Thu, 21 Jul 2011 11:25:07 -0700
7063628: Use cbcond on T4
Summary: Add new short branch instruction to Hotspot sparc assembler.
Reviewed-by: never, twisti, jrose
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_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx];
513 new_exit_value = move_loop_predicates(entry, new_exit_value, !counted_loop);
514 _igvn.hash_delete(head);
515 head->set_req(LoopNode::EntryControl, new_exit_value);
516 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
517 Node* old = head->fast_out(j);
518 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
519 new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
520 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
521 // Then loop body backedge value remains the same.
522 new_exit_value = old->in(LoopNode::LoopBackControl);
523 _igvn.hash_delete(old);
524 old->set_req(LoopNode::EntryControl, new_exit_value);
525 }
526 }
529 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
530 // extra backedge user.
531 Node* new_head = old_new[head->_idx];
532 _igvn.hash_delete(new_head);
533 new_head->set_req(LoopNode::LoopBackControl, C->top());
534 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
535 Node* use = new_head->fast_out(j2);
536 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
537 _igvn.hash_delete(use);
538 use->set_req(LoopNode::LoopBackControl, C->top());
539 }
540 }
543 // Step 4: Correct dom-depth info. Set to loop-head depth.
544 int dd = dom_depth(head);
545 set_idom(head, head->in(1), dd);
546 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
547 Node *old = loop->_body.at(j3);
548 Node *nnn = old_new[old->_idx];
549 if (!has_ctrl(nnn))
550 set_idom(nnn, idom(nnn), dd-1);
551 // While we're at it, remove any SafePoints from the peeled code
552 if (old->Opcode() == Op_SafePoint) {
553 Node *nnn = old_new[old->_idx];
554 lazy_replace(nnn,nnn->in(TypeFunc::Control));
555 }
556 }
558 // Now force out all loop-invariant dominating tests. The optimizer
559 // finds some, but we _know_ they are all useless.
560 peeled_dom_test_elim(loop,old_new);
562 loop->record_for_igvn();
563 }
565 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
567 //------------------------------policy_maximally_unroll------------------------
568 // Calculate exact loop trip count and return true if loop can be maximally
569 // unrolled.
570 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
571 CountedLoopNode *cl = _head->as_CountedLoop();
572 assert(cl->is_normal_loop(), "");
573 if (!cl->is_valid_counted_loop())
574 return false; // Malformed counted loop
576 if (!cl->has_exact_trip_count()) {
577 // Trip count is not exact.
578 return false;
579 }
581 uint trip_count = cl->trip_count();
582 // Note, max_juint is used to indicate unknown trip count.
583 assert(trip_count > 1, "one iteration loop should be optimized out already");
584 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
586 // Real policy: if we maximally unroll, does it get too big?
587 // Allow the unrolled mess to get larger than standard loop
588 // size. After all, it will no longer be a loop.
589 uint body_size = _body.size();
590 uint unroll_limit = (uint)LoopUnrollLimit * 4;
591 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
592 if (trip_count > unroll_limit || body_size > unroll_limit) {
593 return false;
594 }
596 // Fully unroll a loop with few iterations regardless next
597 // conditions since following loop optimizations will split
598 // such loop anyway (pre-main-post).
599 if (trip_count <= 3)
600 return true;
602 // Take into account that after unroll conjoined heads and tails will fold,
603 // otherwise policy_unroll() may allow more unrolling than max unrolling.
604 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
605 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
606 if (body_size != tst_body_size) // Check for int overflow
607 return false;
608 if (new_body_size > unroll_limit ||
609 // Unrolling can result in a large amount of node construction
610 new_body_size >= MaxNodeLimit - phase->C->unique()) {
611 return false;
612 }
614 // Do not unroll a loop with String intrinsics code.
615 // String intrinsics are large and have loops.
616 for (uint k = 0; k < _body.size(); k++) {
617 Node* n = _body.at(k);
618 switch (n->Opcode()) {
619 case Op_StrComp:
620 case Op_StrEquals:
621 case Op_StrIndexOf:
622 case Op_AryEq: {
623 return false;
624 }
625 } // switch
626 }
628 return true; // Do maximally unroll
629 }
632 #define MAX_UNROLL 16 // maximum number of unrolls for main loop
634 //------------------------------policy_unroll----------------------------------
635 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
636 // the loop is a CountedLoop and the body is small enough.
637 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
639 CountedLoopNode *cl = _head->as_CountedLoop();
640 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
642 if (!cl->is_valid_counted_loop())
643 return false; // Malformed counted loop
645 // Protect against over-unrolling.
646 // After split at least one iteration will be executed in pre-loop.
647 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
649 int future_unroll_ct = cl->unrolled_count() * 2;
650 if (future_unroll_ct > MAX_UNROLL) return false;
652 // Check for initial stride being a small enough constant
653 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
655 // Don't unroll if the next round of unrolling would push us
656 // over the expected trip count of the loop. One is subtracted
657 // from the expected trip count because the pre-loop normally
658 // executes 1 iteration.
659 if (UnrollLimitForProfileCheck > 0 &&
660 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
661 future_unroll_ct > UnrollLimitForProfileCheck &&
662 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
663 return false;
664 }
666 // When unroll count is greater than LoopUnrollMin, don't unroll if:
667 // the residual iterations are more than 10% of the trip count
668 // and rounds of "unroll,optimize" are not making significant progress
669 // Progress defined as current size less than 20% larger than previous size.
670 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
671 future_unroll_ct > LoopUnrollMin &&
672 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
673 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
674 return false;
675 }
677 Node *init_n = cl->init_trip();
678 Node *limit_n = cl->limit();
679 int stride_con = cl->stride_con();
680 // Non-constant bounds.
681 // Protect against over-unrolling when init or/and limit are not constant
682 // (so that trip_count's init value is maxint) but iv range is known.
683 if (init_n == NULL || !init_n->is_Con() ||
684 limit_n == NULL || !limit_n->is_Con()) {
685 Node* phi = cl->phi();
686 if (phi != NULL) {
687 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
688 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
689 int next_stride = stride_con * 2; // stride after this unroll
690 if (next_stride > 0) {
691 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
692 iv_type->_lo + next_stride > iv_type->_hi) {
693 return false; // over-unrolling
694 }
695 } else if (next_stride < 0) {
696 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
697 iv_type->_hi + next_stride < iv_type->_lo) {
698 return false; // over-unrolling
699 }
700 }
701 }
702 }
704 // After unroll limit will be adjusted: new_limit = limit-stride.
705 // Bailout if adjustment overflow.
706 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
707 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
708 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
709 return false; // overflow
711 // Adjust body_size to determine if we unroll or not
712 uint body_size = _body.size();
713 // Also count ModL, DivL and MulL which expand mightly
714 for (uint k = 0; k < _body.size(); k++) {
715 Node* n = _body.at(k);
716 switch (n->Opcode()) {
717 case Op_ModL: body_size += 30; break;
718 case Op_DivL: body_size += 30; break;
719 case Op_MulL: body_size += 10; break;
720 case Op_StrComp:
721 case Op_StrEquals:
722 case Op_StrIndexOf:
723 case Op_AryEq: {
724 // Do not unroll a loop with String intrinsics code.
725 // String intrinsics are large and have loops.
726 return false;
727 }
728 } // switch
729 }
731 // Check for being too big
732 if (body_size > (uint)LoopUnrollLimit) {
733 // Normal case: loop too big
734 return false;
735 }
737 // Unroll once! (Each trip will soon do double iterations)
738 return true;
739 }
741 //------------------------------policy_align-----------------------------------
742 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
743 // expression that does the alignment. Note that only one array base can be
744 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
745 // if we vectorize short memory ops into longer memory ops, we may want to
746 // increase alignment.
747 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
748 return false;
749 }
751 //------------------------------policy_range_check-----------------------------
752 // Return TRUE or FALSE if the loop should be range-check-eliminated.
753 // Actually we do iteration-splitting, a more powerful form of RCE.
754 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
755 if (!RangeCheckElimination) return false;
757 CountedLoopNode *cl = _head->as_CountedLoop();
758 // If we unrolled with no intention of doing RCE and we later
759 // changed our minds, we got no pre-loop. Either we need to
760 // make a new pre-loop, or we gotta disallow RCE.
761 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
762 Node *trip_counter = cl->phi();
764 // Check loop body for tests of trip-counter plus loop-invariant vs
765 // loop-invariant.
766 for (uint i = 0; i < _body.size(); i++) {
767 Node *iff = _body[i];
768 if (iff->Opcode() == Op_If) { // Test?
770 // Comparing trip+off vs limit
771 Node *bol = iff->in(1);
772 if (bol->req() != 2) continue; // dead constant test
773 if (!bol->is_Bool()) {
774 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
775 continue;
776 }
777 if (bol->as_Bool()->_test._test == BoolTest::ne)
778 continue; // not RC
780 Node *cmp = bol->in(1);
782 Node *rc_exp = cmp->in(1);
783 Node *limit = cmp->in(2);
785 Node *limit_c = phase->get_ctrl(limit);
786 if( limit_c == phase->C->top() )
787 return false; // Found dead test on live IF? No RCE!
788 if( is_member(phase->get_loop(limit_c) ) ) {
789 // Compare might have operands swapped; commute them
790 rc_exp = cmp->in(2);
791 limit = cmp->in(1);
792 limit_c = phase->get_ctrl(limit);
793 if( is_member(phase->get_loop(limit_c) ) )
794 continue; // Both inputs are loop varying; cannot RCE
795 }
797 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
798 continue;
799 }
800 // Yeah! Found a test like 'trip+off vs limit'
801 // Test is an IfNode, has 2 projections. If BOTH are in the loop
802 // we need loop unswitching instead of iteration splitting.
803 if( is_loop_exit(iff) )
804 return true; // Found reason to split iterations
805 } // End of is IF
806 }
808 return false;
809 }
811 //------------------------------policy_peel_only-------------------------------
812 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
813 // for unrolling loops with NO array accesses.
814 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
816 for( uint i = 0; i < _body.size(); i++ )
817 if( _body[i]->is_Mem() )
818 return false;
820 // No memory accesses at all!
821 return true;
822 }
824 //------------------------------clone_up_backedge_goo--------------------------
825 // If Node n lives in the back_ctrl block and cannot float, we clone a private
826 // version of n in preheader_ctrl block and return that, otherwise return n.
827 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
828 if( get_ctrl(n) != back_ctrl ) return n;
830 // Only visit once
831 if (visited.test_set(n->_idx)) {
832 Node *x = clones.find(n->_idx);
833 if (x != NULL)
834 return x;
835 return n;
836 }
838 Node *x = NULL; // If required, a clone of 'n'
839 // Check for 'n' being pinned in the backedge.
840 if( n->in(0) && n->in(0) == back_ctrl ) {
841 assert(clones.find(n->_idx) == NULL, "dead loop");
842 x = n->clone(); // Clone a copy of 'n' to preheader
843 clones.push(x, n->_idx);
844 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
845 }
847 // Recursive fixup any other input edges into x.
848 // If there are no changes we can just return 'n', otherwise
849 // we need to clone a private copy and change it.
850 for( uint i = 1; i < n->req(); i++ ) {
851 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
852 if( g != n->in(i) ) {
853 if( !x ) {
854 assert(clones.find(n->_idx) == NULL, "dead loop");
855 x = n->clone();
856 clones.push(x, n->_idx);
857 }
858 x->set_req(i, g);
859 }
860 }
861 if( x ) { // x can legally float to pre-header location
862 register_new_node( x, preheader_ctrl );
863 return x;
864 } else { // raise n to cover LCA of uses
865 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
866 }
867 return n;
868 }
870 //------------------------------insert_pre_post_loops--------------------------
871 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
872 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
873 // alignment. Useful to unroll loops that do no array accesses.
874 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
876 #ifndef PRODUCT
877 if (TraceLoopOpts) {
878 if (peel_only)
879 tty->print("PeelMainPost ");
880 else
881 tty->print("PreMainPost ");
882 loop->dump_head();
883 }
884 #endif
885 C->set_major_progress();
887 // Find common pieces of the loop being guarded with pre & post loops
888 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
889 assert( main_head->is_normal_loop(), "" );
890 CountedLoopEndNode *main_end = main_head->loopexit();
891 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
892 uint dd_main_head = dom_depth(main_head);
893 uint max = main_head->outcnt();
895 Node *pre_header= main_head->in(LoopNode::EntryControl);
896 Node *init = main_head->init_trip();
897 Node *incr = main_end ->incr();
898 Node *limit = main_end ->limit();
899 Node *stride = main_end ->stride();
900 Node *cmp = main_end ->cmp_node();
901 BoolTest::mask b_test = main_end->test_trip();
903 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
904 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
905 if( bol->outcnt() != 1 ) {
906 bol = bol->clone();
907 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
908 _igvn.hash_delete(main_end);
909 main_end->set_req(CountedLoopEndNode::TestValue, bol);
910 }
911 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
912 if( cmp->outcnt() != 1 ) {
913 cmp = cmp->clone();
914 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
915 _igvn.hash_delete(bol);
916 bol->set_req(1, cmp);
917 }
919 //------------------------------
920 // Step A: Create Post-Loop.
921 Node* main_exit = main_end->proj_out(false);
922 assert( main_exit->Opcode() == Op_IfFalse, "" );
923 int dd_main_exit = dom_depth(main_exit);
925 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
926 // loop pre-header illegally has 2 control users (old & new loops).
927 clone_loop( loop, old_new, dd_main_exit );
928 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
929 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
930 post_head->set_post_loop(main_head);
932 // Reduce the post-loop trip count.
933 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
934 post_end->_prob = PROB_FAIR;
936 // Build the main-loop normal exit.
937 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
938 _igvn.register_new_node_with_optimizer( new_main_exit );
939 set_idom(new_main_exit, main_end, dd_main_exit );
940 set_loop(new_main_exit, loop->_parent);
942 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
943 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
944 // (the main-loop trip-counter exit value) because we will be changing
945 // the exit value (via unrolling) so we cannot constant-fold away the zero
946 // trip guard until all unrolling is done.
947 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
948 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
949 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
950 register_new_node( zer_opaq, new_main_exit );
951 register_new_node( zer_cmp , new_main_exit );
952 register_new_node( zer_bol , new_main_exit );
954 // Build the IfNode
955 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
956 _igvn.register_new_node_with_optimizer( zer_iff );
957 set_idom(zer_iff, new_main_exit, dd_main_exit);
958 set_loop(zer_iff, loop->_parent);
960 // Plug in the false-path, taken if we need to skip post-loop
961 _igvn.hash_delete( main_exit );
962 main_exit->set_req(0, zer_iff);
963 _igvn._worklist.push(main_exit);
964 set_idom(main_exit, zer_iff, dd_main_exit);
965 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
966 // Make the true-path, must enter the post loop
967 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
968 _igvn.register_new_node_with_optimizer( zer_taken );
969 set_idom(zer_taken, zer_iff, dd_main_exit);
970 set_loop(zer_taken, loop->_parent);
971 // Plug in the true path
972 _igvn.hash_delete( post_head );
973 post_head->set_req(LoopNode::EntryControl, zer_taken);
974 set_idom(post_head, zer_taken, dd_main_exit);
976 Arena *a = Thread::current()->resource_area();
977 VectorSet visited(a);
978 Node_Stack clones(a, main_head->back_control()->outcnt());
979 // Step A3: Make the fall-in values to the post-loop come from the
980 // fall-out values of the main-loop.
981 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
982 Node* main_phi = main_head->fast_out(i);
983 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
984 Node *post_phi = old_new[main_phi->_idx];
985 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
986 post_head->init_control(),
987 main_phi->in(LoopNode::LoopBackControl),
988 visited, clones);
989 _igvn.hash_delete(post_phi);
990 post_phi->set_req( LoopNode::EntryControl, fallmain );
991 }
992 }
994 // Update local caches for next stanza
995 main_exit = new_main_exit;
998 //------------------------------
999 // Step B: Create Pre-Loop.
1001 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
1002 // loop pre-header illegally has 2 control users (old & new loops).
1003 clone_loop( loop, old_new, dd_main_head );
1004 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1005 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1006 pre_head->set_pre_loop(main_head);
1007 Node *pre_incr = old_new[incr->_idx];
1009 // Reduce the pre-loop trip count.
1010 pre_end->_prob = PROB_FAIR;
1012 // Find the pre-loop normal exit.
1013 Node* pre_exit = pre_end->proj_out(false);
1014 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1015 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
1016 _igvn.register_new_node_with_optimizer( new_pre_exit );
1017 set_idom(new_pre_exit, pre_end, dd_main_head);
1018 set_loop(new_pre_exit, loop->_parent);
1020 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1021 // pre-loop, the main-loop may not execute at all. Later in life this
1022 // zero-trip guard will become the minimum-trip guard when we unroll
1023 // the main-loop.
1024 Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
1025 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
1026 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
1027 register_new_node( min_opaq, new_pre_exit );
1028 register_new_node( min_cmp , new_pre_exit );
1029 register_new_node( min_bol , new_pre_exit );
1031 // Build the IfNode (assume the main-loop is executed always).
1032 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1033 _igvn.register_new_node_with_optimizer( min_iff );
1034 set_idom(min_iff, new_pre_exit, dd_main_head);
1035 set_loop(min_iff, loop->_parent);
1037 // Plug in the false-path, taken if we need to skip main-loop
1038 _igvn.hash_delete( pre_exit );
1039 pre_exit->set_req(0, min_iff);
1040 set_idom(pre_exit, min_iff, dd_main_head);
1041 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1042 // Make the true-path, must enter the main loop
1043 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
1044 _igvn.register_new_node_with_optimizer( min_taken );
1045 set_idom(min_taken, min_iff, dd_main_head);
1046 set_loop(min_taken, loop->_parent);
1047 // Plug in the true path
1048 _igvn.hash_delete( main_head );
1049 main_head->set_req(LoopNode::EntryControl, min_taken);
1050 set_idom(main_head, min_taken, dd_main_head);
1052 visited.Clear();
1053 clones.clear();
1054 // Step B3: Make the fall-in values to the main-loop come from the
1055 // fall-out values of the pre-loop.
1056 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1057 Node* main_phi = main_head->fast_out(i2);
1058 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1059 Node *pre_phi = old_new[main_phi->_idx];
1060 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1061 main_head->init_control(),
1062 pre_phi->in(LoopNode::LoopBackControl),
1063 visited, clones);
1064 _igvn.hash_delete(main_phi);
1065 main_phi->set_req( LoopNode::EntryControl, fallpre );
1066 }
1067 }
1069 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1070 // RCE and alignment may change this later.
1071 Node *cmp_end = pre_end->cmp_node();
1072 assert( cmp_end->in(2) == limit, "" );
1073 Node *pre_limit = new (C, 3) AddINode( init, stride );
1075 // Save the original loop limit in this Opaque1 node for
1076 // use by range check elimination.
1077 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
1079 register_new_node( pre_limit, pre_head->in(0) );
1080 register_new_node( pre_opaq , pre_head->in(0) );
1082 // Since no other users of pre-loop compare, I can hack limit directly
1083 assert( cmp_end->outcnt() == 1, "no other users" );
1084 _igvn.hash_delete(cmp_end);
1085 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1087 // Special case for not-equal loop bounds:
1088 // Change pre loop test, main loop test, and the
1089 // main loop guard test to use lt or gt depending on stride
1090 // direction:
1091 // positive stride use <
1092 // negative stride use >
1093 //
1094 // not-equal test is kept for post loop to handle case
1095 // when init > limit when stride > 0 (and reverse).
1097 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1099 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1100 // Modify pre loop end condition
1101 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1102 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
1103 register_new_node( new_bol0, pre_head->in(0) );
1104 _igvn.hash_delete(pre_end);
1105 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
1106 // Modify main loop guard condition
1107 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1108 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
1109 register_new_node( new_bol1, new_pre_exit );
1110 _igvn.hash_delete(min_iff);
1111 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1112 // Modify main loop end condition
1113 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1114 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
1115 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1116 _igvn.hash_delete(main_end);
1117 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
1118 }
1120 // Flag main loop
1121 main_head->set_main_loop();
1122 if( peel_only ) main_head->set_main_no_pre_loop();
1124 // Subtract a trip count for the pre-loop.
1125 main_head->set_trip_count(main_head->trip_count() - 1);
1127 // It's difficult to be precise about the trip-counts
1128 // for the pre/post loops. They are usually very short,
1129 // so guess that 4 trips is a reasonable value.
1130 post_head->set_profile_trip_cnt(4.0);
1131 pre_head->set_profile_trip_cnt(4.0);
1133 // Now force out all loop-invariant dominating tests. The optimizer
1134 // finds some, but we _know_ they are all useless.
1135 peeled_dom_test_elim(loop,old_new);
1136 }
1138 //------------------------------is_invariant-----------------------------
1139 // Return true if n is invariant
1140 bool IdealLoopTree::is_invariant(Node* n) const {
1141 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1142 if (n_c->is_top()) return false;
1143 return !is_member(_phase->get_loop(n_c));
1144 }
1147 //------------------------------do_unroll--------------------------------------
1148 // Unroll the loop body one step - make each trip do 2 iterations.
1149 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1150 assert(LoopUnrollLimit, "");
1151 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1152 CountedLoopEndNode *loop_end = loop_head->loopexit();
1153 assert(loop_end, "");
1154 #ifndef PRODUCT
1155 if (PrintOpto && VerifyLoopOptimizations) {
1156 tty->print("Unrolling ");
1157 loop->dump_head();
1158 } else if (TraceLoopOpts) {
1159 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1160 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1161 } else {
1162 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1163 }
1164 loop->dump_head();
1165 }
1166 #endif
1168 // Remember loop node count before unrolling to detect
1169 // if rounds of unroll,optimize are making progress
1170 loop_head->set_node_count_before_unroll(loop->_body.size());
1172 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1173 Node *limit = loop_head->limit();
1174 Node *init = loop_head->init_trip();
1175 Node *stride = loop_head->stride();
1177 Node *opaq = NULL;
1178 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1179 // Search for zero-trip guard.
1180 assert( loop_head->is_main_loop(), "" );
1181 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1182 Node *iff = ctrl->in(0);
1183 assert( iff->Opcode() == Op_If, "" );
1184 Node *bol = iff->in(1);
1185 assert( bol->Opcode() == Op_Bool, "" );
1186 Node *cmp = bol->in(1);
1187 assert( cmp->Opcode() == Op_CmpI, "" );
1188 opaq = cmp->in(2);
1189 // Occasionally it's possible for a zero-trip guard Opaque1 node to be
1190 // optimized away and then another round of loop opts attempted.
1191 // We can not optimize this particular loop in that case.
1192 if (opaq->Opcode() != Op_Opaque1)
1193 return; // Cannot find zero-trip guard! Bail out!
1194 // Zero-trip test uses an 'opaque' node which is not shared.
1195 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1196 }
1198 C->set_major_progress();
1200 Node* new_limit = NULL;
1201 if (UnrollLimitCheck) {
1202 int stride_con = stride->get_int();
1203 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1204 uint old_trip_count = loop_head->trip_count();
1205 // Verify that unroll policy result is still valid.
1206 assert(old_trip_count > 1 &&
1207 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1209 // Adjust loop limit to keep valid iterations number after unroll.
1210 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1211 // which may overflow.
1212 if (!adjust_min_trip) {
1213 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1214 "odd trip count for maximally unroll");
1215 // Don't need to adjust limit for maximally unroll since trip count is even.
1216 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1217 // Loop's limit is constant. Loop's init could be constant when pre-loop
1218 // become peeled iteration.
1219 long init_con = init->get_int();
1220 // We can keep old loop limit if iterations count stays the same:
1221 // old_trip_count == new_trip_count * 2
1222 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1223 // so we also don't need to adjust zero trip test.
1224 long limit_con = limit->get_int();
1225 // (stride_con*2) not overflow since stride_con <= 8.
1226 int new_stride_con = stride_con * 2;
1227 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1228 long trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1229 // New trip count should satisfy next conditions.
1230 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1231 uint new_trip_count = (uint)trip_count;
1232 adjust_min_trip = (old_trip_count != new_trip_count*2);
1233 }
1235 if (adjust_min_trip) {
1236 // Step 2: Adjust the trip limit if it is called for.
1237 // The adjustment amount is -stride. Need to make sure if the
1238 // adjustment underflows or overflows, then the main loop is skipped.
1239 Node* cmp = loop_end->cmp_node();
1240 assert(cmp->in(2) == limit, "sanity");
1241 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1243 // Verify that policy_unroll result is still valid.
1244 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1245 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1246 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1248 if (limit->is_Con()) {
1249 // The check in policy_unroll and the assert above guarantee
1250 // no underflow if limit is constant.
1251 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1252 set_ctrl(new_limit, C->root());
1253 } else {
1254 // Limit is not constant.
1255 if (loop_head->unrolled_count() == 1) { // only for first unroll
1256 // Separate limit by Opaque node in case it is an incremented
1257 // variable from previous loop to avoid using pre-incremented
1258 // value which could increase register pressure.
1259 // Otherwise reorg_offsets() optimization will create a separate
1260 // Opaque node for each use of trip-counter and as result
1261 // zero trip guard limit will be different from loop limit.
1262 assert(has_ctrl(opaq), "should have it");
1263 Node* opaq_ctrl = get_ctrl(opaq);
1264 limit = new (C, 2) Opaque2Node( C, limit );
1265 register_new_node( limit, opaq_ctrl );
1266 }
1267 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
1268 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
1269 // No underflow.
1270 new_limit = new (C, 3) SubINode(limit, stride);
1271 } else {
1272 // (limit - stride) may underflow.
1273 // Clamp the adjustment value with MININT or MAXINT:
1274 //
1275 // new_limit = limit-stride
1276 // if (stride > 0)
1277 // new_limit = (limit < new_limit) ? MININT : new_limit;
1278 // else
1279 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1280 //
1281 BoolTest::mask bt = loop_end->test_trip();
1282 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1283 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1284 set_ctrl(adj_max, C->root());
1285 Node* old_limit = NULL;
1286 Node* adj_limit = NULL;
1287 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1288 if (loop_head->unrolled_count() > 1 &&
1289 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1290 limit->in(CMoveNode::IfTrue) == adj_max &&
1291 bol->as_Bool()->_test._test == bt &&
1292 bol->in(1)->Opcode() == Op_CmpI &&
1293 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1294 // Loop was unrolled before.
1295 // Optimize the limit to avoid nested CMove:
1296 // use original limit as old limit.
1297 old_limit = bol->in(1)->in(1);
1298 // Adjust previous adjusted limit.
1299 adj_limit = limit->in(CMoveNode::IfFalse);
1300 adj_limit = new (C, 3) SubINode(adj_limit, stride);
1301 } else {
1302 old_limit = limit;
1303 adj_limit = new (C, 3) SubINode(limit, stride);
1304 }
1305 assert(old_limit != NULL && adj_limit != NULL, "");
1306 register_new_node( adj_limit, ctrl ); // adjust amount
1307 Node* adj_cmp = new (C, 3) CmpINode(old_limit, adj_limit);
1308 register_new_node( adj_cmp, ctrl );
1309 Node* adj_bool = new (C, 2) BoolNode(adj_cmp, bt);
1310 register_new_node( adj_bool, ctrl );
1311 new_limit = new (C, 4) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1312 }
1313 register_new_node(new_limit, ctrl);
1314 }
1315 assert(new_limit != NULL, "");
1316 // Replace in loop test.
1317 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1318 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1319 // Don't need to create new test since only one user.
1320 _igvn.hash_delete(cmp);
1321 cmp->set_req(2, new_limit);
1322 } else {
1323 // Create new test since it is shared.
1324 Node* ctrl2 = loop_end->in(0);
1325 Node* cmp2 = cmp->clone();
1326 cmp2->set_req(2, new_limit);
1327 register_new_node(cmp2, ctrl2);
1328 Node* bol2 = loop_end->in(1)->clone();
1329 bol2->set_req(1, cmp2);
1330 register_new_node(bol2, ctrl2);
1331 _igvn.hash_delete(loop_end);
1332 loop_end->set_req(1, bol2);
1333 }
1334 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1335 // Make it a 1-trip test (means at least 2 trips).
1337 // Guard test uses an 'opaque' node which is not shared. Hence I
1338 // can edit it's inputs directly. Hammer in the new limit for the
1339 // minimum-trip guard.
1340 assert(opaq->outcnt() == 1, "");
1341 _igvn.hash_delete(opaq);
1342 opaq->set_req(1, new_limit);
1343 }
1345 // Adjust max trip count. The trip count is intentionally rounded
1346 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1347 // the main, unrolled, part of the loop will never execute as it is protected
1348 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1349 // and later determined that part of the unrolled loop was dead.
1350 loop_head->set_trip_count(old_trip_count / 2);
1352 // Double the count of original iterations in the unrolled loop body.
1353 loop_head->double_unrolled_count();
1355 } else { // LoopLimitCheck
1357 // Adjust max trip count. The trip count is intentionally rounded
1358 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1359 // the main, unrolled, part of the loop will never execute as it is protected
1360 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1361 // and later determined that part of the unrolled loop was dead.
1362 loop_head->set_trip_count(loop_head->trip_count() / 2);
1364 // Double the count of original iterations in the unrolled loop body.
1365 loop_head->double_unrolled_count();
1367 // -----------
1368 // Step 2: Cut back the trip counter for an unroll amount of 2.
1369 // Loop will normally trip (limit - init)/stride_con. Since it's a
1370 // CountedLoop this is exact (stride divides limit-init exactly).
1371 // We are going to double the loop body, so we want to knock off any
1372 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1373 Node *span = new (C, 3) SubINode( limit, init );
1374 register_new_node( span, ctrl );
1375 Node *trip = new (C, 3) DivINode( 0, span, stride );
1376 register_new_node( trip, ctrl );
1377 Node *mtwo = _igvn.intcon(-2);
1378 set_ctrl(mtwo, C->root());
1379 Node *rond = new (C, 3) AndINode( trip, mtwo );
1380 register_new_node( rond, ctrl );
1381 Node *spn2 = new (C, 3) MulINode( rond, stride );
1382 register_new_node( spn2, ctrl );
1383 new_limit = new (C, 3) AddINode( spn2, init );
1384 register_new_node( new_limit, ctrl );
1386 // Hammer in the new limit
1387 Node *ctrl2 = loop_end->in(0);
1388 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), new_limit );
1389 register_new_node( cmp2, ctrl2 );
1390 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
1391 register_new_node( bol2, ctrl2 );
1392 _igvn.hash_delete(loop_end);
1393 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
1395 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1396 // Make it a 1-trip test (means at least 2 trips).
1397 if( adjust_min_trip ) {
1398 assert( new_limit != NULL, "" );
1399 // Guard test uses an 'opaque' node which is not shared. Hence I
1400 // can edit it's inputs directly. Hammer in the new limit for the
1401 // minimum-trip guard.
1402 assert( opaq->outcnt() == 1, "" );
1403 _igvn.hash_delete(opaq);
1404 opaq->set_req(1, new_limit);
1405 }
1406 } // LoopLimitCheck
1408 // ---------
1409 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1410 // represents the odd iterations; since the loop trips an even number of
1411 // times its backedge is never taken. Kill the backedge.
1412 uint dd = dom_depth(loop_head);
1413 clone_loop( loop, old_new, dd );
1415 // Make backedges of the clone equal to backedges of the original.
1416 // Make the fall-in from the original come from the fall-out of the clone.
1417 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1418 Node* phi = loop_head->fast_out(j);
1419 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1420 Node *newphi = old_new[phi->_idx];
1421 _igvn.hash_delete( phi );
1422 _igvn.hash_delete( newphi );
1424 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1425 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1426 phi ->set_req(LoopNode::LoopBackControl, C->top());
1427 }
1428 }
1429 Node *clone_head = old_new[loop_head->_idx];
1430 _igvn.hash_delete( clone_head );
1431 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1432 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1433 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1434 loop->_head = clone_head; // New loop header
1436 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1437 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1439 // Kill the clone's backedge
1440 Node *newcle = old_new[loop_end->_idx];
1441 _igvn.hash_delete( newcle );
1442 Node *one = _igvn.intcon(1);
1443 set_ctrl(one, C->root());
1444 newcle->set_req(1, one);
1445 // Force clone into same loop body
1446 uint max = loop->_body.size();
1447 for( uint k = 0; k < max; k++ ) {
1448 Node *old = loop->_body.at(k);
1449 Node *nnn = old_new[old->_idx];
1450 loop->_body.push(nnn);
1451 if (!has_ctrl(old))
1452 set_loop(nnn, loop);
1453 }
1455 loop->record_for_igvn();
1456 }
1458 //------------------------------do_maximally_unroll----------------------------
1460 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1461 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1462 assert(cl->has_exact_trip_count(), "trip count is not exact");
1463 assert(cl->trip_count() > 0, "");
1464 #ifndef PRODUCT
1465 if (TraceLoopOpts) {
1466 tty->print("MaxUnroll %d ", cl->trip_count());
1467 loop->dump_head();
1468 }
1469 #endif
1471 // If loop is tripping an odd number of times, peel odd iteration
1472 if ((cl->trip_count() & 1) == 1) {
1473 do_peeling(loop, old_new);
1474 }
1476 // Now its tripping an even number of times remaining. Double loop body.
1477 // Do not adjust pre-guards; they are not needed and do not exist.
1478 if (cl->trip_count() > 0) {
1479 assert((cl->trip_count() & 1) == 0, "missed peeling");
1480 do_unroll(loop, old_new, false);
1481 }
1482 }
1484 //------------------------------dominates_backedge---------------------------------
1485 // Returns true if ctrl is executed on every complete iteration
1486 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1487 assert(ctrl->is_CFG(), "must be control");
1488 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1489 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1490 }
1492 //------------------------------adjust_limit-----------------------------------
1493 // Helper function for add_constraint().
1494 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1495 // Compute "I :: (limit-offset)/scale"
1496 Node *con = new (C, 3) SubINode(rc_limit, offset);
1497 register_new_node(con, pre_ctrl);
1498 Node *X = new (C, 3) DivINode(0, con, scale);
1499 register_new_node(X, pre_ctrl);
1501 // Adjust loop limit
1502 loop_limit = (stride_con > 0)
1503 ? (Node*)(new (C, 3) MinINode(loop_limit, X))
1504 : (Node*)(new (C, 3) MaxINode(loop_limit, X));
1505 register_new_node(loop_limit, pre_ctrl);
1506 return loop_limit;
1507 }
1509 //------------------------------add_constraint---------------------------------
1510 // Constrain the main loop iterations so the conditions:
1511 // low_limit <= scale_con * I + offset < upper_limit
1512 // always holds true. That is, either increase the number of iterations in
1513 // the pre-loop or the post-loop until the condition holds true in the main
1514 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1515 // stride and scale are constants (offset and limit often are).
1516 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 ) {
1517 // For positive stride, the pre-loop limit always uses a MAX function
1518 // and the main loop a MIN function. For negative stride these are
1519 // reversed.
1521 // Also for positive stride*scale the affine function is increasing, so the
1522 // pre-loop must check for underflow and the post-loop for overflow.
1523 // Negative stride*scale reverses this; pre-loop checks for overflow and
1524 // post-loop for underflow.
1526 Node *scale = _igvn.intcon(scale_con);
1527 set_ctrl(scale, C->root());
1529 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1530 // The overflow limit: scale*I+offset < upper_limit
1531 // For main-loop compute
1532 // ( if (scale > 0) /* and stride > 0 */
1533 // I < (upper_limit-offset)/scale
1534 // else /* scale < 0 and stride < 0 */
1535 // I > (upper_limit-offset)/scale
1536 // )
1537 //
1538 // (upper_limit-offset) may overflow or underflow.
1539 // But it is fine since main loop will either have
1540 // less iterations or will be skipped in such case.
1541 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1543 // The underflow limit: low_limit <= scale*I+offset.
1544 // For pre-loop compute
1545 // NOT(scale*I+offset >= low_limit)
1546 // scale*I+offset < low_limit
1547 // ( if (scale > 0) /* and stride > 0 */
1548 // I < (low_limit-offset)/scale
1549 // else /* scale < 0 and stride < 0 */
1550 // I > (low_limit-offset)/scale
1551 // )
1553 if (low_limit->get_int() == -max_jint) {
1554 if (!RangeLimitCheck) return;
1555 // We need this guard when scale*pre_limit+offset >= limit
1556 // due to underflow. So we need execute pre-loop until
1557 // scale*I+offset >= min_int. But (min_int-offset) will
1558 // underflow when offset > 0 and X will be > original_limit
1559 // when stride > 0. To avoid it we replace positive offset with 0.
1560 //
1561 // Also (min_int+1 == -max_int) is used instead of min_int here
1562 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1563 Node* shift = _igvn.intcon(31);
1564 set_ctrl(shift, C->root());
1565 Node* sign = new (C, 3) RShiftINode(offset, shift);
1566 register_new_node(sign, pre_ctrl);
1567 offset = new (C, 3) AndINode(offset, sign);
1568 register_new_node(offset, pre_ctrl);
1569 } else {
1570 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1571 // The only problem we have here when offset == min_int
1572 // since (0-min_int) == min_int. It may be fine for stride > 0
1573 // but for stride < 0 X will be < original_limit. To avoid it
1574 // max(pre_limit, original_limit) is used in do_range_check().
1575 }
1576 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1577 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1579 } else { // stride_con*scale_con < 0
1580 // For negative stride*scale pre-loop checks for overflow and
1581 // post-loop for underflow.
1582 //
1583 // The overflow limit: scale*I+offset < upper_limit
1584 // For pre-loop compute
1585 // NOT(scale*I+offset < upper_limit)
1586 // scale*I+offset >= upper_limit
1587 // scale*I+offset+1 > upper_limit
1588 // ( if (scale < 0) /* and stride > 0 */
1589 // I < (upper_limit-(offset+1))/scale
1590 // else /* scale > 0 and stride < 0 */
1591 // I > (upper_limit-(offset+1))/scale
1592 // )
1593 //
1594 // (upper_limit-offset-1) may underflow or overflow.
1595 // To avoid it min(pre_limit, original_limit) is used
1596 // in do_range_check() for stride > 0 and max() for < 0.
1597 Node *one = _igvn.intcon(1);
1598 set_ctrl(one, C->root());
1600 Node *plus_one = new (C, 3) AddINode(offset, one);
1601 register_new_node( plus_one, pre_ctrl );
1602 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1603 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1605 if (low_limit->get_int() == -max_jint) {
1606 if (!RangeLimitCheck) return;
1607 // We need this guard when scale*main_limit+offset >= limit
1608 // due to underflow. So we need execute main-loop while
1609 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1610 // underflow when (offset+1) > 0 and X will be < main_limit
1611 // when scale < 0 (and stride > 0). To avoid it we replace
1612 // positive (offset+1) with 0.
1613 //
1614 // Also (min_int+1 == -max_int) is used instead of min_int here
1615 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1616 Node* shift = _igvn.intcon(31);
1617 set_ctrl(shift, C->root());
1618 Node* sign = new (C, 3) RShiftINode(plus_one, shift);
1619 register_new_node(sign, pre_ctrl);
1620 plus_one = new (C, 3) AndINode(plus_one, sign);
1621 register_new_node(plus_one, pre_ctrl);
1622 } else {
1623 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1624 // The only problem we have here when offset == max_int
1625 // since (max_int+1) == min_int and (0-min_int) == min_int.
1626 // But it is fine since main loop will either have
1627 // less iterations or will be skipped in such case.
1628 }
1629 // The underflow limit: low_limit <= scale*I+offset.
1630 // For main-loop compute
1631 // scale*I+offset+1 > low_limit
1632 // ( if (scale < 0) /* and stride > 0 */
1633 // I < (low_limit-(offset+1))/scale
1634 // else /* scale > 0 and stride < 0 */
1635 // I > (low_limit-(offset+1))/scale
1636 // )
1638 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1639 }
1640 }
1643 //------------------------------is_scaled_iv---------------------------------
1644 // Return true if exp is a constant times an induction var
1645 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1646 if (exp == iv) {
1647 if (p_scale != NULL) {
1648 *p_scale = 1;
1649 }
1650 return true;
1651 }
1652 int opc = exp->Opcode();
1653 if (opc == Op_MulI) {
1654 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1655 if (p_scale != NULL) {
1656 *p_scale = exp->in(2)->get_int();
1657 }
1658 return true;
1659 }
1660 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1661 if (p_scale != NULL) {
1662 *p_scale = exp->in(1)->get_int();
1663 }
1664 return true;
1665 }
1666 } else if (opc == Op_LShiftI) {
1667 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1668 if (p_scale != NULL) {
1669 *p_scale = 1 << exp->in(2)->get_int();
1670 }
1671 return true;
1672 }
1673 }
1674 return false;
1675 }
1677 //-----------------------------is_scaled_iv_plus_offset------------------------------
1678 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1679 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1680 if (is_scaled_iv(exp, iv, p_scale)) {
1681 if (p_offset != NULL) {
1682 Node *zero = _igvn.intcon(0);
1683 set_ctrl(zero, C->root());
1684 *p_offset = zero;
1685 }
1686 return true;
1687 }
1688 int opc = exp->Opcode();
1689 if (opc == Op_AddI) {
1690 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1691 if (p_offset != NULL) {
1692 *p_offset = exp->in(2);
1693 }
1694 return true;
1695 }
1696 if (exp->in(2)->is_Con()) {
1697 Node* offset2 = NULL;
1698 if (depth < 2 &&
1699 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1700 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1701 if (p_offset != NULL) {
1702 Node *ctrl_off2 = get_ctrl(offset2);
1703 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
1704 register_new_node(offset, ctrl_off2);
1705 *p_offset = offset;
1706 }
1707 return true;
1708 }
1709 }
1710 } else if (opc == Op_SubI) {
1711 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1712 if (p_offset != NULL) {
1713 Node *zero = _igvn.intcon(0);
1714 set_ctrl(zero, C->root());
1715 Node *ctrl_off = get_ctrl(exp->in(2));
1716 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
1717 register_new_node(offset, ctrl_off);
1718 *p_offset = offset;
1719 }
1720 return true;
1721 }
1722 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1723 if (p_offset != NULL) {
1724 *p_scale *= -1;
1725 *p_offset = exp->in(1);
1726 }
1727 return true;
1728 }
1729 }
1730 return false;
1731 }
1733 //------------------------------do_range_check---------------------------------
1734 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1735 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1736 #ifndef PRODUCT
1737 if (PrintOpto && VerifyLoopOptimizations) {
1738 tty->print("Range Check Elimination ");
1739 loop->dump_head();
1740 } else if (TraceLoopOpts) {
1741 tty->print("RangeCheck ");
1742 loop->dump_head();
1743 }
1744 #endif
1745 assert(RangeCheckElimination, "");
1746 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1747 assert(cl->is_main_loop(), "");
1749 // protect against stride not being a constant
1750 if (!cl->stride_is_con())
1751 return;
1753 // Find the trip counter; we are iteration splitting based on it
1754 Node *trip_counter = cl->phi();
1755 // Find the main loop limit; we will trim it's iterations
1756 // to not ever trip end tests
1757 Node *main_limit = cl->limit();
1759 // Need to find the main-loop zero-trip guard
1760 Node *ctrl = cl->in(LoopNode::EntryControl);
1761 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1762 Node *iffm = ctrl->in(0);
1763 assert(iffm->Opcode() == Op_If, "");
1764 Node *bolzm = iffm->in(1);
1765 assert(bolzm->Opcode() == Op_Bool, "");
1766 Node *cmpzm = bolzm->in(1);
1767 assert(cmpzm->is_Cmp(), "");
1768 Node *opqzm = cmpzm->in(2);
1769 // Can not optimize a loop if zero-trip Opaque1 node is optimized
1770 // away and then another round of loop opts attempted.
1771 if (opqzm->Opcode() != Op_Opaque1)
1772 return;
1773 assert(opqzm->in(1) == main_limit, "do not understand situation");
1775 // Find the pre-loop limit; we will expand it's iterations to
1776 // not ever trip low tests.
1777 Node *p_f = iffm->in(0);
1778 assert(p_f->Opcode() == Op_IfFalse, "");
1779 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1780 assert(pre_end->loopnode()->is_pre_loop(), "");
1781 Node *pre_opaq1 = pre_end->limit();
1782 // Occasionally it's possible for a pre-loop Opaque1 node to be
1783 // optimized away and then another round of loop opts attempted.
1784 // We can not optimize this particular loop in that case.
1785 if (pre_opaq1->Opcode() != Op_Opaque1)
1786 return;
1787 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1788 Node *pre_limit = pre_opaq->in(1);
1790 // Where do we put new limit calculations
1791 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1793 // Ensure the original loop limit is available from the
1794 // pre-loop Opaque1 node.
1795 Node *orig_limit = pre_opaq->original_loop_limit();
1796 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1797 return;
1799 // Must know if its a count-up or count-down loop
1801 int stride_con = cl->stride_con();
1802 Node *zero = _igvn.intcon(0);
1803 Node *one = _igvn.intcon(1);
1804 // Use symmetrical int range [-max_jint,max_jint]
1805 Node *mini = _igvn.intcon(-max_jint);
1806 set_ctrl(zero, C->root());
1807 set_ctrl(one, C->root());
1808 set_ctrl(mini, C->root());
1810 // Range checks that do not dominate the loop backedge (ie.
1811 // conditionally executed) can lengthen the pre loop limit beyond
1812 // the original loop limit. To prevent this, the pre limit is
1813 // (for stride > 0) MINed with the original loop limit (MAXed
1814 // stride < 0) when some range_check (rc) is conditionally
1815 // executed.
1816 bool conditional_rc = false;
1818 // Check loop body for tests of trip-counter plus loop-invariant vs
1819 // loop-invariant.
1820 for( uint i = 0; i < loop->_body.size(); i++ ) {
1821 Node *iff = loop->_body[i];
1822 if( iff->Opcode() == Op_If ) { // Test?
1824 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1825 // we need loop unswitching instead of iteration splitting.
1826 Node *exit = loop->is_loop_exit(iff);
1827 if( !exit ) continue;
1828 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1830 // Get boolean condition to test
1831 Node *i1 = iff->in(1);
1832 if( !i1->is_Bool() ) continue;
1833 BoolNode *bol = i1->as_Bool();
1834 BoolTest b_test = bol->_test;
1835 // Flip sense of test if exit condition is flipped
1836 if( flip )
1837 b_test = b_test.negate();
1839 // Get compare
1840 Node *cmp = bol->in(1);
1842 // Look for trip_counter + offset vs limit
1843 Node *rc_exp = cmp->in(1);
1844 Node *limit = cmp->in(2);
1845 jint scale_con= 1; // Assume trip counter not scaled
1847 Node *limit_c = get_ctrl(limit);
1848 if( loop->is_member(get_loop(limit_c) ) ) {
1849 // Compare might have operands swapped; commute them
1850 b_test = b_test.commute();
1851 rc_exp = cmp->in(2);
1852 limit = cmp->in(1);
1853 limit_c = get_ctrl(limit);
1854 if( loop->is_member(get_loop(limit_c) ) )
1855 continue; // Both inputs are loop varying; cannot RCE
1856 }
1857 // Here we know 'limit' is loop invariant
1859 // 'limit' maybe pinned below the zero trip test (probably from a
1860 // previous round of rce), in which case, it can't be used in the
1861 // zero trip test expression which must occur before the zero test's if.
1862 if( limit_c == ctrl ) {
1863 continue; // Don't rce this check but continue looking for other candidates.
1864 }
1866 // Check for scaled induction variable plus an offset
1867 Node *offset = NULL;
1869 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1870 continue;
1871 }
1873 Node *offset_c = get_ctrl(offset);
1874 if( loop->is_member( get_loop(offset_c) ) )
1875 continue; // Offset is not really loop invariant
1876 // Here we know 'offset' is loop invariant.
1878 // As above for the 'limit', the 'offset' maybe pinned below the
1879 // zero trip test.
1880 if( offset_c == ctrl ) {
1881 continue; // Don't rce this check but continue looking for other candidates.
1882 }
1883 #ifdef ASSERT
1884 if (TraceRangeLimitCheck) {
1885 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
1886 bol->dump(2);
1887 }
1888 #endif
1889 // At this point we have the expression as:
1890 // scale_con * trip_counter + offset :: limit
1891 // where scale_con, offset and limit are loop invariant. Trip_counter
1892 // monotonically increases by stride_con, a constant. Both (or either)
1893 // stride_con and scale_con can be negative which will flip about the
1894 // sense of the test.
1896 // Adjust pre and main loop limits to guard the correct iteration set
1897 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1898 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1899 // The underflow and overflow limits: 0 <= scale*I+offset < limit
1900 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
1901 if (!conditional_rc) {
1902 // (0-offset)/scale could be outside of loop iterations range.
1903 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1904 }
1905 } else {
1906 #ifndef PRODUCT
1907 if( PrintOpto )
1908 tty->print_cr("missed RCE opportunity");
1909 #endif
1910 continue; // In release mode, ignore it
1911 }
1912 } else { // Otherwise work on normal compares
1913 switch( b_test._test ) {
1914 case BoolTest::gt:
1915 // Fall into GE case
1916 case BoolTest::ge:
1917 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
1918 scale_con = -scale_con;
1919 offset = new (C, 3) SubINode( zero, offset );
1920 register_new_node( offset, pre_ctrl );
1921 limit = new (C, 3) SubINode( zero, limit );
1922 register_new_node( limit, pre_ctrl );
1923 // Fall into LE case
1924 case BoolTest::le:
1925 if (b_test._test != BoolTest::gt) {
1926 // Convert X <= Y to X < Y+1
1927 limit = new (C, 3) AddINode( limit, one );
1928 register_new_node( limit, pre_ctrl );
1929 }
1930 // Fall into LT case
1931 case BoolTest::lt:
1932 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
1933 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
1934 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
1935 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
1936 if (!conditional_rc) {
1937 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
1938 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
1939 // still be outside of loop range.
1940 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1941 }
1942 break;
1943 default:
1944 #ifndef PRODUCT
1945 if( PrintOpto )
1946 tty->print_cr("missed RCE opportunity");
1947 #endif
1948 continue; // Unhandled case
1949 }
1950 }
1952 // Kill the eliminated test
1953 C->set_major_progress();
1954 Node *kill_con = _igvn.intcon( 1-flip );
1955 set_ctrl(kill_con, C->root());
1956 _igvn.hash_delete(iff);
1957 iff->set_req(1, kill_con);
1958 _igvn._worklist.push(iff);
1959 // Find surviving projection
1960 assert(iff->is_If(), "");
1961 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1962 // Find loads off the surviving projection; remove their control edge
1963 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1964 Node* cd = dp->fast_out(i); // Control-dependent node
1965 if( cd->is_Load() ) { // Loads can now float around in the loop
1966 _igvn.hash_delete(cd);
1967 // Allow the load to float around in the loop, or before it
1968 // but NOT before the pre-loop.
1969 cd->set_req(0, ctrl); // ctrl, not NULL
1970 _igvn._worklist.push(cd);
1971 --i;
1972 --imax;
1973 }
1974 }
1976 } // End of is IF
1978 }
1980 // Update loop limits
1981 if (conditional_rc) {
1982 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
1983 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
1984 register_new_node(pre_limit, pre_ctrl);
1985 }
1986 _igvn.hash_delete(pre_opaq);
1987 pre_opaq->set_req(1, pre_limit);
1989 // Note:: we are making the main loop limit no longer precise;
1990 // need to round up based on stride.
1991 cl->set_nonexact_trip_count();
1992 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
1993 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1994 // Hopefully, compiler will optimize for powers of 2.
1995 Node *ctrl = get_ctrl(main_limit);
1996 Node *stride = cl->stride();
1997 Node *init = cl->init_trip();
1998 Node *span = new (C, 3) SubINode(main_limit,init);
1999 register_new_node(span,ctrl);
2000 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
2001 Node *add = new (C, 3) AddINode(span,rndup);
2002 register_new_node(add,ctrl);
2003 Node *div = new (C, 3) DivINode(0,add,stride);
2004 register_new_node(div,ctrl);
2005 Node *mul = new (C, 3) MulINode(div,stride);
2006 register_new_node(mul,ctrl);
2007 Node *newlim = new (C, 3) AddINode(mul,init);
2008 register_new_node(newlim,ctrl);
2009 main_limit = newlim;
2010 }
2012 Node *main_cle = cl->loopexit();
2013 Node *main_bol = main_cle->in(1);
2014 // Hacking loop bounds; need private copies of exit test
2015 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2016 _igvn.hash_delete(main_cle);
2017 main_bol = main_bol->clone();// Clone a private BoolNode
2018 register_new_node( main_bol, main_cle->in(0) );
2019 main_cle->set_req(1,main_bol);
2020 }
2021 Node *main_cmp = main_bol->in(1);
2022 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2023 _igvn.hash_delete(main_bol);
2024 main_cmp = main_cmp->clone();// Clone a private CmpNode
2025 register_new_node( main_cmp, main_cle->in(0) );
2026 main_bol->set_req(1,main_cmp);
2027 }
2028 // Hack the now-private loop bounds
2029 _igvn.hash_delete(main_cmp);
2030 main_cmp->set_req(2, main_limit);
2031 _igvn._worklist.push(main_cmp);
2032 // The OpaqueNode is unshared by design
2033 _igvn.hash_delete(opqzm);
2034 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2035 opqzm->set_req(1,main_limit);
2036 _igvn._worklist.push(opqzm);
2037 }
2039 //------------------------------DCE_loop_body----------------------------------
2040 // Remove simplistic dead code from loop body
2041 void IdealLoopTree::DCE_loop_body() {
2042 for( uint i = 0; i < _body.size(); i++ )
2043 if( _body.at(i)->outcnt() == 0 )
2044 _body.map( i--, _body.pop() );
2045 }
2048 //------------------------------adjust_loop_exit_prob--------------------------
2049 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2050 // Replace with a 1-in-10 exit guess.
2051 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2052 Node *test = tail();
2053 while( test != _head ) {
2054 uint top = test->Opcode();
2055 if( top == Op_IfTrue || top == Op_IfFalse ) {
2056 int test_con = ((ProjNode*)test)->_con;
2057 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2058 IfNode *iff = test->in(0)->as_If();
2059 if( iff->outcnt() == 2 ) { // Ignore dead tests
2060 Node *bol = iff->in(1);
2061 if( bol && bol->req() > 1 && bol->in(1) &&
2062 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2063 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2064 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2065 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2066 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2067 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2068 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2069 return; // Allocation loops RARELY take backedge
2070 // Find the OTHER exit path from the IF
2071 Node* ex = iff->proj_out(1-test_con);
2072 float p = iff->_prob;
2073 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2074 if( top == Op_IfTrue ) {
2075 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2076 iff->_prob = PROB_STATIC_FREQUENT;
2077 }
2078 } else {
2079 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2080 iff->_prob = PROB_STATIC_INFREQUENT;
2081 }
2082 }
2083 }
2084 }
2085 }
2086 test = phase->idom(test);
2087 }
2088 }
2091 //------------------------------policy_do_remove_empty_loop--------------------
2092 // Micro-benchmark spamming. Policy is to always remove empty loops.
2093 // The 'DO' part is to replace the trip counter with the value it will
2094 // have on the last iteration. This will break the loop.
2095 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2096 // Minimum size must be empty loop
2097 if (_body.size() > EMPTY_LOOP_SIZE)
2098 return false;
2100 if (!_head->is_CountedLoop())
2101 return false; // Dead loop
2102 CountedLoopNode *cl = _head->as_CountedLoop();
2103 if (!cl->loopexit())
2104 return false; // Malformed loop
2105 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2106 return false; // Infinite loop
2108 #ifdef ASSERT
2109 // Ensure only one phi which is the iv.
2110 Node* iv = NULL;
2111 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2112 Node* n = cl->fast_out(i);
2113 if (n->Opcode() == Op_Phi) {
2114 assert(iv == NULL, "Too many phis" );
2115 iv = n;
2116 }
2117 }
2118 assert(iv == cl->phi(), "Wrong phi" );
2119 #endif
2121 // main and post loops have explicitly created zero trip guard
2122 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2123 if (needs_guard) {
2124 // Skip guard if values not overlap.
2125 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2126 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2127 int stride_con = cl->stride_con();
2128 if (stride_con > 0) {
2129 needs_guard = (init_t->_hi >= limit_t->_lo);
2130 } else {
2131 needs_guard = (init_t->_lo <= limit_t->_hi);
2132 }
2133 }
2134 if (needs_guard) {
2135 // Check for an obvious zero trip guard.
2136 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2137 if (inctrl->Opcode() == Op_IfTrue) {
2138 // The test should look like just the backedge of a CountedLoop
2139 Node* iff = inctrl->in(0);
2140 if (iff->is_If()) {
2141 Node* bol = iff->in(1);
2142 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2143 Node* cmp = bol->in(1);
2144 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2145 needs_guard = false;
2146 }
2147 }
2148 }
2149 }
2150 }
2152 #ifndef PRODUCT
2153 if (PrintOpto) {
2154 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2155 this->dump_head();
2156 } else if (TraceLoopOpts) {
2157 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2158 this->dump_head();
2159 }
2160 #endif
2162 if (needs_guard) {
2163 // Peel the loop to ensure there's a zero trip guard
2164 Node_List old_new;
2165 phase->do_peeling(this, old_new);
2166 }
2168 // Replace the phi at loop head with the final value of the last
2169 // iteration. Then the CountedLoopEnd will collapse (backedge never
2170 // taken) and all loop-invariant uses of the exit values will be correct.
2171 Node *phi = cl->phi();
2172 Node *exact_limit = phase->exact_limit(this);
2173 if (exact_limit != cl->limit()) {
2174 // We also need to replace the original limit to collapse loop exit.
2175 Node* cmp = cl->loopexit()->cmp_node();
2176 assert(cl->limit() == cmp->in(2), "sanity");
2177 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2178 phase->_igvn.hash_delete(cmp);
2179 cmp->set_req(2, exact_limit);
2180 phase->_igvn._worklist.push(cmp); // put cmp on worklist
2181 }
2182 // Note: the final value after increment should not overflow since
2183 // counted loop has limit check predicate.
2184 Node *final = new (phase->C, 3) SubINode( exact_limit, cl->stride() );
2185 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2186 phase->_igvn.replace_node(phi,final);
2187 phase->C->set_major_progress();
2188 return true;
2189 }
2191 //------------------------------policy_do_one_iteration_loop-------------------
2192 // Convert one iteration loop into normal code.
2193 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2194 if (!_head->as_Loop()->is_valid_counted_loop())
2195 return false; // Only for counted loop
2197 CountedLoopNode *cl = _head->as_CountedLoop();
2198 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2199 return false;
2200 }
2202 #ifndef PRODUCT
2203 if(TraceLoopOpts) {
2204 tty->print("OneIteration ");
2205 this->dump_head();
2206 }
2207 #endif
2209 Node *init_n = cl->init_trip();
2210 #ifdef ASSERT
2211 // Loop boundaries should be constant since trip count is exact.
2212 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2213 #endif
2214 // Replace the phi at loop head with the value of the init_trip.
2215 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2216 // and all loop-invariant uses of the exit values will be correct.
2217 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2218 phase->C->set_major_progress();
2219 return true;
2220 }
2222 //=============================================================================
2223 //------------------------------iteration_split_impl---------------------------
2224 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2225 // Compute exact loop trip count if possible.
2226 compute_exact_trip_count(phase);
2228 // Convert one iteration loop into normal code.
2229 if (policy_do_one_iteration_loop(phase))
2230 return true;
2232 // Check and remove empty loops (spam micro-benchmarks)
2233 if (policy_do_remove_empty_loop(phase))
2234 return true; // Here we removed an empty loop
2236 bool should_peel = policy_peeling(phase); // Should we peel?
2238 bool should_unswitch = policy_unswitching(phase);
2240 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2241 // This removes loop-invariant tests (usually null checks).
2242 if (!_head->is_CountedLoop()) { // Non-counted loop
2243 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2244 // Partial peel succeeded so terminate this round of loop opts
2245 return false;
2246 }
2247 if (should_peel) { // Should we peel?
2248 #ifndef PRODUCT
2249 if (PrintOpto) tty->print_cr("should_peel");
2250 #endif
2251 phase->do_peeling(this,old_new);
2252 } else if (should_unswitch) {
2253 phase->do_unswitching(this, old_new);
2254 }
2255 return true;
2256 }
2257 CountedLoopNode *cl = _head->as_CountedLoop();
2259 if (!cl->loopexit()) return true; // Ignore various kinds of broken loops
2261 // Do nothing special to pre- and post- loops
2262 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2264 // Compute loop trip count from profile data
2265 compute_profile_trip_cnt(phase);
2267 // Before attempting fancy unrolling, RCE or alignment, see if we want
2268 // to completely unroll this loop or do loop unswitching.
2269 if (cl->is_normal_loop()) {
2270 if (should_unswitch) {
2271 phase->do_unswitching(this, old_new);
2272 return true;
2273 }
2274 bool should_maximally_unroll = policy_maximally_unroll(phase);
2275 if (should_maximally_unroll) {
2276 // Here we did some unrolling and peeling. Eventually we will
2277 // completely unroll this loop and it will no longer be a loop.
2278 phase->do_maximally_unroll(this,old_new);
2279 return true;
2280 }
2281 }
2283 // Skip next optimizations if running low on nodes. Note that
2284 // policy_unswitching and policy_maximally_unroll have this check.
2285 uint nodes_left = MaxNodeLimit - phase->C->unique();
2286 if ((2 * _body.size()) > nodes_left) {
2287 return true;
2288 }
2290 // Counted loops may be peeled, may need some iterations run up
2291 // front for RCE, and may want to align loop refs to a cache
2292 // line. Thus we clone a full loop up front whose trip count is
2293 // at least 1 (if peeling), but may be several more.
2295 // The main loop will start cache-line aligned with at least 1
2296 // iteration of the unrolled body (zero-trip test required) and
2297 // will have some range checks removed.
2299 // A post-loop will finish any odd iterations (leftover after
2300 // unrolling), plus any needed for RCE purposes.
2302 bool should_unroll = policy_unroll(phase);
2304 bool should_rce = policy_range_check(phase);
2306 bool should_align = policy_align(phase);
2308 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2309 // need a pre-loop. We may still need to peel an initial iteration but
2310 // we will not be needing an unknown number of pre-iterations.
2311 //
2312 // Basically, if may_rce_align reports FALSE first time through,
2313 // we will not be able to later do RCE or Aligning on this loop.
2314 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2316 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2317 // we switch to the pre-/main-/post-loop model. This model also covers
2318 // peeling.
2319 if (should_rce || should_align || should_unroll) {
2320 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2321 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2323 // Adjust the pre- and main-loop limits to let the pre and post loops run
2324 // with full checks, but the main-loop with no checks. Remove said
2325 // checks from the main body.
2326 if (should_rce)
2327 phase->do_range_check(this,old_new);
2329 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2330 // twice as many iterations as before) and the main body limit (only do
2331 // an even number of trips). If we are peeling, we might enable some RCE
2332 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2333 // peeling.
2334 if (should_unroll && !should_peel)
2335 phase->do_unroll(this,old_new, true);
2337 // Adjust the pre-loop limits to align the main body
2338 // iterations.
2339 if (should_align)
2340 Unimplemented();
2342 } else { // Else we have an unchanged counted loop
2343 if (should_peel) // Might want to peel but do nothing else
2344 phase->do_peeling(this,old_new);
2345 }
2346 return true;
2347 }
2350 //=============================================================================
2351 //------------------------------iteration_split--------------------------------
2352 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2353 // Recursively iteration split nested loops
2354 if (_child && !_child->iteration_split(phase, old_new))
2355 return false;
2357 // Clean out prior deadwood
2358 DCE_loop_body();
2361 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2362 // Replace with a 1-in-10 exit guess.
2363 if (_parent /*not the root loop*/ &&
2364 !_irreducible &&
2365 // Also ignore the occasional dead backedge
2366 !tail()->is_top()) {
2367 adjust_loop_exit_prob(phase);
2368 }
2370 // Gate unrolling, RCE and peeling efforts.
2371 if (!_child && // If not an inner loop, do not split
2372 !_irreducible &&
2373 _allow_optimizations &&
2374 !tail()->is_top()) { // Also ignore the occasional dead backedge
2375 if (!_has_call) {
2376 if (!iteration_split_impl(phase, old_new)) {
2377 return false;
2378 }
2379 } else if (policy_unswitching(phase)) {
2380 phase->do_unswitching(this, old_new);
2381 }
2382 }
2384 // Minor offset re-organization to remove loop-fallout uses of
2385 // trip counter when there was no major reshaping.
2386 phase->reorg_offsets(this);
2388 if (_next && !_next->iteration_split(phase, old_new))
2389 return false;
2390 return true;
2391 }
2394 //=============================================================================
2395 // Process all the loops in the loop tree and replace any fill
2396 // patterns with an intrisc version.
2397 bool PhaseIdealLoop::do_intrinsify_fill() {
2398 bool changed = false;
2399 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2400 IdealLoopTree* lpt = iter.current();
2401 changed |= intrinsify_fill(lpt);
2402 }
2403 return changed;
2404 }
2407 // Examine an inner loop looking for a a single store of an invariant
2408 // value in a unit stride loop,
2409 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2410 Node*& shift, Node*& con) {
2411 const char* msg = NULL;
2412 Node* msg_node = NULL;
2414 store_value = NULL;
2415 con = NULL;
2416 shift = NULL;
2418 // Process the loop looking for stores. If there are multiple
2419 // stores or extra control flow give at this point.
2420 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2421 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2422 Node* n = lpt->_body.at(i);
2423 if (n->outcnt() == 0) continue; // Ignore dead
2424 if (n->is_Store()) {
2425 if (store != NULL) {
2426 msg = "multiple stores";
2427 break;
2428 }
2429 int opc = n->Opcode();
2430 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) {
2431 msg = "oop fills not handled";
2432 break;
2433 }
2434 Node* value = n->in(MemNode::ValueIn);
2435 if (!lpt->is_invariant(value)) {
2436 msg = "variant store value";
2437 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2438 msg = "not array address";
2439 }
2440 store = n;
2441 store_value = value;
2442 } else if (n->is_If() && n != head->loopexit()) {
2443 msg = "extra control flow";
2444 msg_node = n;
2445 }
2446 }
2448 if (store == NULL) {
2449 // No store in loop
2450 return false;
2451 }
2453 if (msg == NULL && head->stride_con() != 1) {
2454 // could handle negative strides too
2455 if (head->stride_con() < 0) {
2456 msg = "negative stride";
2457 } else {
2458 msg = "non-unit stride";
2459 }
2460 }
2462 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2463 msg = "can't handle store address";
2464 msg_node = store->in(MemNode::Address);
2465 }
2467 if (msg == NULL &&
2468 (!store->in(MemNode::Memory)->is_Phi() ||
2469 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2470 msg = "store memory isn't proper phi";
2471 msg_node = store->in(MemNode::Memory);
2472 }
2474 // Make sure there is an appropriate fill routine
2475 BasicType t = store->as_Mem()->memory_type();
2476 const char* fill_name;
2477 if (msg == NULL &&
2478 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2479 msg = "unsupported store";
2480 msg_node = store;
2481 }
2483 if (msg != NULL) {
2484 #ifndef PRODUCT
2485 if (TraceOptimizeFill) {
2486 tty->print_cr("not fill intrinsic candidate: %s", msg);
2487 if (msg_node != NULL) msg_node->dump();
2488 }
2489 #endif
2490 return false;
2491 }
2493 // Make sure the address expression can be handled. It should be
2494 // head->phi * elsize + con. head->phi might have a ConvI2L.
2495 Node* elements[4];
2496 Node* conv = NULL;
2497 bool found_index = false;
2498 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2499 for (int e = 0; e < count; e++) {
2500 Node* n = elements[e];
2501 if (n->is_Con() && con == NULL) {
2502 con = n;
2503 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2504 Node* value = n->in(1);
2505 #ifdef _LP64
2506 if (value->Opcode() == Op_ConvI2L) {
2507 conv = value;
2508 value = value->in(1);
2509 }
2510 #endif
2511 if (value != head->phi()) {
2512 msg = "unhandled shift in address";
2513 } else {
2514 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2515 msg = "scale doesn't match";
2516 } else {
2517 found_index = true;
2518 shift = n;
2519 }
2520 }
2521 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2522 if (n->in(1) == head->phi()) {
2523 found_index = true;
2524 conv = n;
2525 } else {
2526 msg = "unhandled input to ConvI2L";
2527 }
2528 } else if (n == head->phi()) {
2529 // no shift, check below for allowed cases
2530 found_index = true;
2531 } else {
2532 msg = "unhandled node in address";
2533 msg_node = n;
2534 }
2535 }
2537 if (count == -1) {
2538 msg = "malformed address expression";
2539 msg_node = store;
2540 }
2542 if (!found_index) {
2543 msg = "missing use of index";
2544 }
2546 // byte sized items won't have a shift
2547 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2548 msg = "can't find shift";
2549 msg_node = store;
2550 }
2552 if (msg != NULL) {
2553 #ifndef PRODUCT
2554 if (TraceOptimizeFill) {
2555 tty->print_cr("not fill intrinsic: %s", msg);
2556 if (msg_node != NULL) msg_node->dump();
2557 }
2558 #endif
2559 return false;
2560 }
2562 // No make sure all the other nodes in the loop can be handled
2563 VectorSet ok(Thread::current()->resource_area());
2565 // store related values are ok
2566 ok.set(store->_idx);
2567 ok.set(store->in(MemNode::Memory)->_idx);
2569 // Loop structure is ok
2570 ok.set(head->_idx);
2571 ok.set(head->loopexit()->_idx);
2572 ok.set(head->phi()->_idx);
2573 ok.set(head->incr()->_idx);
2574 ok.set(head->loopexit()->cmp_node()->_idx);
2575 ok.set(head->loopexit()->in(1)->_idx);
2577 // Address elements are ok
2578 if (con) ok.set(con->_idx);
2579 if (shift) ok.set(shift->_idx);
2580 if (conv) ok.set(conv->_idx);
2582 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2583 Node* n = lpt->_body.at(i);
2584 if (n->outcnt() == 0) continue; // Ignore dead
2585 if (ok.test(n->_idx)) continue;
2586 // Backedge projection is ok
2587 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue;
2588 if (!n->is_AddP()) {
2589 msg = "unhandled node";
2590 msg_node = n;
2591 break;
2592 }
2593 }
2595 // Make sure no unexpected values are used outside the loop
2596 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2597 Node* n = lpt->_body.at(i);
2598 // These values can be replaced with other nodes if they are used
2599 // outside the loop.
2600 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2601 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2602 Node* use = iter.get();
2603 if (!lpt->_body.contains(use)) {
2604 msg = "node is used outside loop";
2605 // lpt->_body.dump();
2606 msg_node = n;
2607 break;
2608 }
2609 }
2610 }
2612 #ifdef ASSERT
2613 if (TraceOptimizeFill) {
2614 if (msg != NULL) {
2615 tty->print_cr("no fill intrinsic: %s", msg);
2616 if (msg_node != NULL) msg_node->dump();
2617 } else {
2618 tty->print_cr("fill intrinsic for:");
2619 }
2620 store->dump();
2621 if (Verbose) {
2622 lpt->_body.dump();
2623 }
2624 }
2625 #endif
2627 return msg == NULL;
2628 }
2632 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2633 // Only for counted inner loops
2634 if (!lpt->is_counted() || !lpt->is_inner()) {
2635 return false;
2636 }
2638 // Must have constant stride
2639 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2640 if (!head->stride_is_con() || !head->is_normal_loop()) {
2641 return false;
2642 }
2644 // Check that the body only contains a store of a loop invariant
2645 // value that is indexed by the loop phi.
2646 Node* store = NULL;
2647 Node* store_value = NULL;
2648 Node* shift = NULL;
2649 Node* offset = NULL;
2650 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2651 return false;
2652 }
2654 #ifndef PRODUCT
2655 if (TraceLoopOpts) {
2656 tty->print("ArrayFill ");
2657 lpt->dump_head();
2658 }
2659 #endif
2661 // Now replace the whole loop body by a call to a fill routine that
2662 // covers the same region as the loop.
2663 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2665 // Build an expression for the beginning of the copy region
2666 Node* index = head->init_trip();
2667 #ifdef _LP64
2668 index = new (C, 2) ConvI2LNode(index);
2669 _igvn.register_new_node_with_optimizer(index);
2670 #endif
2671 if (shift != NULL) {
2672 // byte arrays don't require a shift but others do.
2673 index = new (C, 3) LShiftXNode(index, shift->in(2));
2674 _igvn.register_new_node_with_optimizer(index);
2675 }
2676 index = new (C, 4) AddPNode(base, base, index);
2677 _igvn.register_new_node_with_optimizer(index);
2678 Node* from = new (C, 4) AddPNode(base, index, offset);
2679 _igvn.register_new_node_with_optimizer(from);
2680 // Compute the number of elements to copy
2681 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip());
2682 _igvn.register_new_node_with_optimizer(len);
2684 BasicType t = store->as_Mem()->memory_type();
2685 bool aligned = false;
2686 if (offset != NULL && head->init_trip()->is_Con()) {
2687 int element_size = type2aelembytes(t);
2688 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2689 }
2691 // Build a call to the fill routine
2692 const char* fill_name;
2693 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2694 assert(fill != NULL, "what?");
2696 // Convert float/double to int/long for fill routines
2697 if (t == T_FLOAT) {
2698 store_value = new (C, 2) MoveF2INode(store_value);
2699 _igvn.register_new_node_with_optimizer(store_value);
2700 } else if (t == T_DOUBLE) {
2701 store_value = new (C, 2) MoveD2LNode(store_value);
2702 _igvn.register_new_node_with_optimizer(store_value);
2703 }
2705 Node* mem_phi = store->in(MemNode::Memory);
2706 Node* result_ctrl;
2707 Node* result_mem;
2708 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2709 int size = call_type->domain()->cnt();
2710 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill,
2711 fill_name, TypeAryPtr::get_array_body_type(t));
2712 call->init_req(TypeFunc::Parms+0, from);
2713 call->init_req(TypeFunc::Parms+1, store_value);
2714 #ifdef _LP64
2715 len = new (C, 2) ConvI2LNode(len);
2716 _igvn.register_new_node_with_optimizer(len);
2717 #endif
2718 call->init_req(TypeFunc::Parms+2, len);
2719 #ifdef _LP64
2720 call->init_req(TypeFunc::Parms+3, C->top());
2721 #endif
2722 call->init_req( TypeFunc::Control, head->init_control());
2723 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o
2724 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) );
2725 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) );
2726 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) );
2727 _igvn.register_new_node_with_optimizer(call);
2728 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
2729 _igvn.register_new_node_with_optimizer(result_ctrl);
2730 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory);
2731 _igvn.register_new_node_with_optimizer(result_mem);
2733 // If this fill is tightly coupled to an allocation and overwrites
2734 // the whole body, allow it to take over the zeroing.
2735 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2736 if (alloc != NULL && alloc->is_AllocateArray()) {
2737 Node* length = alloc->as_AllocateArray()->Ideal_length();
2738 if (head->limit() == length &&
2739 head->init_trip() == _igvn.intcon(0)) {
2740 if (TraceOptimizeFill) {
2741 tty->print_cr("Eliminated zeroing in allocation");
2742 }
2743 alloc->maybe_set_complete(&_igvn);
2744 } else {
2745 #ifdef ASSERT
2746 if (TraceOptimizeFill) {
2747 tty->print_cr("filling array but bounds don't match");
2748 alloc->dump();
2749 head->init_trip()->dump();
2750 head->limit()->dump();
2751 length->dump();
2752 }
2753 #endif
2754 }
2755 }
2757 // Redirect the old control and memory edges that are outside the loop.
2758 Node* exit = head->loopexit()->proj_out(0);
2759 // Sometimes the memory phi of the head is used as the outgoing
2760 // state of the loop. It's safe in this case to replace it with the
2761 // result_mem.
2762 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2763 _igvn.replace_node(exit, result_ctrl);
2764 _igvn.replace_node(store, result_mem);
2765 // Any uses the increment outside of the loop become the loop limit.
2766 _igvn.replace_node(head->incr(), head->limit());
2768 // Disconnect the head from the loop.
2769 for (uint i = 0; i < lpt->_body.size(); i++) {
2770 Node* n = lpt->_body.at(i);
2771 _igvn.replace_node(n, C->top());
2772 }
2774 return true;
2775 }