }
3810 if (n == head->phi()) {
3811 found_index = true;
3812 } else {
3813 msg = "unhandled input to ConvI2L";
3814 }
3815 } else if (n == head->phi()) {
3816 // no shift, check below for allowed cases
3817 found_index = true;
3818 } else {
3819 msg = "unhandled node in address";
3820 msg_node = n;
3821 }
3822 }
3823
3824 if (count == -1) {
3825 msg = "malformed address expression";
3826 msg_node = store;
3827 }
3828
3829 if (!found_index) {
3830 msg = "missing use of index";
3831 }
3832
3833 // byte sized items won't have a shift
3834 if (msg == nullptr && shift == nullptr && t != T_BYTE && t != T_BOOLEAN) {
3835 msg = "can't find shift";
3836 msg_node = store;
3837 }
3838
3839 if (msg != nullptr) {
3840 #ifndef PRODUCT
3841 if (TraceOptimizeFill) {
3842 tty->print_cr("not fill intrinsic: %s", msg);
3843 if (msg_node != nullptr) msg_node->dump();
3844 }
3845 #endif
3846 return false;
3847 }
3848
3849 // No make sure all the other nodes in the loop can be handled
3850 VectorSet ok;
3851
3852 // store related values are ok
3853 ok.set(store->_idx);
3854 ok.set(store->in(MemNode::Memory)->_idx);
3855
3856 CountedLoopEndNode* loop_exit = head->loopexit();
3857
3858 // Loop structure is ok
3859 ok.set(head->_idx);
3860 ok.set(loop_exit->_idx);
3861 ok.set(head->phi()->_idx);
3862 ok.set(head->incr()->_idx);
3863 ok.set(loop_exit->cmp_node()->_idx);
3864 ok.set(loop_exit->in(1)->_idx);
3865
3866 // Address elements are ok
3867 if (con) ok.set(con->_idx);
3868 if (shift) ok.set(shift->_idx);
3869 if (cast) ok.set(cast->_idx);
3870 if (conv) ok.set(conv->_idx);
3871
3872 for (uint i = 0; msg == nullptr && i < lpt->_body.size(); i++) {
3873 Node* n = lpt->_body.at(i);
3874 if (n->outcnt() == 0) continue; // Ignore dead
3875 if (ok.test(n->_idx)) continue;
3876 // Backedge projection is ok
3877 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
3878 if (!n->is_AddP()) {
3879 msg = "unhandled node";
3880 msg_node = n;
3881 break;
3882 }
3883 }
3884
3885 // Make sure no unexpected values are used outside the loop
3886 for (uint i = 0; msg == nullptr && i < lpt->_body.size(); i++) {
3887 Node* n = lpt->_body.at(i);
3888 // These values can be replaced with other nodes if they are used
3889 // outside the loop.
3890 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
3891 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
3892 Node* use = iter.get();
3893 if (!lpt->_body.contains(use)) {
3894 msg = "node is used outside loop";
3895 msg_node = n;
3896 break;
3897 }
3898 }
3899 }
3900
3901 #ifdef ASSERT
3902 if (TraceOptimizeFill) {
3903 if (msg != nullptr) {
3904 tty->print_cr("no fill intrinsic: %s", msg);
3905 if (msg_node != nullptr) msg_node->dump();
3906 } else {
3907 tty->print_cr("fill intrinsic for:");
3908 }
3909 store->dump();
3910 if (Verbose) {
3911 lpt->_body.dump();
3912 }
3913 }
3914 #endif
3915
3916 return msg == nullptr;
3917 }
3918
3919
3920
3921 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
3922 // Only for counted inner loops
3923 if (!lpt->is_counted() || !lpt->is_innermost()) {
3924 return false;
3925 }
3926
3927 // Must have constant stride
3928 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3929 if (!head->is_valid_counted_loop(T_INT) || !head->is_normal_loop()) {
3930 return false;
3931 }
3932
3933 head->verify_strip_mined(1);
3934
3935 // Check that the body only contains a store of a loop invariant
3936 // value that is indexed by the loop phi.
3937 Node* store = nullptr;
3938 Node* store_value = nullptr;
3939 Node* shift = nullptr;
3940 Node* offset = nullptr;
3941 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
3942 return false;
3943 }
3944
3945 IfFalseNode* exit = head->loopexit()->false_proj_or_null();
3946 if (exit == nullptr) {
3947 return false;
3948 }
3949
3950 #ifndef PRODUCT
3951 if (TraceLoopOpts) {
3952 tty->print("ArrayFill ");
3953 lpt->dump_head();
3954 }
3955 #endif
3956
3957 // Now replace the whole loop body by a call to a fill routine that
3958 // covers the same region as the loop.
3959 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
3960
3961 // Build an expression for the beginning of the copy region
3962 Node* index = head->init_trip();
3963 #ifdef _LP64
3964 index = new ConvI2LNode(index);
3965 _igvn.register_new_node_with_optimizer(index);
3966 #endif
3967 if (shift != nullptr) {
3968 // byte arrays don't require a shift but others do.
3969 index = new LShiftXNode(index, shift->in(2));
3970 _igvn.register_new_node_with_optimizer(index);
3971 }
3972 Node* from = new AddPNode(base, base, index);
3973 _igvn.register_new_node_with_optimizer(from);
3974 // For normal array fills, C2 uses two AddP nodes for array element
3975 // addressing. But for array fills with Unsafe call, there's only one
3976 // AddP node adding an absolute offset, so we do a null check here.
3977 assert(offset != nullptr || C->has_unsafe_access(),
3978 "Only array fills with unsafe have no extra offset");
3979 if (offset != nullptr) {
3980 from = new AddPNode(base, from, offset);
3981 _igvn.register_new_node_with_optimizer(from);
3982 }
3983 // Compute the number of elements to copy
3984 Node* len = new SubINode(head->limit(), head->init_trip());
3985 _igvn.register_new_node_with_optimizer(len);
3986
3987 // If the store is on the backedge, it is not executed in the last
3988 // iteration, and we must subtract 1 from the len.
3989 IfTrueNode* backedge = head->loopexit()->true_proj();
3990 if (store->in(0) == backedge) {
3991 len = new SubINode(len, _igvn.intcon(1));
3992 _igvn.register_new_node_with_optimizer(len);
3993 #ifndef PRODUCT
3994 if (TraceOptimizeFill) {
3995 tty->print_cr("ArrayFill store on backedge, subtract 1 from len.");
3996 }
3997 #endif
3998 }
3999
4000 BasicType t = store->adr_type()->isa_aryptr()->elem()->array_element_basic_type();
4001 bool aligned = false;
4002 if (offset != nullptr && head->init_trip()->is_Con()) {
4003 int element_size = type2aelembytes(t);
4004 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
4005 }
4006
4007 // Build a call to the fill routine
4008 const char* fill_name;
4009 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
4010 assert(fill != nullptr, "what?");
4011
4012 // Convert float/double to int/long for fill routines
4013 if (t == T_FLOAT) {
4014 store_value = new MoveF2INode(store_value);
4015 _igvn.register_new_node_with_optimizer(store_value);
4016 } else if (t == T_DOUBLE) {
4017 store_value = new MoveD2LNode(store_value);
4018 _igvn.register_new_node_with_optimizer(store_value);
4019 }
4020
4021 Node* mem_phi = store->in(MemNode::Memory);
4022 Node* result_ctrl;
4023 Node* result_mem;
4024 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
4025 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
4026 fill_name, TypeAryPtr::get_array_body_type(t));
4027 uint cnt = 0;
4028 call->init_req(TypeFunc::Parms + cnt++, from);
4029 call->init_req(TypeFunc::Parms + cnt++, store_value);
4030 #ifdef _LP64
4031 len = new ConvI2LNode(len);
4032 _igvn.register_new_node_with_optimizer(len);
4033 #endif
4034 call->init_req(TypeFunc::Parms + cnt++, len);
4035 #ifdef _LP64
4036 call->init_req(TypeFunc::Parms + cnt++, C->top());
4037 #endif
4038 call->init_req(TypeFunc::Control, head->init_control());
4039 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O.
4040 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl));
4041 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out_or_null(TypeFunc::ReturnAdr));
4042 Node* frame = new ParmNode(C->start(), TypeFunc::FramePtr);
4043 _igvn.register_new_node_with_optimizer(frame);
4044 call->init_req(TypeFunc::FramePtr, frame);
4045 _igvn.register_new_node_with_optimizer(call);
4046 result_ctrl = new ProjNode(call,TypeFunc::Control);
4047 _igvn.register_new_node_with_optimizer(result_ctrl);
4048 result_mem = new ProjNode(call,TypeFunc::Memory);
4049 _igvn.register_new_node_with_optimizer(result_mem);
4050
4051 /* Disable following optimization until proper fix (add missing checks).
4052
4053 // If this fill is tightly coupled to an allocation and overwrites
4054 // the whole body, allow it to take over the zeroing.
4055 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
4056 if (alloc != nullptr && alloc->is_AllocateArray()) {
4057 Node* length = alloc->as_AllocateArray()->Ideal_length();
4058 if (head->limit() == length &&
4059 head->init_trip() == _igvn.intcon(0)) {
4060 if (TraceOptimizeFill) {
4061 tty->print_cr("Eliminated zeroing in allocation");
4062 }
4063 alloc->maybe_set_complete(&_igvn);
4064 } else {
4065 #ifdef ASSERT
4066 if (TraceOptimizeFill) {
4067 tty->print_cr("filling array but bounds don't match");
4068 alloc->dump();
4069 head->init_trip()->dump();
4070 head->limit()->dump();
4071 length->dump();
4072 }
4073 #endif
4074 }
4075 }
4076 */
4077
4078 if (head->is_strip_mined()) {
4079 // Inner strip mined loop goes away so get rid of outer strip
4080 // mined loop
4081 Node* outer_sfpt = head->outer_safepoint();
4082 Node* in = outer_sfpt->in(0);
4083 Node* outer_out = head->outer_loop_exit();
4084 replace_node_and_forward_ctrl(outer_out, in);
4085 _igvn.replace_input_of(outer_sfpt, 0, C->top());
4086 }
4087
4088 // Redirect the old control and memory edges that are outside the loop.
4089 // Sometimes the memory phi of the head is used as the outgoing
4090 // state of the loop. It's safe in this case to replace it with the
4091 // result_mem.
4092 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
4093 replace_node_and_forward_ctrl(exit, result_ctrl);
4094 _igvn.replace_node(store, result_mem);
4095 // Any uses the increment outside of the loop become the loop limit.
4096 _igvn.replace_node(head->incr(), head->limit());
4097
4098 // Disconnect the head from the loop.
4099 for (uint i = 0; i < lpt->_body.size(); i++) {
4100 Node* n = lpt->_body.at(i);
4101 _igvn.replace_node(n, C->top());
4102 }
4103
4104 #ifndef PRODUCT
4105 if (TraceOptimizeFill) {
4106 tty->print("ArrayFill call ");
4107 call->dump();
4108 }
4109 #endif
4110
4111 return true;
4112 }