1 /*
2 * Copyright (c) 2000, 2021, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/castnode.hpp"
31 #include "opto/connode.hpp"
32 #include "opto/convertnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/movenode.hpp"
37 #include "opto/opaquenode.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "opto/subnode.hpp"
41 #include "opto/superword.hpp"
42 #include "opto/vectornode.hpp"
43 #include "runtime/globals_extension.hpp"
44
45 //------------------------------is_loop_exit-----------------------------------
46 // Given an IfNode, return the loop-exiting projection or NULL if both
47 // arms remain in the loop.
is_loop_exit(Node * iff) const48 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
49 if (iff->outcnt() != 2) return NULL; // Ignore partially dead tests
50 PhaseIdealLoop *phase = _phase;
51 // Test is an IfNode, has 2 projections. If BOTH are in the loop
52 // we need loop unswitching instead of peeling.
53 if (!is_member(phase->get_loop(iff->raw_out(0))))
54 return iff->raw_out(0);
55 if (!is_member(phase->get_loop(iff->raw_out(1))))
56 return iff->raw_out(1);
57 return NULL;
58 }
59
60
61 //=============================================================================
62
63
64 //------------------------------record_for_igvn----------------------------
65 // Put loop body on igvn work list
record_for_igvn()66 void IdealLoopTree::record_for_igvn() {
67 for (uint i = 0; i < _body.size(); i++) {
68 Node *n = _body.at(i);
69 _phase->_igvn._worklist.push(n);
70 }
71 // put body of outer strip mined loop on igvn work list as well
72 if (_head->is_CountedLoop() && _head->as_Loop()->is_strip_mined()) {
73 CountedLoopNode* l = _head->as_CountedLoop();
74 Node* outer_loop = l->outer_loop();
75 assert(outer_loop != NULL, "missing piece of strip mined loop");
76 _phase->_igvn._worklist.push(outer_loop);
77 Node* outer_loop_tail = l->outer_loop_tail();
78 assert(outer_loop_tail != NULL, "missing piece of strip mined loop");
79 _phase->_igvn._worklist.push(outer_loop_tail);
80 Node* outer_loop_end = l->outer_loop_end();
81 assert(outer_loop_end != NULL, "missing piece of strip mined loop");
82 _phase->_igvn._worklist.push(outer_loop_end);
83 Node* outer_safepoint = l->outer_safepoint();
84 assert(outer_safepoint != NULL, "missing piece of strip mined loop");
85 _phase->_igvn._worklist.push(outer_safepoint);
86 Node* cle_out = _head->as_CountedLoop()->loopexit()->proj_out(false);
87 assert(cle_out != NULL, "missing piece of strip mined loop");
88 _phase->_igvn._worklist.push(cle_out);
89 }
90 }
91
92 //------------------------------compute_exact_trip_count-----------------------
93 // Compute loop trip count if possible. Do not recalculate trip count for
94 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
compute_trip_count(PhaseIdealLoop * phase)95 void IdealLoopTree::compute_trip_count(PhaseIdealLoop* phase) {
96 if (!_head->as_Loop()->is_valid_counted_loop(T_INT)) {
97 return;
98 }
99 CountedLoopNode* cl = _head->as_CountedLoop();
100 // Trip count may become nonexact for iteration split loops since
101 // RCE modifies limits. Note, _trip_count value is not reset since
102 // it is used to limit unrolling of main loop.
103 cl->set_nonexact_trip_count();
104
105 // Loop's test should be part of loop.
106 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
107 return; // Infinite loop
108
109 #ifdef ASSERT
110 BoolTest::mask bt = cl->loopexit()->test_trip();
111 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
112 bt == BoolTest::ne, "canonical test is expected");
113 #endif
114
115 Node* init_n = cl->init_trip();
116 Node* limit_n = cl->limit();
117 if (init_n != NULL && limit_n != NULL) {
118 // Use longs to avoid integer overflow.
119 int stride_con = cl->stride_con();
120 const TypeInt* init_type = phase->_igvn.type(init_n)->is_int();
121 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
122 jlong init_con = (stride_con > 0) ? init_type->_lo : init_type->_hi;
123 jlong limit_con = (stride_con > 0) ? limit_type->_hi : limit_type->_lo;
124 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
125 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
126 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
127 if (init_n->is_Con() && limit_n->is_Con()) {
128 // Set exact trip count.
129 cl->set_exact_trip_count((uint)trip_count);
130 } else if (cl->unrolled_count() == 1) {
131 // Set maximum trip count before unrolling.
132 cl->set_trip_count((uint)trip_count);
133 }
134 }
135 }
136 }
137
138 //------------------------------compute_profile_trip_cnt----------------------------
139 // Compute loop trip count from profile data as
140 // (backedge_count + loop_exit_count) / loop_exit_count
141
compute_profile_trip_cnt_helper(Node * n)142 float IdealLoopTree::compute_profile_trip_cnt_helper(Node* n) {
143 if (n->is_If()) {
144 IfNode *iff = n->as_If();
145 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) {
146 Node *exit = is_loop_exit(iff);
147 if (exit) {
148 float exit_prob = iff->_prob;
149 if (exit->Opcode() == Op_IfFalse) {
150 exit_prob = 1.0 - exit_prob;
151 }
152 if (exit_prob > PROB_MIN) {
153 float exit_cnt = iff->_fcnt * exit_prob;
154 return exit_cnt;
155 }
156 }
157 }
158 }
159 if (n->is_Jump()) {
160 JumpNode *jmp = n->as_Jump();
161 if (jmp->_fcnt != COUNT_UNKNOWN) {
162 float* probs = jmp->_probs;
163 float exit_prob = 0;
164 PhaseIdealLoop *phase = _phase;
165 for (DUIterator_Fast imax, i = jmp->fast_outs(imax); i < imax; i++) {
166 JumpProjNode* u = jmp->fast_out(i)->as_JumpProj();
167 if (!is_member(_phase->get_loop(u))) {
168 exit_prob += probs[u->_con];
169 }
170 }
171 return exit_prob * jmp->_fcnt;
172 }
173 }
174 return 0;
175 }
176
compute_profile_trip_cnt(PhaseIdealLoop * phase)177 void IdealLoopTree::compute_profile_trip_cnt(PhaseIdealLoop *phase) {
178 if (!_head->is_Loop()) {
179 return;
180 }
181 LoopNode* head = _head->as_Loop();
182 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
183 return; // Already computed
184 }
185 float trip_cnt = (float)max_jint; // default is big
186
187 Node* back = head->in(LoopNode::LoopBackControl);
188 while (back != head) {
189 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
190 back->in(0) &&
191 back->in(0)->is_If() &&
192 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
193 back->in(0)->as_If()->_prob != PROB_UNKNOWN &&
194 (back->Opcode() == Op_IfTrue ? 1-back->in(0)->as_If()->_prob : back->in(0)->as_If()->_prob) > PROB_MIN) {
195 break;
196 }
197 back = phase->idom(back);
198 }
199 if (back != head) {
200 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
201 back->in(0), "if-projection exists");
202 IfNode* back_if = back->in(0)->as_If();
203 float loop_back_cnt = back_if->_fcnt * (back->Opcode() == Op_IfTrue ? back_if->_prob : (1 - back_if->_prob));
204
205 // Now compute a loop exit count
206 float loop_exit_cnt = 0.0f;
207 if (_child == NULL) {
208 for (uint i = 0; i < _body.size(); i++) {
209 Node *n = _body[i];
210 loop_exit_cnt += compute_profile_trip_cnt_helper(n);
211 }
212 } else {
213 ResourceMark rm;
214 Unique_Node_List wq;
215 wq.push(back);
216 for (uint i = 0; i < wq.size(); i++) {
217 Node *n = wq.at(i);
218 assert(n->is_CFG(), "only control nodes");
219 if (n != head) {
220 if (n->is_Region()) {
221 for (uint j = 1; j < n->req(); j++) {
222 wq.push(n->in(j));
223 }
224 } else {
225 loop_exit_cnt += compute_profile_trip_cnt_helper(n);
226 wq.push(n->in(0));
227 }
228 }
229 }
230
231 }
232 if (loop_exit_cnt > 0.0f) {
233 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
234 } else {
235 // No exit count so use
236 trip_cnt = loop_back_cnt;
237 }
238 } else {
239 head->mark_profile_trip_failed();
240 }
241 #ifndef PRODUCT
242 if (TraceProfileTripCount) {
243 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
244 }
245 #endif
246 head->set_profile_trip_cnt(trip_cnt);
247 }
248
249 //---------------------find_invariant-----------------------------
250 // Return nonzero index of invariant operand for an associative
251 // binary operation of (nonconstant) invariant and variant values.
252 // Helper for reassociate_invariants.
find_invariant(Node * n,PhaseIdealLoop * phase)253 int IdealLoopTree::find_invariant(Node* n, PhaseIdealLoop *phase) {
254 bool in1_invar = this->is_invariant(n->in(1));
255 bool in2_invar = this->is_invariant(n->in(2));
256 if (in1_invar && !in2_invar) return 1;
257 if (!in1_invar && in2_invar) return 2;
258 return 0;
259 }
260
261 //---------------------is_associative-----------------------------
262 // Return TRUE if "n" is an associative binary node. If "base" is
263 // not NULL, "n" must be re-associative with it.
is_associative(Node * n,Node * base)264 bool IdealLoopTree::is_associative(Node* n, Node* base) {
265 int op = n->Opcode();
266 if (base != NULL) {
267 assert(is_associative(base), "Base node should be associative");
268 int base_op = base->Opcode();
269 if (base_op == Op_AddI || base_op == Op_SubI) {
270 return op == Op_AddI || op == Op_SubI;
271 }
272 if (base_op == Op_AddL || base_op == Op_SubL) {
273 return op == Op_AddL || op == Op_SubL;
274 }
275 return op == base_op;
276 } else {
277 // Integer "add/sub/mul/and/or/xor" operations are associative.
278 return op == Op_AddI || op == Op_AddL
279 || op == Op_SubI || op == Op_SubL
280 || op == Op_MulI || op == Op_MulL
281 || op == Op_AndI || op == Op_AndL
282 || op == Op_OrI || op == Op_OrL
283 || op == Op_XorI || op == Op_XorL;
284 }
285 }
286
287 //---------------------reassociate_add_sub------------------------
288 // Reassociate invariant add and subtract expressions:
289 //
290 // inv1 + (x + inv2) => ( inv1 + inv2) + x
291 // (x + inv2) + inv1 => ( inv1 + inv2) + x
292 // inv1 + (x - inv2) => ( inv1 - inv2) + x
293 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
294 // (x + inv2) - inv1 => (-inv1 + inv2) + x
295 // (x - inv2) + inv1 => ( inv1 - inv2) + x
296 // (x - inv2) - inv1 => (-inv1 - inv2) + x
297 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
298 // inv1 - (x - inv2) => ( inv1 + inv2) - x
299 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
300 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
301 // inv1 - (x + inv2) => ( inv1 - inv2) - x
302 //
reassociate_add_sub(Node * n1,int inv1_idx,int inv2_idx,PhaseIdealLoop * phase)303 Node* IdealLoopTree::reassociate_add_sub(Node* n1, int inv1_idx, int inv2_idx, PhaseIdealLoop *phase) {
304 assert(n1->is_Add() || n1->is_Sub(), "Target node should be add or subtract");
305 Node* n2 = n1->in(3 - inv1_idx);
306 Node* inv1 = n1->in(inv1_idx);
307 Node* inv2 = n2->in(inv2_idx);
308 Node* x = n2->in(3 - inv2_idx);
309
310 bool neg_x = n2->is_Sub() && inv2_idx == 1;
311 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
312 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
313 if (n1->is_Sub() && inv1_idx == 1) {
314 neg_x = !neg_x;
315 neg_inv2 = !neg_inv2;
316 }
317
318 bool is_int = n1->bottom_type()->isa_int() != NULL;
319 Node* inv1_c = phase->get_ctrl(inv1);
320 Node* n_inv1;
321 if (neg_inv1) {
322 Node* zero;
323 if (is_int) {
324 zero = phase->_igvn.intcon(0);
325 n_inv1 = new SubINode(zero, inv1);
326 } else {
327 zero = phase->_igvn.longcon(0L);
328 n_inv1 = new SubLNode(zero, inv1);
329 }
330 phase->set_ctrl(zero, phase->C->root());
331 phase->register_new_node(n_inv1, inv1_c);
332 } else {
333 n_inv1 = inv1;
334 }
335
336 Node* inv;
337 if (is_int) {
338 if (neg_inv2) {
339 inv = new SubINode(n_inv1, inv2);
340 } else {
341 inv = new AddINode(n_inv1, inv2);
342 }
343 phase->register_new_node(inv, phase->get_early_ctrl(inv));
344 if (neg_x) {
345 return new SubINode(inv, x);
346 } else {
347 return new AddINode(x, inv);
348 }
349 } else {
350 if (neg_inv2) {
351 inv = new SubLNode(n_inv1, inv2);
352 } else {
353 inv = new AddLNode(n_inv1, inv2);
354 }
355 phase->register_new_node(inv, phase->get_early_ctrl(inv));
356 if (neg_x) {
357 return new SubLNode(inv, x);
358 } else {
359 return new AddLNode(x, inv);
360 }
361 }
362 }
363
364 //---------------------reassociate-----------------------------
365 // Reassociate invariant binary expressions with add/sub/mul/
366 // and/or/xor operators.
367 // For add/sub expressions: see "reassociate_add_sub"
368 //
369 // For mul/and/or/xor expressions:
370 //
371 // inv1 op (x op inv2) => (inv1 op inv2) op x
372 //
reassociate(Node * n1,PhaseIdealLoop * phase)373 Node* IdealLoopTree::reassociate(Node* n1, PhaseIdealLoop *phase) {
374 if (!is_associative(n1) || n1->outcnt() == 0) return NULL;
375 if (is_invariant(n1)) return NULL;
376 // Don't mess with add of constant (igvn moves them to expression tree root.)
377 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
378
379 int inv1_idx = find_invariant(n1, phase);
380 if (!inv1_idx) return NULL;
381 Node* n2 = n1->in(3 - inv1_idx);
382 if (!is_associative(n2, n1)) return NULL;
383 int inv2_idx = find_invariant(n2, phase);
384 if (!inv2_idx) return NULL;
385
386 if (!phase->may_require_nodes(10, 10)) return NULL;
387
388 Node* result = NULL;
389 switch (n1->Opcode()) {
390 case Op_AddI:
391 case Op_AddL:
392 case Op_SubI:
393 case Op_SubL:
394 result = reassociate_add_sub(n1, inv1_idx, inv2_idx, phase);
395 break;
396 case Op_MulI:
397 case Op_MulL:
398 case Op_AndI:
399 case Op_AndL:
400 case Op_OrI:
401 case Op_OrL:
402 case Op_XorI:
403 case Op_XorL: {
404 Node* inv1 = n1->in(inv1_idx);
405 Node* inv2 = n2->in(inv2_idx);
406 Node* x = n2->in(3 - inv2_idx);
407 Node* inv = n2->clone_with_data_edge(inv1, inv2);
408 phase->register_new_node(inv, phase->get_early_ctrl(inv));
409 result = n1->clone_with_data_edge(x, inv);
410 break;
411 }
412 default:
413 ShouldNotReachHere();
414 }
415
416 assert(result != NULL, "");
417 phase->register_new_node(result, phase->get_ctrl(n1));
418 phase->_igvn.replace_node(n1, result);
419 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
420 _body.yank(n1);
421 return result;
422 }
423
424 //---------------------reassociate_invariants-----------------------------
425 // Reassociate invariant expressions:
reassociate_invariants(PhaseIdealLoop * phase)426 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
427 for (int i = _body.size() - 1; i >= 0; i--) {
428 Node *n = _body.at(i);
429 for (int j = 0; j < 5; j++) {
430 Node* nn = reassociate(n, phase);
431 if (nn == NULL) break;
432 n = nn; // again
433 }
434 }
435 }
436
437 //------------------------------policy_peeling---------------------------------
438 // Return TRUE if the loop should be peeled, otherwise return FALSE. Peeling
439 // is applicable if we can make a loop-invariant test (usually a null-check)
440 // execute before we enter the loop. When TRUE, the estimated node budget is
441 // also requested.
policy_peeling(PhaseIdealLoop * phase)442 bool IdealLoopTree::policy_peeling(PhaseIdealLoop *phase) {
443 uint estimate = estimate_peeling(phase);
444
445 return estimate == 0 ? false : phase->may_require_nodes(estimate);
446 }
447
448 // Perform actual policy and size estimate for the loop peeling transform, and
449 // return the estimated loop size if peeling is applicable, otherwise return
450 // zero. No node budget is allocated.
estimate_peeling(PhaseIdealLoop * phase)451 uint IdealLoopTree::estimate_peeling(PhaseIdealLoop *phase) {
452
453 // If nodes are depleted, some transform has miscalculated its needs.
454 assert(!phase->exceeding_node_budget(), "sanity");
455
456 // Peeling does loop cloning which can result in O(N^2) node construction.
457 if (_body.size() > 255) {
458 return 0; // Suppress too large body size.
459 }
460 // Optimistic estimate that approximates loop body complexity via data and
461 // control flow fan-out (instead of using the more pessimistic: BodySize^2).
462 uint estimate = est_loop_clone_sz(2);
463
464 if (phase->exceeding_node_budget(estimate)) {
465 return 0; // Too large to safely clone.
466 }
467
468 // Check for vectorized loops, any peeling done was already applied.
469 if (_head->is_CountedLoop()) {
470 CountedLoopNode* cl = _head->as_CountedLoop();
471 if (cl->is_unroll_only() || cl->trip_count() == 1) {
472 return 0;
473 }
474 }
475
476 Node* test = tail();
477
478 while (test != _head) { // Scan till run off top of loop
479 if (test->is_If()) { // Test?
480 Node *ctrl = phase->get_ctrl(test->in(1));
481 if (ctrl->is_top()) {
482 return 0; // Found dead test on live IF? No peeling!
483 }
484 // Standard IF only has one input value to check for loop invariance.
485 assert(test->Opcode() == Op_If ||
486 test->Opcode() == Op_CountedLoopEnd ||
487 test->Opcode() == Op_LongCountedLoopEnd ||
488 test->Opcode() == Op_RangeCheck,
489 "Check this code when new subtype is added");
490 // Condition is not a member of this loop?
491 if (!is_member(phase->get_loop(ctrl)) && is_loop_exit(test)) {
492 return estimate; // Found reason to peel!
493 }
494 }
495 // Walk up dominators to loop _head looking for test which is executed on
496 // every path through the loop.
497 test = phase->idom(test);
498 }
499 return 0;
500 }
501
502 //------------------------------peeled_dom_test_elim---------------------------
503 // If we got the effect of peeling, either by actually peeling or by making
504 // a pre-loop which must execute at least once, we can remove all
505 // loop-invariant dominated tests in the main body.
peeled_dom_test_elim(IdealLoopTree * loop,Node_List & old_new)506 void PhaseIdealLoop::peeled_dom_test_elim(IdealLoopTree *loop, Node_List &old_new) {
507 bool progress = true;
508 while (progress) {
509 progress = false; // Reset for next iteration
510 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
511 Node *test = prev->in(0);
512 while (test != loop->_head) { // Scan till run off top of loop
513
514 int p_op = prev->Opcode();
515 if ((p_op == Op_IfFalse || p_op == Op_IfTrue) &&
516 test->is_If() && // Test?
517 !test->in(1)->is_Con() && // And not already obvious?
518 // Condition is not a member of this loop?
519 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
520 // Walk loop body looking for instances of this test
521 for (uint i = 0; i < loop->_body.size(); i++) {
522 Node *n = loop->_body.at(i);
523 if (n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/) {
524 // IfNode was dominated by version in peeled loop body
525 progress = true;
526 dominated_by(old_new[prev->_idx], n);
527 }
528 }
529 }
530 prev = test;
531 test = idom(test);
532 } // End of scan tests in loop
533
534 } // End of while (progress)
535 }
536
537 //------------------------------do_peeling-------------------------------------
538 // Peel the first iteration of the given loop.
539 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
540 // The pre-loop illegally has 2 control users (old & new loops).
541 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
542 // Do this by making the old-loop fall-in edges act as if they came
543 // around the loopback from the prior iteration (follow the old-loop
544 // backedges) and then map to the new peeled iteration. This leaves
545 // the pre-loop with only 1 user (the new peeled iteration), but the
546 // peeled-loop backedge has 2 users.
547 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
548 // extra backedge user.
549 //
550 // orig
551 //
552 // stmt1
553 // |
554 // v
555 // loop predicate
556 // |
557 // v
558 // loop<----+
559 // | |
560 // stmt2 |
561 // | |
562 // v |
563 // if ^
564 // / \ |
565 // / \ |
566 // v v |
567 // false true |
568 // / \ |
569 // / ----+
570 // |
571 // v
572 // exit
573 //
574 //
575 // after clone loop
576 //
577 // stmt1
578 // |
579 // v
580 // loop predicate
581 // / \
582 // clone / \ orig
583 // / \
584 // / \
585 // v v
586 // +---->loop clone loop<----+
587 // | | | |
588 // | stmt2 clone stmt2 |
589 // | | | |
590 // | v v |
591 // ^ if clone If ^
592 // | / \ / \ |
593 // | / \ / \ |
594 // | v v v v |
595 // | true false false true |
596 // | / \ / \ |
597 // +---- \ / ----+
598 // \ /
599 // 1v v2
600 // region
601 // |
602 // v
603 // exit
604 //
605 //
606 // after peel and predicate move
607 //
608 // stmt1
609 // /
610 // /
611 // clone / orig
612 // /
613 // / +----------+
614 // / | |
615 // / loop predicate |
616 // / | |
617 // v v |
618 // TOP-->loop clone loop<----+ |
619 // | | | |
620 // stmt2 clone stmt2 | |
621 // | | | ^
622 // v v | |
623 // if clone If ^ |
624 // / \ / \ | |
625 // / \ / \ | |
626 // v v v v | |
627 // true false false true | |
628 // | \ / \ | |
629 // | \ / ----+ ^
630 // | \ / |
631 // | 1v v2 |
632 // v region |
633 // | | |
634 // | v |
635 // | exit |
636 // | |
637 // +--------------->-----------------+
638 //
639 //
640 // final graph
641 //
642 // stmt1
643 // |
644 // v
645 // stmt2 clone
646 // |
647 // v
648 // if clone
649 // / |
650 // / |
651 // v v
652 // false true
653 // | |
654 // | v
655 // | loop predicate
656 // | |
657 // | v
658 // | loop<----+
659 // | | |
660 // | stmt2 |
661 // | | |
662 // | v |
663 // v if ^
664 // | / \ |
665 // | / \ |
666 // | v v |
667 // | false true |
668 // | | \ |
669 // v v --+
670 // region
671 // |
672 // v
673 // exit
674 //
do_peeling(IdealLoopTree * loop,Node_List & old_new)675 void PhaseIdealLoop::do_peeling(IdealLoopTree *loop, Node_List &old_new) {
676
677 C->set_major_progress();
678 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
679 // 'pre' loop from the main and the 'pre' can no longer have its
680 // iterations adjusted. Therefore, we need to declare this loop as
681 // no longer a 'main' loop; it will need new pre and post loops before
682 // we can do further RCE.
683 #ifndef PRODUCT
684 if (TraceLoopOpts) {
685 tty->print("Peel ");
686 loop->dump_head();
687 }
688 #endif
689 LoopNode* head = loop->_head->as_Loop();
690 bool counted_loop = head->is_CountedLoop();
691 if (counted_loop) {
692 CountedLoopNode *cl = head->as_CountedLoop();
693 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
694 cl->set_trip_count(cl->trip_count() - 1);
695 if (cl->is_main_loop()) {
696 cl->set_normal_loop();
697 #ifndef PRODUCT
698 if (PrintOpto && VerifyLoopOptimizations) {
699 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
700 loop->dump_head();
701 }
702 #endif
703 }
704 }
705 Node* entry = head->in(LoopNode::EntryControl);
706
707 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
708 // The pre-loop illegally has 2 control users (old & new loops).
709 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), ControlAroundStripMined);
710
711 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
712 // Do this by making the old-loop fall-in edges act as if they came
713 // around the loopback from the prior iteration (follow the old-loop
714 // backedges) and then map to the new peeled iteration. This leaves
715 // the pre-loop with only 1 user (the new peeled iteration), but the
716 // peeled-loop backedge has 2 users.
717 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
718 _igvn.hash_delete(head->skip_strip_mined());
719 head->skip_strip_mined()->set_req(LoopNode::EntryControl, new_entry);
720 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
721 Node* old = head->fast_out(j);
722 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
723 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
724 if (!new_exit_value) // Backedge value is ALSO loop invariant?
725 // Then loop body backedge value remains the same.
726 new_exit_value = old->in(LoopNode::LoopBackControl);
727 _igvn.hash_delete(old);
728 old->set_req(LoopNode::EntryControl, new_exit_value);
729 }
730 }
731
732
733 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
734 // extra backedge user.
735 Node* new_head = old_new[head->_idx];
736 _igvn.hash_delete(new_head);
737 new_head->set_req(LoopNode::LoopBackControl, C->top());
738 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
739 Node* use = new_head->fast_out(j2);
740 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
741 _igvn.hash_delete(use);
742 use->set_req(LoopNode::LoopBackControl, C->top());
743 }
744 }
745
746 // Step 4: Correct dom-depth info. Set to loop-head depth.
747
748 int dd = dom_depth(head->skip_strip_mined());
749 set_idom(head->skip_strip_mined(), head->skip_strip_mined()->in(LoopNode::EntryControl), dd);
750 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
751 Node *old = loop->_body.at(j3);
752 Node *nnn = old_new[old->_idx];
753 if (!has_ctrl(nnn)) {
754 set_idom(nnn, idom(nnn), dd-1);
755 }
756 }
757
758 // Now force out all loop-invariant dominating tests. The optimizer
759 // finds some, but we _know_ they are all useless.
760 peeled_dom_test_elim(loop,old_new);
761
762 loop->record_for_igvn();
763 }
764
765 //------------------------------policy_maximally_unroll------------------------
766 // Calculate the exact loop trip-count and return TRUE if loop can be fully,
767 // i.e. maximally, unrolled, otherwise return FALSE. When TRUE, the estimated
768 // node budget is also requested.
policy_maximally_unroll(PhaseIdealLoop * phase) const769 bool IdealLoopTree::policy_maximally_unroll(PhaseIdealLoop* phase) const {
770 CountedLoopNode* cl = _head->as_CountedLoop();
771 assert(cl->is_normal_loop(), "");
772 if (!cl->is_valid_counted_loop(T_INT)) {
773 return false; // Malformed counted loop.
774 }
775 if (!cl->has_exact_trip_count()) {
776 return false; // Trip count is not exact.
777 }
778
779 uint trip_count = cl->trip_count();
780 // Note, max_juint is used to indicate unknown trip count.
781 assert(trip_count > 1, "one iteration loop should be optimized out already");
782 assert(trip_count < max_juint, "exact trip_count should be less than max_juint.");
783
784 // If nodes are depleted, some transform has miscalculated its needs.
785 assert(!phase->exceeding_node_budget(), "sanity");
786
787 // Allow the unrolled body to get larger than the standard loop size limit.
788 uint unroll_limit = (uint)LoopUnrollLimit * 4;
789 assert((intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
790 if (trip_count > unroll_limit || _body.size() > unroll_limit) {
791 return false;
792 }
793
794 uint new_body_size = est_loop_unroll_sz(trip_count);
795
796 if (new_body_size == UINT_MAX) { // Check for bad estimate (overflow).
797 return false;
798 }
799
800 // Fully unroll a loop with few iterations, regardless of other conditions,
801 // since the following (general) loop optimizations will split such loop in
802 // any case (into pre-main-post).
803 if (trip_count <= 3) {
804 return phase->may_require_nodes(new_body_size);
805 }
806
807 // Reject if unrolling will result in too much node construction.
808 if (new_body_size > unroll_limit || phase->exceeding_node_budget(new_body_size)) {
809 return false;
810 }
811
812 // Do not unroll a loop with String intrinsics code.
813 // String intrinsics are large and have loops.
814 for (uint k = 0; k < _body.size(); k++) {
815 Node* n = _body.at(k);
816 switch (n->Opcode()) {
817 case Op_StrComp:
818 case Op_StrEquals:
819 case Op_StrIndexOf:
820 case Op_StrIndexOfChar:
821 case Op_EncodeISOArray:
822 case Op_AryEq:
823 case Op_HasNegatives: {
824 return false;
825 }
826 #if INCLUDE_RTM_OPT
827 case Op_FastLock:
828 case Op_FastUnlock: {
829 // Don't unroll RTM locking code because it is large.
830 if (UseRTMLocking) {
831 return false;
832 }
833 }
834 #endif
835 } // switch
836 }
837
838 return phase->may_require_nodes(new_body_size);
839 }
840
841
842 //------------------------------policy_unroll----------------------------------
843 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll if
844 // the loop is a counted loop and the loop body is small enough. When TRUE,
845 // the estimated node budget is also requested.
policy_unroll(PhaseIdealLoop * phase)846 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) {
847
848 CountedLoopNode *cl = _head->as_CountedLoop();
849 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
850
851 if (!cl->is_valid_counted_loop(T_INT)) {
852 return false; // Malformed counted loop
853 }
854
855 // If nodes are depleted, some transform has miscalculated its needs.
856 assert(!phase->exceeding_node_budget(), "sanity");
857
858 // Protect against over-unrolling.
859 // After split at least one iteration will be executed in pre-loop.
860 if (cl->trip_count() <= (cl->is_normal_loop() ? 2u : 1u)) {
861 return false;
862 }
863 _local_loop_unroll_limit = LoopUnrollLimit;
864 _local_loop_unroll_factor = 4;
865 int future_unroll_cnt = cl->unrolled_count() * 2;
866 if (!cl->is_vectorized_loop()) {
867 if (future_unroll_cnt > LoopMaxUnroll) return false;
868 } else {
869 // obey user constraints on vector mapped loops with additional unrolling applied
870 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1;
871 if ((future_unroll_cnt / unroll_constraint) > LoopMaxUnroll) return false;
872 }
873
874 // Check for initial stride being a small enough constant
875 if (abs(cl->stride_con()) > (1<<2)*future_unroll_cnt) return false;
876
877 // Don't unroll if the next round of unrolling would push us
878 // over the expected trip count of the loop. One is subtracted
879 // from the expected trip count because the pre-loop normally
880 // executes 1 iteration.
881 if (UnrollLimitForProfileCheck > 0 &&
882 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
883 future_unroll_cnt > UnrollLimitForProfileCheck &&
884 (float)future_unroll_cnt > cl->profile_trip_cnt() - 1.0) {
885 return false;
886 }
887
888 // When unroll count is greater than LoopUnrollMin, don't unroll if:
889 // the residual iterations are more than 10% of the trip count
890 // and rounds of "unroll,optimize" are not making significant progress
891 // Progress defined as current size less than 20% larger than previous size.
892 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
893 future_unroll_cnt > LoopUnrollMin &&
894 (future_unroll_cnt - 1) * (100 / LoopPercentProfileLimit) > cl->profile_trip_cnt() &&
895 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
896 return false;
897 }
898
899 Node *init_n = cl->init_trip();
900 Node *limit_n = cl->limit();
901 int stride_con = cl->stride_con();
902 if (limit_n == NULL) return false; // We will dereference it below.
903
904 // Non-constant bounds.
905 // Protect against over-unrolling when init or/and limit are not constant
906 // (so that trip_count's init value is maxint) but iv range is known.
907 if (init_n == NULL || !init_n->is_Con() || !limit_n->is_Con()) {
908 Node* phi = cl->phi();
909 if (phi != NULL) {
910 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
911 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
912 int next_stride = stride_con * 2; // stride after this unroll
913 if (next_stride > 0) {
914 if (iv_type->_lo > max_jint - next_stride || // overflow
915 iv_type->_lo + next_stride > iv_type->_hi) {
916 return false; // over-unrolling
917 }
918 } else if (next_stride < 0) {
919 if (iv_type->_hi < min_jint - next_stride || // overflow
920 iv_type->_hi + next_stride < iv_type->_lo) {
921 return false; // over-unrolling
922 }
923 }
924 }
925 }
926
927 // After unroll limit will be adjusted: new_limit = limit-stride.
928 // Bailout if adjustment overflow.
929 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
930 if ((stride_con > 0 && ((min_jint + stride_con) > limit_type->_hi)) ||
931 (stride_con < 0 && ((max_jint + stride_con) < limit_type->_lo)))
932 return false; // overflow
933
934 // Adjust body_size to determine if we unroll or not
935 uint body_size = _body.size();
936 // Key test to unroll loop in CRC32 java code
937 int xors_in_loop = 0;
938 // Also count ModL, DivL and MulL which expand mightly
939 for (uint k = 0; k < _body.size(); k++) {
940 Node* n = _body.at(k);
941 switch (n->Opcode()) {
942 case Op_XorI: xors_in_loop++; break; // CRC32 java code
943 case Op_ModL: body_size += 30; break;
944 case Op_DivL: body_size += 30; break;
945 case Op_MulL: body_size += 10; break;
946 case Op_StrComp:
947 case Op_StrEquals:
948 case Op_StrIndexOf:
949 case Op_StrIndexOfChar:
950 case Op_EncodeISOArray:
951 case Op_AryEq:
952 case Op_HasNegatives: {
953 // Do not unroll a loop with String intrinsics code.
954 // String intrinsics are large and have loops.
955 return false;
956 }
957 #if INCLUDE_RTM_OPT
958 case Op_FastLock:
959 case Op_FastUnlock: {
960 // Don't unroll RTM locking code because it is large.
961 if (UseRTMLocking) {
962 return false;
963 }
964 }
965 #endif
966 } // switch
967 }
968
969 if (UseSuperWord) {
970 if (!cl->is_reduction_loop()) {
971 phase->mark_reductions(this);
972 }
973
974 // Only attempt slp analysis when user controls do not prohibit it
975 if (LoopMaxUnroll > _local_loop_unroll_factor) {
976 // Once policy_slp_analysis succeeds, mark the loop with the
977 // maximal unroll factor so that we minimize analysis passes
978 if (future_unroll_cnt >= _local_loop_unroll_factor) {
979 policy_unroll_slp_analysis(cl, phase, future_unroll_cnt);
980 }
981 }
982 }
983
984 int slp_max_unroll_factor = cl->slp_max_unroll();
985 if ((LoopMaxUnroll < slp_max_unroll_factor) && FLAG_IS_DEFAULT(LoopMaxUnroll) && UseSubwordForMaxVector) {
986 LoopMaxUnroll = slp_max_unroll_factor;
987 }
988
989 uint estimate = est_loop_clone_sz(2);
990
991 if (cl->has_passed_slp()) {
992 if (slp_max_unroll_factor >= future_unroll_cnt) {
993 return phase->may_require_nodes(estimate);
994 }
995 return false; // Loop too big.
996 }
997
998 // Check for being too big
999 if (body_size > (uint)_local_loop_unroll_limit) {
1000 if ((cl->is_subword_loop() || xors_in_loop >= 4) && body_size < 4u * LoopUnrollLimit) {
1001 return phase->may_require_nodes(estimate);
1002 }
1003 return false; // Loop too big.
1004 }
1005
1006 if (cl->is_unroll_only()) {
1007 if (TraceSuperWordLoopUnrollAnalysis) {
1008 tty->print_cr("policy_unroll passed vector loop(vlen=%d, factor=%d)\n",
1009 slp_max_unroll_factor, future_unroll_cnt);
1010 }
1011 }
1012
1013 // Unroll once! (Each trip will soon do double iterations)
1014 return phase->may_require_nodes(estimate);
1015 }
1016
policy_unroll_slp_analysis(CountedLoopNode * cl,PhaseIdealLoop * phase,int future_unroll_cnt)1017 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_cnt) {
1018
1019 // If nodes are depleted, some transform has miscalculated its needs.
1020 assert(!phase->exceeding_node_budget(), "sanity");
1021
1022 // Enable this functionality target by target as needed
1023 if (SuperWordLoopUnrollAnalysis) {
1024 if (!cl->was_slp_analyzed()) {
1025 SuperWord sw(phase);
1026 sw.transform_loop(this, false);
1027
1028 // If the loop is slp canonical analyze it
1029 if (sw.early_return() == false) {
1030 sw.unrolling_analysis(_local_loop_unroll_factor);
1031 }
1032 }
1033
1034 if (cl->has_passed_slp()) {
1035 int slp_max_unroll_factor = cl->slp_max_unroll();
1036 if (slp_max_unroll_factor >= future_unroll_cnt) {
1037 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor;
1038 if (new_limit > LoopUnrollLimit) {
1039 if (TraceSuperWordLoopUnrollAnalysis) {
1040 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit);
1041 }
1042 _local_loop_unroll_limit = new_limit;
1043 }
1044 }
1045 }
1046 }
1047 }
1048
1049 //------------------------------policy_range_check-----------------------------
1050 // Return TRUE or FALSE if the loop should be range-check-eliminated or not.
1051 // When TRUE, the estimated node budget is also requested.
1052 //
1053 // We will actually perform iteration-splitting, a more powerful form of RCE.
policy_range_check(PhaseIdealLoop * phase) const1054 bool IdealLoopTree::policy_range_check(PhaseIdealLoop *phase) const {
1055 if (!RangeCheckElimination) return false;
1056
1057 // If nodes are depleted, some transform has miscalculated its needs.
1058 assert(!phase->exceeding_node_budget(), "sanity");
1059
1060 CountedLoopNode *cl = _head->as_CountedLoop();
1061 // If we unrolled with no intention of doing RCE and we later changed our
1062 // minds, we got no pre-loop. Either we need to make a new pre-loop, or we
1063 // have to disallow RCE.
1064 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
1065 Node *trip_counter = cl->phi();
1066
1067 // check for vectorized loops, some opts are no longer needed
1068 // RCE needs pre/main/post loops. Don't apply it on a single iteration loop.
1069 if (cl->is_unroll_only() || (cl->is_normal_loop() && cl->trip_count() == 1)) return false;
1070
1071 // Check loop body for tests of trip-counter plus loop-invariant vs
1072 // loop-invariant.
1073 for (uint i = 0; i < _body.size(); i++) {
1074 Node *iff = _body[i];
1075 if (iff->Opcode() == Op_If ||
1076 iff->Opcode() == Op_RangeCheck) { // Test?
1077
1078 // Comparing trip+off vs limit
1079 Node *bol = iff->in(1);
1080 if (bol->req() != 2) {
1081 continue; // dead constant test
1082 }
1083 if (!bol->is_Bool()) {
1084 assert(bol->Opcode() == Op_Conv2B, "predicate check only");
1085 continue;
1086 }
1087 if (bol->as_Bool()->_test._test == BoolTest::ne) {
1088 continue; // not RC
1089 }
1090 Node *cmp = bol->in(1);
1091 Node *rc_exp = cmp->in(1);
1092 Node *limit = cmp->in(2);
1093
1094 Node *limit_c = phase->get_ctrl(limit);
1095 if (limit_c == phase->C->top()) {
1096 return false; // Found dead test on live IF? No RCE!
1097 }
1098 if (is_member(phase->get_loop(limit_c))) {
1099 // Compare might have operands swapped; commute them
1100 rc_exp = cmp->in(2);
1101 limit = cmp->in(1);
1102 limit_c = phase->get_ctrl(limit);
1103 if (is_member(phase->get_loop(limit_c))) {
1104 continue; // Both inputs are loop varying; cannot RCE
1105 }
1106 }
1107
1108 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
1109 continue;
1110 }
1111 // Found a test like 'trip+off vs limit'. Test is an IfNode, has two (2)
1112 // projections. If BOTH are in the loop we need loop unswitching instead
1113 // of iteration splitting.
1114 if (is_loop_exit(iff)) {
1115 // Found valid reason to split iterations (if there is room).
1116 // NOTE: Usually a gross overestimate.
1117 return phase->may_require_nodes(est_loop_clone_sz(2));
1118 }
1119 } // End of is IF
1120 }
1121
1122 return false;
1123 }
1124
1125 //------------------------------policy_peel_only-------------------------------
1126 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
1127 // for unrolling loops with NO array accesses.
policy_peel_only(PhaseIdealLoop * phase) const1128 bool IdealLoopTree::policy_peel_only(PhaseIdealLoop *phase) const {
1129
1130 // If nodes are depleted, some transform has miscalculated its needs.
1131 assert(!phase->exceeding_node_budget(), "sanity");
1132
1133 // check for vectorized loops, any peeling done was already applied
1134 if (_head->is_CountedLoop() && _head->as_CountedLoop()->is_unroll_only()) {
1135 return false;
1136 }
1137
1138 for (uint i = 0; i < _body.size(); i++) {
1139 if (_body[i]->is_Mem()) {
1140 return false;
1141 }
1142 }
1143 // No memory accesses at all!
1144 return true;
1145 }
1146
1147 //------------------------------clone_up_backedge_goo--------------------------
1148 // If Node n lives in the back_ctrl block and cannot float, we clone a private
1149 // version of n in preheader_ctrl block and return that, otherwise return n.
clone_up_backedge_goo(Node * back_ctrl,Node * preheader_ctrl,Node * n,VectorSet & visited,Node_Stack & clones)1150 Node *PhaseIdealLoop::clone_up_backedge_goo(Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones) {
1151 if (get_ctrl(n) != back_ctrl) return n;
1152
1153 // Only visit once
1154 if (visited.test_set(n->_idx)) {
1155 Node *x = clones.find(n->_idx);
1156 return (x != NULL) ? x : n;
1157 }
1158
1159 Node *x = NULL; // If required, a clone of 'n'
1160 // Check for 'n' being pinned in the backedge.
1161 if (n->in(0) && n->in(0) == back_ctrl) {
1162 assert(clones.find(n->_idx) == NULL, "dead loop");
1163 x = n->clone(); // Clone a copy of 'n' to preheader
1164 clones.push(x, n->_idx);
1165 x->set_req(0, preheader_ctrl); // Fix x's control input to preheader
1166 }
1167
1168 // Recursive fixup any other input edges into x.
1169 // If there are no changes we can just return 'n', otherwise
1170 // we need to clone a private copy and change it.
1171 for (uint i = 1; i < n->req(); i++) {
1172 Node *g = clone_up_backedge_goo(back_ctrl, preheader_ctrl, n->in(i), visited, clones);
1173 if (g != n->in(i)) {
1174 if (!x) {
1175 assert(clones.find(n->_idx) == NULL, "dead loop");
1176 x = n->clone();
1177 clones.push(x, n->_idx);
1178 }
1179 x->set_req(i, g);
1180 }
1181 }
1182 if (x) { // x can legally float to pre-header location
1183 register_new_node(x, preheader_ctrl);
1184 return x;
1185 } else { // raise n to cover LCA of uses
1186 set_ctrl(n, find_non_split_ctrl(back_ctrl->in(0)));
1187 }
1188 return n;
1189 }
1190
cast_incr_before_loop(Node * incr,Node * ctrl,Node * loop)1191 Node* PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) {
1192 Node* castii = new CastIINode(incr, TypeInt::INT, true);
1193 castii->set_req(0, ctrl);
1194 register_new_node(castii, ctrl);
1195 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) {
1196 Node* n = incr->fast_out(i);
1197 if (n->is_Phi() && n->in(0) == loop) {
1198 int nrep = n->replace_edge(incr, castii);
1199 return castii;
1200 }
1201 }
1202 return NULL;
1203 }
1204
1205 #ifdef ASSERT
ensure_zero_trip_guard_proj(Node * node,bool is_main_loop)1206 void PhaseIdealLoop::ensure_zero_trip_guard_proj(Node* node, bool is_main_loop) {
1207 assert(node->is_IfProj(), "must be the zero trip guard If node");
1208 Node* zer_bol = node->in(0)->in(1);
1209 assert(zer_bol != NULL && zer_bol->is_Bool(), "must be Bool");
1210 Node* zer_cmp = zer_bol->in(1);
1211 assert(zer_cmp != NULL && zer_cmp->Opcode() == Op_CmpI, "must be CmpI");
1212 // For the main loop, the opaque node is the second input to zer_cmp, for the post loop it's the first input node
1213 Node* zer_opaq = zer_cmp->in(is_main_loop ? 2 : 1);
1214 assert(zer_opaq != NULL && zer_opaq->Opcode() == Op_Opaque1, "must be Opaque1");
1215 }
1216 #endif
1217
1218 // Make a copy of the skeleton range check predicates before the main
1219 // loop and set the initial value of loop as input. After unrolling,
1220 // the range of values for the induction variable in the main loop can
1221 // fall outside the allowed range of values by the array access (main
1222 // loop is never executed). When that happens, range check
1223 // CastII/ConvI2L nodes cause some data paths to die. For consistency,
1224 // the control paths must die too but the range checks were removed by
1225 // predication. The range checks that we add here guarantee that they do.
copy_skeleton_predicates_to_main_loop_helper(Node * predicate,Node * init,Node * stride,IdealLoopTree * outer_loop,LoopNode * outer_main_head,uint dd_main_head,const uint idx_before_pre_post,const uint idx_after_post_before_pre,Node * zero_trip_guard_proj_main,Node * zero_trip_guard_proj_post,const Node_List & old_new)1226 void PhaseIdealLoop::copy_skeleton_predicates_to_main_loop_helper(Node* predicate, Node* init, Node* stride,
1227 IdealLoopTree* outer_loop, LoopNode* outer_main_head,
1228 uint dd_main_head, const uint idx_before_pre_post,
1229 const uint idx_after_post_before_pre, Node* zero_trip_guard_proj_main,
1230 Node* zero_trip_guard_proj_post, const Node_List &old_new) {
1231 if (predicate != NULL) {
1232 #ifdef ASSERT
1233 ensure_zero_trip_guard_proj(zero_trip_guard_proj_main, true);
1234 ensure_zero_trip_guard_proj(zero_trip_guard_proj_post, false);
1235 #endif
1236 IfNode* iff = predicate->in(0)->as_If();
1237 ProjNode* uncommon_proj = iff->proj_out(1 - predicate->as_Proj()->_con);
1238 Node* rgn = uncommon_proj->unique_ctrl_out();
1239 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
1240 assert(iff->in(1)->in(1)->Opcode() == Op_Opaque1, "unexpected predicate shape");
1241 predicate = iff->in(0);
1242 Node* current_proj = outer_main_head->in(LoopNode::EntryControl);
1243 Node* prev_proj = current_proj;
1244 Node* opaque_init = new OpaqueLoopInitNode(C, init);
1245 register_new_node(opaque_init, outer_main_head->in(LoopNode::EntryControl));
1246 Node* opaque_stride = new OpaqueLoopStrideNode(C, stride);
1247 register_new_node(opaque_stride, outer_main_head->in(LoopNode::EntryControl));
1248
1249 while (predicate != NULL && predicate->is_Proj() && predicate->in(0)->is_If()) {
1250 iff = predicate->in(0)->as_If();
1251 uncommon_proj = iff->proj_out(1 - predicate->as_Proj()->_con);
1252 if (uncommon_proj->unique_ctrl_out() != rgn)
1253 break;
1254 if (iff->in(1)->Opcode() == Op_Opaque4) {
1255 assert(skeleton_predicate_has_opaque(iff), "unexpected");
1256 // Clone the skeleton predicate twice and initialize one with the initial
1257 // value of the loop induction variable. Leave the other predicate
1258 // to be initialized when increasing the stride during loop unrolling.
1259 prev_proj = clone_skeleton_predicate_for_main_loop(iff, opaque_init, NULL, predicate, uncommon_proj, current_proj, outer_loop, prev_proj);
1260 assert(skeleton_predicate_has_opaque(prev_proj->in(0)->as_If()), "");
1261
1262 prev_proj = clone_skeleton_predicate_for_main_loop(iff, init, stride, predicate, uncommon_proj, current_proj, outer_loop, prev_proj);
1263 assert(!skeleton_predicate_has_opaque(prev_proj->in(0)->as_If()), "");
1264
1265 // Rewire any control inputs from the cloned skeleton predicates down to the main and post loop for data nodes that are part of the
1266 // main loop (and were cloned to the pre and post loop).
1267 for (DUIterator i = predicate->outs(); predicate->has_out(i); i++) {
1268 Node* loop_node = predicate->out(i);
1269 Node* pre_loop_node = old_new[loop_node->_idx];
1270 // Change the control if 'loop_node' is part of the main loop. If there is an old->new mapping and the index of
1271 // 'pre_loop_node' is greater than idx_before_pre_post, then we know that 'loop_node' was cloned and is part of
1272 // the main loop (and 'pre_loop_node' is part of the pre loop).
1273 if (!loop_node->is_CFG() && (pre_loop_node != NULL && pre_loop_node->_idx > idx_after_post_before_pre)) {
1274 // 'loop_node' is a data node and part of the main loop. Rewire the control to the projection of the zero-trip guard if node
1275 // of the main loop that is immediately preceding the cloned predicates.
1276 _igvn.replace_input_of(loop_node, 0, zero_trip_guard_proj_main);
1277 --i;
1278 } else if (loop_node->_idx > idx_before_pre_post && loop_node->_idx < idx_after_post_before_pre) {
1279 // 'loop_node' is a data node and part of the post loop. Rewire the control to the projection of the zero-trip guard if node
1280 // of the post loop that is immediately preceding the post loop header node (there are no cloned predicates for the post loop).
1281 assert(pre_loop_node == NULL, "a node belonging to the post loop should not have an old_new mapping at this stage");
1282 _igvn.replace_input_of(loop_node, 0, zero_trip_guard_proj_post);
1283 --i;
1284 }
1285 }
1286
1287 // Remove the skeleton predicate from the pre-loop
1288 _igvn.replace_input_of(iff, 1, _igvn.intcon(1));
1289 }
1290 predicate = predicate->in(0)->in(0);
1291 }
1292 _igvn.replace_input_of(outer_main_head, LoopNode::EntryControl, prev_proj);
1293 set_idom(outer_main_head, prev_proj, dd_main_head);
1294 }
1295 }
1296
skeleton_follow_inputs(Node * n,int op)1297 static bool skeleton_follow_inputs(Node* n, int op) {
1298 return (n->is_Bool() ||
1299 n->is_Cmp() ||
1300 op == Op_AndL ||
1301 op == Op_OrL ||
1302 op == Op_RShiftL ||
1303 op == Op_LShiftL ||
1304 op == Op_AddL ||
1305 op == Op_AddI ||
1306 op == Op_MulL ||
1307 op == Op_MulI ||
1308 op == Op_SubL ||
1309 op == Op_SubI ||
1310 op == Op_ConvI2L);
1311 }
1312
skeleton_predicate_has_opaque(IfNode * iff)1313 bool PhaseIdealLoop::skeleton_predicate_has_opaque(IfNode* iff) {
1314 ResourceMark rm;
1315 Unique_Node_List wq;
1316 wq.push(iff->in(1)->in(1));
1317 for (uint i = 0; i < wq.size(); i++) {
1318 Node* n = wq.at(i);
1319 int op = n->Opcode();
1320 if (skeleton_follow_inputs(n, op)) {
1321 for (uint j = 1; j < n->req(); j++) {
1322 Node* m = n->in(j);
1323 if (m != NULL) {
1324 wq.push(m);
1325 }
1326 }
1327 continue;
1328 }
1329 if (n->is_Opaque1()) {
1330 return true;
1331 }
1332 }
1333 return false;
1334 }
1335
1336 // Clone the skeleton predicate bool for a main or unswitched loop:
1337 // Main loop: Set new_init and new_stride nodes as new inputs.
1338 // Unswitched loop: new_init and new_stride are both NULL. Clone OpaqueLoopInit and OpaqueLoopStride instead.
clone_skeleton_predicate_bool(Node * iff,Node * new_init,Node * new_stride,Node * predicate,Node * uncommon_proj,Node * control,IdealLoopTree * outer_loop)1339 Node* PhaseIdealLoop::clone_skeleton_predicate_bool(Node* iff, Node* new_init, Node* new_stride, Node* predicate, Node* uncommon_proj,
1340 Node* control, IdealLoopTree* outer_loop) {
1341 Node_Stack to_clone(2);
1342 to_clone.push(iff->in(1), 1);
1343 uint current = C->unique();
1344 Node* result = NULL;
1345 bool is_unswitched_loop = new_init == NULL && new_stride == NULL;
1346 assert(new_init != NULL || is_unswitched_loop, "new_init must be set when new_stride is non-null");
1347 // Look for the opaque node to replace with the new value
1348 // and clone everything in between. We keep the Opaque4 node
1349 // so the duplicated predicates are eliminated once loop
1350 // opts are over: they are here only to keep the IR graph
1351 // consistent.
1352 do {
1353 Node* n = to_clone.node();
1354 uint i = to_clone.index();
1355 Node* m = n->in(i);
1356 int op = m->Opcode();
1357 if (skeleton_follow_inputs(m, op)) {
1358 to_clone.push(m, 1);
1359 continue;
1360 }
1361 if (m->is_Opaque1()) {
1362 if (n->_idx < current) {
1363 n = n->clone();
1364 register_new_node(n, control);
1365 }
1366 if (op == Op_OpaqueLoopInit) {
1367 if (is_unswitched_loop && m->_idx < current && new_init == NULL) {
1368 new_init = m->clone();
1369 register_new_node(new_init, control);
1370 }
1371 n->set_req(i, new_init);
1372 } else {
1373 assert(op == Op_OpaqueLoopStride, "unexpected opaque node");
1374 if (is_unswitched_loop && m->_idx < current && new_stride == NULL) {
1375 new_stride = m->clone();
1376 register_new_node(new_stride, control);
1377 }
1378 if (new_stride != NULL) {
1379 n->set_req(i, new_stride);
1380 }
1381 }
1382 to_clone.set_node(n);
1383 }
1384 while (true) {
1385 Node* cur = to_clone.node();
1386 uint j = to_clone.index();
1387 if (j+1 < cur->req()) {
1388 to_clone.set_index(j+1);
1389 break;
1390 }
1391 to_clone.pop();
1392 if (to_clone.size() == 0) {
1393 result = cur;
1394 break;
1395 }
1396 Node* next = to_clone.node();
1397 j = to_clone.index();
1398 if (next->in(j) != cur) {
1399 assert(cur->_idx >= current || next->in(j)->Opcode() == Op_Opaque1, "new node or Opaque1 being replaced");
1400 if (next->_idx < current) {
1401 next = next->clone();
1402 register_new_node(next, control);
1403 to_clone.set_node(next);
1404 }
1405 next->set_req(j, cur);
1406 }
1407 }
1408 } while (result == NULL);
1409 assert(result->_idx >= current, "new node expected");
1410 assert(!is_unswitched_loop || new_init != NULL, "new_init must always be found and cloned");
1411 return result;
1412 }
1413
1414 // Clone a skeleton predicate for the main loop. new_init and new_stride are set as new inputs. Since the predicates cannot fail at runtime,
1415 // Halt nodes are inserted instead of uncommon traps.
clone_skeleton_predicate_for_main_loop(Node * iff,Node * new_init,Node * new_stride,Node * predicate,Node * uncommon_proj,Node * control,IdealLoopTree * outer_loop,Node * input_proj)1416 Node* PhaseIdealLoop::clone_skeleton_predicate_for_main_loop(Node* iff, Node* new_init, Node* new_stride, Node* predicate, Node* uncommon_proj,
1417 Node* control, IdealLoopTree* outer_loop, Node* input_proj) {
1418 Node* result = clone_skeleton_predicate_bool(iff, new_init, new_stride, predicate, uncommon_proj, control, outer_loop);
1419 Node* proj = predicate->clone();
1420 Node* other_proj = uncommon_proj->clone();
1421 Node* new_iff = iff->clone();
1422 new_iff->set_req(1, result);
1423 proj->set_req(0, new_iff);
1424 other_proj->set_req(0, new_iff);
1425 Node* frame = new ParmNode(C->start(), TypeFunc::FramePtr);
1426 register_new_node(frame, C->start());
1427 // It's impossible for the predicate to fail at runtime. Use an Halt node.
1428 Node* halt = new HaltNode(other_proj, frame, "duplicated predicate failed which is impossible");
1429 C->root()->add_req(halt);
1430 new_iff->set_req(0, input_proj);
1431
1432 register_control(new_iff, outer_loop->_parent, input_proj);
1433 register_control(proj, outer_loop->_parent, new_iff);
1434 register_control(other_proj, _ltree_root, new_iff);
1435 register_control(halt, _ltree_root, other_proj);
1436 return proj;
1437 }
1438
copy_skeleton_predicates_to_main_loop(CountedLoopNode * pre_head,Node * init,Node * stride,IdealLoopTree * outer_loop,LoopNode * outer_main_head,uint dd_main_head,const uint idx_before_pre_post,const uint idx_after_post_before_pre,Node * zero_trip_guard_proj_main,Node * zero_trip_guard_proj_post,const Node_List & old_new)1439 void PhaseIdealLoop::copy_skeleton_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride,
1440 IdealLoopTree* outer_loop, LoopNode* outer_main_head,
1441 uint dd_main_head, const uint idx_before_pre_post,
1442 const uint idx_after_post_before_pre, Node* zero_trip_guard_proj_main,
1443 Node* zero_trip_guard_proj_post, const Node_List &old_new) {
1444 if (UseLoopPredicate) {
1445 Node* entry = pre_head->in(LoopNode::EntryControl);
1446 Node* predicate = NULL;
1447 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1448 if (predicate != NULL) {
1449 entry = skip_loop_predicates(entry);
1450 }
1451 Node* profile_predicate = NULL;
1452 if (UseProfiledLoopPredicate) {
1453 profile_predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
1454 if (profile_predicate != NULL) {
1455 entry = skip_loop_predicates(entry);
1456 }
1457 }
1458 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1459 copy_skeleton_predicates_to_main_loop_helper(predicate, init, stride, outer_loop, outer_main_head, dd_main_head,
1460 idx_before_pre_post, idx_after_post_before_pre, zero_trip_guard_proj_main,
1461 zero_trip_guard_proj_post, old_new);
1462 copy_skeleton_predicates_to_main_loop_helper(profile_predicate, init, stride, outer_loop, outer_main_head, dd_main_head,
1463 idx_before_pre_post, idx_after_post_before_pre, zero_trip_guard_proj_main,
1464 zero_trip_guard_proj_post, old_new);
1465 }
1466 }
1467
1468 //------------------------------insert_pre_post_loops--------------------------
1469 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
1470 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
1471 // alignment. Useful to unroll loops that do no array accesses.
insert_pre_post_loops(IdealLoopTree * loop,Node_List & old_new,bool peel_only)1472 void PhaseIdealLoop::insert_pre_post_loops(IdealLoopTree *loop, Node_List &old_new, bool peel_only) {
1473
1474 #ifndef PRODUCT
1475 if (TraceLoopOpts) {
1476 if (peel_only)
1477 tty->print("PeelMainPost ");
1478 else
1479 tty->print("PreMainPost ");
1480 loop->dump_head();
1481 }
1482 #endif
1483 C->set_major_progress();
1484
1485 // Find common pieces of the loop being guarded with pre & post loops
1486 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1487 assert(main_head->is_normal_loop(), "");
1488 CountedLoopEndNode *main_end = main_head->loopexit();
1489 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1490
1491 Node *pre_header= main_head->in(LoopNode::EntryControl);
1492 Node *init = main_head->init_trip();
1493 Node *incr = main_end ->incr();
1494 Node *limit = main_end ->limit();
1495 Node *stride = main_end ->stride();
1496 Node *cmp = main_end ->cmp_node();
1497 BoolTest::mask b_test = main_end->test_trip();
1498
1499 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
1500 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
1501 if (bol->outcnt() != 1) {
1502 bol = bol->clone();
1503 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
1504 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol);
1505 }
1506 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
1507 if (cmp->outcnt() != 1) {
1508 cmp = cmp->clone();
1509 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
1510 _igvn.replace_input_of(bol, 1, cmp);
1511 }
1512
1513 // Add the post loop
1514 const uint idx_before_pre_post = Compile::current()->unique();
1515 CountedLoopNode *post_head = NULL;
1516 Node *main_exit = insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1517 const uint idx_after_post_before_pre = Compile::current()->unique();
1518
1519 //------------------------------
1520 // Step B: Create Pre-Loop.
1521
1522 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1523 // loop pre-header illegally has 2 control users (old & new loops).
1524 LoopNode* outer_main_head = main_head;
1525 IdealLoopTree* outer_loop = loop;
1526 if (main_head->is_strip_mined()) {
1527 main_head->verify_strip_mined(1);
1528 outer_main_head = main_head->outer_loop();
1529 outer_loop = loop->_parent;
1530 assert(outer_loop->_head == outer_main_head, "broken loop tree");
1531 }
1532 uint dd_main_head = dom_depth(outer_main_head);
1533 clone_loop(loop, old_new, dd_main_head, ControlAroundStripMined);
1534 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1535 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1536 pre_head->set_pre_loop(main_head);
1537 Node *pre_incr = old_new[incr->_idx];
1538
1539 // Reduce the pre-loop trip count.
1540 pre_end->_prob = PROB_FAIR;
1541
1542 // Find the pre-loop normal exit.
1543 Node* pre_exit = pre_end->proj_out(false);
1544 assert(pre_exit->Opcode() == Op_IfFalse, "");
1545 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end);
1546 _igvn.register_new_node_with_optimizer(new_pre_exit);
1547 set_idom(new_pre_exit, pre_end, dd_main_head);
1548 set_loop(new_pre_exit, outer_loop->_parent);
1549
1550 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1551 // pre-loop, the main-loop may not execute at all. Later in life this
1552 // zero-trip guard will become the minimum-trip guard when we unroll
1553 // the main-loop.
1554 Node *min_opaq = new Opaque1Node(C, limit);
1555 Node *min_cmp = new CmpINode(pre_incr, min_opaq);
1556 Node *min_bol = new BoolNode(min_cmp, b_test);
1557 register_new_node(min_opaq, new_pre_exit);
1558 register_new_node(min_cmp , new_pre_exit);
1559 register_new_node(min_bol , new_pre_exit);
1560
1561 // Build the IfNode (assume the main-loop is executed always).
1562 IfNode *min_iff = new IfNode(new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN);
1563 _igvn.register_new_node_with_optimizer(min_iff);
1564 set_idom(min_iff, new_pre_exit, dd_main_head);
1565 set_loop(min_iff, outer_loop->_parent);
1566
1567 // Plug in the false-path, taken if we need to skip main-loop
1568 _igvn.hash_delete(pre_exit);
1569 pre_exit->set_req(0, min_iff);
1570 set_idom(pre_exit, min_iff, dd_main_head);
1571 set_idom(pre_exit->unique_ctrl_out(), min_iff, dd_main_head);
1572 // Make the true-path, must enter the main loop
1573 Node *min_taken = new IfTrueNode(min_iff);
1574 _igvn.register_new_node_with_optimizer(min_taken);
1575 set_idom(min_taken, min_iff, dd_main_head);
1576 set_loop(min_taken, outer_loop->_parent);
1577 // Plug in the true path
1578 _igvn.hash_delete(outer_main_head);
1579 outer_main_head->set_req(LoopNode::EntryControl, min_taken);
1580 set_idom(outer_main_head, min_taken, dd_main_head);
1581
1582 VectorSet visited;
1583 Node_Stack clones(main_head->back_control()->outcnt());
1584 // Step B3: Make the fall-in values to the main-loop come from the
1585 // fall-out values of the pre-loop.
1586 for (DUIterator i2 = main_head->outs(); main_head->has_out(i2); i2++) {
1587 Node* main_phi = main_head->out(i2);
1588 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0) {
1589 Node* pre_phi = old_new[main_phi->_idx];
1590 Node* fallpre = clone_up_backedge_goo(pre_head->back_control(),
1591 main_head->skip_strip_mined()->in(LoopNode::EntryControl),
1592 pre_phi->in(LoopNode::LoopBackControl),
1593 visited, clones);
1594 _igvn.hash_delete(main_phi);
1595 main_phi->set_req(LoopNode::EntryControl, fallpre);
1596 }
1597 }
1598
1599 // Nodes inside the loop may be control dependent on a predicate
1600 // that was moved before the preloop. If the back branch of the main
1601 // or post loops becomes dead, those nodes won't be dependent on the
1602 // test that guards that loop nest anymore which could lead to an
1603 // incorrect array access because it executes independently of the
1604 // test that was guarding the loop nest. We add a special CastII on
1605 // the if branch that enters the loop, between the input induction
1606 // variable value and the induction variable Phi to preserve correct
1607 // dependencies.
1608
1609 // CastII for the main loop:
1610 Node* castii = cast_incr_before_loop(pre_incr, min_taken, main_head);
1611 assert(castii != NULL, "no castII inserted");
1612 assert(post_head->in(1)->is_IfProj(), "must be zero-trip guard If node projection of the post loop");
1613 copy_skeleton_predicates_to_main_loop(pre_head, castii, stride, outer_loop, outer_main_head, dd_main_head,
1614 idx_before_pre_post, idx_after_post_before_pre, min_taken, post_head->in(1), old_new);
1615
1616 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1617 // RCE and alignment may change this later.
1618 Node *cmp_end = pre_end->cmp_node();
1619 assert(cmp_end->in(2) == limit, "");
1620 Node *pre_limit = new AddINode(init, stride);
1621
1622 // Save the original loop limit in this Opaque1 node for
1623 // use by range check elimination.
1624 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit);
1625
1626 register_new_node(pre_limit, pre_head->in(0));
1627 register_new_node(pre_opaq , pre_head->in(0));
1628
1629 // Since no other users of pre-loop compare, I can hack limit directly
1630 assert(cmp_end->outcnt() == 1, "no other users");
1631 _igvn.hash_delete(cmp_end);
1632 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1633
1634 // Special case for not-equal loop bounds:
1635 // Change pre loop test, main loop test, and the
1636 // main loop guard test to use lt or gt depending on stride
1637 // direction:
1638 // positive stride use <
1639 // negative stride use >
1640 //
1641 // not-equal test is kept for post loop to handle case
1642 // when init > limit when stride > 0 (and reverse).
1643
1644 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1645
1646 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1647 // Modify pre loop end condition
1648 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1649 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test);
1650 register_new_node(new_bol0, pre_head->in(0));
1651 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0);
1652 // Modify main loop guard condition
1653 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1654 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test);
1655 register_new_node(new_bol1, new_pre_exit);
1656 _igvn.hash_delete(min_iff);
1657 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1658 // Modify main loop end condition
1659 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1660 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test);
1661 register_new_node(new_bol2, main_end->in(CountedLoopEndNode::TestControl));
1662 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2);
1663 }
1664
1665 // Flag main loop
1666 main_head->set_main_loop();
1667 if (peel_only) {
1668 main_head->set_main_no_pre_loop();
1669 }
1670
1671 // Subtract a trip count for the pre-loop.
1672 main_head->set_trip_count(main_head->trip_count() - 1);
1673
1674 // It's difficult to be precise about the trip-counts
1675 // for the pre/post loops. They are usually very short,
1676 // so guess that 4 trips is a reasonable value.
1677 post_head->set_profile_trip_cnt(4.0);
1678 pre_head->set_profile_trip_cnt(4.0);
1679
1680 // Now force out all loop-invariant dominating tests. The optimizer
1681 // finds some, but we _know_ they are all useless.
1682 peeled_dom_test_elim(loop,old_new);
1683 loop->record_for_igvn();
1684 }
1685
1686 //------------------------------insert_vector_post_loop------------------------
1687 // Insert a copy of the atomic unrolled vectorized main loop as a post loop,
1688 // unroll_policy has already informed us that more unrolling is about to
1689 // happen to the main loop. The resultant post loop will serve as a
1690 // vectorized drain loop.
insert_vector_post_loop(IdealLoopTree * loop,Node_List & old_new)1691 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1692 if (!loop->_head->is_CountedLoop()) return;
1693
1694 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1695
1696 // only process vectorized main loops
1697 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return;
1698
1699 int slp_max_unroll_factor = cl->slp_max_unroll();
1700 int cur_unroll = cl->unrolled_count();
1701
1702 if (slp_max_unroll_factor == 0) return;
1703
1704 // only process atomic unroll vector loops (not super unrolled after vectorization)
1705 if (cur_unroll != slp_max_unroll_factor) return;
1706
1707 // we only ever process this one time
1708 if (cl->has_atomic_post_loop()) return;
1709
1710 if (!may_require_nodes(loop->est_loop_clone_sz(2))) {
1711 return;
1712 }
1713
1714 #ifndef PRODUCT
1715 if (TraceLoopOpts) {
1716 tty->print("PostVector ");
1717 loop->dump_head();
1718 }
1719 #endif
1720 C->set_major_progress();
1721
1722 // Find common pieces of the loop being guarded with pre & post loops
1723 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1724 CountedLoopEndNode *main_end = main_head->loopexit();
1725 // diagnostic to show loop end is not properly formed
1726 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1727
1728 // mark this loop as processed
1729 main_head->mark_has_atomic_post_loop();
1730
1731 Node *incr = main_end->incr();
1732 Node *limit = main_end->limit();
1733
1734 // In this case we throw away the result as we are not using it to connect anything else.
1735 CountedLoopNode *post_head = NULL;
1736 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1737
1738 // It's difficult to be precise about the trip-counts
1739 // for post loops. They are usually very short,
1740 // so guess that unit vector trips is a reasonable value.
1741 post_head->set_profile_trip_cnt(cur_unroll);
1742
1743 // Now force out all loop-invariant dominating tests. The optimizer
1744 // finds some, but we _know_ they are all useless.
1745 peeled_dom_test_elim(loop, old_new);
1746 loop->record_for_igvn();
1747 }
1748
1749
1750 //-------------------------insert_scalar_rced_post_loop------------------------
1751 // Insert a copy of the rce'd main loop as a post loop,
1752 // We have not unrolled the main loop, so this is the right time to inject this.
1753 // Later we will examine the partner of this post loop pair which still has range checks
1754 // to see inject code which tests at runtime if the range checks are applicable.
insert_scalar_rced_post_loop(IdealLoopTree * loop,Node_List & old_new)1755 void PhaseIdealLoop::insert_scalar_rced_post_loop(IdealLoopTree *loop, Node_List &old_new) {
1756 if (!loop->_head->is_CountedLoop()) return;
1757
1758 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1759
1760 // only process RCE'd main loops
1761 if (!cl->is_main_loop() || cl->range_checks_present()) return;
1762
1763 #ifndef PRODUCT
1764 if (TraceLoopOpts) {
1765 tty->print("PostScalarRce ");
1766 loop->dump_head();
1767 }
1768 #endif
1769 C->set_major_progress();
1770
1771 // Find common pieces of the loop being guarded with pre & post loops
1772 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
1773 CountedLoopEndNode *main_end = main_head->loopexit();
1774 // diagnostic to show loop end is not properly formed
1775 assert(main_end->outcnt() == 2, "1 true, 1 false path only");
1776
1777 Node *incr = main_end->incr();
1778 Node *limit = main_end->limit();
1779
1780 // In this case we throw away the result as we are not using it to connect anything else.
1781 CountedLoopNode *post_head = NULL;
1782 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head);
1783
1784 // It's difficult to be precise about the trip-counts
1785 // for post loops. They are usually very short,
1786 // so guess that unit vector trips is a reasonable value.
1787 post_head->set_profile_trip_cnt(4.0);
1788 post_head->set_is_rce_post_loop();
1789
1790 // Now force out all loop-invariant dominating tests. The optimizer
1791 // finds some, but we _know_ they are all useless.
1792 peeled_dom_test_elim(loop, old_new);
1793 loop->record_for_igvn();
1794 }
1795
1796
1797 //------------------------------insert_post_loop-------------------------------
1798 // Insert post loops. Add a post loop to the given loop passed.
insert_post_loop(IdealLoopTree * loop,Node_List & old_new,CountedLoopNode * main_head,CountedLoopEndNode * main_end,Node * incr,Node * limit,CountedLoopNode * & post_head)1799 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1800 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1801 Node *incr, Node *limit, CountedLoopNode *&post_head) {
1802 IfNode* outer_main_end = main_end;
1803 IdealLoopTree* outer_loop = loop;
1804 if (main_head->is_strip_mined()) {
1805 main_head->verify_strip_mined(1);
1806 outer_main_end = main_head->outer_loop_end();
1807 outer_loop = loop->_parent;
1808 assert(outer_loop->_head == main_head->in(LoopNode::EntryControl), "broken loop tree");
1809 }
1810
1811 //------------------------------
1812 // Step A: Create a new post-Loop.
1813 Node* main_exit = outer_main_end->proj_out(false);
1814 assert(main_exit->Opcode() == Op_IfFalse, "");
1815 int dd_main_exit = dom_depth(main_exit);
1816
1817 // Step A1: Clone the loop body of main. The clone becomes the post-loop.
1818 // The main loop pre-header illegally has 2 control users (old & new loops).
1819 clone_loop(loop, old_new, dd_main_exit, ControlAroundStripMined);
1820 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, "");
1821 post_head = old_new[main_head->_idx]->as_CountedLoop();
1822 post_head->set_normal_loop();
1823 post_head->set_post_loop(main_head);
1824
1825 // Reduce the post-loop trip count.
1826 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd();
1827 post_end->_prob = PROB_FAIR;
1828
1829 // Build the main-loop normal exit.
1830 IfFalseNode *new_main_exit = new IfFalseNode(outer_main_end);
1831 _igvn.register_new_node_with_optimizer(new_main_exit);
1832 set_idom(new_main_exit, outer_main_end, dd_main_exit);
1833 set_loop(new_main_exit, outer_loop->_parent);
1834
1835 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
1836 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
1837 // (the previous loop trip-counter exit value) because we will be changing
1838 // the exit value (via additional unrolling) so we cannot constant-fold away the zero
1839 // trip guard until all unrolling is done.
1840 Node *zer_opaq = new Opaque1Node(C, incr);
1841 Node *zer_cmp = new CmpINode(zer_opaq, limit);
1842 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip());
1843 register_new_node(zer_opaq, new_main_exit);
1844 register_new_node(zer_cmp, new_main_exit);
1845 register_new_node(zer_bol, new_main_exit);
1846
1847 // Build the IfNode
1848 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN);
1849 _igvn.register_new_node_with_optimizer(zer_iff);
1850 set_idom(zer_iff, new_main_exit, dd_main_exit);
1851 set_loop(zer_iff, outer_loop->_parent);
1852
1853 // Plug in the false-path, taken if we need to skip this post-loop
1854 _igvn.replace_input_of(main_exit, 0, zer_iff);
1855 set_idom(main_exit, zer_iff, dd_main_exit);
1856 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
1857 // Make the true-path, must enter this post loop
1858 Node *zer_taken = new IfTrueNode(zer_iff);
1859 _igvn.register_new_node_with_optimizer(zer_taken);
1860 set_idom(zer_taken, zer_iff, dd_main_exit);
1861 set_loop(zer_taken, outer_loop->_parent);
1862 // Plug in the true path
1863 _igvn.hash_delete(post_head);
1864 post_head->set_req(LoopNode::EntryControl, zer_taken);
1865 set_idom(post_head, zer_taken, dd_main_exit);
1866
1867 VectorSet visited;
1868 Node_Stack clones(main_head->back_control()->outcnt());
1869 // Step A3: Make the fall-in values to the post-loop come from the
1870 // fall-out values of the main-loop.
1871 for (DUIterator i = main_head->outs(); main_head->has_out(i); i++) {
1872 Node* main_phi = main_head->out(i);
1873 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0) {
1874 Node* cur_phi = old_new[main_phi->_idx];
1875 Node* fallnew = clone_up_backedge_goo(main_head->back_control(),
1876 post_head->init_control(),
1877 main_phi->in(LoopNode::LoopBackControl),
1878 visited, clones);
1879 _igvn.hash_delete(cur_phi);
1880 cur_phi->set_req(LoopNode::EntryControl, fallnew);
1881 }
1882 }
1883
1884 // CastII for the new post loop:
1885 Node* castii = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head);
1886 assert(castii != NULL, "no castII inserted");
1887
1888 return new_main_exit;
1889 }
1890
1891 //------------------------------is_invariant-----------------------------
1892 // Return true if n is invariant
is_invariant(Node * n) const1893 bool IdealLoopTree::is_invariant(Node* n) const {
1894 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1895 if (n_c->is_top()) return false;
1896 return !is_member(_phase->get_loop(n_c));
1897 }
1898
update_main_loop_skeleton_predicates(Node * ctrl,CountedLoopNode * loop_head,Node * init,int stride_con)1899 void PhaseIdealLoop::update_main_loop_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con) {
1900 // Search for skeleton predicates and update them according to the new stride
1901 Node* entry = ctrl;
1902 Node* prev_proj = ctrl;
1903 LoopNode* outer_loop_head = loop_head->skip_strip_mined();
1904 IdealLoopTree* outer_loop = get_loop(outer_loop_head);
1905
1906 // Compute the value of the loop induction variable at the end of the
1907 // first iteration of the unrolled loop: init + new_stride_con - init_inc
1908 int new_stride_con = stride_con * 2;
1909 Node* max_value = _igvn.intcon(new_stride_con);
1910 set_ctrl(max_value, C->root());
1911
1912 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
1913 IfNode* iff = entry->in(0)->as_If();
1914 ProjNode* proj = iff->proj_out(1 - entry->as_Proj()->_con);
1915 if (proj->unique_ctrl_out()->Opcode() != Op_Halt) {
1916 break;
1917 }
1918 if (iff->in(1)->Opcode() == Op_Opaque4) {
1919 // Look for predicate with an Opaque1 node that can be used as a template
1920 if (!skeleton_predicate_has_opaque(iff)) {
1921 // No Opaque1 node? It's either the check for the first value
1922 // of the first iteration or the check for the last value of
1923 // the first iteration of an unrolled loop. We can't
1924 // tell. Kill it in any case.
1925 _igvn.replace_input_of(iff, 1, iff->in(1)->in(2));
1926 } else {
1927 // Add back predicates updated for the new stride.
1928 prev_proj = clone_skeleton_predicate_for_main_loop(iff, init, max_value, entry, proj, ctrl, outer_loop, prev_proj);
1929 assert(!skeleton_predicate_has_opaque(prev_proj->in(0)->as_If()), "unexpected");
1930 }
1931 }
1932 entry = entry->in(0)->in(0);
1933 }
1934 if (prev_proj != ctrl) {
1935 _igvn.replace_input_of(outer_loop_head, LoopNode::EntryControl, prev_proj);
1936 set_idom(outer_loop_head, prev_proj, dom_depth(outer_loop_head));
1937 }
1938 }
1939
1940 //------------------------------do_unroll--------------------------------------
1941 // Unroll the loop body one step - make each trip do 2 iterations.
do_unroll(IdealLoopTree * loop,Node_List & old_new,bool adjust_min_trip)1942 void PhaseIdealLoop::do_unroll(IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip) {
1943 assert(LoopUnrollLimit, "");
1944 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1945 CountedLoopEndNode *loop_end = loop_head->loopexit();
1946 #ifndef PRODUCT
1947 if (PrintOpto && VerifyLoopOptimizations) {
1948 tty->print("Unrolling ");
1949 loop->dump_head();
1950 } else if (TraceLoopOpts) {
1951 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1952 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1953 } else {
1954 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1955 }
1956 loop->dump_head();
1957 }
1958
1959 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
1960 Node_Stack stack(C->live_nodes() >> 2);
1961 Node_List rpo_list;
1962 VectorSet visited;
1963 visited.set(loop_head->_idx);
1964 rpo(loop_head, stack, visited, rpo_list);
1965 dump(loop, rpo_list.size(), rpo_list);
1966 }
1967 #endif
1968
1969 // Remember loop node count before unrolling to detect
1970 // if rounds of unroll,optimize are making progress
1971 loop_head->set_node_count_before_unroll(loop->_body.size());
1972
1973 Node *ctrl = loop_head->skip_strip_mined()->in(LoopNode::EntryControl);
1974 Node *limit = loop_head->limit();
1975 Node *init = loop_head->init_trip();
1976 Node *stride = loop_head->stride();
1977
1978 Node *opaq = NULL;
1979 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1980 // Search for zero-trip guard.
1981
1982 // Check the shape of the graph at the loop entry. If an inappropriate
1983 // graph shape is encountered, the compiler bails out loop unrolling;
1984 // compilation of the method will still succeed.
1985 if (!is_canonical_loop_entry(loop_head)) {
1986 return;
1987 }
1988 opaq = loop_head->skip_predicates()->in(0)->in(1)->in(1)->in(2);
1989 // Zero-trip test uses an 'opaque' node which is not shared.
1990 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1991 }
1992
1993 C->set_major_progress();
1994
1995 Node* new_limit = NULL;
1996 int stride_con = stride->get_int();
1997 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1998 uint old_trip_count = loop_head->trip_count();
1999 // Verify that unroll policy result is still valid.
2000 assert(old_trip_count > 1 &&
2001 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
2002
2003 update_main_loop_skeleton_predicates(ctrl, loop_head, init, stride_con);
2004
2005 // Adjust loop limit to keep valid iterations number after unroll.
2006 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
2007 // which may overflow.
2008 if (!adjust_min_trip) {
2009 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
2010 "odd trip count for maximally unroll");
2011 // Don't need to adjust limit for maximally unroll since trip count is even.
2012 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
2013 // Loop's limit is constant. Loop's init could be constant when pre-loop
2014 // become peeled iteration.
2015 jlong init_con = init->get_int();
2016 // We can keep old loop limit if iterations count stays the same:
2017 // old_trip_count == new_trip_count * 2
2018 // Note: since old_trip_count >= 2 then new_trip_count >= 1
2019 // so we also don't need to adjust zero trip test.
2020 jlong limit_con = limit->get_int();
2021 // (stride_con*2) not overflow since stride_con <= 8.
2022 int new_stride_con = stride_con * 2;
2023 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
2024 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
2025 // New trip count should satisfy next conditions.
2026 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
2027 uint new_trip_count = (uint)trip_count;
2028 adjust_min_trip = (old_trip_count != new_trip_count*2);
2029 }
2030
2031 if (adjust_min_trip) {
2032 // Step 2: Adjust the trip limit if it is called for.
2033 // The adjustment amount is -stride. Need to make sure if the
2034 // adjustment underflows or overflows, then the main loop is skipped.
2035 Node* cmp = loop_end->cmp_node();
2036 assert(cmp->in(2) == limit, "sanity");
2037 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
2038
2039 // Verify that policy_unroll result is still valid.
2040 const TypeInt* limit_type = _igvn.type(limit)->is_int();
2041 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
2042 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo),
2043 "sanity");
2044
2045 if (limit->is_Con()) {
2046 // The check in policy_unroll and the assert above guarantee
2047 // no underflow if limit is constant.
2048 new_limit = _igvn.intcon(limit->get_int() - stride_con);
2049 set_ctrl(new_limit, C->root());
2050 } else {
2051 // Limit is not constant.
2052 if (loop_head->unrolled_count() == 1) { // only for first unroll
2053 // Separate limit by Opaque node in case it is an incremented
2054 // variable from previous loop to avoid using pre-incremented
2055 // value which could increase register pressure.
2056 // Otherwise reorg_offsets() optimization will create a separate
2057 // Opaque node for each use of trip-counter and as result
2058 // zero trip guard limit will be different from loop limit.
2059 assert(has_ctrl(opaq), "should have it");
2060 Node* opaq_ctrl = get_ctrl(opaq);
2061 limit = new Opaque2Node(C, limit);
2062 register_new_node(limit, opaq_ctrl);
2063 }
2064 if ((stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo)) ||
2065 (stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi))) {
2066 // No underflow.
2067 new_limit = new SubINode(limit, stride);
2068 } else {
2069 // (limit - stride) may underflow.
2070 // Clamp the adjustment value with MININT or MAXINT:
2071 //
2072 // new_limit = limit-stride
2073 // if (stride > 0)
2074 // new_limit = (limit < new_limit) ? MININT : new_limit;
2075 // else
2076 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
2077 //
2078 BoolTest::mask bt = loop_end->test_trip();
2079 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2080 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
2081 set_ctrl(adj_max, C->root());
2082 Node* old_limit = NULL;
2083 Node* adj_limit = NULL;
2084 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
2085 if (loop_head->unrolled_count() > 1 &&
2086 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
2087 limit->in(CMoveNode::IfTrue) == adj_max &&
2088 bol->as_Bool()->_test._test == bt &&
2089 bol->in(1)->Opcode() == Op_CmpI &&
2090 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
2091 // Loop was unrolled before.
2092 // Optimize the limit to avoid nested CMove:
2093 // use original limit as old limit.
2094 old_limit = bol->in(1)->in(1);
2095 // Adjust previous adjusted limit.
2096 adj_limit = limit->in(CMoveNode::IfFalse);
2097 adj_limit = new SubINode(adj_limit, stride);
2098 } else {
2099 old_limit = limit;
2100 adj_limit = new SubINode(limit, stride);
2101 }
2102 assert(old_limit != NULL && adj_limit != NULL, "");
2103 register_new_node(adj_limit, ctrl); // adjust amount
2104 Node* adj_cmp = new CmpINode(old_limit, adj_limit);
2105 register_new_node(adj_cmp, ctrl);
2106 Node* adj_bool = new BoolNode(adj_cmp, bt);
2107 register_new_node(adj_bool, ctrl);
2108 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
2109 }
2110 register_new_node(new_limit, ctrl);
2111 }
2112
2113 assert(new_limit != NULL, "");
2114 // Replace in loop test.
2115 assert(loop_end->in(1)->in(1) == cmp, "sanity");
2116 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
2117 // Don't need to create new test since only one user.
2118 _igvn.hash_delete(cmp);
2119 cmp->set_req(2, new_limit);
2120 } else {
2121 // Create new test since it is shared.
2122 Node* ctrl2 = loop_end->in(0);
2123 Node* cmp2 = cmp->clone();
2124 cmp2->set_req(2, new_limit);
2125 register_new_node(cmp2, ctrl2);
2126 Node* bol2 = loop_end->in(1)->clone();
2127 bol2->set_req(1, cmp2);
2128 register_new_node(bol2, ctrl2);
2129 _igvn.replace_input_of(loop_end, 1, bol2);
2130 }
2131 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
2132 // Make it a 1-trip test (means at least 2 trips).
2133
2134 // Guard test uses an 'opaque' node which is not shared. Hence I
2135 // can edit it's inputs directly. Hammer in the new limit for the
2136 // minimum-trip guard.
2137 assert(opaq->outcnt() == 1, "");
2138 _igvn.replace_input_of(opaq, 1, new_limit);
2139 }
2140
2141 // Adjust max trip count. The trip count is intentionally rounded
2142 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
2143 // the main, unrolled, part of the loop will never execute as it is protected
2144 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
2145 // and later determined that part of the unrolled loop was dead.
2146 loop_head->set_trip_count(old_trip_count / 2);
2147
2148 // Double the count of original iterations in the unrolled loop body.
2149 loop_head->double_unrolled_count();
2150
2151 // ---------
2152 // Step 4: Clone the loop body. Move it inside the loop. This loop body
2153 // represents the odd iterations; since the loop trips an even number of
2154 // times its backedge is never taken. Kill the backedge.
2155 uint dd = dom_depth(loop_head);
2156 clone_loop(loop, old_new, dd, IgnoreStripMined);
2157
2158 // Make backedges of the clone equal to backedges of the original.
2159 // Make the fall-in from the original come from the fall-out of the clone.
2160 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
2161 Node* phi = loop_head->fast_out(j);
2162 if (phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0) {
2163 Node *newphi = old_new[phi->_idx];
2164 _igvn.hash_delete(phi);
2165 _igvn.hash_delete(newphi);
2166
2167 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
2168 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
2169 phi ->set_req(LoopNode::LoopBackControl, C->top());
2170 }
2171 }
2172 Node *clone_head = old_new[loop_head->_idx];
2173 _igvn.hash_delete(clone_head);
2174 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
2175 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
2176 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
2177 loop->_head = clone_head; // New loop header
2178
2179 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
2180 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
2181
2182 // Kill the clone's backedge
2183 Node *newcle = old_new[loop_end->_idx];
2184 _igvn.hash_delete(newcle);
2185 Node *one = _igvn.intcon(1);
2186 set_ctrl(one, C->root());
2187 newcle->set_req(1, one);
2188 // Force clone into same loop body
2189 uint max = loop->_body.size();
2190 for (uint k = 0; k < max; k++) {
2191 Node *old = loop->_body.at(k);
2192 Node *nnn = old_new[old->_idx];
2193 loop->_body.push(nnn);
2194 if (!has_ctrl(old)) {
2195 set_loop(nnn, loop);
2196 }
2197 }
2198
2199 loop->record_for_igvn();
2200 loop_head->clear_strip_mined();
2201
2202 #ifndef PRODUCT
2203 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) {
2204 tty->print("\nnew loop after unroll\n"); loop->dump_head();
2205 for (uint i = 0; i < loop->_body.size(); i++) {
2206 loop->_body.at(i)->dump();
2207 }
2208 if (C->clone_map().is_debug()) {
2209 tty->print("\nCloneMap\n");
2210 Dict* dict = C->clone_map().dict();
2211 DictI i(dict);
2212 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size());
2213 for (int ii = 0; i.test(); ++i, ++ii) {
2214 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key));
2215 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen());
2216 if (ii % 10 == 9) {
2217 tty->print_cr(" ");
2218 }
2219 }
2220 tty->print_cr(" ");
2221 }
2222 }
2223 #endif
2224 }
2225
2226 //------------------------------do_maximally_unroll----------------------------
2227
do_maximally_unroll(IdealLoopTree * loop,Node_List & old_new)2228 void PhaseIdealLoop::do_maximally_unroll(IdealLoopTree *loop, Node_List &old_new) {
2229 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2230 assert(cl->has_exact_trip_count(), "trip count is not exact");
2231 assert(cl->trip_count() > 0, "");
2232 #ifndef PRODUCT
2233 if (TraceLoopOpts) {
2234 tty->print("MaxUnroll %d ", cl->trip_count());
2235 loop->dump_head();
2236 }
2237 #endif
2238
2239 // If loop is tripping an odd number of times, peel odd iteration
2240 if ((cl->trip_count() & 1) == 1) {
2241 do_peeling(loop, old_new);
2242 }
2243
2244 // Now its tripping an even number of times remaining. Double loop body.
2245 // Do not adjust pre-guards; they are not needed and do not exist.
2246 if (cl->trip_count() > 0) {
2247 assert((cl->trip_count() & 1) == 0, "missed peeling");
2248 do_unroll(loop, old_new, false);
2249 }
2250 }
2251
mark_reductions(IdealLoopTree * loop)2252 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) {
2253 if (SuperWordReductions == false) return;
2254
2255 CountedLoopNode* loop_head = loop->_head->as_CountedLoop();
2256 if (loop_head->unrolled_count() > 1) {
2257 return;
2258 }
2259
2260 Node* trip_phi = loop_head->phi();
2261 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) {
2262 Node* phi = loop_head->fast_out(i);
2263 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) {
2264 // For definitions which are loop inclusive and not tripcounts.
2265 Node* def_node = phi->in(LoopNode::LoopBackControl);
2266
2267 if (def_node != NULL) {
2268 Node* n_ctrl = get_ctrl(def_node);
2269 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) {
2270 // Now test it to see if it fits the standard pattern for a reduction operator.
2271 int opc = def_node->Opcode();
2272 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())
2273 || opc == Op_MinD || opc == Op_MinF || opc == Op_MaxD || opc == Op_MaxF) {
2274 if (!def_node->is_reduction()) { // Not marked yet
2275 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it
2276 bool ok = false;
2277 for (unsigned j = 1; j < def_node->req(); j++) {
2278 Node* in = def_node->in(j);
2279 if (in == phi) {
2280 ok = true;
2281 break;
2282 }
2283 }
2284
2285 // do nothing if we did not match the initial criteria
2286 if (ok == false) {
2287 continue;
2288 }
2289
2290 // The result of the reduction must not be used in the loop
2291 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) {
2292 Node* u = def_node->fast_out(i);
2293 if (!loop->is_member(get_loop(ctrl_or_self(u)))) {
2294 continue;
2295 }
2296 if (u == phi) {
2297 continue;
2298 }
2299 ok = false;
2300 }
2301
2302 // iff the uses conform
2303 if (ok) {
2304 def_node->add_flag(Node::Flag_is_reduction);
2305 loop_head->mark_has_reductions();
2306 }
2307 }
2308 }
2309 }
2310 }
2311 }
2312 }
2313 }
2314
2315 //------------------------------adjust_limit-----------------------------------
2316 // Helper function that computes new loop limit as (rc_limit-offset)/scale
adjust_limit(bool is_positive_stride,Node * scale,Node * offset,Node * rc_limit,Node * old_limit,Node * pre_ctrl,bool round)2317 Node* PhaseIdealLoop::adjust_limit(bool is_positive_stride, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round) {
2318 Node* sub = new SubLNode(rc_limit, offset);
2319 register_new_node(sub, pre_ctrl);
2320 Node* limit = new DivLNode(NULL, sub, scale);
2321 register_new_node(limit, pre_ctrl);
2322
2323 // When the absolute value of scale is greater than one, the division
2324 // may round limit down/up, so add/sub one to/from the limit.
2325 if (round) {
2326 limit = new AddLNode(limit, _igvn.longcon(is_positive_stride ? -1 : 1));
2327 register_new_node(limit, pre_ctrl);
2328 }
2329
2330 // Clamp the limit to handle integer under-/overflows.
2331 // When reducing the limit, clamp to [min_jint, old_limit]:
2332 // MIN(old_limit, MAX(limit, min_jint))
2333 // When increasing the limit, clamp to [old_limit, max_jint]:
2334 // MAX(old_limit, MIN(limit, max_jint))
2335 Node* cmp = new CmpLNode(limit, _igvn.longcon(is_positive_stride ? min_jint : max_jint));
2336 register_new_node(cmp, pre_ctrl);
2337 Node* bol = new BoolNode(cmp, is_positive_stride ? BoolTest::lt : BoolTest::gt);
2338 register_new_node(bol, pre_ctrl);
2339 limit = new ConvL2INode(limit);
2340 register_new_node(limit, pre_ctrl);
2341 limit = new CMoveINode(bol, limit, _igvn.intcon(is_positive_stride ? min_jint : max_jint), TypeInt::INT);
2342 register_new_node(limit, pre_ctrl);
2343
2344 limit = is_positive_stride ? (Node*)(new MinINode(old_limit, limit))
2345 : (Node*)(new MaxINode(old_limit, limit));
2346 register_new_node(limit, pre_ctrl);
2347 return limit;
2348 }
2349
2350 //------------------------------add_constraint---------------------------------
2351 // Constrain the main loop iterations so the conditions:
2352 // low_limit <= scale_con*I + offset < upper_limit
2353 // always hold true. That is, either increase the number of iterations in the
2354 // pre-loop or reduce the number of iterations in the main-loop until the condition
2355 // holds true in the main-loop. Stride, scale, offset and limit are all loop
2356 // invariant. Further, stride and scale are constants (offset and limit often are).
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)2357 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) {
2358 assert(_igvn.type(offset)->isa_long() != NULL && _igvn.type(low_limit)->isa_long() != NULL &&
2359 _igvn.type(upper_limit)->isa_long() != NULL, "arguments should be long values");
2360
2361 // For a positive stride, we need to reduce the main-loop limit and
2362 // increase the pre-loop limit. This is reversed for a negative stride.
2363 bool is_positive_stride = (stride_con > 0);
2364
2365 // If the absolute scale value is greater one, division in 'adjust_limit' may require
2366 // rounding. Make sure the ABS method correctly handles min_jint.
2367 // Only do this for the pre-loop, one less iteration of the main loop doesn't hurt.
2368 bool round = ABS(scale_con) > 1;
2369
2370 Node* scale = _igvn.longcon(scale_con);
2371 set_ctrl(scale, C->root());
2372
2373 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
2374 // Positive stride*scale: the affine function is increasing,
2375 // the pre-loop checks for underflow and the post-loop for overflow.
2376
2377 // The overflow limit: scale*I+offset < upper_limit
2378 // For the main-loop limit compute:
2379 // ( if (scale > 0) /* and stride > 0 */
2380 // I < (upper_limit-offset)/scale
2381 // else /* scale < 0 and stride < 0 */
2382 // I > (upper_limit-offset)/scale
2383 // )
2384 *main_limit = adjust_limit(is_positive_stride, scale, offset, upper_limit, *main_limit, pre_ctrl, false);
2385
2386 // The underflow limit: low_limit <= scale*I+offset
2387 // For the pre-loop limit compute:
2388 // NOT(scale*I+offset >= low_limit)
2389 // scale*I+offset < low_limit
2390 // ( if (scale > 0) /* and stride > 0 */
2391 // I < (low_limit-offset)/scale
2392 // else /* scale < 0 and stride < 0 */
2393 // I > (low_limit-offset)/scale
2394 // )
2395 *pre_limit = adjust_limit(!is_positive_stride, scale, offset, low_limit, *pre_limit, pre_ctrl, round);
2396 } else {
2397 // Negative stride*scale: the affine function is decreasing,
2398 // the pre-loop checks for overflow and the post-loop for underflow.
2399
2400 // The overflow limit: scale*I+offset < upper_limit
2401 // For the pre-loop limit compute:
2402 // NOT(scale*I+offset < upper_limit)
2403 // scale*I+offset >= upper_limit
2404 // scale*I+offset+1 > upper_limit
2405 // ( if (scale < 0) /* and stride > 0 */
2406 // I < (upper_limit-(offset+1))/scale
2407 // else /* scale > 0 and stride < 0 */
2408 // I > (upper_limit-(offset+1))/scale
2409 // )
2410 Node* one = _igvn.longcon(1);
2411 set_ctrl(one, C->root());
2412 Node* plus_one = new AddLNode(offset, one);
2413 register_new_node(plus_one, pre_ctrl);
2414 *pre_limit = adjust_limit(!is_positive_stride, scale, plus_one, upper_limit, *pre_limit, pre_ctrl, round);
2415
2416 // The underflow limit: low_limit <= scale*I+offset
2417 // For the main-loop limit compute:
2418 // scale*I+offset+1 > low_limit
2419 // ( if (scale < 0) /* and stride > 0 */
2420 // I < (low_limit-(offset+1))/scale
2421 // else /* scale > 0 and stride < 0 */
2422 // I > (low_limit-(offset+1))/scale
2423 // )
2424 *main_limit = adjust_limit(is_positive_stride, scale, plus_one, low_limit, *main_limit, pre_ctrl, false);
2425 }
2426 }
2427
2428 //------------------------------is_scaled_iv---------------------------------
2429 // Return true if exp is a constant times an induction var
is_scaled_iv(Node * exp,Node * iv,int * p_scale)2430 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
2431 exp = exp->uncast();
2432 if (exp == iv) {
2433 if (p_scale != NULL) {
2434 *p_scale = 1;
2435 }
2436 return true;
2437 }
2438 int opc = exp->Opcode();
2439 if (opc == Op_MulI) {
2440 if (exp->in(1)->uncast() == iv && exp->in(2)->is_Con()) {
2441 if (p_scale != NULL) {
2442 *p_scale = exp->in(2)->get_int();
2443 }
2444 return true;
2445 }
2446 if (exp->in(2)->uncast() == iv && exp->in(1)->is_Con()) {
2447 if (p_scale != NULL) {
2448 *p_scale = exp->in(1)->get_int();
2449 }
2450 return true;
2451 }
2452 } else if (opc == Op_LShiftI) {
2453 if (exp->in(1)->uncast() == iv && exp->in(2)->is_Con()) {
2454 if (p_scale != NULL) {
2455 *p_scale = 1 << exp->in(2)->get_int();
2456 }
2457 return true;
2458 }
2459 }
2460 return false;
2461 }
2462
2463 //-----------------------------is_scaled_iv_plus_offset------------------------------
2464 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
is_scaled_iv_plus_offset(Node * exp,Node * iv,int * p_scale,Node ** p_offset,int depth)2465 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
2466 if (is_scaled_iv(exp, iv, p_scale)) {
2467 if (p_offset != NULL) {
2468 Node *zero = _igvn.intcon(0);
2469 set_ctrl(zero, C->root());
2470 *p_offset = zero;
2471 }
2472 return true;
2473 }
2474 exp = exp->uncast();
2475 int opc = exp->Opcode();
2476 if (opc == Op_AddI) {
2477 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
2478 if (p_offset != NULL) {
2479 *p_offset = exp->in(2);
2480 }
2481 return true;
2482 }
2483 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
2484 if (p_offset != NULL) {
2485 *p_offset = exp->in(1);
2486 }
2487 return true;
2488 }
2489 if (exp->in(2)->is_Con()) {
2490 Node* offset2 = NULL;
2491 if (depth < 2 &&
2492 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
2493 p_offset != NULL ? &offset2 : NULL, depth+1)) {
2494 if (p_offset != NULL) {
2495 Node *ctrl_off2 = get_ctrl(offset2);
2496 Node* offset = new AddINode(offset2, exp->in(2));
2497 register_new_node(offset, ctrl_off2);
2498 *p_offset = offset;
2499 }
2500 return true;
2501 }
2502 }
2503 } else if (opc == Op_SubI) {
2504 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
2505 if (p_offset != NULL) {
2506 Node *zero = _igvn.intcon(0);
2507 set_ctrl(zero, C->root());
2508 Node *ctrl_off = get_ctrl(exp->in(2));
2509 Node* offset = new SubINode(zero, exp->in(2));
2510 register_new_node(offset, ctrl_off);
2511 *p_offset = offset;
2512 }
2513 return true;
2514 }
2515 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
2516 if (p_offset != NULL) {
2517 *p_scale *= -1;
2518 *p_offset = exp->in(1);
2519 }
2520 return true;
2521 }
2522 }
2523 return false;
2524 }
2525
2526 // Same as PhaseIdealLoop::duplicate_predicates() but for range checks
2527 // eliminated by iteration splitting.
add_range_check_predicate(IdealLoopTree * loop,CountedLoopNode * cl,Node * predicate_proj,int scale_con,Node * offset,Node * limit,jint stride_con,Node * value)2528 Node* PhaseIdealLoop::add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
2529 Node* predicate_proj, int scale_con, Node* offset,
2530 Node* limit, jint stride_con, Node* value) {
2531 bool overflow = false;
2532 BoolNode* bol = rc_predicate(loop, predicate_proj, scale_con, offset, value, NULL, stride_con, limit, (stride_con > 0) != (scale_con > 0), overflow);
2533 Node* opaque_bol = new Opaque4Node(C, bol, _igvn.intcon(1));
2534 register_new_node(opaque_bol, predicate_proj);
2535 IfNode* new_iff = NULL;
2536 if (overflow) {
2537 new_iff = new IfNode(predicate_proj, opaque_bol, PROB_MAX, COUNT_UNKNOWN);
2538 } else {
2539 new_iff = new RangeCheckNode(predicate_proj, opaque_bol, PROB_MAX, COUNT_UNKNOWN);
2540 }
2541 register_control(new_iff, loop->_parent, predicate_proj);
2542 Node* iffalse = new IfFalseNode(new_iff);
2543 register_control(iffalse, _ltree_root, new_iff);
2544 ProjNode* iftrue = new IfTrueNode(new_iff);
2545 register_control(iftrue, loop->_parent, new_iff);
2546 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
2547 register_new_node(frame, C->start());
2548 Node* halt = new HaltNode(iffalse, frame, "range check predicate failed which is impossible");
2549 register_control(halt, _ltree_root, iffalse);
2550 C->root()->add_req(halt);
2551 return iftrue;
2552 }
2553
2554 //------------------------------do_range_check---------------------------------
2555 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
do_range_check(IdealLoopTree * loop,Node_List & old_new)2556 int PhaseIdealLoop::do_range_check(IdealLoopTree *loop, Node_List &old_new) {
2557 #ifndef PRODUCT
2558 if (PrintOpto && VerifyLoopOptimizations) {
2559 tty->print("Range Check Elimination ");
2560 loop->dump_head();
2561 } else if (TraceLoopOpts) {
2562 tty->print("RangeCheck ");
2563 loop->dump_head();
2564 }
2565 #endif
2566
2567 assert(RangeCheckElimination, "");
2568 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2569 // If we fail before trying to eliminate range checks, set multiversion state
2570 int closed_range_checks = 1;
2571
2572 // protect against stride not being a constant
2573 if (!cl->stride_is_con()) {
2574 return closed_range_checks;
2575 }
2576 // Find the trip counter; we are iteration splitting based on it
2577 Node *trip_counter = cl->phi();
2578 // Find the main loop limit; we will trim it's iterations
2579 // to not ever trip end tests
2580 Node *main_limit = cl->limit();
2581
2582 // Check graph shape. Cannot optimize a loop if zero-trip
2583 // Opaque1 node is optimized away and then another round
2584 // of loop opts attempted.
2585 if (!is_canonical_loop_entry(cl)) {
2586 return closed_range_checks;
2587 }
2588
2589 // Need to find the main-loop zero-trip guard
2590 Node *ctrl = cl->skip_predicates();
2591 Node *iffm = ctrl->in(0);
2592 Node *opqzm = iffm->in(1)->in(1)->in(2);
2593 assert(opqzm->in(1) == main_limit, "do not understand situation");
2594
2595 // Find the pre-loop limit; we will expand its iterations to
2596 // not ever trip low tests.
2597 Node *p_f = iffm->in(0);
2598 // pre loop may have been optimized out
2599 if (p_f->Opcode() != Op_IfFalse) {
2600 return closed_range_checks;
2601 }
2602 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
2603 assert(pre_end->loopnode()->is_pre_loop(), "");
2604 Node *pre_opaq1 = pre_end->limit();
2605 // Occasionally it's possible for a pre-loop Opaque1 node to be
2606 // optimized away and then another round of loop opts attempted.
2607 // We can not optimize this particular loop in that case.
2608 if (pre_opaq1->Opcode() != Op_Opaque1) {
2609 return closed_range_checks;
2610 }
2611 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
2612 Node *pre_limit = pre_opaq->in(1);
2613
2614 // Where do we put new limit calculations
2615 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
2616
2617 // Ensure the original loop limit is available from the
2618 // pre-loop Opaque1 node.
2619 Node *orig_limit = pre_opaq->original_loop_limit();
2620 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) {
2621 return closed_range_checks;
2622 }
2623 // Must know if its a count-up or count-down loop
2624
2625 int stride_con = cl->stride_con();
2626 Node* zero = _igvn.longcon(0);
2627 Node* one = _igvn.longcon(1);
2628 // Use symmetrical int range [-max_jint,max_jint]
2629 Node* mini = _igvn.longcon(-max_jint);
2630 set_ctrl(zero, C->root());
2631 set_ctrl(one, C->root());
2632 set_ctrl(mini, C->root());
2633
2634 // Count number of range checks and reduce by load range limits, if zero,
2635 // the loop is in canonical form to multiversion.
2636 closed_range_checks = 0;
2637
2638 Node* predicate_proj = cl->skip_strip_mined()->in(LoopNode::EntryControl);
2639 assert(predicate_proj->is_Proj() && predicate_proj->in(0)->is_If(), "if projection only");
2640
2641 // Check loop body for tests of trip-counter plus loop-invariant vs loop-variant.
2642 for (uint i = 0; i < loop->_body.size(); i++) {
2643 Node *iff = loop->_body[i];
2644 if (iff->Opcode() == Op_If ||
2645 iff->Opcode() == Op_RangeCheck) { // Test?
2646 // Test is an IfNode, has 2 projections. If BOTH are in the loop
2647 // we need loop unswitching instead of iteration splitting.
2648 closed_range_checks++;
2649 Node *exit = loop->is_loop_exit(iff);
2650 if (!exit) continue;
2651 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
2652
2653 // Get boolean condition to test
2654 Node *i1 = iff->in(1);
2655 if (!i1->is_Bool()) continue;
2656 BoolNode *bol = i1->as_Bool();
2657 BoolTest b_test = bol->_test;
2658 // Flip sense of test if exit condition is flipped
2659 if (flip) {
2660 b_test = b_test.negate();
2661 }
2662 // Get compare
2663 Node *cmp = bol->in(1);
2664
2665 // Look for trip_counter + offset vs limit
2666 Node *rc_exp = cmp->in(1);
2667 Node *limit = cmp->in(2);
2668 int scale_con= 1; // Assume trip counter not scaled
2669
2670 Node *limit_c = get_ctrl(limit);
2671 if (loop->is_member(get_loop(limit_c))) {
2672 // Compare might have operands swapped; commute them
2673 b_test = b_test.commute();
2674 rc_exp = cmp->in(2);
2675 limit = cmp->in(1);
2676 limit_c = get_ctrl(limit);
2677 if (loop->is_member(get_loop(limit_c))) {
2678 continue; // Both inputs are loop varying; cannot RCE
2679 }
2680 }
2681 // Here we know 'limit' is loop invariant
2682
2683 // 'limit' maybe pinned below the zero trip test (probably from a
2684 // previous round of rce), in which case, it can't be used in the
2685 // zero trip test expression which must occur before the zero test's if.
2686 if (is_dominator(ctrl, limit_c)) {
2687 continue; // Don't rce this check but continue looking for other candidates.
2688 }
2689
2690 // Check for scaled induction variable plus an offset
2691 Node *offset = NULL;
2692
2693 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
2694 continue;
2695 }
2696
2697 Node *offset_c = get_ctrl(offset);
2698 if (loop->is_member(get_loop(offset_c))) {
2699 continue; // Offset is not really loop invariant
2700 }
2701 // Here we know 'offset' is loop invariant.
2702
2703 // As above for the 'limit', the 'offset' maybe pinned below the
2704 // zero trip test.
2705 if (is_dominator(ctrl, offset_c)) {
2706 continue; // Don't rce this check but continue looking for other candidates.
2707 }
2708 #ifdef ASSERT
2709 if (TraceRangeLimitCheck) {
2710 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
2711 bol->dump(2);
2712 }
2713 #endif
2714 // At this point we have the expression as:
2715 // scale_con * trip_counter + offset :: limit
2716 // where scale_con, offset and limit are loop invariant. Trip_counter
2717 // monotonically increases by stride_con, a constant. Both (or either)
2718 // stride_con and scale_con can be negative which will flip about the
2719 // sense of the test.
2720
2721 // Perform the limit computations in jlong to avoid overflow
2722 jlong lscale_con = scale_con;
2723 Node* int_offset = offset;
2724 offset = new ConvI2LNode(offset);
2725 register_new_node(offset, pre_ctrl);
2726 Node* int_limit = limit;
2727 limit = new ConvI2LNode(limit);
2728 register_new_node(limit, pre_ctrl);
2729
2730 // Adjust pre and main loop limits to guard the correct iteration set
2731 if (cmp->Opcode() == Op_CmpU) { // Unsigned compare is really 2 tests
2732 if (b_test._test == BoolTest::lt) { // Range checks always use lt
2733 // The underflow and overflow limits: 0 <= scale*I+offset < limit
2734 add_constraint(stride_con, lscale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit);
2735 Node* init = cl->init_trip();
2736 Node* opaque_init = new OpaqueLoopInitNode(C, init);
2737 register_new_node(opaque_init, predicate_proj);
2738
2739 // predicate on first value of first iteration
2740 predicate_proj = add_range_check_predicate(loop, cl, predicate_proj, scale_con, int_offset, int_limit, stride_con, init);
2741 assert(!skeleton_predicate_has_opaque(predicate_proj->in(0)->as_If()), "unexpected");
2742
2743 // template predicate so it can be updated on next unrolling
2744 predicate_proj = add_range_check_predicate(loop, cl, predicate_proj, scale_con, int_offset, int_limit, stride_con, opaque_init);
2745 assert(skeleton_predicate_has_opaque(predicate_proj->in(0)->as_If()), "unexpected");
2746
2747 Node* opaque_stride = new OpaqueLoopStrideNode(C, cl->stride());
2748 register_new_node(opaque_stride, predicate_proj);
2749 Node* max_value = new SubINode(opaque_stride, cl->stride());
2750 register_new_node(max_value, predicate_proj);
2751 max_value = new AddINode(opaque_init, max_value);
2752 register_new_node(max_value, predicate_proj);
2753 predicate_proj = add_range_check_predicate(loop, cl, predicate_proj, scale_con, int_offset, int_limit, stride_con, max_value);
2754 assert(skeleton_predicate_has_opaque(predicate_proj->in(0)->as_If()), "unexpected");
2755
2756 } else {
2757 if (PrintOpto) {
2758 tty->print_cr("missed RCE opportunity");
2759 }
2760 continue; // In release mode, ignore it
2761 }
2762 } else { // Otherwise work on normal compares
2763 switch(b_test._test) {
2764 case BoolTest::gt:
2765 // Fall into GE case
2766 case BoolTest::ge:
2767 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
2768 lscale_con = -lscale_con;
2769 offset = new SubLNode(zero, offset);
2770 register_new_node(offset, pre_ctrl);
2771 limit = new SubLNode(zero, limit);
2772 register_new_node(limit, pre_ctrl);
2773 // Fall into LE case
2774 case BoolTest::le:
2775 if (b_test._test != BoolTest::gt) {
2776 // Convert X <= Y to X < Y+1
2777 limit = new AddLNode(limit, one);
2778 register_new_node(limit, pre_ctrl);
2779 }
2780 // Fall into LT case
2781 case BoolTest::lt:
2782 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
2783 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
2784 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
2785 add_constraint(stride_con, lscale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit);
2786 break;
2787 default:
2788 if (PrintOpto) {
2789 tty->print_cr("missed RCE opportunity");
2790 }
2791 continue; // Unhandled case
2792 }
2793 }
2794
2795 // Kill the eliminated test
2796 C->set_major_progress();
2797 Node *kill_con = _igvn.intcon(1-flip);
2798 set_ctrl(kill_con, C->root());
2799 _igvn.replace_input_of(iff, 1, kill_con);
2800 // Find surviving projection
2801 assert(iff->is_If(), "");
2802 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
2803 // Find loads off the surviving projection; remove their control edge
2804 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
2805 Node* cd = dp->fast_out(i); // Control-dependent node
2806 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
2807 // Allow the load to float around in the loop, or before it
2808 // but NOT before the pre-loop.
2809 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
2810 --i;
2811 --imax;
2812 }
2813 }
2814 if (int_limit->Opcode() == Op_LoadRange) {
2815 closed_range_checks--;
2816 }
2817 } // End of is IF
2818 }
2819 if (predicate_proj != cl->skip_strip_mined()->in(LoopNode::EntryControl)) {
2820 _igvn.replace_input_of(cl->skip_strip_mined(), LoopNode::EntryControl, predicate_proj);
2821 set_idom(cl->skip_strip_mined(), predicate_proj, dom_depth(cl->skip_strip_mined()));
2822 }
2823
2824 // Update loop limits
2825 if (pre_limit != orig_limit) {
2826 // Computed pre-loop limit can be outside of loop iterations range.
2827 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit)
2828 : (Node*)new MaxINode(pre_limit, orig_limit);
2829 register_new_node(pre_limit, pre_ctrl);
2830 }
2831 _igvn.replace_input_of(pre_opaq, 1, pre_limit);
2832
2833 // Note:: we are making the main loop limit no longer precise;
2834 // need to round up based on stride.
2835 cl->set_nonexact_trip_count();
2836 Node *main_cle = cl->loopexit();
2837 Node *main_bol = main_cle->in(1);
2838 // Hacking loop bounds; need private copies of exit test
2839 if (main_bol->outcnt() > 1) { // BoolNode shared?
2840 main_bol = main_bol->clone(); // Clone a private BoolNode
2841 register_new_node(main_bol, main_cle->in(0));
2842 _igvn.replace_input_of(main_cle, 1, main_bol);
2843 }
2844 Node *main_cmp = main_bol->in(1);
2845 if (main_cmp->outcnt() > 1) { // CmpNode shared?
2846 main_cmp = main_cmp->clone(); // Clone a private CmpNode
2847 register_new_node(main_cmp, main_cle->in(0));
2848 _igvn.replace_input_of(main_bol, 1, main_cmp);
2849 }
2850 // Hack the now-private loop bounds
2851 _igvn.replace_input_of(main_cmp, 2, main_limit);
2852 // The OpaqueNode is unshared by design
2853 assert(opqzm->outcnt() == 1, "cannot hack shared node");
2854 _igvn.replace_input_of(opqzm, 1, main_limit);
2855
2856 return closed_range_checks;
2857 }
2858
2859 //------------------------------has_range_checks-------------------------------
2860 // Check to see if RCE cleaned the current loop of range-checks.
has_range_checks(IdealLoopTree * loop)2861 void PhaseIdealLoop::has_range_checks(IdealLoopTree *loop) {
2862 assert(RangeCheckElimination, "");
2863
2864 // skip if not a counted loop
2865 if (!loop->is_counted()) return;
2866
2867 CountedLoopNode *cl = loop->_head->as_CountedLoop();
2868
2869 // skip this loop if it is already checked
2870 if (cl->has_been_range_checked()) return;
2871
2872 // Now check for existence of range checks
2873 for (uint i = 0; i < loop->_body.size(); i++) {
2874 Node *iff = loop->_body[i];
2875 int iff_opc = iff->Opcode();
2876 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2877 cl->mark_has_range_checks();
2878 break;
2879 }
2880 }
2881 cl->set_has_been_range_checked();
2882 }
2883
2884 //-------------------------multi_version_post_loops----------------------------
2885 // Check the range checks that remain, if simple, use the bounds to guard
2886 // which version to a post loop we execute, one with range checks or one without
multi_version_post_loops(IdealLoopTree * rce_loop,IdealLoopTree * legacy_loop)2887 bool PhaseIdealLoop::multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop) {
2888 bool multi_version_succeeded = false;
2889 assert(RangeCheckElimination, "");
2890 CountedLoopNode *legacy_cl = legacy_loop->_head->as_CountedLoop();
2891 assert(legacy_cl->is_post_loop(), "");
2892
2893 // Check for existence of range checks using the unique instance to make a guard with
2894 Unique_Node_List worklist;
2895 for (uint i = 0; i < legacy_loop->_body.size(); i++) {
2896 Node *iff = legacy_loop->_body[i];
2897 int iff_opc = iff->Opcode();
2898 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) {
2899 worklist.push(iff);
2900 }
2901 }
2902
2903 // Find RCE'd post loop so that we can stage its guard.
2904 if (!is_canonical_loop_entry(legacy_cl)) return multi_version_succeeded;
2905 Node* ctrl = legacy_cl->in(LoopNode::EntryControl);
2906 Node* iffm = ctrl->in(0);
2907
2908 // Now we test that both the post loops are connected
2909 Node* post_loop_region = iffm->in(0);
2910 if (post_loop_region == NULL) return multi_version_succeeded;
2911 if (!post_loop_region->is_Region()) return multi_version_succeeded;
2912 Node* covering_region = post_loop_region->in(RegionNode::Control+1);
2913 if (covering_region == NULL) return multi_version_succeeded;
2914 if (!covering_region->is_Region()) return multi_version_succeeded;
2915 Node* p_f = covering_region->in(RegionNode::Control);
2916 if (p_f == NULL) return multi_version_succeeded;
2917 if (!p_f->is_IfFalse()) return multi_version_succeeded;
2918 if (!p_f->in(0)->is_CountedLoopEnd()) return multi_version_succeeded;
2919 CountedLoopEndNode* rce_loop_end = p_f->in(0)->as_CountedLoopEnd();
2920 if (rce_loop_end == NULL) return multi_version_succeeded;
2921 CountedLoopNode* rce_cl = rce_loop_end->loopnode();
2922 if (rce_cl == NULL || !rce_cl->is_post_loop()) return multi_version_succeeded;
2923 CountedLoopNode *known_rce_cl = rce_loop->_head->as_CountedLoop();
2924 if (rce_cl != known_rce_cl) return multi_version_succeeded;
2925
2926 // Then we fetch the cover entry test
2927 ctrl = rce_cl->in(LoopNode::EntryControl);
2928 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return multi_version_succeeded;
2929
2930 #ifndef PRODUCT
2931 if (TraceLoopOpts) {
2932 tty->print("PostMultiVersion\n");
2933 rce_loop->dump_head();
2934 legacy_loop->dump_head();
2935 }
2936 #endif
2937
2938 // Now fetch the limit we want to compare against
2939 Node *limit = rce_cl->limit();
2940 bool first_time = true;
2941
2942 // If we got this far, we identified the post loop which has been RCE'd and
2943 // we have a work list. Now we will try to transform the if guard to cause
2944 // the loop pair to be multi version executed with the determination left to runtime
2945 // or the optimizer if full information is known about the given arrays at compile time.
2946 Node *last_min = NULL;
2947 multi_version_succeeded = true;
2948 while (worklist.size()) {
2949 Node* rc_iffm = worklist.pop();
2950 if (rc_iffm->is_If()) {
2951 Node *rc_bolzm = rc_iffm->in(1);
2952 if (rc_bolzm->is_Bool()) {
2953 Node *rc_cmpzm = rc_bolzm->in(1);
2954 if (rc_cmpzm->is_Cmp()) {
2955 Node *rc_left = rc_cmpzm->in(2);
2956 if (rc_left->Opcode() != Op_LoadRange) {
2957 multi_version_succeeded = false;
2958 break;
2959 }
2960 if (first_time) {
2961 last_min = rc_left;
2962 first_time = false;
2963 } else {
2964 Node *cur_min = new MinINode(last_min, rc_left);
2965 last_min = cur_min;
2966 _igvn.register_new_node_with_optimizer(last_min);
2967 }
2968 }
2969 }
2970 }
2971 }
2972
2973 // All we have to do is update the limit of the rce loop
2974 // with the min of our expression and the current limit.
2975 // We will use this expression to replace the current limit.
2976 if (last_min && multi_version_succeeded) {
2977 Node *cur_min = new MinINode(last_min, limit);
2978 _igvn.register_new_node_with_optimizer(cur_min);
2979 Node *cmp_node = rce_loop_end->cmp_node();
2980 _igvn.replace_input_of(cmp_node, 2, cur_min);
2981 set_ctrl(cur_min, ctrl);
2982 set_loop(cur_min, rce_loop->_parent);
2983
2984 legacy_cl->mark_is_multiversioned();
2985 rce_cl->mark_is_multiversioned();
2986 multi_version_succeeded = true;
2987
2988 C->set_major_progress();
2989 }
2990
2991 return multi_version_succeeded;
2992 }
2993
2994 //-------------------------poison_rce_post_loop--------------------------------
2995 // Causes the rce'd post loop to be optimized away if multiversioning fails
poison_rce_post_loop(IdealLoopTree * rce_loop)2996 void PhaseIdealLoop::poison_rce_post_loop(IdealLoopTree *rce_loop) {
2997 CountedLoopNode *rce_cl = rce_loop->_head->as_CountedLoop();
2998 Node* ctrl = rce_cl->in(LoopNode::EntryControl);
2999 if (ctrl->is_IfTrue() || ctrl->is_IfFalse()) {
3000 Node* iffm = ctrl->in(0);
3001 if (iffm->is_If()) {
3002 Node* cur_bool = iffm->in(1);
3003 if (cur_bool->is_Bool()) {
3004 Node* cur_cmp = cur_bool->in(1);
3005 if (cur_cmp->is_Cmp()) {
3006 BoolTest::mask new_test = BoolTest::gt;
3007 BoolNode *new_bool = new BoolNode(cur_cmp, new_test);
3008 _igvn.replace_node(cur_bool, new_bool);
3009 _igvn._worklist.push(new_bool);
3010 Node* left_op = cur_cmp->in(1);
3011 _igvn.replace_input_of(cur_cmp, 2, left_op);
3012 C->set_major_progress();
3013 }
3014 }
3015 }
3016 }
3017 }
3018
3019 //------------------------------DCE_loop_body----------------------------------
3020 // Remove simplistic dead code from loop body
DCE_loop_body()3021 void IdealLoopTree::DCE_loop_body() {
3022 for (uint i = 0; i < _body.size(); i++) {
3023 if (_body.at(i)->outcnt() == 0) {
3024 _body.map(i, _body.pop());
3025 i--; // Ensure we revisit the updated index.
3026 }
3027 }
3028 }
3029
3030
3031 //------------------------------adjust_loop_exit_prob--------------------------
3032 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
3033 // Replace with a 1-in-10 exit guess.
adjust_loop_exit_prob(PhaseIdealLoop * phase)3034 void IdealLoopTree::adjust_loop_exit_prob(PhaseIdealLoop *phase) {
3035 Node *test = tail();
3036 while (test != _head) {
3037 uint top = test->Opcode();
3038 if (top == Op_IfTrue || top == Op_IfFalse) {
3039 int test_con = ((ProjNode*)test)->_con;
3040 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
3041 IfNode *iff = test->in(0)->as_If();
3042 if (iff->outcnt() == 2) { // Ignore dead tests
3043 Node *bol = iff->in(1);
3044 if (bol && bol->req() > 1 && bol->in(1) &&
3045 ((bol->in(1)->Opcode() == Op_StorePConditional) ||
3046 (bol->in(1)->Opcode() == Op_StoreIConditional) ||
3047 (bol->in(1)->Opcode() == Op_StoreLConditional) ||
3048 (bol->in(1)->Opcode() == Op_CompareAndExchangeB) ||
3049 (bol->in(1)->Opcode() == Op_CompareAndExchangeS) ||
3050 (bol->in(1)->Opcode() == Op_CompareAndExchangeI) ||
3051 (bol->in(1)->Opcode() == Op_CompareAndExchangeL) ||
3052 (bol->in(1)->Opcode() == Op_CompareAndExchangeP) ||
3053 (bol->in(1)->Opcode() == Op_CompareAndExchangeN) ||
3054 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapB) ||
3055 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapS) ||
3056 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI) ||
3057 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL) ||
3058 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP) ||
3059 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN) ||
3060 (bol->in(1)->Opcode() == Op_CompareAndSwapB) ||
3061 (bol->in(1)->Opcode() == Op_CompareAndSwapS) ||
3062 (bol->in(1)->Opcode() == Op_CompareAndSwapI) ||
3063 (bol->in(1)->Opcode() == Op_CompareAndSwapL) ||
3064 (bol->in(1)->Opcode() == Op_CompareAndSwapP) ||
3065 (bol->in(1)->Opcode() == Op_CompareAndSwapN) ||
3066 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndExchangeP) ||
3067 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndExchangeN) ||
3068 (bol->in(1)->Opcode() == Op_ShenandoahWeakCompareAndSwapP) ||
3069 (bol->in(1)->Opcode() == Op_ShenandoahWeakCompareAndSwapN) ||
3070 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndSwapP) ||
3071 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndSwapN)))
3072 return; // Allocation loops RARELY take backedge
3073 // Find the OTHER exit path from the IF
3074 Node* ex = iff->proj_out(1-test_con);
3075 float p = iff->_prob;
3076 if (!phase->is_member(this, ex) && iff->_fcnt == COUNT_UNKNOWN) {
3077 if (top == Op_IfTrue) {
3078 if (p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
3079 iff->_prob = PROB_STATIC_FREQUENT;
3080 }
3081 } else {
3082 if (p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
3083 iff->_prob = PROB_STATIC_INFREQUENT;
3084 }
3085 }
3086 }
3087 }
3088 }
3089 test = phase->idom(test);
3090 }
3091 }
3092
3093 #ifdef ASSERT
locate_pre_from_main(CountedLoopNode * main_loop)3094 static CountedLoopNode* locate_pre_from_main(CountedLoopNode* main_loop) {
3095 assert(!main_loop->is_main_no_pre_loop(), "Does not have a pre loop");
3096 Node* ctrl = main_loop->skip_predicates();
3097 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
3098 Node* iffm = ctrl->in(0);
3099 assert(iffm->Opcode() == Op_If, "");
3100 Node* p_f = iffm->in(0);
3101 assert(p_f->Opcode() == Op_IfFalse, "");
3102 CountedLoopNode* pre_loop = p_f->in(0)->as_CountedLoopEnd()->loopnode();
3103 assert(pre_loop->is_pre_loop(), "No pre loop found");
3104 return pre_loop;
3105 }
3106 #endif
3107
3108 // Remove the main and post loops and make the pre loop execute all
3109 // iterations. Useful when the pre loop is found empty.
remove_main_post_loops(CountedLoopNode * cl,PhaseIdealLoop * phase)3110 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) {
3111 CountedLoopEndNode* pre_end = cl->loopexit();
3112 Node* pre_cmp = pre_end->cmp_node();
3113 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) {
3114 // Only safe to remove the main loop if the compiler optimized it
3115 // out based on an unknown number of iterations
3116 return;
3117 }
3118
3119 // Can we find the main loop?
3120 if (_next == NULL) {
3121 return;
3122 }
3123
3124 Node* next_head = _next->_head;
3125 if (!next_head->is_CountedLoop()) {
3126 return;
3127 }
3128
3129 CountedLoopNode* main_head = next_head->as_CountedLoop();
3130 if (!main_head->is_main_loop() || main_head->is_main_no_pre_loop()) {
3131 return;
3132 }
3133
3134 assert(locate_pre_from_main(main_head) == cl, "bad main loop");
3135 Node* main_iff = main_head->skip_predicates()->in(0);
3136
3137 // Remove the Opaque1Node of the pre loop and make it execute all iterations
3138 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2));
3139 // Remove the Opaque1Node of the main loop so it can be optimized out
3140 Node* main_cmp = main_iff->in(1)->in(1);
3141 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?");
3142 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1));
3143 }
3144
3145 //------------------------------do_remove_empty_loop---------------------------
3146 // We always attempt remove empty loops. The approach is to replace the trip
3147 // counter with the value it will have on the last iteration. This will break
3148 // the loop.
do_remove_empty_loop(PhaseIdealLoop * phase)3149 bool IdealLoopTree::do_remove_empty_loop(PhaseIdealLoop *phase) {
3150 // Minimum size must be empty loop
3151 if (_body.size() > EMPTY_LOOP_SIZE) {
3152 return false;
3153 }
3154 if (!_head->is_CountedLoop()) {
3155 return false; // Dead loop
3156 }
3157 CountedLoopNode *cl = _head->as_CountedLoop();
3158 if (!cl->is_valid_counted_loop(T_INT)) {
3159 return false; // Malformed loop
3160 }
3161 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) {
3162 return false; // Infinite loop
3163 }
3164 if (cl->is_pre_loop()) {
3165 // If the loop we are removing is a pre-loop then the main and post loop
3166 // can be removed as well.
3167 remove_main_post_loops(cl, phase);
3168 }
3169
3170 #ifdef ASSERT
3171 // Ensure only one phi which is the iv.
3172 Node* iv = NULL;
3173 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
3174 Node* n = cl->fast_out(i);
3175 if (n->Opcode() == Op_Phi) {
3176 assert(iv == NULL, "Too many phis");
3177 iv = n;
3178 }
3179 }
3180 assert(iv == cl->phi(), "Wrong phi");
3181 #endif
3182
3183 // main and post loops have explicitly created zero trip guard
3184 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
3185 if (needs_guard) {
3186 // Skip guard if values not overlap.
3187 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
3188 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
3189 int stride_con = cl->stride_con();
3190 if (stride_con > 0) {
3191 needs_guard = (init_t->_hi >= limit_t->_lo);
3192 } else {
3193 needs_guard = (init_t->_lo <= limit_t->_hi);
3194 }
3195 }
3196 if (needs_guard) {
3197 // Check for an obvious zero trip guard.
3198 Node* inctrl = PhaseIdealLoop::skip_all_loop_predicates(cl->skip_predicates());
3199 if (inctrl->Opcode() == Op_IfTrue || inctrl->Opcode() == Op_IfFalse) {
3200 bool maybe_swapped = (inctrl->Opcode() == Op_IfFalse);
3201 // The test should look like just the backedge of a CountedLoop
3202 Node* iff = inctrl->in(0);
3203 if (iff->is_If()) {
3204 Node* bol = iff->in(1);
3205 if (bol->is_Bool()) {
3206 BoolTest test = bol->as_Bool()->_test;
3207 if (maybe_swapped) {
3208 test._test = test.commute();
3209 test._test = test.negate();
3210 }
3211 if (test._test == cl->loopexit()->test_trip()) {
3212 Node* cmp = bol->in(1);
3213 int init_idx = maybe_swapped ? 2 : 1;
3214 int limit_idx = maybe_swapped ? 1 : 2;
3215 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) {
3216 needs_guard = false;
3217 }
3218 }
3219 }
3220 }
3221 }
3222 }
3223
3224 #ifndef PRODUCT
3225 if (PrintOpto) {
3226 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
3227 this->dump_head();
3228 } else if (TraceLoopOpts) {
3229 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
3230 this->dump_head();
3231 }
3232 #endif
3233
3234 if (needs_guard) {
3235 // Peel the loop to ensure there's a zero trip guard
3236 Node_List old_new;
3237 phase->do_peeling(this, old_new);
3238 }
3239
3240 // Replace the phi at loop head with the final value of the last
3241 // iteration. Then the CountedLoopEnd will collapse (backedge never
3242 // taken) and all loop-invariant uses of the exit values will be correct.
3243 Node *phi = cl->phi();
3244 Node *exact_limit = phase->exact_limit(this);
3245 if (exact_limit != cl->limit()) {
3246 // We also need to replace the original limit to collapse loop exit.
3247 Node* cmp = cl->loopexit()->cmp_node();
3248 assert(cl->limit() == cmp->in(2), "sanity");
3249 // Duplicate cmp node if it has other users
3250 if (cmp->outcnt() > 1) {
3251 cmp = cmp->clone();
3252 cmp = phase->_igvn.register_new_node_with_optimizer(cmp);
3253 BoolNode *bol = cl->loopexit()->in(CountedLoopEndNode::TestValue)->as_Bool();
3254 phase->_igvn.replace_input_of(bol, 1, cmp); // put bol on worklist
3255 }
3256 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
3257 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
3258 }
3259 // Note: the final value after increment should not overflow since
3260 // counted loop has limit check predicate.
3261 Node *final = new SubINode(exact_limit, cl->stride());
3262 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
3263 phase->_igvn.replace_node(phi,final);
3264 phase->C->set_major_progress();
3265 return true;
3266 }
3267
3268 //------------------------------do_one_iteration_loop--------------------------
3269 // Convert one iteration loop into normal code.
do_one_iteration_loop(PhaseIdealLoop * phase)3270 bool IdealLoopTree::do_one_iteration_loop(PhaseIdealLoop *phase) {
3271 if (!_head->as_Loop()->is_valid_counted_loop(T_INT)) {
3272 return false; // Only for counted loop
3273 }
3274 CountedLoopNode *cl = _head->as_CountedLoop();
3275 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
3276 return false;
3277 }
3278
3279 #ifndef PRODUCT
3280 if (TraceLoopOpts) {
3281 tty->print("OneIteration ");
3282 this->dump_head();
3283 }
3284 #endif
3285
3286 Node *init_n = cl->init_trip();
3287 // Loop boundaries should be constant since trip count is exact.
3288 assert((cl->stride_con() > 0 && init_n->get_int() + cl->stride_con() >= cl->limit()->get_int()) ||
3289 (cl->stride_con() < 0 && init_n->get_int() + cl->stride_con() <= cl->limit()->get_int()), "should be one iteration");
3290 // Replace the phi at loop head with the value of the init_trip.
3291 // Then the CountedLoopEnd will collapse (backedge will not be taken)
3292 // and all loop-invariant uses of the exit values will be correct.
3293 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
3294 phase->C->set_major_progress();
3295 return true;
3296 }
3297
3298 //=============================================================================
3299 //------------------------------iteration_split_impl---------------------------
iteration_split_impl(PhaseIdealLoop * phase,Node_List & old_new)3300 bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_new) {
3301 // Compute loop trip count if possible.
3302 compute_trip_count(phase);
3303
3304 // Convert one iteration loop into normal code.
3305 if (do_one_iteration_loop(phase)) {
3306 return true;
3307 }
3308 // Check and remove empty loops (spam micro-benchmarks)
3309 if (do_remove_empty_loop(phase)) {
3310 return true; // Here we removed an empty loop
3311 }
3312
3313 AutoNodeBudget node_budget(phase);
3314
3315 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
3316 // This removes loop-invariant tests (usually null checks).
3317 if (!_head->is_CountedLoop()) { // Non-counted loop
3318 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
3319 // Partial peel succeeded so terminate this round of loop opts
3320 return false;
3321 }
3322 if (policy_peeling(phase)) { // Should we peel?
3323 if (PrintOpto) { tty->print_cr("should_peel"); }
3324 phase->do_peeling(this, old_new);
3325 } else if (policy_unswitching(phase)) {
3326 phase->do_unswitching(this, old_new);
3327 }
3328 return true;
3329 }
3330 CountedLoopNode *cl = _head->as_CountedLoop();
3331
3332 if (!cl->is_valid_counted_loop(T_INT)) return true; // Ignore various kinds of broken loops
3333
3334 // Do nothing special to pre- and post- loops
3335 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
3336
3337 // Compute loop trip count from profile data
3338 compute_profile_trip_cnt(phase);
3339
3340 // Before attempting fancy unrolling, RCE or alignment, see if we want
3341 // to completely unroll this loop or do loop unswitching.
3342 if (cl->is_normal_loop()) {
3343 if (policy_unswitching(phase)) {
3344 phase->do_unswitching(this, old_new);
3345 return true;
3346 }
3347 if (policy_maximally_unroll(phase)) {
3348 // Here we did some unrolling and peeling. Eventually we will
3349 // completely unroll this loop and it will no longer be a loop.
3350 phase->do_maximally_unroll(this, old_new);
3351 return true;
3352 }
3353 }
3354
3355 uint est_peeling = estimate_peeling(phase);
3356 bool should_peel = 0 < est_peeling;
3357
3358 // Counted loops may be peeled, or may need some iterations run up
3359 // front for RCE. Thus we clone a full loop up front whose trip count is
3360 // at least 1 (if peeling), but may be several more.
3361
3362 // The main loop will start cache-line aligned with at least 1
3363 // iteration of the unrolled body (zero-trip test required) and
3364 // will have some range checks removed.
3365
3366 // A post-loop will finish any odd iterations (leftover after
3367 // unrolling), plus any needed for RCE purposes.
3368
3369 bool should_unroll = policy_unroll(phase);
3370 bool should_rce = policy_range_check(phase);
3371
3372 // If not RCE'ing (iteration splitting), then we do not need a pre-loop.
3373 // We may still need to peel an initial iteration but we will not
3374 // be needing an unknown number of pre-iterations.
3375 //
3376 // Basically, if peel_only reports TRUE first time through, we will not
3377 // be able to later do RCE on this loop.
3378 bool peel_only = policy_peel_only(phase) && !should_rce;
3379
3380 // If we have any of these conditions (RCE, unrolling) met, then
3381 // we switch to the pre-/main-/post-loop model. This model also covers
3382 // peeling.
3383 if (should_rce || should_unroll) {
3384 if (cl->is_normal_loop()) { // Convert to 'pre/main/post' loops
3385 uint estimate = est_loop_clone_sz(3);
3386 if (!phase->may_require_nodes(estimate)) {
3387 return false;
3388 }
3389 phase->insert_pre_post_loops(this, old_new, peel_only);
3390 }
3391 // Adjust the pre- and main-loop limits to let the pre and post loops run
3392 // with full checks, but the main-loop with no checks. Remove said checks
3393 // from the main body.
3394 if (should_rce) {
3395 if (phase->do_range_check(this, old_new) != 0) {
3396 cl->mark_has_range_checks();
3397 }
3398 } else if (PostLoopMultiversioning) {
3399 phase->has_range_checks(this);
3400 }
3401
3402 if (should_unroll && !should_peel && PostLoopMultiversioning) {
3403 // Try to setup multiversioning on main loops before they are unrolled
3404 if (cl->is_main_loop() && (cl->unrolled_count() == 1)) {
3405 phase->insert_scalar_rced_post_loop(this, old_new);
3406 }
3407 }
3408
3409 // Double loop body for unrolling. Adjust the minimum-trip test (will do
3410 // twice as many iterations as before) and the main body limit (only do
3411 // an even number of trips). If we are peeling, we might enable some RCE
3412 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
3413 // peeling.
3414 if (should_unroll && !should_peel) {
3415 if (SuperWordLoopUnrollAnalysis) {
3416 phase->insert_vector_post_loop(this, old_new);
3417 }
3418 phase->do_unroll(this, old_new, true);
3419 }
3420 } else { // Else we have an unchanged counted loop
3421 if (should_peel) { // Might want to peel but do nothing else
3422 if (phase->may_require_nodes(est_peeling)) {
3423 phase->do_peeling(this, old_new);
3424 }
3425 }
3426 }
3427 return true;
3428 }
3429
3430
3431 //=============================================================================
3432 //------------------------------iteration_split--------------------------------
iteration_split(PhaseIdealLoop * phase,Node_List & old_new)3433 bool IdealLoopTree::iteration_split(PhaseIdealLoop* phase, Node_List &old_new) {
3434 // Recursively iteration split nested loops
3435 if (_child && !_child->iteration_split(phase, old_new)) {
3436 return false;
3437 }
3438
3439 // Clean out prior deadwood
3440 DCE_loop_body();
3441
3442 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
3443 // Replace with a 1-in-10 exit guess.
3444 if (!is_root() && is_loop()) {
3445 adjust_loop_exit_prob(phase);
3446 }
3447
3448 // Unrolling, RCE and peeling efforts, iff innermost loop.
3449 if (_allow_optimizations && is_innermost()) {
3450 if (!_has_call) {
3451 if (!iteration_split_impl(phase, old_new)) {
3452 return false;
3453 }
3454 } else {
3455 AutoNodeBudget node_budget(phase);
3456 if (policy_unswitching(phase)) {
3457 phase->do_unswitching(this, old_new);
3458 }
3459 }
3460 }
3461
3462 // Minor offset re-organization to remove loop-fallout uses of
3463 // trip counter when there was no major reshaping.
3464 phase->reorg_offsets(this);
3465
3466 if (_next && !_next->iteration_split(phase, old_new)) {
3467 return false;
3468 }
3469 return true;
3470 }
3471
3472
3473 //=============================================================================
3474 // Process all the loops in the loop tree and replace any fill
3475 // patterns with an intrinsic version.
do_intrinsify_fill()3476 bool PhaseIdealLoop::do_intrinsify_fill() {
3477 bool changed = false;
3478 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3479 IdealLoopTree* lpt = iter.current();
3480 changed |= intrinsify_fill(lpt);
3481 }
3482 return changed;
3483 }
3484
3485
3486 // Examine an inner loop looking for a a single store of an invariant
3487 // value in a unit stride loop,
match_fill_loop(IdealLoopTree * lpt,Node * & store,Node * & store_value,Node * & shift,Node * & con)3488 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
3489 Node*& shift, Node*& con) {
3490 const char* msg = NULL;
3491 Node* msg_node = NULL;
3492
3493 store_value = NULL;
3494 con = NULL;
3495 shift = NULL;
3496
3497 // Process the loop looking for stores. If there are multiple
3498 // stores or extra control flow give at this point.
3499 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3500 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3501 Node* n = lpt->_body.at(i);
3502 if (n->outcnt() == 0) continue; // Ignore dead
3503 if (n->is_Store()) {
3504 if (store != NULL) {
3505 msg = "multiple stores";
3506 break;
3507 }
3508 int opc = n->Opcode();
3509 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
3510 msg = "oop fills not handled";
3511 break;
3512 }
3513 Node* value = n->in(MemNode::ValueIn);
3514 if (!lpt->is_invariant(value)) {
3515 msg = "variant store value";
3516 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
3517 msg = "not array address";
3518 }
3519 store = n;
3520 store_value = value;
3521 } else if (n->is_If() && n != head->loopexit_or_null()) {
3522 msg = "extra control flow";
3523 msg_node = n;
3524 }
3525 }
3526
3527 if (store == NULL) {
3528 // No store in loop
3529 return false;
3530 }
3531
3532 if (msg == NULL && head->stride_con() != 1) {
3533 // could handle negative strides too
3534 if (head->stride_con() < 0) {
3535 msg = "negative stride";
3536 } else {
3537 msg = "non-unit stride";
3538 }
3539 }
3540
3541 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
3542 msg = "can't handle store address";
3543 msg_node = store->in(MemNode::Address);
3544 }
3545
3546 if (msg == NULL &&
3547 (!store->in(MemNode::Memory)->is_Phi() ||
3548 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
3549 msg = "store memory isn't proper phi";
3550 msg_node = store->in(MemNode::Memory);
3551 }
3552
3553 // Make sure there is an appropriate fill routine
3554 BasicType t = store->as_Mem()->memory_type();
3555 const char* fill_name;
3556 if (msg == NULL &&
3557 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
3558 msg = "unsupported store";
3559 msg_node = store;
3560 }
3561
3562 if (msg != NULL) {
3563 #ifndef PRODUCT
3564 if (TraceOptimizeFill) {
3565 tty->print_cr("not fill intrinsic candidate: %s", msg);
3566 if (msg_node != NULL) msg_node->dump();
3567 }
3568 #endif
3569 return false;
3570 }
3571
3572 // Make sure the address expression can be handled. It should be
3573 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()).
3574 Node* elements[4];
3575 Node* cast = NULL;
3576 Node* conv = NULL;
3577 bool found_index = false;
3578 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
3579 for (int e = 0; e < count; e++) {
3580 Node* n = elements[e];
3581 if (n->is_Con() && con == NULL) {
3582 con = n;
3583 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
3584 Node* value = n->in(1);
3585 #ifdef _LP64
3586 if (value->Opcode() == Op_ConvI2L) {
3587 conv = value;
3588 value = value->in(1);
3589 }
3590 if (value->Opcode() == Op_CastII &&
3591 value->as_CastII()->has_range_check()) {
3592 // Skip range check dependent CastII nodes
3593 cast = value;
3594 value = value->in(1);
3595 }
3596 #endif
3597 if (value != head->phi()) {
3598 msg = "unhandled shift in address";
3599 } else {
3600 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
3601 msg = "scale doesn't match";
3602 } else {
3603 found_index = true;
3604 shift = n;
3605 }
3606 }
3607 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
3608 conv = n;
3609 n = n->in(1);
3610 if (n->Opcode() == Op_CastII &&
3611 n->as_CastII()->has_range_check()) {
3612 // Skip range check dependent CastII nodes
3613 cast = n;
3614 n = n->in(1);
3615 }
3616 if (n == head->phi()) {
3617 found_index = true;
3618 } else {
3619 msg = "unhandled input to ConvI2L";
3620 }
3621 } else if (n == head->phi()) {
3622 // no shift, check below for allowed cases
3623 found_index = true;
3624 } else {
3625 msg = "unhandled node in address";
3626 msg_node = n;
3627 }
3628 }
3629
3630 if (count == -1) {
3631 msg = "malformed address expression";
3632 msg_node = store;
3633 }
3634
3635 if (!found_index) {
3636 msg = "missing use of index";
3637 }
3638
3639 // byte sized items won't have a shift
3640 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
3641 msg = "can't find shift";
3642 msg_node = store;
3643 }
3644
3645 if (msg != NULL) {
3646 #ifndef PRODUCT
3647 if (TraceOptimizeFill) {
3648 tty->print_cr("not fill intrinsic: %s", msg);
3649 if (msg_node != NULL) msg_node->dump();
3650 }
3651 #endif
3652 return false;
3653 }
3654
3655 // No make sure all the other nodes in the loop can be handled
3656 VectorSet ok;
3657
3658 // store related values are ok
3659 ok.set(store->_idx);
3660 ok.set(store->in(MemNode::Memory)->_idx);
3661
3662 CountedLoopEndNode* loop_exit = head->loopexit();
3663
3664 // Loop structure is ok
3665 ok.set(head->_idx);
3666 ok.set(loop_exit->_idx);
3667 ok.set(head->phi()->_idx);
3668 ok.set(head->incr()->_idx);
3669 ok.set(loop_exit->cmp_node()->_idx);
3670 ok.set(loop_exit->in(1)->_idx);
3671
3672 // Address elements are ok
3673 if (con) ok.set(con->_idx);
3674 if (shift) ok.set(shift->_idx);
3675 if (cast) ok.set(cast->_idx);
3676 if (conv) ok.set(conv->_idx);
3677
3678 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3679 Node* n = lpt->_body.at(i);
3680 if (n->outcnt() == 0) continue; // Ignore dead
3681 if (ok.test(n->_idx)) continue;
3682 // Backedge projection is ok
3683 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
3684 if (!n->is_AddP()) {
3685 msg = "unhandled node";
3686 msg_node = n;
3687 break;
3688 }
3689 }
3690
3691 // Make sure no unexpected values are used outside the loop
3692 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
3693 Node* n = lpt->_body.at(i);
3694 // These values can be replaced with other nodes if they are used
3695 // outside the loop.
3696 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
3697 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
3698 Node* use = iter.get();
3699 if (!lpt->_body.contains(use)) {
3700 if (n->is_CountedLoop() && n->as_CountedLoop()->is_strip_mined()) {
3701 // In strip-mined counted loops, the CountedLoopNode may be
3702 // used by the address polling node of the outer safepoint.
3703 // Skip this use because it's safe.
3704 #ifdef ASSERT
3705 Node* sfpt = n->as_CountedLoop()->outer_safepoint();
3706 Node* polladr = sfpt->in(TypeFunc::Parms+0);
3707 assert(use == polladr, "the use should be a safepoint polling");
3708 #endif
3709 continue;
3710 } else {
3711 msg = "node is used outside loop";
3712 msg_node = n;
3713 break;
3714 }
3715 }
3716 }
3717 }
3718
3719 #ifdef ASSERT
3720 if (TraceOptimizeFill) {
3721 if (msg != NULL) {
3722 tty->print_cr("no fill intrinsic: %s", msg);
3723 if (msg_node != NULL) msg_node->dump();
3724 } else {
3725 tty->print_cr("fill intrinsic for:");
3726 }
3727 store->dump();
3728 if (Verbose) {
3729 lpt->_body.dump();
3730 }
3731 }
3732 #endif
3733
3734 return msg == NULL;
3735 }
3736
3737
3738
intrinsify_fill(IdealLoopTree * lpt)3739 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
3740 // Only for counted inner loops
3741 if (!lpt->is_counted() || !lpt->is_innermost()) {
3742 return false;
3743 }
3744
3745 // Must have constant stride
3746 CountedLoopNode* head = lpt->_head->as_CountedLoop();
3747 if (!head->is_valid_counted_loop(T_INT) || !head->is_normal_loop()) {
3748 return false;
3749 }
3750
3751 head->verify_strip_mined(1);
3752
3753 // Check that the body only contains a store of a loop invariant
3754 // value that is indexed by the loop phi.
3755 Node* store = NULL;
3756 Node* store_value = NULL;
3757 Node* shift = NULL;
3758 Node* offset = NULL;
3759 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
3760 return false;
3761 }
3762
3763 Node* exit = head->loopexit()->proj_out_or_null(0);
3764 if (exit == NULL) {
3765 return false;
3766 }
3767
3768 #ifndef PRODUCT
3769 if (TraceLoopOpts) {
3770 tty->print("ArrayFill ");
3771 lpt->dump_head();
3772 }
3773 #endif
3774
3775 // Now replace the whole loop body by a call to a fill routine that
3776 // covers the same region as the loop.
3777 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
3778
3779 // Build an expression for the beginning of the copy region
3780 Node* index = head->init_trip();
3781 #ifdef _LP64
3782 index = new ConvI2LNode(index);
3783 _igvn.register_new_node_with_optimizer(index);
3784 #endif
3785 if (shift != NULL) {
3786 // byte arrays don't require a shift but others do.
3787 index = new LShiftXNode(index, shift->in(2));
3788 _igvn.register_new_node_with_optimizer(index);
3789 }
3790 index = new AddPNode(base, base, index);
3791 _igvn.register_new_node_with_optimizer(index);
3792 Node* from = new AddPNode(base, index, offset);
3793 _igvn.register_new_node_with_optimizer(from);
3794 // Compute the number of elements to copy
3795 Node* len = new SubINode(head->limit(), head->init_trip());
3796 _igvn.register_new_node_with_optimizer(len);
3797
3798 BasicType t = store->as_Mem()->memory_type();
3799 bool aligned = false;
3800 if (offset != NULL && head->init_trip()->is_Con()) {
3801 int element_size = type2aelembytes(t);
3802 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
3803 }
3804
3805 // Build a call to the fill routine
3806 const char* fill_name;
3807 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
3808 assert(fill != NULL, "what?");
3809
3810 // Convert float/double to int/long for fill routines
3811 if (t == T_FLOAT) {
3812 store_value = new MoveF2INode(store_value);
3813 _igvn.register_new_node_with_optimizer(store_value);
3814 } else if (t == T_DOUBLE) {
3815 store_value = new MoveD2LNode(store_value);
3816 _igvn.register_new_node_with_optimizer(store_value);
3817 }
3818
3819 Node* mem_phi = store->in(MemNode::Memory);
3820 Node* result_ctrl;
3821 Node* result_mem;
3822 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
3823 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill,
3824 fill_name, TypeAryPtr::get_array_body_type(t));
3825 uint cnt = 0;
3826 call->init_req(TypeFunc::Parms + cnt++, from);
3827 call->init_req(TypeFunc::Parms + cnt++, store_value);
3828 #ifdef _LP64
3829 len = new ConvI2LNode(len);
3830 _igvn.register_new_node_with_optimizer(len);
3831 #endif
3832 call->init_req(TypeFunc::Parms + cnt++, len);
3833 #ifdef _LP64
3834 call->init_req(TypeFunc::Parms + cnt++, C->top());
3835 #endif
3836 call->init_req(TypeFunc::Control, head->init_control());
3837 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O.
3838 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl));
3839 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out_or_null(TypeFunc::ReturnAdr));
3840 call->init_req(TypeFunc::FramePtr, C->start()->proj_out_or_null(TypeFunc::FramePtr));
3841 _igvn.register_new_node_with_optimizer(call);
3842 result_ctrl = new ProjNode(call,TypeFunc::Control);
3843 _igvn.register_new_node_with_optimizer(result_ctrl);
3844 result_mem = new ProjNode(call,TypeFunc::Memory);
3845 _igvn.register_new_node_with_optimizer(result_mem);
3846
3847 /* Disable following optimization until proper fix (add missing checks).
3848
3849 // If this fill is tightly coupled to an allocation and overwrites
3850 // the whole body, allow it to take over the zeroing.
3851 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
3852 if (alloc != NULL && alloc->is_AllocateArray()) {
3853 Node* length = alloc->as_AllocateArray()->Ideal_length();
3854 if (head->limit() == length &&
3855 head->init_trip() == _igvn.intcon(0)) {
3856 if (TraceOptimizeFill) {
3857 tty->print_cr("Eliminated zeroing in allocation");
3858 }
3859 alloc->maybe_set_complete(&_igvn);
3860 } else {
3861 #ifdef ASSERT
3862 if (TraceOptimizeFill) {
3863 tty->print_cr("filling array but bounds don't match");
3864 alloc->dump();
3865 head->init_trip()->dump();
3866 head->limit()->dump();
3867 length->dump();
3868 }
3869 #endif
3870 }
3871 }
3872 */
3873
3874 if (head->is_strip_mined()) {
3875 // Inner strip mined loop goes away so get rid of outer strip
3876 // mined loop
3877 Node* outer_sfpt = head->outer_safepoint();
3878 Node* in = outer_sfpt->in(0);
3879 Node* outer_out = head->outer_loop_exit();
3880 lazy_replace(outer_out, in);
3881 _igvn.replace_input_of(outer_sfpt, 0, C->top());
3882 }
3883
3884 // Redirect the old control and memory edges that are outside the loop.
3885 // Sometimes the memory phi of the head is used as the outgoing
3886 // state of the loop. It's safe in this case to replace it with the
3887 // result_mem.
3888 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
3889 lazy_replace(exit, result_ctrl);
3890 _igvn.replace_node(store, result_mem);
3891 // Any uses the increment outside of the loop become the loop limit.
3892 _igvn.replace_node(head->incr(), head->limit());
3893
3894 // Disconnect the head from the loop.
3895 for (uint i = 0; i < lpt->_body.size(); i++) {
3896 Node* n = lpt->_body.at(i);
3897 _igvn.replace_node(n, C->top());
3898 }
3899
3900 return true;
3901 }
3902