Mon, 26 Sep 2011 10:24:05 -0700
7081933: Use zeroing elimination optimization for large array
Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it.
Reviewed-by: twisti
1 /*
2 * Copyright (c) 2000, 2011, 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 */
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/connode.hpp"
31 #include "opto/divnode.hpp"
32 #include "opto/loopnode.hpp"
33 #include "opto/mulnode.hpp"
34 #include "opto/rootnode.hpp"
35 #include "opto/runtime.hpp"
36 #include "opto/subnode.hpp"
38 //------------------------------is_loop_exit-----------------------------------
39 // Given an IfNode, return the loop-exiting projection or NULL if both
40 // arms remain in the loop.
41 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
43 PhaseIdealLoop *phase = _phase;
44 // Test is an IfNode, has 2 projections. If BOTH are in the loop
45 // we need loop unswitching instead of peeling.
46 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
47 return iff->raw_out(0);
48 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
49 return iff->raw_out(1);
50 return NULL;
51 }
54 //=============================================================================
57 //------------------------------record_for_igvn----------------------------
58 // Put loop body on igvn work list
59 void IdealLoopTree::record_for_igvn() {
60 for( uint i = 0; i < _body.size(); i++ ) {
61 Node *n = _body.at(i);
62 _phase->_igvn._worklist.push(n);
63 }
64 }
66 //------------------------------compute_exact_trip_count-----------------------
67 // Compute loop exact trip count if possible. Do not recalculate trip count for
68 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
69 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
70 if (!_head->as_Loop()->is_valid_counted_loop()) {
71 return;
72 }
73 CountedLoopNode* cl = _head->as_CountedLoop();
74 // Trip count may become nonexact for iteration split loops since
75 // RCE modifies limits. Note, _trip_count value is not reset since
76 // it is used to limit unrolling of main loop.
77 cl->set_nonexact_trip_count();
79 // Loop's test should be part of loop.
80 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
81 return; // Infinite loop
83 #ifdef ASSERT
84 BoolTest::mask bt = cl->loopexit()->test_trip();
85 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
86 bt == BoolTest::ne, "canonical test is expected");
87 #endif
89 Node* init_n = cl->init_trip();
90 Node* limit_n = cl->limit();
91 if (init_n != NULL && init_n->is_Con() &&
92 limit_n != NULL && limit_n->is_Con()) {
93 // Use longs to avoid integer overflow.
94 int stride_con = cl->stride_con();
95 long init_con = cl->init_trip()->get_int();
96 long limit_con = cl->limit()->get_int();
97 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
98 long trip_count = (limit_con - init_con + stride_m)/stride_con;
99 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
100 // Set exact trip count.
101 cl->set_exact_trip_count((uint)trip_count);
102 }
103 }
104 }
106 //------------------------------compute_profile_trip_cnt----------------------------
107 // Compute loop trip count from profile data as
108 // (backedge_count + loop_exit_count) / loop_exit_count
109 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
110 if (!_head->is_CountedLoop()) {
111 return;
112 }
113 CountedLoopNode* head = _head->as_CountedLoop();
114 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
115 return; // Already computed
116 }
117 float trip_cnt = (float)max_jint; // default is big
119 Node* back = head->in(LoopNode::LoopBackControl);
120 while (back != head) {
121 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
122 back->in(0) &&
123 back->in(0)->is_If() &&
124 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
125 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
126 break;
127 }
128 back = phase->idom(back);
129 }
130 if (back != head) {
131 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
132 back->in(0), "if-projection exists");
133 IfNode* back_if = back->in(0)->as_If();
134 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
136 // Now compute a loop exit count
137 float loop_exit_cnt = 0.0f;
138 for( uint i = 0; i < _body.size(); i++ ) {
139 Node *n = _body[i];
140 if( n->is_If() ) {
141 IfNode *iff = n->as_If();
142 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
143 Node *exit = is_loop_exit(iff);
144 if( exit ) {
145 float exit_prob = iff->_prob;
146 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
147 if (exit_prob > PROB_MIN) {
148 float exit_cnt = iff->_fcnt * exit_prob;
149 loop_exit_cnt += exit_cnt;
150 }
151 }
152 }
153 }
154 }
155 if (loop_exit_cnt > 0.0f) {
156 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
157 } else {
158 // No exit count so use
159 trip_cnt = loop_back_cnt;
160 }
161 }
162 #ifndef PRODUCT
163 if (TraceProfileTripCount) {
164 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
165 }
166 #endif
167 head->set_profile_trip_cnt(trip_cnt);
168 }
170 //---------------------is_invariant_addition-----------------------------
171 // Return nonzero index of invariant operand for an Add or Sub
172 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
173 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
174 int op = n->Opcode();
175 if (op == Op_AddI || op == Op_SubI) {
176 bool in1_invar = this->is_invariant(n->in(1));
177 bool in2_invar = this->is_invariant(n->in(2));
178 if (in1_invar && !in2_invar) return 1;
179 if (!in1_invar && in2_invar) return 2;
180 }
181 return 0;
182 }
184 //---------------------reassociate_add_sub-----------------------------
185 // Reassociate invariant add and subtract expressions:
186 //
187 // inv1 + (x + inv2) => ( inv1 + inv2) + x
188 // (x + inv2) + inv1 => ( inv1 + inv2) + x
189 // inv1 + (x - inv2) => ( inv1 - inv2) + x
190 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
191 // (x + inv2) - inv1 => (-inv1 + inv2) + x
192 // (x - inv2) + inv1 => ( inv1 - inv2) + x
193 // (x - inv2) - inv1 => (-inv1 - inv2) + x
194 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
195 // inv1 - (x - inv2) => ( inv1 + inv2) - x
196 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
197 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
198 // inv1 - (x + inv2) => ( inv1 - inv2) - x
199 //
200 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
201 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
202 if (is_invariant(n1)) return NULL;
203 int inv1_idx = is_invariant_addition(n1, phase);
204 if (!inv1_idx) return NULL;
205 // Don't mess with add of constant (igvn moves them to expression tree root.)
206 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
207 Node* inv1 = n1->in(inv1_idx);
208 Node* n2 = n1->in(3 - inv1_idx);
209 int inv2_idx = is_invariant_addition(n2, phase);
210 if (!inv2_idx) return NULL;
211 Node* x = n2->in(3 - inv2_idx);
212 Node* inv2 = n2->in(inv2_idx);
214 bool neg_x = n2->is_Sub() && inv2_idx == 1;
215 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
216 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
217 if (n1->is_Sub() && inv1_idx == 1) {
218 neg_x = !neg_x;
219 neg_inv2 = !neg_inv2;
220 }
221 Node* inv1_c = phase->get_ctrl(inv1);
222 Node* inv2_c = phase->get_ctrl(inv2);
223 Node* n_inv1;
224 if (neg_inv1) {
225 Node *zero = phase->_igvn.intcon(0);
226 phase->set_ctrl(zero, phase->C->root());
227 n_inv1 = new (phase->C, 3) SubINode(zero, inv1);
228 phase->register_new_node(n_inv1, inv1_c);
229 } else {
230 n_inv1 = inv1;
231 }
232 Node* inv;
233 if (neg_inv2) {
234 inv = new (phase->C, 3) SubINode(n_inv1, inv2);
235 } else {
236 inv = new (phase->C, 3) AddINode(n_inv1, inv2);
237 }
238 phase->register_new_node(inv, phase->get_early_ctrl(inv));
240 Node* addx;
241 if (neg_x) {
242 addx = new (phase->C, 3) SubINode(inv, x);
243 } else {
244 addx = new (phase->C, 3) AddINode(x, inv);
245 }
246 phase->register_new_node(addx, phase->get_ctrl(x));
247 phase->_igvn.replace_node(n1, addx);
248 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
249 _body.yank(n1);
250 return addx;
251 }
253 //---------------------reassociate_invariants-----------------------------
254 // Reassociate invariant expressions:
255 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
256 for (int i = _body.size() - 1; i >= 0; i--) {
257 Node *n = _body.at(i);
258 for (int j = 0; j < 5; j++) {
259 Node* nn = reassociate_add_sub(n, phase);
260 if (nn == NULL) break;
261 n = nn; // again
262 };
263 }
264 }
266 //------------------------------policy_peeling---------------------------------
267 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
268 // make some loop-invariant test (usually a null-check) happen before the loop.
269 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
270 Node *test = ((IdealLoopTree*)this)->tail();
271 int body_size = ((IdealLoopTree*)this)->_body.size();
272 int uniq = phase->C->unique();
273 // Peeling does loop cloning which can result in O(N^2) node construction
274 if( body_size > 255 /* Prevent overflow for large body_size */
275 || (body_size * body_size + uniq > MaxNodeLimit) ) {
276 return false; // too large to safely clone
277 }
278 while( test != _head ) { // Scan till run off top of loop
279 if( test->is_If() ) { // Test?
280 Node *ctrl = phase->get_ctrl(test->in(1));
281 if (ctrl->is_top())
282 return false; // Found dead test on live IF? No peeling!
283 // Standard IF only has one input value to check for loop invariance
284 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
285 // Condition is not a member of this loop?
286 if( !is_member(phase->get_loop(ctrl)) &&
287 is_loop_exit(test) )
288 return true; // Found reason to peel!
289 }
290 // Walk up dominators to loop _head looking for test which is
291 // executed on every path thru loop.
292 test = phase->idom(test);
293 }
294 return false;
295 }
297 //------------------------------peeled_dom_test_elim---------------------------
298 // If we got the effect of peeling, either by actually peeling or by making
299 // a pre-loop which must execute at least once, we can remove all
300 // loop-invariant dominated tests in the main body.
301 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
302 bool progress = true;
303 while( progress ) {
304 progress = false; // Reset for next iteration
305 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
306 Node *test = prev->in(0);
307 while( test != loop->_head ) { // Scan till run off top of loop
309 int p_op = prev->Opcode();
310 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
311 test->is_If() && // Test?
312 !test->in(1)->is_Con() && // And not already obvious?
313 // Condition is not a member of this loop?
314 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
315 // Walk loop body looking for instances of this test
316 for( uint i = 0; i < loop->_body.size(); i++ ) {
317 Node *n = loop->_body.at(i);
318 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
319 // IfNode was dominated by version in peeled loop body
320 progress = true;
321 dominated_by( old_new[prev->_idx], n );
322 }
323 }
324 }
325 prev = test;
326 test = idom(test);
327 } // End of scan tests in loop
329 } // End of while( progress )
330 }
332 //------------------------------do_peeling-------------------------------------
333 // Peel the first iteration of the given loop.
334 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
335 // The pre-loop illegally has 2 control users (old & new loops).
336 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
337 // Do this by making the old-loop fall-in edges act as if they came
338 // around the loopback from the prior iteration (follow the old-loop
339 // backedges) and then map to the new peeled iteration. This leaves
340 // the pre-loop with only 1 user (the new peeled iteration), but the
341 // peeled-loop backedge has 2 users.
342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
343 // extra backedge user.
344 //
345 // orig
346 //
347 // stmt1
348 // |
349 // v
350 // loop predicate
351 // |
352 // v
353 // loop<----+
354 // | |
355 // stmt2 |
356 // | |
357 // v |
358 // if ^
359 // / \ |
360 // / \ |
361 // v v |
362 // false true |
363 // / \ |
364 // / ----+
365 // |
366 // v
367 // exit
368 //
369 //
370 // after clone loop
371 //
372 // stmt1
373 // |
374 // v
375 // loop predicate
376 // / \
377 // clone / \ orig
378 // / \
379 // / \
380 // v v
381 // +---->loop clone loop<----+
382 // | | | |
383 // | stmt2 clone stmt2 |
384 // | | | |
385 // | v v |
386 // ^ if clone If ^
387 // | / \ / \ |
388 // | / \ / \ |
389 // | v v v v |
390 // | true false false true |
391 // | / \ / \ |
392 // +---- \ / ----+
393 // \ /
394 // 1v v2
395 // region
396 // |
397 // v
398 // exit
399 //
400 //
401 // after peel and predicate move
402 //
403 // stmt1
404 // /
405 // /
406 // clone / orig
407 // /
408 // / +----------+
409 // / | |
410 // / loop predicate |
411 // / | |
412 // v v |
413 // TOP-->loop clone loop<----+ |
414 // | | | |
415 // stmt2 clone stmt2 | |
416 // | | | ^
417 // v v | |
418 // if clone If ^ |
419 // / \ / \ | |
420 // / \ / \ | |
421 // v v v v | |
422 // true false false true | |
423 // | \ / \ | |
424 // | \ / ----+ ^
425 // | \ / |
426 // | 1v v2 |
427 // v region |
428 // | | |
429 // | v |
430 // | exit |
431 // | |
432 // +--------------->-----------------+
433 //
434 //
435 // final graph
436 //
437 // stmt1
438 // |
439 // v
440 // stmt2 clone
441 // |
442 // v
443 // if clone
444 // / |
445 // / |
446 // v v
447 // false true
448 // | |
449 // | v
450 // | loop predicate
451 // | |
452 // | v
453 // | loop<----+
454 // | | |
455 // | stmt2 |
456 // | | |
457 // | v |
458 // v if ^
459 // | / \ |
460 // | / \ |
461 // | v v |
462 // | false true |
463 // | | \ |
464 // v v --+
465 // region
466 // |
467 // v
468 // exit
469 //
470 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
472 C->set_major_progress();
473 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
474 // 'pre' loop from the main and the 'pre' can no longer have it's
475 // iterations adjusted. Therefore, we need to declare this loop as
476 // no longer a 'main' loop; it will need new pre and post loops before
477 // we can do further RCE.
478 #ifndef PRODUCT
479 if (TraceLoopOpts) {
480 tty->print("Peel ");
481 loop->dump_head();
482 }
483 #endif
484 Node* head = loop->_head;
485 bool counted_loop = head->is_CountedLoop();
486 if (counted_loop) {
487 CountedLoopNode *cl = head->as_CountedLoop();
488 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
489 cl->set_trip_count(cl->trip_count() - 1);
490 if (cl->is_main_loop()) {
491 cl->set_normal_loop();
492 #ifndef PRODUCT
493 if (PrintOpto && VerifyLoopOptimizations) {
494 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
495 loop->dump_head();
496 }
497 #endif
498 }
499 }
500 Node* entry = head->in(LoopNode::EntryControl);
502 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
503 // The pre-loop illegally has 2 control users (old & new loops).
504 clone_loop( loop, old_new, dom_depth(head) );
506 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
507 // Do this by making the old-loop fall-in edges act as if they came
508 // around the loopback from the prior iteration (follow the old-loop
509 // backedges) and then map to the new peeled iteration. This leaves
510 // the pre-loop with only 1 user (the new peeled iteration), but the
511 // peeled-loop backedge has 2 users.
512 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
513 _igvn.hash_delete(head);
514 head->set_req(LoopNode::EntryControl, new_entry);
515 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
516 Node* old = head->fast_out(j);
517 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
518 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
519 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
520 // Then loop body backedge value remains the same.
521 new_exit_value = old->in(LoopNode::LoopBackControl);
522 _igvn.hash_delete(old);
523 old->set_req(LoopNode::EntryControl, new_exit_value);
524 }
525 }
528 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
529 // extra backedge user.
530 Node* new_head = old_new[head->_idx];
531 _igvn.hash_delete(new_head);
532 new_head->set_req(LoopNode::LoopBackControl, C->top());
533 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
534 Node* use = new_head->fast_out(j2);
535 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
536 _igvn.hash_delete(use);
537 use->set_req(LoopNode::LoopBackControl, C->top());
538 }
539 }
542 // Step 4: Correct dom-depth info. Set to loop-head depth.
543 int dd = dom_depth(head);
544 set_idom(head, head->in(1), dd);
545 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
546 Node *old = loop->_body.at(j3);
547 Node *nnn = old_new[old->_idx];
548 if (!has_ctrl(nnn))
549 set_idom(nnn, idom(nnn), dd-1);
550 // While we're at it, remove any SafePoints from the peeled code
551 if (old->Opcode() == Op_SafePoint) {
552 Node *nnn = old_new[old->_idx];
553 lazy_replace(nnn,nnn->in(TypeFunc::Control));
554 }
555 }
557 // Now force out all loop-invariant dominating tests. The optimizer
558 // finds some, but we _know_ they are all useless.
559 peeled_dom_test_elim(loop,old_new);
561 loop->record_for_igvn();
562 }
564 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
566 //------------------------------policy_maximally_unroll------------------------
567 // Calculate exact loop trip count and return true if loop can be maximally
568 // unrolled.
569 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
570 CountedLoopNode *cl = _head->as_CountedLoop();
571 assert(cl->is_normal_loop(), "");
572 if (!cl->is_valid_counted_loop())
573 return false; // Malformed counted loop
575 if (!cl->has_exact_trip_count()) {
576 // Trip count is not exact.
577 return false;
578 }
580 uint trip_count = cl->trip_count();
581 // Note, max_juint is used to indicate unknown trip count.
582 assert(trip_count > 1, "one iteration loop should be optimized out already");
583 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
585 // Real policy: if we maximally unroll, does it get too big?
586 // Allow the unrolled mess to get larger than standard loop
587 // size. After all, it will no longer be a loop.
588 uint body_size = _body.size();
589 uint unroll_limit = (uint)LoopUnrollLimit * 4;
590 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
591 if (trip_count > unroll_limit || body_size > unroll_limit) {
592 return false;
593 }
595 // Fully unroll a loop with few iterations regardless next
596 // conditions since following loop optimizations will split
597 // such loop anyway (pre-main-post).
598 if (trip_count <= 3)
599 return true;
601 // Take into account that after unroll conjoined heads and tails will fold,
602 // otherwise policy_unroll() may allow more unrolling than max unrolling.
603 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
604 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
605 if (body_size != tst_body_size) // Check for int overflow
606 return false;
607 if (new_body_size > unroll_limit ||
608 // Unrolling can result in a large amount of node construction
609 new_body_size >= MaxNodeLimit - phase->C->unique()) {
610 return false;
611 }
613 // Do not unroll a loop with String intrinsics code.
614 // String intrinsics are large and have loops.
615 for (uint k = 0; k < _body.size(); k++) {
616 Node* n = _body.at(k);
617 switch (n->Opcode()) {
618 case Op_StrComp:
619 case Op_StrEquals:
620 case Op_StrIndexOf:
621 case Op_AryEq: {
622 return false;
623 }
624 } // switch
625 }
627 return true; // Do maximally unroll
628 }
631 #define MAX_UNROLL 16 // maximum number of unrolls for main loop
633 //------------------------------policy_unroll----------------------------------
634 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
635 // the loop is a CountedLoop and the body is small enough.
636 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
638 CountedLoopNode *cl = _head->as_CountedLoop();
639 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
641 if (!cl->is_valid_counted_loop())
642 return false; // Malformed counted loop
644 // Protect against over-unrolling.
645 // After split at least one iteration will be executed in pre-loop.
646 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
648 int future_unroll_ct = cl->unrolled_count() * 2;
649 if (future_unroll_ct > MAX_UNROLL) return false;
651 // Check for initial stride being a small enough constant
652 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
654 // Don't unroll if the next round of unrolling would push us
655 // over the expected trip count of the loop. One is subtracted
656 // from the expected trip count because the pre-loop normally
657 // executes 1 iteration.
658 if (UnrollLimitForProfileCheck > 0 &&
659 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
660 future_unroll_ct > UnrollLimitForProfileCheck &&
661 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
662 return false;
663 }
665 // When unroll count is greater than LoopUnrollMin, don't unroll if:
666 // the residual iterations are more than 10% of the trip count
667 // and rounds of "unroll,optimize" are not making significant progress
668 // Progress defined as current size less than 20% larger than previous size.
669 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
670 future_unroll_ct > LoopUnrollMin &&
671 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
672 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
673 return false;
674 }
676 Node *init_n = cl->init_trip();
677 Node *limit_n = cl->limit();
678 int stride_con = cl->stride_con();
679 // Non-constant bounds.
680 // Protect against over-unrolling when init or/and limit are not constant
681 // (so that trip_count's init value is maxint) but iv range is known.
682 if (init_n == NULL || !init_n->is_Con() ||
683 limit_n == NULL || !limit_n->is_Con()) {
684 Node* phi = cl->phi();
685 if (phi != NULL) {
686 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
687 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
688 int next_stride = stride_con * 2; // stride after this unroll
689 if (next_stride > 0) {
690 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
691 iv_type->_lo + next_stride > iv_type->_hi) {
692 return false; // over-unrolling
693 }
694 } else if (next_stride < 0) {
695 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
696 iv_type->_hi + next_stride < iv_type->_lo) {
697 return false; // over-unrolling
698 }
699 }
700 }
701 }
703 // After unroll limit will be adjusted: new_limit = limit-stride.
704 // Bailout if adjustment overflow.
705 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
706 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
707 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
708 return false; // overflow
710 // Adjust body_size to determine if we unroll or not
711 uint body_size = _body.size();
712 // Key test to unroll loop in CRC32 java code
713 int xors_in_loop = 0;
714 // Also count ModL, DivL and MulL which expand mightly
715 for (uint k = 0; k < _body.size(); k++) {
716 Node* n = _body.at(k);
717 switch (n->Opcode()) {
718 case Op_XorI: xors_in_loop++; break; // CRC32 java code
719 case Op_ModL: body_size += 30; break;
720 case Op_DivL: body_size += 30; break;
721 case Op_MulL: body_size += 10; break;
722 case Op_StrComp:
723 case Op_StrEquals:
724 case Op_StrIndexOf:
725 case Op_AryEq: {
726 // Do not unroll a loop with String intrinsics code.
727 // String intrinsics are large and have loops.
728 return false;
729 }
730 } // switch
731 }
733 // Check for being too big
734 if (body_size > (uint)LoopUnrollLimit) {
735 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
736 // Normal case: loop too big
737 return false;
738 }
740 // Unroll once! (Each trip will soon do double iterations)
741 return true;
742 }
744 //------------------------------policy_align-----------------------------------
745 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
746 // expression that does the alignment. Note that only one array base can be
747 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
748 // if we vectorize short memory ops into longer memory ops, we may want to
749 // increase alignment.
750 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
751 return false;
752 }
754 //------------------------------policy_range_check-----------------------------
755 // Return TRUE or FALSE if the loop should be range-check-eliminated.
756 // Actually we do iteration-splitting, a more powerful form of RCE.
757 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
758 if (!RangeCheckElimination) return false;
760 CountedLoopNode *cl = _head->as_CountedLoop();
761 // If we unrolled with no intention of doing RCE and we later
762 // changed our minds, we got no pre-loop. Either we need to
763 // make a new pre-loop, or we gotta disallow RCE.
764 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
765 Node *trip_counter = cl->phi();
767 // Check loop body for tests of trip-counter plus loop-invariant vs
768 // loop-invariant.
769 for (uint i = 0; i < _body.size(); i++) {
770 Node *iff = _body[i];
771 if (iff->Opcode() == Op_If) { // Test?
773 // Comparing trip+off vs limit
774 Node *bol = iff->in(1);
775 if (bol->req() != 2) continue; // dead constant test
776 if (!bol->is_Bool()) {
777 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
778 continue;
779 }
780 if (bol->as_Bool()->_test._test == BoolTest::ne)
781 continue; // not RC
783 Node *cmp = bol->in(1);
785 Node *rc_exp = cmp->in(1);
786 Node *limit = cmp->in(2);
788 Node *limit_c = phase->get_ctrl(limit);
789 if( limit_c == phase->C->top() )
790 return false; // Found dead test on live IF? No RCE!
791 if( is_member(phase->get_loop(limit_c) ) ) {
792 // Compare might have operands swapped; commute them
793 rc_exp = cmp->in(2);
794 limit = cmp->in(1);
795 limit_c = phase->get_ctrl(limit);
796 if( is_member(phase->get_loop(limit_c) ) )
797 continue; // Both inputs are loop varying; cannot RCE
798 }
800 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
801 continue;
802 }
803 // Yeah! Found a test like 'trip+off vs limit'
804 // Test is an IfNode, has 2 projections. If BOTH are in the loop
805 // we need loop unswitching instead of iteration splitting.
806 if( is_loop_exit(iff) )
807 return true; // Found reason to split iterations
808 } // End of is IF
809 }
811 return false;
812 }
814 //------------------------------policy_peel_only-------------------------------
815 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
816 // for unrolling loops with NO array accesses.
817 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
819 for( uint i = 0; i < _body.size(); i++ )
820 if( _body[i]->is_Mem() )
821 return false;
823 // No memory accesses at all!
824 return true;
825 }
827 //------------------------------clone_up_backedge_goo--------------------------
828 // If Node n lives in the back_ctrl block and cannot float, we clone a private
829 // version of n in preheader_ctrl block and return that, otherwise return n.
830 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
831 if( get_ctrl(n) != back_ctrl ) return n;
833 // Only visit once
834 if (visited.test_set(n->_idx)) {
835 Node *x = clones.find(n->_idx);
836 if (x != NULL)
837 return x;
838 return n;
839 }
841 Node *x = NULL; // If required, a clone of 'n'
842 // Check for 'n' being pinned in the backedge.
843 if( n->in(0) && n->in(0) == back_ctrl ) {
844 assert(clones.find(n->_idx) == NULL, "dead loop");
845 x = n->clone(); // Clone a copy of 'n' to preheader
846 clones.push(x, n->_idx);
847 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
848 }
850 // Recursive fixup any other input edges into x.
851 // If there are no changes we can just return 'n', otherwise
852 // we need to clone a private copy and change it.
853 for( uint i = 1; i < n->req(); i++ ) {
854 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
855 if( g != n->in(i) ) {
856 if( !x ) {
857 assert(clones.find(n->_idx) == NULL, "dead loop");
858 x = n->clone();
859 clones.push(x, n->_idx);
860 }
861 x->set_req(i, g);
862 }
863 }
864 if( x ) { // x can legally float to pre-header location
865 register_new_node( x, preheader_ctrl );
866 return x;
867 } else { // raise n to cover LCA of uses
868 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
869 }
870 return n;
871 }
873 //------------------------------insert_pre_post_loops--------------------------
874 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
875 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
876 // alignment. Useful to unroll loops that do no array accesses.
877 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
879 #ifndef PRODUCT
880 if (TraceLoopOpts) {
881 if (peel_only)
882 tty->print("PeelMainPost ");
883 else
884 tty->print("PreMainPost ");
885 loop->dump_head();
886 }
887 #endif
888 C->set_major_progress();
890 // Find common pieces of the loop being guarded with pre & post loops
891 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
892 assert( main_head->is_normal_loop(), "" );
893 CountedLoopEndNode *main_end = main_head->loopexit();
894 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
895 uint dd_main_head = dom_depth(main_head);
896 uint max = main_head->outcnt();
898 Node *pre_header= main_head->in(LoopNode::EntryControl);
899 Node *init = main_head->init_trip();
900 Node *incr = main_end ->incr();
901 Node *limit = main_end ->limit();
902 Node *stride = main_end ->stride();
903 Node *cmp = main_end ->cmp_node();
904 BoolTest::mask b_test = main_end->test_trip();
906 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
907 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
908 if( bol->outcnt() != 1 ) {
909 bol = bol->clone();
910 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
911 _igvn.hash_delete(main_end);
912 main_end->set_req(CountedLoopEndNode::TestValue, bol);
913 }
914 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
915 if( cmp->outcnt() != 1 ) {
916 cmp = cmp->clone();
917 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
918 _igvn.hash_delete(bol);
919 bol->set_req(1, cmp);
920 }
922 //------------------------------
923 // Step A: Create Post-Loop.
924 Node* main_exit = main_end->proj_out(false);
925 assert( main_exit->Opcode() == Op_IfFalse, "" );
926 int dd_main_exit = dom_depth(main_exit);
928 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
929 // loop pre-header illegally has 2 control users (old & new loops).
930 clone_loop( loop, old_new, dd_main_exit );
931 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
932 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
933 post_head->set_post_loop(main_head);
935 // Reduce the post-loop trip count.
936 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
937 post_end->_prob = PROB_FAIR;
939 // Build the main-loop normal exit.
940 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
941 _igvn.register_new_node_with_optimizer( new_main_exit );
942 set_idom(new_main_exit, main_end, dd_main_exit );
943 set_loop(new_main_exit, loop->_parent);
945 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
946 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
947 // (the main-loop trip-counter exit value) because we will be changing
948 // the exit value (via unrolling) so we cannot constant-fold away the zero
949 // trip guard until all unrolling is done.
950 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
951 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
952 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
953 register_new_node( zer_opaq, new_main_exit );
954 register_new_node( zer_cmp , new_main_exit );
955 register_new_node( zer_bol , new_main_exit );
957 // Build the IfNode
958 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
959 _igvn.register_new_node_with_optimizer( zer_iff );
960 set_idom(zer_iff, new_main_exit, dd_main_exit);
961 set_loop(zer_iff, loop->_parent);
963 // Plug in the false-path, taken if we need to skip post-loop
964 _igvn.hash_delete( main_exit );
965 main_exit->set_req(0, zer_iff);
966 _igvn._worklist.push(main_exit);
967 set_idom(main_exit, zer_iff, dd_main_exit);
968 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
969 // Make the true-path, must enter the post loop
970 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
971 _igvn.register_new_node_with_optimizer( zer_taken );
972 set_idom(zer_taken, zer_iff, dd_main_exit);
973 set_loop(zer_taken, loop->_parent);
974 // Plug in the true path
975 _igvn.hash_delete( post_head );
976 post_head->set_req(LoopNode::EntryControl, zer_taken);
977 set_idom(post_head, zer_taken, dd_main_exit);
979 Arena *a = Thread::current()->resource_area();
980 VectorSet visited(a);
981 Node_Stack clones(a, main_head->back_control()->outcnt());
982 // Step A3: Make the fall-in values to the post-loop come from the
983 // fall-out values of the main-loop.
984 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
985 Node* main_phi = main_head->fast_out(i);
986 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
987 Node *post_phi = old_new[main_phi->_idx];
988 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
989 post_head->init_control(),
990 main_phi->in(LoopNode::LoopBackControl),
991 visited, clones);
992 _igvn.hash_delete(post_phi);
993 post_phi->set_req( LoopNode::EntryControl, fallmain );
994 }
995 }
997 // Update local caches for next stanza
998 main_exit = new_main_exit;
1001 //------------------------------
1002 // Step B: Create Pre-Loop.
1004 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1005 // loop pre-header illegally has 2 control users (old & new loops).
1006 clone_loop( loop, old_new, dd_main_head );
1007 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1008 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1009 pre_head->set_pre_loop(main_head);
1010 Node *pre_incr = old_new[incr->_idx];
1012 // Reduce the pre-loop trip count.
1013 pre_end->_prob = PROB_FAIR;
1015 // Find the pre-loop normal exit.
1016 Node* pre_exit = pre_end->proj_out(false);
1017 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1018 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
1019 _igvn.register_new_node_with_optimizer( new_pre_exit );
1020 set_idom(new_pre_exit, pre_end, dd_main_head);
1021 set_loop(new_pre_exit, loop->_parent);
1023 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1024 // pre-loop, the main-loop may not execute at all. Later in life this
1025 // zero-trip guard will become the minimum-trip guard when we unroll
1026 // the main-loop.
1027 Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
1028 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
1029 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
1030 register_new_node( min_opaq, new_pre_exit );
1031 register_new_node( min_cmp , new_pre_exit );
1032 register_new_node( min_bol , new_pre_exit );
1034 // Build the IfNode (assume the main-loop is executed always).
1035 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1036 _igvn.register_new_node_with_optimizer( min_iff );
1037 set_idom(min_iff, new_pre_exit, dd_main_head);
1038 set_loop(min_iff, loop->_parent);
1040 // Plug in the false-path, taken if we need to skip main-loop
1041 _igvn.hash_delete( pre_exit );
1042 pre_exit->set_req(0, min_iff);
1043 set_idom(pre_exit, min_iff, dd_main_head);
1044 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1045 // Make the true-path, must enter the main loop
1046 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
1047 _igvn.register_new_node_with_optimizer( min_taken );
1048 set_idom(min_taken, min_iff, dd_main_head);
1049 set_loop(min_taken, loop->_parent);
1050 // Plug in the true path
1051 _igvn.hash_delete( main_head );
1052 main_head->set_req(LoopNode::EntryControl, min_taken);
1053 set_idom(main_head, min_taken, dd_main_head);
1055 visited.Clear();
1056 clones.clear();
1057 // Step B3: Make the fall-in values to the main-loop come from the
1058 // fall-out values of the pre-loop.
1059 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1060 Node* main_phi = main_head->fast_out(i2);
1061 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1062 Node *pre_phi = old_new[main_phi->_idx];
1063 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1064 main_head->init_control(),
1065 pre_phi->in(LoopNode::LoopBackControl),
1066 visited, clones);
1067 _igvn.hash_delete(main_phi);
1068 main_phi->set_req( LoopNode::EntryControl, fallpre );
1069 }
1070 }
1072 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1073 // RCE and alignment may change this later.
1074 Node *cmp_end = pre_end->cmp_node();
1075 assert( cmp_end->in(2) == limit, "" );
1076 Node *pre_limit = new (C, 3) AddINode( init, stride );
1078 // Save the original loop limit in this Opaque1 node for
1079 // use by range check elimination.
1080 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
1082 register_new_node( pre_limit, pre_head->in(0) );
1083 register_new_node( pre_opaq , pre_head->in(0) );
1085 // Since no other users of pre-loop compare, I can hack limit directly
1086 assert( cmp_end->outcnt() == 1, "no other users" );
1087 _igvn.hash_delete(cmp_end);
1088 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1090 // Special case for not-equal loop bounds:
1091 // Change pre loop test, main loop test, and the
1092 // main loop guard test to use lt or gt depending on stride
1093 // direction:
1094 // positive stride use <
1095 // negative stride use >
1096 //
1097 // not-equal test is kept for post loop to handle case
1098 // when init > limit when stride > 0 (and reverse).
1100 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1102 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1103 // Modify pre loop end condition
1104 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1105 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
1106 register_new_node( new_bol0, pre_head->in(0) );
1107 _igvn.hash_delete(pre_end);
1108 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
1109 // Modify main loop guard condition
1110 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1111 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
1112 register_new_node( new_bol1, new_pre_exit );
1113 _igvn.hash_delete(min_iff);
1114 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1115 // Modify main loop end condition
1116 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1117 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
1118 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1119 _igvn.hash_delete(main_end);
1120 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
1121 }
1123 // Flag main loop
1124 main_head->set_main_loop();
1125 if( peel_only ) main_head->set_main_no_pre_loop();
1127 // Subtract a trip count for the pre-loop.
1128 main_head->set_trip_count(main_head->trip_count() - 1);
1130 // It's difficult to be precise about the trip-counts
1131 // for the pre/post loops. They are usually very short,
1132 // so guess that 4 trips is a reasonable value.
1133 post_head->set_profile_trip_cnt(4.0);
1134 pre_head->set_profile_trip_cnt(4.0);
1136 // Now force out all loop-invariant dominating tests. The optimizer
1137 // finds some, but we _know_ they are all useless.
1138 peeled_dom_test_elim(loop,old_new);
1139 }
1141 //------------------------------is_invariant-----------------------------
1142 // Return true if n is invariant
1143 bool IdealLoopTree::is_invariant(Node* n) const {
1144 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1145 if (n_c->is_top()) return false;
1146 return !is_member(_phase->get_loop(n_c));
1147 }
1150 //------------------------------do_unroll--------------------------------------
1151 // Unroll the loop body one step - make each trip do 2 iterations.
1152 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1153 assert(LoopUnrollLimit, "");
1154 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1155 CountedLoopEndNode *loop_end = loop_head->loopexit();
1156 assert(loop_end, "");
1157 #ifndef PRODUCT
1158 if (PrintOpto && VerifyLoopOptimizations) {
1159 tty->print("Unrolling ");
1160 loop->dump_head();
1161 } else if (TraceLoopOpts) {
1162 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1163 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1164 } else {
1165 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1166 }
1167 loop->dump_head();
1168 }
1169 #endif
1171 // Remember loop node count before unrolling to detect
1172 // if rounds of unroll,optimize are making progress
1173 loop_head->set_node_count_before_unroll(loop->_body.size());
1175 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1176 Node *limit = loop_head->limit();
1177 Node *init = loop_head->init_trip();
1178 Node *stride = loop_head->stride();
1180 Node *opaq = NULL;
1181 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1182 // Search for zero-trip guard.
1183 assert( loop_head->is_main_loop(), "" );
1184 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1185 Node *iff = ctrl->in(0);
1186 assert( iff->Opcode() == Op_If, "" );
1187 Node *bol = iff->in(1);
1188 assert( bol->Opcode() == Op_Bool, "" );
1189 Node *cmp = bol->in(1);
1190 assert( cmp->Opcode() == Op_CmpI, "" );
1191 opaq = cmp->in(2);
1192 // Occasionally it's possible for a zero-trip guard Opaque1 node to be
1193 // optimized away and then another round of loop opts attempted.
1194 // We can not optimize this particular loop in that case.
1195 if (opaq->Opcode() != Op_Opaque1)
1196 return; // Cannot find zero-trip guard! Bail out!
1197 // Zero-trip test uses an 'opaque' node which is not shared.
1198 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1199 }
1201 C->set_major_progress();
1203 Node* new_limit = NULL;
1204 if (UnrollLimitCheck) {
1205 int stride_con = stride->get_int();
1206 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1207 uint old_trip_count = loop_head->trip_count();
1208 // Verify that unroll policy result is still valid.
1209 assert(old_trip_count > 1 &&
1210 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1212 // Adjust loop limit to keep valid iterations number after unroll.
1213 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1214 // which may overflow.
1215 if (!adjust_min_trip) {
1216 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1217 "odd trip count for maximally unroll");
1218 // Don't need to adjust limit for maximally unroll since trip count is even.
1219 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1220 // Loop's limit is constant. Loop's init could be constant when pre-loop
1221 // become peeled iteration.
1222 long init_con = init->get_int();
1223 // We can keep old loop limit if iterations count stays the same:
1224 // old_trip_count == new_trip_count * 2
1225 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1226 // so we also don't need to adjust zero trip test.
1227 long limit_con = limit->get_int();
1228 // (stride_con*2) not overflow since stride_con <= 8.
1229 int new_stride_con = stride_con * 2;
1230 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1231 long trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1232 // New trip count should satisfy next conditions.
1233 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1234 uint new_trip_count = (uint)trip_count;
1235 adjust_min_trip = (old_trip_count != new_trip_count*2);
1236 }
1238 if (adjust_min_trip) {
1239 // Step 2: Adjust the trip limit if it is called for.
1240 // The adjustment amount is -stride. Need to make sure if the
1241 // adjustment underflows or overflows, then the main loop is skipped.
1242 Node* cmp = loop_end->cmp_node();
1243 assert(cmp->in(2) == limit, "sanity");
1244 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1246 // Verify that policy_unroll result is still valid.
1247 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1248 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1249 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1251 if (limit->is_Con()) {
1252 // The check in policy_unroll and the assert above guarantee
1253 // no underflow if limit is constant.
1254 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1255 set_ctrl(new_limit, C->root());
1256 } else {
1257 // Limit is not constant.
1258 if (loop_head->unrolled_count() == 1) { // only for first unroll
1259 // Separate limit by Opaque node in case it is an incremented
1260 // variable from previous loop to avoid using pre-incremented
1261 // value which could increase register pressure.
1262 // Otherwise reorg_offsets() optimization will create a separate
1263 // Opaque node for each use of trip-counter and as result
1264 // zero trip guard limit will be different from loop limit.
1265 assert(has_ctrl(opaq), "should have it");
1266 Node* opaq_ctrl = get_ctrl(opaq);
1267 limit = new (C, 2) Opaque2Node( C, limit );
1268 register_new_node( limit, opaq_ctrl );
1269 }
1270 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
1271 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
1272 // No underflow.
1273 new_limit = new (C, 3) SubINode(limit, stride);
1274 } else {
1275 // (limit - stride) may underflow.
1276 // Clamp the adjustment value with MININT or MAXINT:
1277 //
1278 // new_limit = limit-stride
1279 // if (stride > 0)
1280 // new_limit = (limit < new_limit) ? MININT : new_limit;
1281 // else
1282 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1283 //
1284 BoolTest::mask bt = loop_end->test_trip();
1285 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1286 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1287 set_ctrl(adj_max, C->root());
1288 Node* old_limit = NULL;
1289 Node* adj_limit = NULL;
1290 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1291 if (loop_head->unrolled_count() > 1 &&
1292 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1293 limit->in(CMoveNode::IfTrue) == adj_max &&
1294 bol->as_Bool()->_test._test == bt &&
1295 bol->in(1)->Opcode() == Op_CmpI &&
1296 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1297 // Loop was unrolled before.
1298 // Optimize the limit to avoid nested CMove:
1299 // use original limit as old limit.
1300 old_limit = bol->in(1)->in(1);
1301 // Adjust previous adjusted limit.
1302 adj_limit = limit->in(CMoveNode::IfFalse);
1303 adj_limit = new (C, 3) SubINode(adj_limit, stride);
1304 } else {
1305 old_limit = limit;
1306 adj_limit = new (C, 3) SubINode(limit, stride);
1307 }
1308 assert(old_limit != NULL && adj_limit != NULL, "");
1309 register_new_node( adj_limit, ctrl ); // adjust amount
1310 Node* adj_cmp = new (C, 3) CmpINode(old_limit, adj_limit);
1311 register_new_node( adj_cmp, ctrl );
1312 Node* adj_bool = new (C, 2) BoolNode(adj_cmp, bt);
1313 register_new_node( adj_bool, ctrl );
1314 new_limit = new (C, 4) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1315 }
1316 register_new_node(new_limit, ctrl);
1317 }
1318 assert(new_limit != NULL, "");
1319 // Replace in loop test.
1320 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1321 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1322 // Don't need to create new test since only one user.
1323 _igvn.hash_delete(cmp);
1324 cmp->set_req(2, new_limit);
1325 } else {
1326 // Create new test since it is shared.
1327 Node* ctrl2 = loop_end->in(0);
1328 Node* cmp2 = cmp->clone();
1329 cmp2->set_req(2, new_limit);
1330 register_new_node(cmp2, ctrl2);
1331 Node* bol2 = loop_end->in(1)->clone();
1332 bol2->set_req(1, cmp2);
1333 register_new_node(bol2, ctrl2);
1334 _igvn.hash_delete(loop_end);
1335 loop_end->set_req(1, bol2);
1336 }
1337 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1338 // Make it a 1-trip test (means at least 2 trips).
1340 // Guard test uses an 'opaque' node which is not shared. Hence I
1341 // can edit it's inputs directly. Hammer in the new limit for the
1342 // minimum-trip guard.
1343 assert(opaq->outcnt() == 1, "");
1344 _igvn.hash_delete(opaq);
1345 opaq->set_req(1, new_limit);
1346 }
1348 // Adjust max trip count. The trip count is intentionally rounded
1349 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1350 // the main, unrolled, part of the loop will never execute as it is protected
1351 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1352 // and later determined that part of the unrolled loop was dead.
1353 loop_head->set_trip_count(old_trip_count / 2);
1355 // Double the count of original iterations in the unrolled loop body.
1356 loop_head->double_unrolled_count();
1358 } else { // LoopLimitCheck
1360 // Adjust max trip count. The trip count is intentionally rounded
1361 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1362 // the main, unrolled, part of the loop will never execute as it is protected
1363 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1364 // and later determined that part of the unrolled loop was dead.
1365 loop_head->set_trip_count(loop_head->trip_count() / 2);
1367 // Double the count of original iterations in the unrolled loop body.
1368 loop_head->double_unrolled_count();
1370 // -----------
1371 // Step 2: Cut back the trip counter for an unroll amount of 2.
1372 // Loop will normally trip (limit - init)/stride_con. Since it's a
1373 // CountedLoop this is exact (stride divides limit-init exactly).
1374 // We are going to double the loop body, so we want to knock off any
1375 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1376 Node *span = new (C, 3) SubINode( limit, init );
1377 register_new_node( span, ctrl );
1378 Node *trip = new (C, 3) DivINode( 0, span, stride );
1379 register_new_node( trip, ctrl );
1380 Node *mtwo = _igvn.intcon(-2);
1381 set_ctrl(mtwo, C->root());
1382 Node *rond = new (C, 3) AndINode( trip, mtwo );
1383 register_new_node( rond, ctrl );
1384 Node *spn2 = new (C, 3) MulINode( rond, stride );
1385 register_new_node( spn2, ctrl );
1386 new_limit = new (C, 3) AddINode( spn2, init );
1387 register_new_node( new_limit, ctrl );
1389 // Hammer in the new limit
1390 Node *ctrl2 = loop_end->in(0);
1391 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), new_limit );
1392 register_new_node( cmp2, ctrl2 );
1393 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
1394 register_new_node( bol2, ctrl2 );
1395 _igvn.hash_delete(loop_end);
1396 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
1398 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1399 // Make it a 1-trip test (means at least 2 trips).
1400 if( adjust_min_trip ) {
1401 assert( new_limit != NULL, "" );
1402 // Guard test uses an 'opaque' node which is not shared. Hence I
1403 // can edit it's inputs directly. Hammer in the new limit for the
1404 // minimum-trip guard.
1405 assert( opaq->outcnt() == 1, "" );
1406 _igvn.hash_delete(opaq);
1407 opaq->set_req(1, new_limit);
1408 }
1409 } // LoopLimitCheck
1411 // ---------
1412 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1413 // represents the odd iterations; since the loop trips an even number of
1414 // times its backedge is never taken. Kill the backedge.
1415 uint dd = dom_depth(loop_head);
1416 clone_loop( loop, old_new, dd );
1418 // Make backedges of the clone equal to backedges of the original.
1419 // Make the fall-in from the original come from the fall-out of the clone.
1420 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1421 Node* phi = loop_head->fast_out(j);
1422 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1423 Node *newphi = old_new[phi->_idx];
1424 _igvn.hash_delete( phi );
1425 _igvn.hash_delete( newphi );
1427 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1428 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1429 phi ->set_req(LoopNode::LoopBackControl, C->top());
1430 }
1431 }
1432 Node *clone_head = old_new[loop_head->_idx];
1433 _igvn.hash_delete( clone_head );
1434 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1435 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1436 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1437 loop->_head = clone_head; // New loop header
1439 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1440 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1442 // Kill the clone's backedge
1443 Node *newcle = old_new[loop_end->_idx];
1444 _igvn.hash_delete( newcle );
1445 Node *one = _igvn.intcon(1);
1446 set_ctrl(one, C->root());
1447 newcle->set_req(1, one);
1448 // Force clone into same loop body
1449 uint max = loop->_body.size();
1450 for( uint k = 0; k < max; k++ ) {
1451 Node *old = loop->_body.at(k);
1452 Node *nnn = old_new[old->_idx];
1453 loop->_body.push(nnn);
1454 if (!has_ctrl(old))
1455 set_loop(nnn, loop);
1456 }
1458 loop->record_for_igvn();
1459 }
1461 //------------------------------do_maximally_unroll----------------------------
1463 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1464 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1465 assert(cl->has_exact_trip_count(), "trip count is not exact");
1466 assert(cl->trip_count() > 0, "");
1467 #ifndef PRODUCT
1468 if (TraceLoopOpts) {
1469 tty->print("MaxUnroll %d ", cl->trip_count());
1470 loop->dump_head();
1471 }
1472 #endif
1474 // If loop is tripping an odd number of times, peel odd iteration
1475 if ((cl->trip_count() & 1) == 1) {
1476 do_peeling(loop, old_new);
1477 }
1479 // Now its tripping an even number of times remaining. Double loop body.
1480 // Do not adjust pre-guards; they are not needed and do not exist.
1481 if (cl->trip_count() > 0) {
1482 assert((cl->trip_count() & 1) == 0, "missed peeling");
1483 do_unroll(loop, old_new, false);
1484 }
1485 }
1487 //------------------------------dominates_backedge---------------------------------
1488 // Returns true if ctrl is executed on every complete iteration
1489 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1490 assert(ctrl->is_CFG(), "must be control");
1491 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1492 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1493 }
1495 //------------------------------adjust_limit-----------------------------------
1496 // Helper function for add_constraint().
1497 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1498 // Compute "I :: (limit-offset)/scale"
1499 Node *con = new (C, 3) SubINode(rc_limit, offset);
1500 register_new_node(con, pre_ctrl);
1501 Node *X = new (C, 3) DivINode(0, con, scale);
1502 register_new_node(X, pre_ctrl);
1504 // Adjust loop limit
1505 loop_limit = (stride_con > 0)
1506 ? (Node*)(new (C, 3) MinINode(loop_limit, X))
1507 : (Node*)(new (C, 3) MaxINode(loop_limit, X));
1508 register_new_node(loop_limit, pre_ctrl);
1509 return loop_limit;
1510 }
1512 //------------------------------add_constraint---------------------------------
1513 // Constrain the main loop iterations so the conditions:
1514 // low_limit <= scale_con * I + offset < upper_limit
1515 // always holds true. That is, either increase the number of iterations in
1516 // the pre-loop or the post-loop until the condition holds true in the main
1517 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1518 // stride and scale are constants (offset and limit often are).
1519 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
1520 // For positive stride, the pre-loop limit always uses a MAX function
1521 // and the main loop a MIN function. For negative stride these are
1522 // reversed.
1524 // Also for positive stride*scale the affine function is increasing, so the
1525 // pre-loop must check for underflow and the post-loop for overflow.
1526 // Negative stride*scale reverses this; pre-loop checks for overflow and
1527 // post-loop for underflow.
1529 Node *scale = _igvn.intcon(scale_con);
1530 set_ctrl(scale, C->root());
1532 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1533 // The overflow limit: scale*I+offset < upper_limit
1534 // For main-loop compute
1535 // ( if (scale > 0) /* and stride > 0 */
1536 // I < (upper_limit-offset)/scale
1537 // else /* scale < 0 and stride < 0 */
1538 // I > (upper_limit-offset)/scale
1539 // )
1540 //
1541 // (upper_limit-offset) may overflow or underflow.
1542 // But it is fine since main loop will either have
1543 // less iterations or will be skipped in such case.
1544 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1546 // The underflow limit: low_limit <= scale*I+offset.
1547 // For pre-loop compute
1548 // NOT(scale*I+offset >= low_limit)
1549 // scale*I+offset < low_limit
1550 // ( if (scale > 0) /* and stride > 0 */
1551 // I < (low_limit-offset)/scale
1552 // else /* scale < 0 and stride < 0 */
1553 // I > (low_limit-offset)/scale
1554 // )
1556 if (low_limit->get_int() == -max_jint) {
1557 if (!RangeLimitCheck) return;
1558 // We need this guard when scale*pre_limit+offset >= limit
1559 // due to underflow. So we need execute pre-loop until
1560 // scale*I+offset >= min_int. But (min_int-offset) will
1561 // underflow when offset > 0 and X will be > original_limit
1562 // when stride > 0. To avoid it we replace positive offset with 0.
1563 //
1564 // Also (min_int+1 == -max_int) is used instead of min_int here
1565 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1566 Node* shift = _igvn.intcon(31);
1567 set_ctrl(shift, C->root());
1568 Node* sign = new (C, 3) RShiftINode(offset, shift);
1569 register_new_node(sign, pre_ctrl);
1570 offset = new (C, 3) AndINode(offset, sign);
1571 register_new_node(offset, pre_ctrl);
1572 } else {
1573 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1574 // The only problem we have here when offset == min_int
1575 // since (0-min_int) == min_int. It may be fine for stride > 0
1576 // but for stride < 0 X will be < original_limit. To avoid it
1577 // max(pre_limit, original_limit) is used in do_range_check().
1578 }
1579 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1580 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1582 } else { // stride_con*scale_con < 0
1583 // For negative stride*scale pre-loop checks for overflow and
1584 // post-loop for underflow.
1585 //
1586 // The overflow limit: scale*I+offset < upper_limit
1587 // For pre-loop compute
1588 // NOT(scale*I+offset < upper_limit)
1589 // scale*I+offset >= upper_limit
1590 // scale*I+offset+1 > upper_limit
1591 // ( if (scale < 0) /* and stride > 0 */
1592 // I < (upper_limit-(offset+1))/scale
1593 // else /* scale > 0 and stride < 0 */
1594 // I > (upper_limit-(offset+1))/scale
1595 // )
1596 //
1597 // (upper_limit-offset-1) may underflow or overflow.
1598 // To avoid it min(pre_limit, original_limit) is used
1599 // in do_range_check() for stride > 0 and max() for < 0.
1600 Node *one = _igvn.intcon(1);
1601 set_ctrl(one, C->root());
1603 Node *plus_one = new (C, 3) AddINode(offset, one);
1604 register_new_node( plus_one, pre_ctrl );
1605 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1606 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1608 if (low_limit->get_int() == -max_jint) {
1609 if (!RangeLimitCheck) return;
1610 // We need this guard when scale*main_limit+offset >= limit
1611 // due to underflow. So we need execute main-loop while
1612 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1613 // underflow when (offset+1) > 0 and X will be < main_limit
1614 // when scale < 0 (and stride > 0). To avoid it we replace
1615 // positive (offset+1) with 0.
1616 //
1617 // Also (min_int+1 == -max_int) is used instead of min_int here
1618 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1619 Node* shift = _igvn.intcon(31);
1620 set_ctrl(shift, C->root());
1621 Node* sign = new (C, 3) RShiftINode(plus_one, shift);
1622 register_new_node(sign, pre_ctrl);
1623 plus_one = new (C, 3) AndINode(plus_one, sign);
1624 register_new_node(plus_one, pre_ctrl);
1625 } else {
1626 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1627 // The only problem we have here when offset == max_int
1628 // since (max_int+1) == min_int and (0-min_int) == min_int.
1629 // But it is fine since main loop will either have
1630 // less iterations or will be skipped in such case.
1631 }
1632 // The underflow limit: low_limit <= scale*I+offset.
1633 // For main-loop compute
1634 // scale*I+offset+1 > low_limit
1635 // ( if (scale < 0) /* and stride > 0 */
1636 // I < (low_limit-(offset+1))/scale
1637 // else /* scale > 0 and stride < 0 */
1638 // I > (low_limit-(offset+1))/scale
1639 // )
1641 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1642 }
1643 }
1646 //------------------------------is_scaled_iv---------------------------------
1647 // Return true if exp is a constant times an induction var
1648 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1649 if (exp == iv) {
1650 if (p_scale != NULL) {
1651 *p_scale = 1;
1652 }
1653 return true;
1654 }
1655 int opc = exp->Opcode();
1656 if (opc == Op_MulI) {
1657 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1658 if (p_scale != NULL) {
1659 *p_scale = exp->in(2)->get_int();
1660 }
1661 return true;
1662 }
1663 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1664 if (p_scale != NULL) {
1665 *p_scale = exp->in(1)->get_int();
1666 }
1667 return true;
1668 }
1669 } else if (opc == Op_LShiftI) {
1670 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1671 if (p_scale != NULL) {
1672 *p_scale = 1 << exp->in(2)->get_int();
1673 }
1674 return true;
1675 }
1676 }
1677 return false;
1678 }
1680 //-----------------------------is_scaled_iv_plus_offset------------------------------
1681 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1682 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1683 if (is_scaled_iv(exp, iv, p_scale)) {
1684 if (p_offset != NULL) {
1685 Node *zero = _igvn.intcon(0);
1686 set_ctrl(zero, C->root());
1687 *p_offset = zero;
1688 }
1689 return true;
1690 }
1691 int opc = exp->Opcode();
1692 if (opc == Op_AddI) {
1693 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1694 if (p_offset != NULL) {
1695 *p_offset = exp->in(2);
1696 }
1697 return true;
1698 }
1699 if (exp->in(2)->is_Con()) {
1700 Node* offset2 = NULL;
1701 if (depth < 2 &&
1702 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1703 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1704 if (p_offset != NULL) {
1705 Node *ctrl_off2 = get_ctrl(offset2);
1706 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
1707 register_new_node(offset, ctrl_off2);
1708 *p_offset = offset;
1709 }
1710 return true;
1711 }
1712 }
1713 } else if (opc == Op_SubI) {
1714 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1715 if (p_offset != NULL) {
1716 Node *zero = _igvn.intcon(0);
1717 set_ctrl(zero, C->root());
1718 Node *ctrl_off = get_ctrl(exp->in(2));
1719 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
1720 register_new_node(offset, ctrl_off);
1721 *p_offset = offset;
1722 }
1723 return true;
1724 }
1725 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1726 if (p_offset != NULL) {
1727 *p_scale *= -1;
1728 *p_offset = exp->in(1);
1729 }
1730 return true;
1731 }
1732 }
1733 return false;
1734 }
1736 //------------------------------do_range_check---------------------------------
1737 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1738 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1739 #ifndef PRODUCT
1740 if (PrintOpto && VerifyLoopOptimizations) {
1741 tty->print("Range Check Elimination ");
1742 loop->dump_head();
1743 } else if (TraceLoopOpts) {
1744 tty->print("RangeCheck ");
1745 loop->dump_head();
1746 }
1747 #endif
1748 assert(RangeCheckElimination, "");
1749 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1750 assert(cl->is_main_loop(), "");
1752 // protect against stride not being a constant
1753 if (!cl->stride_is_con())
1754 return;
1756 // Find the trip counter; we are iteration splitting based on it
1757 Node *trip_counter = cl->phi();
1758 // Find the main loop limit; we will trim it's iterations
1759 // to not ever trip end tests
1760 Node *main_limit = cl->limit();
1762 // Need to find the main-loop zero-trip guard
1763 Node *ctrl = cl->in(LoopNode::EntryControl);
1764 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1765 Node *iffm = ctrl->in(0);
1766 assert(iffm->Opcode() == Op_If, "");
1767 Node *bolzm = iffm->in(1);
1768 assert(bolzm->Opcode() == Op_Bool, "");
1769 Node *cmpzm = bolzm->in(1);
1770 assert(cmpzm->is_Cmp(), "");
1771 Node *opqzm = cmpzm->in(2);
1772 // Can not optimize a loop if zero-trip Opaque1 node is optimized
1773 // away and then another round of loop opts attempted.
1774 if (opqzm->Opcode() != Op_Opaque1)
1775 return;
1776 assert(opqzm->in(1) == main_limit, "do not understand situation");
1778 // Find the pre-loop limit; we will expand it's iterations to
1779 // not ever trip low tests.
1780 Node *p_f = iffm->in(0);
1781 assert(p_f->Opcode() == Op_IfFalse, "");
1782 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1783 assert(pre_end->loopnode()->is_pre_loop(), "");
1784 Node *pre_opaq1 = pre_end->limit();
1785 // Occasionally it's possible for a pre-loop Opaque1 node to be
1786 // optimized away and then another round of loop opts attempted.
1787 // We can not optimize this particular loop in that case.
1788 if (pre_opaq1->Opcode() != Op_Opaque1)
1789 return;
1790 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1791 Node *pre_limit = pre_opaq->in(1);
1793 // Where do we put new limit calculations
1794 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1796 // Ensure the original loop limit is available from the
1797 // pre-loop Opaque1 node.
1798 Node *orig_limit = pre_opaq->original_loop_limit();
1799 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1800 return;
1802 // Must know if its a count-up or count-down loop
1804 int stride_con = cl->stride_con();
1805 Node *zero = _igvn.intcon(0);
1806 Node *one = _igvn.intcon(1);
1807 // Use symmetrical int range [-max_jint,max_jint]
1808 Node *mini = _igvn.intcon(-max_jint);
1809 set_ctrl(zero, C->root());
1810 set_ctrl(one, C->root());
1811 set_ctrl(mini, C->root());
1813 // Range checks that do not dominate the loop backedge (ie.
1814 // conditionally executed) can lengthen the pre loop limit beyond
1815 // the original loop limit. To prevent this, the pre limit is
1816 // (for stride > 0) MINed with the original loop limit (MAXed
1817 // stride < 0) when some range_check (rc) is conditionally
1818 // executed.
1819 bool conditional_rc = false;
1821 // Check loop body for tests of trip-counter plus loop-invariant vs
1822 // loop-invariant.
1823 for( uint i = 0; i < loop->_body.size(); i++ ) {
1824 Node *iff = loop->_body[i];
1825 if( iff->Opcode() == Op_If ) { // Test?
1827 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1828 // we need loop unswitching instead of iteration splitting.
1829 Node *exit = loop->is_loop_exit(iff);
1830 if( !exit ) continue;
1831 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1833 // Get boolean condition to test
1834 Node *i1 = iff->in(1);
1835 if( !i1->is_Bool() ) continue;
1836 BoolNode *bol = i1->as_Bool();
1837 BoolTest b_test = bol->_test;
1838 // Flip sense of test if exit condition is flipped
1839 if( flip )
1840 b_test = b_test.negate();
1842 // Get compare
1843 Node *cmp = bol->in(1);
1845 // Look for trip_counter + offset vs limit
1846 Node *rc_exp = cmp->in(1);
1847 Node *limit = cmp->in(2);
1848 jint scale_con= 1; // Assume trip counter not scaled
1850 Node *limit_c = get_ctrl(limit);
1851 if( loop->is_member(get_loop(limit_c) ) ) {
1852 // Compare might have operands swapped; commute them
1853 b_test = b_test.commute();
1854 rc_exp = cmp->in(2);
1855 limit = cmp->in(1);
1856 limit_c = get_ctrl(limit);
1857 if( loop->is_member(get_loop(limit_c) ) )
1858 continue; // Both inputs are loop varying; cannot RCE
1859 }
1860 // Here we know 'limit' is loop invariant
1862 // 'limit' maybe pinned below the zero trip test (probably from a
1863 // previous round of rce), in which case, it can't be used in the
1864 // zero trip test expression which must occur before the zero test's if.
1865 if( limit_c == ctrl ) {
1866 continue; // Don't rce this check but continue looking for other candidates.
1867 }
1869 // Check for scaled induction variable plus an offset
1870 Node *offset = NULL;
1872 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1873 continue;
1874 }
1876 Node *offset_c = get_ctrl(offset);
1877 if( loop->is_member( get_loop(offset_c) ) )
1878 continue; // Offset is not really loop invariant
1879 // Here we know 'offset' is loop invariant.
1881 // As above for the 'limit', the 'offset' maybe pinned below the
1882 // zero trip test.
1883 if( offset_c == ctrl ) {
1884 continue; // Don't rce this check but continue looking for other candidates.
1885 }
1886 #ifdef ASSERT
1887 if (TraceRangeLimitCheck) {
1888 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
1889 bol->dump(2);
1890 }
1891 #endif
1892 // At this point we have the expression as:
1893 // scale_con * trip_counter + offset :: limit
1894 // where scale_con, offset and limit are loop invariant. Trip_counter
1895 // monotonically increases by stride_con, a constant. Both (or either)
1896 // stride_con and scale_con can be negative which will flip about the
1897 // sense of the test.
1899 // Adjust pre and main loop limits to guard the correct iteration set
1900 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1901 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1902 // The underflow and overflow limits: 0 <= scale*I+offset < limit
1903 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
1904 if (!conditional_rc) {
1905 // (0-offset)/scale could be outside of loop iterations range.
1906 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1907 }
1908 } else {
1909 #ifndef PRODUCT
1910 if( PrintOpto )
1911 tty->print_cr("missed RCE opportunity");
1912 #endif
1913 continue; // In release mode, ignore it
1914 }
1915 } else { // Otherwise work on normal compares
1916 switch( b_test._test ) {
1917 case BoolTest::gt:
1918 // Fall into GE case
1919 case BoolTest::ge:
1920 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
1921 scale_con = -scale_con;
1922 offset = new (C, 3) SubINode( zero, offset );
1923 register_new_node( offset, pre_ctrl );
1924 limit = new (C, 3) SubINode( zero, limit );
1925 register_new_node( limit, pre_ctrl );
1926 // Fall into LE case
1927 case BoolTest::le:
1928 if (b_test._test != BoolTest::gt) {
1929 // Convert X <= Y to X < Y+1
1930 limit = new (C, 3) AddINode( limit, one );
1931 register_new_node( limit, pre_ctrl );
1932 }
1933 // Fall into LT case
1934 case BoolTest::lt:
1935 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
1936 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
1937 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
1938 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
1939 if (!conditional_rc) {
1940 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
1941 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
1942 // still be outside of loop range.
1943 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1944 }
1945 break;
1946 default:
1947 #ifndef PRODUCT
1948 if( PrintOpto )
1949 tty->print_cr("missed RCE opportunity");
1950 #endif
1951 continue; // Unhandled case
1952 }
1953 }
1955 // Kill the eliminated test
1956 C->set_major_progress();
1957 Node *kill_con = _igvn.intcon( 1-flip );
1958 set_ctrl(kill_con, C->root());
1959 _igvn.hash_delete(iff);
1960 iff->set_req(1, kill_con);
1961 _igvn._worklist.push(iff);
1962 // Find surviving projection
1963 assert(iff->is_If(), "");
1964 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1965 // Find loads off the surviving projection; remove their control edge
1966 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1967 Node* cd = dp->fast_out(i); // Control-dependent node
1968 if( cd->is_Load() ) { // Loads can now float around in the loop
1969 _igvn.hash_delete(cd);
1970 // Allow the load to float around in the loop, or before it
1971 // but NOT before the pre-loop.
1972 cd->set_req(0, ctrl); // ctrl, not NULL
1973 _igvn._worklist.push(cd);
1974 --i;
1975 --imax;
1976 }
1977 }
1979 } // End of is IF
1981 }
1983 // Update loop limits
1984 if (conditional_rc) {
1985 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
1986 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
1987 register_new_node(pre_limit, pre_ctrl);
1988 }
1989 _igvn.hash_delete(pre_opaq);
1990 pre_opaq->set_req(1, pre_limit);
1992 // Note:: we are making the main loop limit no longer precise;
1993 // need to round up based on stride.
1994 cl->set_nonexact_trip_count();
1995 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
1996 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1997 // Hopefully, compiler will optimize for powers of 2.
1998 Node *ctrl = get_ctrl(main_limit);
1999 Node *stride = cl->stride();
2000 Node *init = cl->init_trip();
2001 Node *span = new (C, 3) SubINode(main_limit,init);
2002 register_new_node(span,ctrl);
2003 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
2004 Node *add = new (C, 3) AddINode(span,rndup);
2005 register_new_node(add,ctrl);
2006 Node *div = new (C, 3) DivINode(0,add,stride);
2007 register_new_node(div,ctrl);
2008 Node *mul = new (C, 3) MulINode(div,stride);
2009 register_new_node(mul,ctrl);
2010 Node *newlim = new (C, 3) AddINode(mul,init);
2011 register_new_node(newlim,ctrl);
2012 main_limit = newlim;
2013 }
2015 Node *main_cle = cl->loopexit();
2016 Node *main_bol = main_cle->in(1);
2017 // Hacking loop bounds; need private copies of exit test
2018 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2019 _igvn.hash_delete(main_cle);
2020 main_bol = main_bol->clone();// Clone a private BoolNode
2021 register_new_node( main_bol, main_cle->in(0) );
2022 main_cle->set_req(1,main_bol);
2023 }
2024 Node *main_cmp = main_bol->in(1);
2025 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2026 _igvn.hash_delete(main_bol);
2027 main_cmp = main_cmp->clone();// Clone a private CmpNode
2028 register_new_node( main_cmp, main_cle->in(0) );
2029 main_bol->set_req(1,main_cmp);
2030 }
2031 // Hack the now-private loop bounds
2032 _igvn.hash_delete(main_cmp);
2033 main_cmp->set_req(2, main_limit);
2034 _igvn._worklist.push(main_cmp);
2035 // The OpaqueNode is unshared by design
2036 _igvn.hash_delete(opqzm);
2037 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2038 opqzm->set_req(1,main_limit);
2039 _igvn._worklist.push(opqzm);
2040 }
2042 //------------------------------DCE_loop_body----------------------------------
2043 // Remove simplistic dead code from loop body
2044 void IdealLoopTree::DCE_loop_body() {
2045 for( uint i = 0; i < _body.size(); i++ )
2046 if( _body.at(i)->outcnt() == 0 )
2047 _body.map( i--, _body.pop() );
2048 }
2051 //------------------------------adjust_loop_exit_prob--------------------------
2052 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2053 // Replace with a 1-in-10 exit guess.
2054 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2055 Node *test = tail();
2056 while( test != _head ) {
2057 uint top = test->Opcode();
2058 if( top == Op_IfTrue || top == Op_IfFalse ) {
2059 int test_con = ((ProjNode*)test)->_con;
2060 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2061 IfNode *iff = test->in(0)->as_If();
2062 if( iff->outcnt() == 2 ) { // Ignore dead tests
2063 Node *bol = iff->in(1);
2064 if( bol && bol->req() > 1 && bol->in(1) &&
2065 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2066 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2067 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2068 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2069 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2070 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2071 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2072 return; // Allocation loops RARELY take backedge
2073 // Find the OTHER exit path from the IF
2074 Node* ex = iff->proj_out(1-test_con);
2075 float p = iff->_prob;
2076 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2077 if( top == Op_IfTrue ) {
2078 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2079 iff->_prob = PROB_STATIC_FREQUENT;
2080 }
2081 } else {
2082 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2083 iff->_prob = PROB_STATIC_INFREQUENT;
2084 }
2085 }
2086 }
2087 }
2088 }
2089 test = phase->idom(test);
2090 }
2091 }
2094 //------------------------------policy_do_remove_empty_loop--------------------
2095 // Micro-benchmark spamming. Policy is to always remove empty loops.
2096 // The 'DO' part is to replace the trip counter with the value it will
2097 // have on the last iteration. This will break the loop.
2098 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2099 // Minimum size must be empty loop
2100 if (_body.size() > EMPTY_LOOP_SIZE)
2101 return false;
2103 if (!_head->is_CountedLoop())
2104 return false; // Dead loop
2105 CountedLoopNode *cl = _head->as_CountedLoop();
2106 if (!cl->is_valid_counted_loop())
2107 return false; // Malformed loop
2108 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2109 return false; // Infinite loop
2111 #ifdef ASSERT
2112 // Ensure only one phi which is the iv.
2113 Node* iv = NULL;
2114 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2115 Node* n = cl->fast_out(i);
2116 if (n->Opcode() == Op_Phi) {
2117 assert(iv == NULL, "Too many phis" );
2118 iv = n;
2119 }
2120 }
2121 assert(iv == cl->phi(), "Wrong phi" );
2122 #endif
2124 // main and post loops have explicitly created zero trip guard
2125 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2126 if (needs_guard) {
2127 // Skip guard if values not overlap.
2128 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2129 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2130 int stride_con = cl->stride_con();
2131 if (stride_con > 0) {
2132 needs_guard = (init_t->_hi >= limit_t->_lo);
2133 } else {
2134 needs_guard = (init_t->_lo <= limit_t->_hi);
2135 }
2136 }
2137 if (needs_guard) {
2138 // Check for an obvious zero trip guard.
2139 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2140 if (inctrl->Opcode() == Op_IfTrue) {
2141 // The test should look like just the backedge of a CountedLoop
2142 Node* iff = inctrl->in(0);
2143 if (iff->is_If()) {
2144 Node* bol = iff->in(1);
2145 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2146 Node* cmp = bol->in(1);
2147 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2148 needs_guard = false;
2149 }
2150 }
2151 }
2152 }
2153 }
2155 #ifndef PRODUCT
2156 if (PrintOpto) {
2157 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2158 this->dump_head();
2159 } else if (TraceLoopOpts) {
2160 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2161 this->dump_head();
2162 }
2163 #endif
2165 if (needs_guard) {
2166 // Peel the loop to ensure there's a zero trip guard
2167 Node_List old_new;
2168 phase->do_peeling(this, old_new);
2169 }
2171 // Replace the phi at loop head with the final value of the last
2172 // iteration. Then the CountedLoopEnd will collapse (backedge never
2173 // taken) and all loop-invariant uses of the exit values will be correct.
2174 Node *phi = cl->phi();
2175 Node *exact_limit = phase->exact_limit(this);
2176 if (exact_limit != cl->limit()) {
2177 // We also need to replace the original limit to collapse loop exit.
2178 Node* cmp = cl->loopexit()->cmp_node();
2179 assert(cl->limit() == cmp->in(2), "sanity");
2180 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2181 phase->_igvn.hash_delete(cmp);
2182 cmp->set_req(2, exact_limit);
2183 phase->_igvn._worklist.push(cmp); // put cmp on worklist
2184 }
2185 // Note: the final value after increment should not overflow since
2186 // counted loop has limit check predicate.
2187 Node *final = new (phase->C, 3) SubINode( exact_limit, cl->stride() );
2188 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2189 phase->_igvn.replace_node(phi,final);
2190 phase->C->set_major_progress();
2191 return true;
2192 }
2194 //------------------------------policy_do_one_iteration_loop-------------------
2195 // Convert one iteration loop into normal code.
2196 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2197 if (!_head->as_Loop()->is_valid_counted_loop())
2198 return false; // Only for counted loop
2200 CountedLoopNode *cl = _head->as_CountedLoop();
2201 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2202 return false;
2203 }
2205 #ifndef PRODUCT
2206 if(TraceLoopOpts) {
2207 tty->print("OneIteration ");
2208 this->dump_head();
2209 }
2210 #endif
2212 Node *init_n = cl->init_trip();
2213 #ifdef ASSERT
2214 // Loop boundaries should be constant since trip count is exact.
2215 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2216 #endif
2217 // Replace the phi at loop head with the value of the init_trip.
2218 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2219 // and all loop-invariant uses of the exit values will be correct.
2220 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2221 phase->C->set_major_progress();
2222 return true;
2223 }
2225 //=============================================================================
2226 //------------------------------iteration_split_impl---------------------------
2227 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2228 // Compute exact loop trip count if possible.
2229 compute_exact_trip_count(phase);
2231 // Convert one iteration loop into normal code.
2232 if (policy_do_one_iteration_loop(phase))
2233 return true;
2235 // Check and remove empty loops (spam micro-benchmarks)
2236 if (policy_do_remove_empty_loop(phase))
2237 return true; // Here we removed an empty loop
2239 bool should_peel = policy_peeling(phase); // Should we peel?
2241 bool should_unswitch = policy_unswitching(phase);
2243 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2244 // This removes loop-invariant tests (usually null checks).
2245 if (!_head->is_CountedLoop()) { // Non-counted loop
2246 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2247 // Partial peel succeeded so terminate this round of loop opts
2248 return false;
2249 }
2250 if (should_peel) { // Should we peel?
2251 #ifndef PRODUCT
2252 if (PrintOpto) tty->print_cr("should_peel");
2253 #endif
2254 phase->do_peeling(this,old_new);
2255 } else if (should_unswitch) {
2256 phase->do_unswitching(this, old_new);
2257 }
2258 return true;
2259 }
2260 CountedLoopNode *cl = _head->as_CountedLoop();
2262 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2264 // Do nothing special to pre- and post- loops
2265 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2267 // Compute loop trip count from profile data
2268 compute_profile_trip_cnt(phase);
2270 // Before attempting fancy unrolling, RCE or alignment, see if we want
2271 // to completely unroll this loop or do loop unswitching.
2272 if (cl->is_normal_loop()) {
2273 if (should_unswitch) {
2274 phase->do_unswitching(this, old_new);
2275 return true;
2276 }
2277 bool should_maximally_unroll = policy_maximally_unroll(phase);
2278 if (should_maximally_unroll) {
2279 // Here we did some unrolling and peeling. Eventually we will
2280 // completely unroll this loop and it will no longer be a loop.
2281 phase->do_maximally_unroll(this,old_new);
2282 return true;
2283 }
2284 }
2286 // Skip next optimizations if running low on nodes. Note that
2287 // policy_unswitching and policy_maximally_unroll have this check.
2288 uint nodes_left = MaxNodeLimit - phase->C->unique();
2289 if ((2 * _body.size()) > nodes_left) {
2290 return true;
2291 }
2293 // Counted loops may be peeled, may need some iterations run up
2294 // front for RCE, and may want to align loop refs to a cache
2295 // line. Thus we clone a full loop up front whose trip count is
2296 // at least 1 (if peeling), but may be several more.
2298 // The main loop will start cache-line aligned with at least 1
2299 // iteration of the unrolled body (zero-trip test required) and
2300 // will have some range checks removed.
2302 // A post-loop will finish any odd iterations (leftover after
2303 // unrolling), plus any needed for RCE purposes.
2305 bool should_unroll = policy_unroll(phase);
2307 bool should_rce = policy_range_check(phase);
2309 bool should_align = policy_align(phase);
2311 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2312 // need a pre-loop. We may still need to peel an initial iteration but
2313 // we will not be needing an unknown number of pre-iterations.
2314 //
2315 // Basically, if may_rce_align reports FALSE first time through,
2316 // we will not be able to later do RCE or Aligning on this loop.
2317 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2319 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2320 // we switch to the pre-/main-/post-loop model. This model also covers
2321 // peeling.
2322 if (should_rce || should_align || should_unroll) {
2323 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2324 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2326 // Adjust the pre- and main-loop limits to let the pre and post loops run
2327 // with full checks, but the main-loop with no checks. Remove said
2328 // checks from the main body.
2329 if (should_rce)
2330 phase->do_range_check(this,old_new);
2332 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2333 // twice as many iterations as before) and the main body limit (only do
2334 // an even number of trips). If we are peeling, we might enable some RCE
2335 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2336 // peeling.
2337 if (should_unroll && !should_peel)
2338 phase->do_unroll(this,old_new, true);
2340 // Adjust the pre-loop limits to align the main body
2341 // iterations.
2342 if (should_align)
2343 Unimplemented();
2345 } else { // Else we have an unchanged counted loop
2346 if (should_peel) // Might want to peel but do nothing else
2347 phase->do_peeling(this,old_new);
2348 }
2349 return true;
2350 }
2353 //=============================================================================
2354 //------------------------------iteration_split--------------------------------
2355 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2356 // Recursively iteration split nested loops
2357 if (_child && !_child->iteration_split(phase, old_new))
2358 return false;
2360 // Clean out prior deadwood
2361 DCE_loop_body();
2364 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2365 // Replace with a 1-in-10 exit guess.
2366 if (_parent /*not the root loop*/ &&
2367 !_irreducible &&
2368 // Also ignore the occasional dead backedge
2369 !tail()->is_top()) {
2370 adjust_loop_exit_prob(phase);
2371 }
2373 // Gate unrolling, RCE and peeling efforts.
2374 if (!_child && // If not an inner loop, do not split
2375 !_irreducible &&
2376 _allow_optimizations &&
2377 !tail()->is_top()) { // Also ignore the occasional dead backedge
2378 if (!_has_call) {
2379 if (!iteration_split_impl(phase, old_new)) {
2380 return false;
2381 }
2382 } else if (policy_unswitching(phase)) {
2383 phase->do_unswitching(this, old_new);
2384 }
2385 }
2387 // Minor offset re-organization to remove loop-fallout uses of
2388 // trip counter when there was no major reshaping.
2389 phase->reorg_offsets(this);
2391 if (_next && !_next->iteration_split(phase, old_new))
2392 return false;
2393 return true;
2394 }
2397 //=============================================================================
2398 // Process all the loops in the loop tree and replace any fill
2399 // patterns with an intrisc version.
2400 bool PhaseIdealLoop::do_intrinsify_fill() {
2401 bool changed = false;
2402 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2403 IdealLoopTree* lpt = iter.current();
2404 changed |= intrinsify_fill(lpt);
2405 }
2406 return changed;
2407 }
2410 // Examine an inner loop looking for a a single store of an invariant
2411 // value in a unit stride loop,
2412 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2413 Node*& shift, Node*& con) {
2414 const char* msg = NULL;
2415 Node* msg_node = NULL;
2417 store_value = NULL;
2418 con = NULL;
2419 shift = NULL;
2421 // Process the loop looking for stores. If there are multiple
2422 // stores or extra control flow give at this point.
2423 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2424 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2425 Node* n = lpt->_body.at(i);
2426 if (n->outcnt() == 0) continue; // Ignore dead
2427 if (n->is_Store()) {
2428 if (store != NULL) {
2429 msg = "multiple stores";
2430 break;
2431 }
2432 int opc = n->Opcode();
2433 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) {
2434 msg = "oop fills not handled";
2435 break;
2436 }
2437 Node* value = n->in(MemNode::ValueIn);
2438 if (!lpt->is_invariant(value)) {
2439 msg = "variant store value";
2440 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2441 msg = "not array address";
2442 }
2443 store = n;
2444 store_value = value;
2445 } else if (n->is_If() && n != head->loopexit()) {
2446 msg = "extra control flow";
2447 msg_node = n;
2448 }
2449 }
2451 if (store == NULL) {
2452 // No store in loop
2453 return false;
2454 }
2456 if (msg == NULL && head->stride_con() != 1) {
2457 // could handle negative strides too
2458 if (head->stride_con() < 0) {
2459 msg = "negative stride";
2460 } else {
2461 msg = "non-unit stride";
2462 }
2463 }
2465 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2466 msg = "can't handle store address";
2467 msg_node = store->in(MemNode::Address);
2468 }
2470 if (msg == NULL &&
2471 (!store->in(MemNode::Memory)->is_Phi() ||
2472 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2473 msg = "store memory isn't proper phi";
2474 msg_node = store->in(MemNode::Memory);
2475 }
2477 // Make sure there is an appropriate fill routine
2478 BasicType t = store->as_Mem()->memory_type();
2479 const char* fill_name;
2480 if (msg == NULL &&
2481 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2482 msg = "unsupported store";
2483 msg_node = store;
2484 }
2486 if (msg != NULL) {
2487 #ifndef PRODUCT
2488 if (TraceOptimizeFill) {
2489 tty->print_cr("not fill intrinsic candidate: %s", msg);
2490 if (msg_node != NULL) msg_node->dump();
2491 }
2492 #endif
2493 return false;
2494 }
2496 // Make sure the address expression can be handled. It should be
2497 // head->phi * elsize + con. head->phi might have a ConvI2L.
2498 Node* elements[4];
2499 Node* conv = NULL;
2500 bool found_index = false;
2501 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2502 for (int e = 0; e < count; e++) {
2503 Node* n = elements[e];
2504 if (n->is_Con() && con == NULL) {
2505 con = n;
2506 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2507 Node* value = n->in(1);
2508 #ifdef _LP64
2509 if (value->Opcode() == Op_ConvI2L) {
2510 conv = value;
2511 value = value->in(1);
2512 }
2513 #endif
2514 if (value != head->phi()) {
2515 msg = "unhandled shift in address";
2516 } else {
2517 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2518 msg = "scale doesn't match";
2519 } else {
2520 found_index = true;
2521 shift = n;
2522 }
2523 }
2524 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2525 if (n->in(1) == head->phi()) {
2526 found_index = true;
2527 conv = n;
2528 } else {
2529 msg = "unhandled input to ConvI2L";
2530 }
2531 } else if (n == head->phi()) {
2532 // no shift, check below for allowed cases
2533 found_index = true;
2534 } else {
2535 msg = "unhandled node in address";
2536 msg_node = n;
2537 }
2538 }
2540 if (count == -1) {
2541 msg = "malformed address expression";
2542 msg_node = store;
2543 }
2545 if (!found_index) {
2546 msg = "missing use of index";
2547 }
2549 // byte sized items won't have a shift
2550 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2551 msg = "can't find shift";
2552 msg_node = store;
2553 }
2555 if (msg != NULL) {
2556 #ifndef PRODUCT
2557 if (TraceOptimizeFill) {
2558 tty->print_cr("not fill intrinsic: %s", msg);
2559 if (msg_node != NULL) msg_node->dump();
2560 }
2561 #endif
2562 return false;
2563 }
2565 // No make sure all the other nodes in the loop can be handled
2566 VectorSet ok(Thread::current()->resource_area());
2568 // store related values are ok
2569 ok.set(store->_idx);
2570 ok.set(store->in(MemNode::Memory)->_idx);
2572 // Loop structure is ok
2573 ok.set(head->_idx);
2574 ok.set(head->loopexit()->_idx);
2575 ok.set(head->phi()->_idx);
2576 ok.set(head->incr()->_idx);
2577 ok.set(head->loopexit()->cmp_node()->_idx);
2578 ok.set(head->loopexit()->in(1)->_idx);
2580 // Address elements are ok
2581 if (con) ok.set(con->_idx);
2582 if (shift) ok.set(shift->_idx);
2583 if (conv) ok.set(conv->_idx);
2585 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2586 Node* n = lpt->_body.at(i);
2587 if (n->outcnt() == 0) continue; // Ignore dead
2588 if (ok.test(n->_idx)) continue;
2589 // Backedge projection is ok
2590 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue;
2591 if (!n->is_AddP()) {
2592 msg = "unhandled node";
2593 msg_node = n;
2594 break;
2595 }
2596 }
2598 // Make sure no unexpected values are used outside the loop
2599 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2600 Node* n = lpt->_body.at(i);
2601 // These values can be replaced with other nodes if they are used
2602 // outside the loop.
2603 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2604 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2605 Node* use = iter.get();
2606 if (!lpt->_body.contains(use)) {
2607 msg = "node is used outside loop";
2608 // lpt->_body.dump();
2609 msg_node = n;
2610 break;
2611 }
2612 }
2613 }
2615 #ifdef ASSERT
2616 if (TraceOptimizeFill) {
2617 if (msg != NULL) {
2618 tty->print_cr("no fill intrinsic: %s", msg);
2619 if (msg_node != NULL) msg_node->dump();
2620 } else {
2621 tty->print_cr("fill intrinsic for:");
2622 }
2623 store->dump();
2624 if (Verbose) {
2625 lpt->_body.dump();
2626 }
2627 }
2628 #endif
2630 return msg == NULL;
2631 }
2635 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2636 // Only for counted inner loops
2637 if (!lpt->is_counted() || !lpt->is_inner()) {
2638 return false;
2639 }
2641 // Must have constant stride
2642 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2643 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
2644 return false;
2645 }
2647 // Check that the body only contains a store of a loop invariant
2648 // value that is indexed by the loop phi.
2649 Node* store = NULL;
2650 Node* store_value = NULL;
2651 Node* shift = NULL;
2652 Node* offset = NULL;
2653 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2654 return false;
2655 }
2657 #ifndef PRODUCT
2658 if (TraceLoopOpts) {
2659 tty->print("ArrayFill ");
2660 lpt->dump_head();
2661 }
2662 #endif
2664 // Now replace the whole loop body by a call to a fill routine that
2665 // covers the same region as the loop.
2666 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2668 // Build an expression for the beginning of the copy region
2669 Node* index = head->init_trip();
2670 #ifdef _LP64
2671 index = new (C, 2) ConvI2LNode(index);
2672 _igvn.register_new_node_with_optimizer(index);
2673 #endif
2674 if (shift != NULL) {
2675 // byte arrays don't require a shift but others do.
2676 index = new (C, 3) LShiftXNode(index, shift->in(2));
2677 _igvn.register_new_node_with_optimizer(index);
2678 }
2679 index = new (C, 4) AddPNode(base, base, index);
2680 _igvn.register_new_node_with_optimizer(index);
2681 Node* from = new (C, 4) AddPNode(base, index, offset);
2682 _igvn.register_new_node_with_optimizer(from);
2683 // Compute the number of elements to copy
2684 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip());
2685 _igvn.register_new_node_with_optimizer(len);
2687 BasicType t = store->as_Mem()->memory_type();
2688 bool aligned = false;
2689 if (offset != NULL && head->init_trip()->is_Con()) {
2690 int element_size = type2aelembytes(t);
2691 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2692 }
2694 // Build a call to the fill routine
2695 const char* fill_name;
2696 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2697 assert(fill != NULL, "what?");
2699 // Convert float/double to int/long for fill routines
2700 if (t == T_FLOAT) {
2701 store_value = new (C, 2) MoveF2INode(store_value);
2702 _igvn.register_new_node_with_optimizer(store_value);
2703 } else if (t == T_DOUBLE) {
2704 store_value = new (C, 2) MoveD2LNode(store_value);
2705 _igvn.register_new_node_with_optimizer(store_value);
2706 }
2708 Node* mem_phi = store->in(MemNode::Memory);
2709 Node* result_ctrl;
2710 Node* result_mem;
2711 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2712 int size = call_type->domain()->cnt();
2713 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill,
2714 fill_name, TypeAryPtr::get_array_body_type(t));
2715 call->init_req(TypeFunc::Parms+0, from);
2716 call->init_req(TypeFunc::Parms+1, store_value);
2717 #ifdef _LP64
2718 len = new (C, 2) ConvI2LNode(len);
2719 _igvn.register_new_node_with_optimizer(len);
2720 #endif
2721 call->init_req(TypeFunc::Parms+2, len);
2722 #ifdef _LP64
2723 call->init_req(TypeFunc::Parms+3, C->top());
2724 #endif
2725 call->init_req( TypeFunc::Control, head->init_control());
2726 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o
2727 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) );
2728 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) );
2729 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) );
2730 _igvn.register_new_node_with_optimizer(call);
2731 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
2732 _igvn.register_new_node_with_optimizer(result_ctrl);
2733 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory);
2734 _igvn.register_new_node_with_optimizer(result_mem);
2736 // If this fill is tightly coupled to an allocation and overwrites
2737 // the whole body, allow it to take over the zeroing.
2738 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2739 if (alloc != NULL && alloc->is_AllocateArray()) {
2740 Node* length = alloc->as_AllocateArray()->Ideal_length();
2741 if (head->limit() == length &&
2742 head->init_trip() == _igvn.intcon(0)) {
2743 if (TraceOptimizeFill) {
2744 tty->print_cr("Eliminated zeroing in allocation");
2745 }
2746 alloc->maybe_set_complete(&_igvn);
2747 } else {
2748 #ifdef ASSERT
2749 if (TraceOptimizeFill) {
2750 tty->print_cr("filling array but bounds don't match");
2751 alloc->dump();
2752 head->init_trip()->dump();
2753 head->limit()->dump();
2754 length->dump();
2755 }
2756 #endif
2757 }
2758 }
2760 // Redirect the old control and memory edges that are outside the loop.
2761 Node* exit = head->loopexit()->proj_out(0);
2762 // Sometimes the memory phi of the head is used as the outgoing
2763 // state of the loop. It's safe in this case to replace it with the
2764 // result_mem.
2765 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2766 _igvn.replace_node(exit, result_ctrl);
2767 _igvn.replace_node(store, result_mem);
2768 // Any uses the increment outside of the loop become the loop limit.
2769 _igvn.replace_node(head->incr(), head->limit());
2771 // Disconnect the head from the loop.
2772 for (uint i = 0; i < lpt->_body.size(); i++) {
2773 Node* n = lpt->_body.at(i);
2774 _igvn.replace_node(n, C->top());
2775 }
2777 return true;
2778 }