Wed, 18 Sep 2013 14:34:56 -0700
8024342: PPC64 (part 111): Support for C calling conventions that require 64-bit ints.
Summary: Some platforms, as ppc and s390x/zArch require that 32-bit ints are passed as 64-bit values to C functions. This change adds support to adapt the signature and to issue proper casts to c2-compiled stubs. The functions are used in generate_native_wrapper(). Adapt signature used by the compiler as in PhaseIdealLoop::intrinsify_fill().
Reviewed-by: kvn
1 /*
2 * Copyright (c) 2000, 2012, 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 jlong init_con = cl->init_trip()->get_int();
96 jlong limit_con = cl->limit()->get_int();
97 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
98 jlong 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) 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) SubINode(n_inv1, inv2);
235 } else {
236 inv = new (phase->C) 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) SubINode(inv, x);
243 } else {
244 addx = new (phase->C) 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 live_node_count = phase->C->live_nodes();
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 + live_node_count > MaxNodeLimit) ) {
276 return false; // too large to safely clone
277 }
278 while( test != _head ) { // Scan till run off top of loop
279 if( test->is_If() ) { // Test?
280 Node *ctrl = phase->get_ctrl(test->in(1));
281 if (ctrl->is_top())
282 return false; // Found dead test on live IF? No peeling!
283 // Standard IF only has one input value to check for loop invariance
284 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
285 // Condition is not a member of this loop?
286 if( !is_member(phase->get_loop(ctrl)) &&
287 is_loop_exit(test) )
288 return true; // Found reason to peel!
289 }
290 // Walk up dominators to loop _head looking for test which is
291 // executed on every path thru loop.
292 test = phase->idom(test);
293 }
294 return false;
295 }
297 //------------------------------peeled_dom_test_elim---------------------------
298 // If we got the effect of peeling, either by actually peeling or by making
299 // a pre-loop which must execute at least once, we can remove all
300 // loop-invariant dominated tests in the main body.
301 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
302 bool progress = true;
303 while( progress ) {
304 progress = false; // Reset for next iteration
305 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
306 Node *test = prev->in(0);
307 while( test != loop->_head ) { // Scan till run off top of loop
309 int p_op = prev->Opcode();
310 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
311 test->is_If() && // Test?
312 !test->in(1)->is_Con() && // And not already obvious?
313 // Condition is not a member of this loop?
314 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
315 // Walk loop body looking for instances of this test
316 for( uint i = 0; i < loop->_body.size(); i++ ) {
317 Node *n = loop->_body.at(i);
318 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
319 // IfNode was dominated by version in peeled loop body
320 progress = true;
321 dominated_by( old_new[prev->_idx], n );
322 }
323 }
324 }
325 prev = test;
326 test = idom(test);
327 } // End of scan tests in loop
329 } // End of while( progress )
330 }
332 //------------------------------do_peeling-------------------------------------
333 // Peel the first iteration of the given loop.
334 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
335 // The pre-loop illegally has 2 control users (old & new loops).
336 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
337 // Do this by making the old-loop fall-in edges act as if they came
338 // around the loopback from the prior iteration (follow the old-loop
339 // backedges) and then map to the new peeled iteration. This leaves
340 // the pre-loop with only 1 user (the new peeled iteration), but the
341 // peeled-loop backedge has 2 users.
342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
343 // extra backedge user.
344 //
345 // orig
346 //
347 // stmt1
348 // |
349 // v
350 // loop predicate
351 // |
352 // v
353 // loop<----+
354 // | |
355 // stmt2 |
356 // | |
357 // v |
358 // if ^
359 // / \ |
360 // / \ |
361 // v v |
362 // false true |
363 // / \ |
364 // / ----+
365 // |
366 // v
367 // exit
368 //
369 //
370 // after clone loop
371 //
372 // stmt1
373 // |
374 // v
375 // loop predicate
376 // / \
377 // clone / \ orig
378 // / \
379 // / \
380 // v v
381 // +---->loop clone loop<----+
382 // | | | |
383 // | stmt2 clone stmt2 |
384 // | | | |
385 // | v v |
386 // ^ if clone If ^
387 // | / \ / \ |
388 // | / \ / \ |
389 // | v v v v |
390 // | true false false true |
391 // | / \ / \ |
392 // +---- \ / ----+
393 // \ /
394 // 1v v2
395 // region
396 // |
397 // v
398 // exit
399 //
400 //
401 // after peel and predicate move
402 //
403 // stmt1
404 // /
405 // /
406 // clone / orig
407 // /
408 // / +----------+
409 // / | |
410 // / loop predicate |
411 // / | |
412 // v v |
413 // TOP-->loop clone loop<----+ |
414 // | | | |
415 // stmt2 clone stmt2 | |
416 // | | | ^
417 // v v | |
418 // if clone If ^ |
419 // / \ / \ | |
420 // / \ / \ | |
421 // v v v v | |
422 // true false false true | |
423 // | \ / \ | |
424 // | \ / ----+ ^
425 // | \ / |
426 // | 1v v2 |
427 // v region |
428 // | | |
429 // | v |
430 // | exit |
431 // | |
432 // +--------------->-----------------+
433 //
434 //
435 // final graph
436 //
437 // stmt1
438 // |
439 // v
440 // stmt2 clone
441 // |
442 // v
443 // if clone
444 // / |
445 // / |
446 // v v
447 // false true
448 // | |
449 // | v
450 // | loop predicate
451 // | |
452 // | v
453 // | loop<----+
454 // | | |
455 // | stmt2 |
456 // | | |
457 // | v |
458 // v if ^
459 // | / \ |
460 // | / \ |
461 // | v v |
462 // | false true |
463 // | | \ |
464 // v v --+
465 // region
466 // |
467 // v
468 // exit
469 //
470 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
472 C->set_major_progress();
473 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
474 // 'pre' loop from the main and the 'pre' can no longer have it's
475 // iterations adjusted. Therefore, we need to declare this loop as
476 // no longer a 'main' loop; it will need new pre and post loops before
477 // we can do further RCE.
478 #ifndef PRODUCT
479 if (TraceLoopOpts) {
480 tty->print("Peel ");
481 loop->dump_head();
482 }
483 #endif
484 Node* head = loop->_head;
485 bool counted_loop = head->is_CountedLoop();
486 if (counted_loop) {
487 CountedLoopNode *cl = head->as_CountedLoop();
488 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
489 cl->set_trip_count(cl->trip_count() - 1);
490 if (cl->is_main_loop()) {
491 cl->set_normal_loop();
492 #ifndef PRODUCT
493 if (PrintOpto && VerifyLoopOptimizations) {
494 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
495 loop->dump_head();
496 }
497 #endif
498 }
499 }
500 Node* entry = head->in(LoopNode::EntryControl);
502 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
503 // The pre-loop illegally has 2 control users (old & new loops).
504 clone_loop( loop, old_new, dom_depth(head) );
506 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
507 // Do this by making the old-loop fall-in edges act as if they came
508 // around the loopback from the prior iteration (follow the old-loop
509 // backedges) and then map to the new peeled iteration. This leaves
510 // the pre-loop with only 1 user (the new peeled iteration), but the
511 // peeled-loop backedge has 2 users.
512 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
513 _igvn.hash_delete(head);
514 head->set_req(LoopNode::EntryControl, new_entry);
515 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
516 Node* old = head->fast_out(j);
517 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
518 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
519 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
520 // Then loop body backedge value remains the same.
521 new_exit_value = old->in(LoopNode::LoopBackControl);
522 _igvn.hash_delete(old);
523 old->set_req(LoopNode::EntryControl, new_exit_value);
524 }
525 }
528 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
529 // extra backedge user.
530 Node* new_head = old_new[head->_idx];
531 _igvn.hash_delete(new_head);
532 new_head->set_req(LoopNode::LoopBackControl, C->top());
533 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
534 Node* use = new_head->fast_out(j2);
535 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
536 _igvn.hash_delete(use);
537 use->set_req(LoopNode::LoopBackControl, C->top());
538 }
539 }
542 // Step 4: Correct dom-depth info. Set to loop-head depth.
543 int dd = dom_depth(head);
544 set_idom(head, head->in(1), dd);
545 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
546 Node *old = loop->_body.at(j3);
547 Node *nnn = old_new[old->_idx];
548 if (!has_ctrl(nnn))
549 set_idom(nnn, idom(nnn), dd-1);
550 }
552 // Now force out all loop-invariant dominating tests. The optimizer
553 // finds some, but we _know_ they are all useless.
554 peeled_dom_test_elim(loop,old_new);
556 loop->record_for_igvn();
557 }
559 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
561 //------------------------------policy_maximally_unroll------------------------
562 // Calculate exact loop trip count and return true if loop can be maximally
563 // unrolled.
564 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
565 CountedLoopNode *cl = _head->as_CountedLoop();
566 assert(cl->is_normal_loop(), "");
567 if (!cl->is_valid_counted_loop())
568 return false; // Malformed counted loop
570 if (!cl->has_exact_trip_count()) {
571 // Trip count is not exact.
572 return false;
573 }
575 uint trip_count = cl->trip_count();
576 // Note, max_juint is used to indicate unknown trip count.
577 assert(trip_count > 1, "one iteration loop should be optimized out already");
578 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
580 // Real policy: if we maximally unroll, does it get too big?
581 // Allow the unrolled mess to get larger than standard loop
582 // size. After all, it will no longer be a loop.
583 uint body_size = _body.size();
584 uint unroll_limit = (uint)LoopUnrollLimit * 4;
585 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
586 if (trip_count > unroll_limit || body_size > unroll_limit) {
587 return false;
588 }
590 // Fully unroll a loop with few iterations regardless next
591 // conditions since following loop optimizations will split
592 // such loop anyway (pre-main-post).
593 if (trip_count <= 3)
594 return true;
596 // Take into account that after unroll conjoined heads and tails will fold,
597 // otherwise policy_unroll() may allow more unrolling than max unrolling.
598 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
599 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
600 if (body_size != tst_body_size) // Check for int overflow
601 return false;
602 if (new_body_size > unroll_limit ||
603 // Unrolling can result in a large amount of node construction
604 new_body_size >= MaxNodeLimit - (uint) phase->C->live_nodes()) {
605 return false;
606 }
608 // Do not unroll a loop with String intrinsics code.
609 // String intrinsics are large and have loops.
610 for (uint k = 0; k < _body.size(); k++) {
611 Node* n = _body.at(k);
612 switch (n->Opcode()) {
613 case Op_StrComp:
614 case Op_StrEquals:
615 case Op_StrIndexOf:
616 case Op_EncodeISOArray:
617 case Op_AryEq: {
618 return false;
619 }
620 } // switch
621 }
623 return true; // Do maximally unroll
624 }
627 //------------------------------policy_unroll----------------------------------
628 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
629 // the loop is a CountedLoop and the body is small enough.
630 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
632 CountedLoopNode *cl = _head->as_CountedLoop();
633 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
635 if (!cl->is_valid_counted_loop())
636 return false; // Malformed counted loop
638 // Protect against over-unrolling.
639 // After split at least one iteration will be executed in pre-loop.
640 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
642 int future_unroll_ct = cl->unrolled_count() * 2;
643 if (future_unroll_ct > LoopMaxUnroll) return false;
645 // Check for initial stride being a small enough constant
646 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
648 // Don't unroll if the next round of unrolling would push us
649 // over the expected trip count of the loop. One is subtracted
650 // from the expected trip count because the pre-loop normally
651 // executes 1 iteration.
652 if (UnrollLimitForProfileCheck > 0 &&
653 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
654 future_unroll_ct > UnrollLimitForProfileCheck &&
655 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
656 return false;
657 }
659 // When unroll count is greater than LoopUnrollMin, don't unroll if:
660 // the residual iterations are more than 10% of the trip count
661 // and rounds of "unroll,optimize" are not making significant progress
662 // Progress defined as current size less than 20% larger than previous size.
663 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
664 future_unroll_ct > LoopUnrollMin &&
665 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
666 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
667 return false;
668 }
670 Node *init_n = cl->init_trip();
671 Node *limit_n = cl->limit();
672 int stride_con = cl->stride_con();
673 // Non-constant bounds.
674 // Protect against over-unrolling when init or/and limit are not constant
675 // (so that trip_count's init value is maxint) but iv range is known.
676 if (init_n == NULL || !init_n->is_Con() ||
677 limit_n == NULL || !limit_n->is_Con()) {
678 Node* phi = cl->phi();
679 if (phi != NULL) {
680 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
681 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
682 int next_stride = stride_con * 2; // stride after this unroll
683 if (next_stride > 0) {
684 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
685 iv_type->_lo + next_stride > iv_type->_hi) {
686 return false; // over-unrolling
687 }
688 } else if (next_stride < 0) {
689 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
690 iv_type->_hi + next_stride < iv_type->_lo) {
691 return false; // over-unrolling
692 }
693 }
694 }
695 }
697 // After unroll limit will be adjusted: new_limit = limit-stride.
698 // Bailout if adjustment overflow.
699 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
700 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
701 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
702 return false; // overflow
704 // Adjust body_size to determine if we unroll or not
705 uint body_size = _body.size();
706 // Key test to unroll loop in CRC32 java code
707 int xors_in_loop = 0;
708 // Also count ModL, DivL and MulL which expand mightly
709 for (uint k = 0; k < _body.size(); k++) {
710 Node* n = _body.at(k);
711 switch (n->Opcode()) {
712 case Op_XorI: xors_in_loop++; break; // CRC32 java code
713 case Op_ModL: body_size += 30; break;
714 case Op_DivL: body_size += 30; break;
715 case Op_MulL: body_size += 10; break;
716 case Op_StrComp:
717 case Op_StrEquals:
718 case Op_StrIndexOf:
719 case Op_EncodeISOArray:
720 case Op_AryEq: {
721 // Do not unroll a loop with String intrinsics code.
722 // String intrinsics are large and have loops.
723 return false;
724 }
725 } // switch
726 }
728 // Check for being too big
729 if (body_size > (uint)LoopUnrollLimit) {
730 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
731 // Normal case: loop too big
732 return false;
733 }
735 // Unroll once! (Each trip will soon do double iterations)
736 return true;
737 }
739 //------------------------------policy_align-----------------------------------
740 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
741 // expression that does the alignment. Note that only one array base can be
742 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
743 // if we vectorize short memory ops into longer memory ops, we may want to
744 // increase alignment.
745 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
746 return false;
747 }
749 //------------------------------policy_range_check-----------------------------
750 // Return TRUE or FALSE if the loop should be range-check-eliminated.
751 // Actually we do iteration-splitting, a more powerful form of RCE.
752 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
753 if (!RangeCheckElimination) return false;
755 CountedLoopNode *cl = _head->as_CountedLoop();
756 // If we unrolled with no intention of doing RCE and we later
757 // changed our minds, we got no pre-loop. Either we need to
758 // make a new pre-loop, or we gotta disallow RCE.
759 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
760 Node *trip_counter = cl->phi();
762 // Check loop body for tests of trip-counter plus loop-invariant vs
763 // loop-invariant.
764 for (uint i = 0; i < _body.size(); i++) {
765 Node *iff = _body[i];
766 if (iff->Opcode() == Op_If) { // Test?
768 // Comparing trip+off vs limit
769 Node *bol = iff->in(1);
770 if (bol->req() != 2) continue; // dead constant test
771 if (!bol->is_Bool()) {
772 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
773 continue;
774 }
775 if (bol->as_Bool()->_test._test == BoolTest::ne)
776 continue; // not RC
778 Node *cmp = bol->in(1);
780 Node *rc_exp = cmp->in(1);
781 Node *limit = cmp->in(2);
783 Node *limit_c = phase->get_ctrl(limit);
784 if( limit_c == phase->C->top() )
785 return false; // Found dead test on live IF? No RCE!
786 if( is_member(phase->get_loop(limit_c) ) ) {
787 // Compare might have operands swapped; commute them
788 rc_exp = cmp->in(2);
789 limit = cmp->in(1);
790 limit_c = phase->get_ctrl(limit);
791 if( is_member(phase->get_loop(limit_c) ) )
792 continue; // Both inputs are loop varying; cannot RCE
793 }
795 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
796 continue;
797 }
798 // Yeah! Found a test like 'trip+off vs limit'
799 // Test is an IfNode, has 2 projections. If BOTH are in the loop
800 // we need loop unswitching instead of iteration splitting.
801 if( is_loop_exit(iff) )
802 return true; // Found reason to split iterations
803 } // End of is IF
804 }
806 return false;
807 }
809 //------------------------------policy_peel_only-------------------------------
810 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
811 // for unrolling loops with NO array accesses.
812 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
814 for( uint i = 0; i < _body.size(); i++ )
815 if( _body[i]->is_Mem() )
816 return false;
818 // No memory accesses at all!
819 return true;
820 }
822 //------------------------------clone_up_backedge_goo--------------------------
823 // If Node n lives in the back_ctrl block and cannot float, we clone a private
824 // version of n in preheader_ctrl block and return that, otherwise return n.
825 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
826 if( get_ctrl(n) != back_ctrl ) return n;
828 // Only visit once
829 if (visited.test_set(n->_idx)) {
830 Node *x = clones.find(n->_idx);
831 if (x != NULL)
832 return x;
833 return n;
834 }
836 Node *x = NULL; // If required, a clone of 'n'
837 // Check for 'n' being pinned in the backedge.
838 if( n->in(0) && n->in(0) == back_ctrl ) {
839 assert(clones.find(n->_idx) == NULL, "dead loop");
840 x = n->clone(); // Clone a copy of 'n' to preheader
841 clones.push(x, n->_idx);
842 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
843 }
845 // Recursive fixup any other input edges into x.
846 // If there are no changes we can just return 'n', otherwise
847 // we need to clone a private copy and change it.
848 for( uint i = 1; i < n->req(); i++ ) {
849 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
850 if( g != n->in(i) ) {
851 if( !x ) {
852 assert(clones.find(n->_idx) == NULL, "dead loop");
853 x = n->clone();
854 clones.push(x, n->_idx);
855 }
856 x->set_req(i, g);
857 }
858 }
859 if( x ) { // x can legally float to pre-header location
860 register_new_node( x, preheader_ctrl );
861 return x;
862 } else { // raise n to cover LCA of uses
863 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
864 }
865 return n;
866 }
868 //------------------------------insert_pre_post_loops--------------------------
869 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
870 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
871 // alignment. Useful to unroll loops that do no array accesses.
872 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
874 #ifndef PRODUCT
875 if (TraceLoopOpts) {
876 if (peel_only)
877 tty->print("PeelMainPost ");
878 else
879 tty->print("PreMainPost ");
880 loop->dump_head();
881 }
882 #endif
883 C->set_major_progress();
885 // Find common pieces of the loop being guarded with pre & post loops
886 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
887 assert( main_head->is_normal_loop(), "" );
888 CountedLoopEndNode *main_end = main_head->loopexit();
889 guarantee(main_end != NULL, "no loop exit node");
890 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
891 uint dd_main_head = dom_depth(main_head);
892 uint max = main_head->outcnt();
894 Node *pre_header= main_head->in(LoopNode::EntryControl);
895 Node *init = main_head->init_trip();
896 Node *incr = main_end ->incr();
897 Node *limit = main_end ->limit();
898 Node *stride = main_end ->stride();
899 Node *cmp = main_end ->cmp_node();
900 BoolTest::mask b_test = main_end->test_trip();
902 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
903 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
904 if( bol->outcnt() != 1 ) {
905 bol = bol->clone();
906 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
907 _igvn.hash_delete(main_end);
908 main_end->set_req(CountedLoopEndNode::TestValue, bol);
909 }
910 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
911 if( cmp->outcnt() != 1 ) {
912 cmp = cmp->clone();
913 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
914 _igvn.hash_delete(bol);
915 bol->set_req(1, cmp);
916 }
918 //------------------------------
919 // Step A: Create Post-Loop.
920 Node* main_exit = main_end->proj_out(false);
921 assert( main_exit->Opcode() == Op_IfFalse, "" );
922 int dd_main_exit = dom_depth(main_exit);
924 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
925 // loop pre-header illegally has 2 control users (old & new loops).
926 clone_loop( loop, old_new, dd_main_exit );
927 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
928 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
929 post_head->set_post_loop(main_head);
931 // Reduce the post-loop trip count.
932 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
933 post_end->_prob = PROB_FAIR;
935 // Build the main-loop normal exit.
936 IfFalseNode *new_main_exit = new (C) IfFalseNode(main_end);
937 _igvn.register_new_node_with_optimizer( new_main_exit );
938 set_idom(new_main_exit, main_end, dd_main_exit );
939 set_loop(new_main_exit, loop->_parent);
941 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
942 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
943 // (the main-loop trip-counter exit value) because we will be changing
944 // the exit value (via unrolling) so we cannot constant-fold away the zero
945 // trip guard until all unrolling is done.
946 Node *zer_opaq = new (C) Opaque1Node(C, incr);
947 Node *zer_cmp = new (C) CmpINode( zer_opaq, limit );
948 Node *zer_bol = new (C) BoolNode( zer_cmp, b_test );
949 register_new_node( zer_opaq, new_main_exit );
950 register_new_node( zer_cmp , new_main_exit );
951 register_new_node( zer_bol , new_main_exit );
953 // Build the IfNode
954 IfNode *zer_iff = new (C) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
955 _igvn.register_new_node_with_optimizer( zer_iff );
956 set_idom(zer_iff, new_main_exit, dd_main_exit);
957 set_loop(zer_iff, loop->_parent);
959 // Plug in the false-path, taken if we need to skip post-loop
960 _igvn.replace_input_of(main_exit, 0, zer_iff);
961 set_idom(main_exit, zer_iff, dd_main_exit);
962 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
963 // Make the true-path, must enter the post loop
964 Node *zer_taken = new (C) IfTrueNode( zer_iff );
965 _igvn.register_new_node_with_optimizer( zer_taken );
966 set_idom(zer_taken, zer_iff, dd_main_exit);
967 set_loop(zer_taken, loop->_parent);
968 // Plug in the true path
969 _igvn.hash_delete( post_head );
970 post_head->set_req(LoopNode::EntryControl, zer_taken);
971 set_idom(post_head, zer_taken, dd_main_exit);
973 Arena *a = Thread::current()->resource_area();
974 VectorSet visited(a);
975 Node_Stack clones(a, main_head->back_control()->outcnt());
976 // Step A3: Make the fall-in values to the post-loop come from the
977 // fall-out values of the main-loop.
978 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
979 Node* main_phi = main_head->fast_out(i);
980 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
981 Node *post_phi = old_new[main_phi->_idx];
982 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
983 post_head->init_control(),
984 main_phi->in(LoopNode::LoopBackControl),
985 visited, clones);
986 _igvn.hash_delete(post_phi);
987 post_phi->set_req( LoopNode::EntryControl, fallmain );
988 }
989 }
991 // Update local caches for next stanza
992 main_exit = new_main_exit;
995 //------------------------------
996 // Step B: Create Pre-Loop.
998 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
999 // loop pre-header illegally has 2 control users (old & new loops).
1000 clone_loop( loop, old_new, dd_main_head );
1001 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
1002 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
1003 pre_head->set_pre_loop(main_head);
1004 Node *pre_incr = old_new[incr->_idx];
1006 // Reduce the pre-loop trip count.
1007 pre_end->_prob = PROB_FAIR;
1009 // Find the pre-loop normal exit.
1010 Node* pre_exit = pre_end->proj_out(false);
1011 assert( pre_exit->Opcode() == Op_IfFalse, "" );
1012 IfFalseNode *new_pre_exit = new (C) IfFalseNode(pre_end);
1013 _igvn.register_new_node_with_optimizer( new_pre_exit );
1014 set_idom(new_pre_exit, pre_end, dd_main_head);
1015 set_loop(new_pre_exit, loop->_parent);
1017 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
1018 // pre-loop, the main-loop may not execute at all. Later in life this
1019 // zero-trip guard will become the minimum-trip guard when we unroll
1020 // the main-loop.
1021 Node *min_opaq = new (C) Opaque1Node(C, limit);
1022 Node *min_cmp = new (C) CmpINode( pre_incr, min_opaq );
1023 Node *min_bol = new (C) BoolNode( min_cmp, b_test );
1024 register_new_node( min_opaq, new_pre_exit );
1025 register_new_node( min_cmp , new_pre_exit );
1026 register_new_node( min_bol , new_pre_exit );
1028 // Build the IfNode (assume the main-loop is executed always).
1029 IfNode *min_iff = new (C) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
1030 _igvn.register_new_node_with_optimizer( min_iff );
1031 set_idom(min_iff, new_pre_exit, dd_main_head);
1032 set_loop(min_iff, loop->_parent);
1034 // Plug in the false-path, taken if we need to skip main-loop
1035 _igvn.hash_delete( pre_exit );
1036 pre_exit->set_req(0, min_iff);
1037 set_idom(pre_exit, min_iff, dd_main_head);
1038 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
1039 // Make the true-path, must enter the main loop
1040 Node *min_taken = new (C) IfTrueNode( min_iff );
1041 _igvn.register_new_node_with_optimizer( min_taken );
1042 set_idom(min_taken, min_iff, dd_main_head);
1043 set_loop(min_taken, loop->_parent);
1044 // Plug in the true path
1045 _igvn.hash_delete( main_head );
1046 main_head->set_req(LoopNode::EntryControl, min_taken);
1047 set_idom(main_head, min_taken, dd_main_head);
1049 visited.Clear();
1050 clones.clear();
1051 // Step B3: Make the fall-in values to the main-loop come from the
1052 // fall-out values of the pre-loop.
1053 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
1054 Node* main_phi = main_head->fast_out(i2);
1055 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
1056 Node *pre_phi = old_new[main_phi->_idx];
1057 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
1058 main_head->init_control(),
1059 pre_phi->in(LoopNode::LoopBackControl),
1060 visited, clones);
1061 _igvn.hash_delete(main_phi);
1062 main_phi->set_req( LoopNode::EntryControl, fallpre );
1063 }
1064 }
1066 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
1067 // RCE and alignment may change this later.
1068 Node *cmp_end = pre_end->cmp_node();
1069 assert( cmp_end->in(2) == limit, "" );
1070 Node *pre_limit = new (C) AddINode( init, stride );
1072 // Save the original loop limit in this Opaque1 node for
1073 // use by range check elimination.
1074 Node *pre_opaq = new (C) Opaque1Node(C, pre_limit, limit);
1076 register_new_node( pre_limit, pre_head->in(0) );
1077 register_new_node( pre_opaq , pre_head->in(0) );
1079 // Since no other users of pre-loop compare, I can hack limit directly
1080 assert( cmp_end->outcnt() == 1, "no other users" );
1081 _igvn.hash_delete(cmp_end);
1082 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
1084 // Special case for not-equal loop bounds:
1085 // Change pre loop test, main loop test, and the
1086 // main loop guard test to use lt or gt depending on stride
1087 // direction:
1088 // positive stride use <
1089 // negative stride use >
1090 //
1091 // not-equal test is kept for post loop to handle case
1092 // when init > limit when stride > 0 (and reverse).
1094 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
1096 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
1097 // Modify pre loop end condition
1098 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1099 BoolNode* new_bol0 = new (C) BoolNode(pre_bol->in(1), new_test);
1100 register_new_node( new_bol0, pre_head->in(0) );
1101 _igvn.hash_delete(pre_end);
1102 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
1103 // Modify main loop guard condition
1104 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
1105 BoolNode* new_bol1 = new (C) BoolNode(min_bol->in(1), new_test);
1106 register_new_node( new_bol1, new_pre_exit );
1107 _igvn.hash_delete(min_iff);
1108 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
1109 // Modify main loop end condition
1110 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
1111 BoolNode* new_bol2 = new (C) BoolNode(main_bol->in(1), new_test);
1112 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
1113 _igvn.hash_delete(main_end);
1114 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
1115 }
1117 // Flag main loop
1118 main_head->set_main_loop();
1119 if( peel_only ) main_head->set_main_no_pre_loop();
1121 // Subtract a trip count for the pre-loop.
1122 main_head->set_trip_count(main_head->trip_count() - 1);
1124 // It's difficult to be precise about the trip-counts
1125 // for the pre/post loops. They are usually very short,
1126 // so guess that 4 trips is a reasonable value.
1127 post_head->set_profile_trip_cnt(4.0);
1128 pre_head->set_profile_trip_cnt(4.0);
1130 // Now force out all loop-invariant dominating tests. The optimizer
1131 // finds some, but we _know_ they are all useless.
1132 peeled_dom_test_elim(loop,old_new);
1133 }
1135 //------------------------------is_invariant-----------------------------
1136 // Return true if n is invariant
1137 bool IdealLoopTree::is_invariant(Node* n) const {
1138 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
1139 if (n_c->is_top()) return false;
1140 return !is_member(_phase->get_loop(n_c));
1141 }
1144 //------------------------------do_unroll--------------------------------------
1145 // Unroll the loop body one step - make each trip do 2 iterations.
1146 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
1147 assert(LoopUnrollLimit, "");
1148 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
1149 CountedLoopEndNode *loop_end = loop_head->loopexit();
1150 assert(loop_end, "");
1151 #ifndef PRODUCT
1152 if (PrintOpto && VerifyLoopOptimizations) {
1153 tty->print("Unrolling ");
1154 loop->dump_head();
1155 } else if (TraceLoopOpts) {
1156 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
1157 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
1158 } else {
1159 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
1160 }
1161 loop->dump_head();
1162 }
1163 #endif
1165 // Remember loop node count before unrolling to detect
1166 // if rounds of unroll,optimize are making progress
1167 loop_head->set_node_count_before_unroll(loop->_body.size());
1169 Node *ctrl = loop_head->in(LoopNode::EntryControl);
1170 Node *limit = loop_head->limit();
1171 Node *init = loop_head->init_trip();
1172 Node *stride = loop_head->stride();
1174 Node *opaq = NULL;
1175 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
1176 // Search for zero-trip guard.
1177 assert( loop_head->is_main_loop(), "" );
1178 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
1179 Node *iff = ctrl->in(0);
1180 assert( iff->Opcode() == Op_If, "" );
1181 Node *bol = iff->in(1);
1182 assert( bol->Opcode() == Op_Bool, "" );
1183 Node *cmp = bol->in(1);
1184 assert( cmp->Opcode() == Op_CmpI, "" );
1185 opaq = cmp->in(2);
1186 // Occasionally it's possible for a zero-trip guard Opaque1 node to be
1187 // optimized away and then another round of loop opts attempted.
1188 // We can not optimize this particular loop in that case.
1189 if (opaq->Opcode() != Op_Opaque1)
1190 return; // Cannot find zero-trip guard! Bail out!
1191 // Zero-trip test uses an 'opaque' node which is not shared.
1192 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
1193 }
1195 C->set_major_progress();
1197 Node* new_limit = NULL;
1198 if (UnrollLimitCheck) {
1199 int stride_con = stride->get_int();
1200 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
1201 uint old_trip_count = loop_head->trip_count();
1202 // Verify that unroll policy result is still valid.
1203 assert(old_trip_count > 1 &&
1204 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
1206 // Adjust loop limit to keep valid iterations number after unroll.
1207 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
1208 // which may overflow.
1209 if (!adjust_min_trip) {
1210 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
1211 "odd trip count for maximally unroll");
1212 // Don't need to adjust limit for maximally unroll since trip count is even.
1213 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
1214 // Loop's limit is constant. Loop's init could be constant when pre-loop
1215 // become peeled iteration.
1216 jlong init_con = init->get_int();
1217 // We can keep old loop limit if iterations count stays the same:
1218 // old_trip_count == new_trip_count * 2
1219 // Note: since old_trip_count >= 2 then new_trip_count >= 1
1220 // so we also don't need to adjust zero trip test.
1221 jlong limit_con = limit->get_int();
1222 // (stride_con*2) not overflow since stride_con <= 8.
1223 int new_stride_con = stride_con * 2;
1224 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
1225 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
1226 // New trip count should satisfy next conditions.
1227 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
1228 uint new_trip_count = (uint)trip_count;
1229 adjust_min_trip = (old_trip_count != new_trip_count*2);
1230 }
1232 if (adjust_min_trip) {
1233 // Step 2: Adjust the trip limit if it is called for.
1234 // The adjustment amount is -stride. Need to make sure if the
1235 // adjustment underflows or overflows, then the main loop is skipped.
1236 Node* cmp = loop_end->cmp_node();
1237 assert(cmp->in(2) == limit, "sanity");
1238 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
1240 // Verify that policy_unroll result is still valid.
1241 const TypeInt* limit_type = _igvn.type(limit)->is_int();
1242 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
1243 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
1245 if (limit->is_Con()) {
1246 // The check in policy_unroll and the assert above guarantee
1247 // no underflow if limit is constant.
1248 new_limit = _igvn.intcon(limit->get_int() - stride_con);
1249 set_ctrl(new_limit, C->root());
1250 } else {
1251 // Limit is not constant.
1252 if (loop_head->unrolled_count() == 1) { // only for first unroll
1253 // Separate limit by Opaque node in case it is an incremented
1254 // variable from previous loop to avoid using pre-incremented
1255 // value which could increase register pressure.
1256 // Otherwise reorg_offsets() optimization will create a separate
1257 // Opaque node for each use of trip-counter and as result
1258 // zero trip guard limit will be different from loop limit.
1259 assert(has_ctrl(opaq), "should have it");
1260 Node* opaq_ctrl = get_ctrl(opaq);
1261 limit = new (C) Opaque2Node( C, limit );
1262 register_new_node( limit, opaq_ctrl );
1263 }
1264 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
1265 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
1266 // No underflow.
1267 new_limit = new (C) SubINode(limit, stride);
1268 } else {
1269 // (limit - stride) may underflow.
1270 // Clamp the adjustment value with MININT or MAXINT:
1271 //
1272 // new_limit = limit-stride
1273 // if (stride > 0)
1274 // new_limit = (limit < new_limit) ? MININT : new_limit;
1275 // else
1276 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
1277 //
1278 BoolTest::mask bt = loop_end->test_trip();
1279 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
1280 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
1281 set_ctrl(adj_max, C->root());
1282 Node* old_limit = NULL;
1283 Node* adj_limit = NULL;
1284 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
1285 if (loop_head->unrolled_count() > 1 &&
1286 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
1287 limit->in(CMoveNode::IfTrue) == adj_max &&
1288 bol->as_Bool()->_test._test == bt &&
1289 bol->in(1)->Opcode() == Op_CmpI &&
1290 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
1291 // Loop was unrolled before.
1292 // Optimize the limit to avoid nested CMove:
1293 // use original limit as old limit.
1294 old_limit = bol->in(1)->in(1);
1295 // Adjust previous adjusted limit.
1296 adj_limit = limit->in(CMoveNode::IfFalse);
1297 adj_limit = new (C) SubINode(adj_limit, stride);
1298 } else {
1299 old_limit = limit;
1300 adj_limit = new (C) SubINode(limit, stride);
1301 }
1302 assert(old_limit != NULL && adj_limit != NULL, "");
1303 register_new_node( adj_limit, ctrl ); // adjust amount
1304 Node* adj_cmp = new (C) CmpINode(old_limit, adj_limit);
1305 register_new_node( adj_cmp, ctrl );
1306 Node* adj_bool = new (C) BoolNode(adj_cmp, bt);
1307 register_new_node( adj_bool, ctrl );
1308 new_limit = new (C) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
1309 }
1310 register_new_node(new_limit, ctrl);
1311 }
1312 assert(new_limit != NULL, "");
1313 // Replace in loop test.
1314 assert(loop_end->in(1)->in(1) == cmp, "sanity");
1315 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
1316 // Don't need to create new test since only one user.
1317 _igvn.hash_delete(cmp);
1318 cmp->set_req(2, new_limit);
1319 } else {
1320 // Create new test since it is shared.
1321 Node* ctrl2 = loop_end->in(0);
1322 Node* cmp2 = cmp->clone();
1323 cmp2->set_req(2, new_limit);
1324 register_new_node(cmp2, ctrl2);
1325 Node* bol2 = loop_end->in(1)->clone();
1326 bol2->set_req(1, cmp2);
1327 register_new_node(bol2, ctrl2);
1328 _igvn.hash_delete(loop_end);
1329 loop_end->set_req(1, bol2);
1330 }
1331 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1332 // Make it a 1-trip test (means at least 2 trips).
1334 // Guard test uses an 'opaque' node which is not shared. Hence I
1335 // can edit it's inputs directly. Hammer in the new limit for the
1336 // minimum-trip guard.
1337 assert(opaq->outcnt() == 1, "");
1338 _igvn.hash_delete(opaq);
1339 opaq->set_req(1, new_limit);
1340 }
1342 // Adjust max trip count. The trip count is intentionally rounded
1343 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1344 // the main, unrolled, part of the loop will never execute as it is protected
1345 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1346 // and later determined that part of the unrolled loop was dead.
1347 loop_head->set_trip_count(old_trip_count / 2);
1349 // Double the count of original iterations in the unrolled loop body.
1350 loop_head->double_unrolled_count();
1352 } else { // LoopLimitCheck
1354 // Adjust max trip count. The trip count is intentionally rounded
1355 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
1356 // the main, unrolled, part of the loop will never execute as it is protected
1357 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
1358 // and later determined that part of the unrolled loop was dead.
1359 loop_head->set_trip_count(loop_head->trip_count() / 2);
1361 // Double the count of original iterations in the unrolled loop body.
1362 loop_head->double_unrolled_count();
1364 // -----------
1365 // Step 2: Cut back the trip counter for an unroll amount of 2.
1366 // Loop will normally trip (limit - init)/stride_con. Since it's a
1367 // CountedLoop this is exact (stride divides limit-init exactly).
1368 // We are going to double the loop body, so we want to knock off any
1369 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1370 Node *span = new (C) SubINode( limit, init );
1371 register_new_node( span, ctrl );
1372 Node *trip = new (C) DivINode( 0, span, stride );
1373 register_new_node( trip, ctrl );
1374 Node *mtwo = _igvn.intcon(-2);
1375 set_ctrl(mtwo, C->root());
1376 Node *rond = new (C) AndINode( trip, mtwo );
1377 register_new_node( rond, ctrl );
1378 Node *spn2 = new (C) MulINode( rond, stride );
1379 register_new_node( spn2, ctrl );
1380 new_limit = new (C) AddINode( spn2, init );
1381 register_new_node( new_limit, ctrl );
1383 // Hammer in the new limit
1384 Node *ctrl2 = loop_end->in(0);
1385 Node *cmp2 = new (C) CmpINode( loop_head->incr(), new_limit );
1386 register_new_node( cmp2, ctrl2 );
1387 Node *bol2 = new (C) BoolNode( cmp2, loop_end->test_trip() );
1388 register_new_node( bol2, ctrl2 );
1389 _igvn.hash_delete(loop_end);
1390 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
1392 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1393 // Make it a 1-trip test (means at least 2 trips).
1394 if( adjust_min_trip ) {
1395 assert( new_limit != NULL, "" );
1396 // Guard test uses an 'opaque' node which is not shared. Hence I
1397 // can edit it's inputs directly. Hammer in the new limit for the
1398 // minimum-trip guard.
1399 assert( opaq->outcnt() == 1, "" );
1400 _igvn.hash_delete(opaq);
1401 opaq->set_req(1, new_limit);
1402 }
1403 } // LoopLimitCheck
1405 // ---------
1406 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1407 // represents the odd iterations; since the loop trips an even number of
1408 // times its backedge is never taken. Kill the backedge.
1409 uint dd = dom_depth(loop_head);
1410 clone_loop( loop, old_new, dd );
1412 // Make backedges of the clone equal to backedges of the original.
1413 // Make the fall-in from the original come from the fall-out of the clone.
1414 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1415 Node* phi = loop_head->fast_out(j);
1416 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1417 Node *newphi = old_new[phi->_idx];
1418 _igvn.hash_delete( phi );
1419 _igvn.hash_delete( newphi );
1421 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1422 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1423 phi ->set_req(LoopNode::LoopBackControl, C->top());
1424 }
1425 }
1426 Node *clone_head = old_new[loop_head->_idx];
1427 _igvn.hash_delete( clone_head );
1428 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1429 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1430 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1431 loop->_head = clone_head; // New loop header
1433 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1434 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1436 // Kill the clone's backedge
1437 Node *newcle = old_new[loop_end->_idx];
1438 _igvn.hash_delete( newcle );
1439 Node *one = _igvn.intcon(1);
1440 set_ctrl(one, C->root());
1441 newcle->set_req(1, one);
1442 // Force clone into same loop body
1443 uint max = loop->_body.size();
1444 for( uint k = 0; k < max; k++ ) {
1445 Node *old = loop->_body.at(k);
1446 Node *nnn = old_new[old->_idx];
1447 loop->_body.push(nnn);
1448 if (!has_ctrl(old))
1449 set_loop(nnn, loop);
1450 }
1452 loop->record_for_igvn();
1453 }
1455 //------------------------------do_maximally_unroll----------------------------
1457 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1458 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1459 assert(cl->has_exact_trip_count(), "trip count is not exact");
1460 assert(cl->trip_count() > 0, "");
1461 #ifndef PRODUCT
1462 if (TraceLoopOpts) {
1463 tty->print("MaxUnroll %d ", cl->trip_count());
1464 loop->dump_head();
1465 }
1466 #endif
1468 // If loop is tripping an odd number of times, peel odd iteration
1469 if ((cl->trip_count() & 1) == 1) {
1470 do_peeling(loop, old_new);
1471 }
1473 // Now its tripping an even number of times remaining. Double loop body.
1474 // Do not adjust pre-guards; they are not needed and do not exist.
1475 if (cl->trip_count() > 0) {
1476 assert((cl->trip_count() & 1) == 0, "missed peeling");
1477 do_unroll(loop, old_new, false);
1478 }
1479 }
1481 //------------------------------dominates_backedge---------------------------------
1482 // Returns true if ctrl is executed on every complete iteration
1483 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1484 assert(ctrl->is_CFG(), "must be control");
1485 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1486 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1487 }
1489 //------------------------------adjust_limit-----------------------------------
1490 // Helper function for add_constraint().
1491 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
1492 // Compute "I :: (limit-offset)/scale"
1493 Node *con = new (C) SubINode(rc_limit, offset);
1494 register_new_node(con, pre_ctrl);
1495 Node *X = new (C) DivINode(0, con, scale);
1496 register_new_node(X, pre_ctrl);
1498 // Adjust loop limit
1499 loop_limit = (stride_con > 0)
1500 ? (Node*)(new (C) MinINode(loop_limit, X))
1501 : (Node*)(new (C) MaxINode(loop_limit, X));
1502 register_new_node(loop_limit, pre_ctrl);
1503 return loop_limit;
1504 }
1506 //------------------------------add_constraint---------------------------------
1507 // Constrain the main loop iterations so the conditions:
1508 // low_limit <= scale_con * I + offset < upper_limit
1509 // always holds true. That is, either increase the number of iterations in
1510 // the pre-loop or the post-loop until the condition holds true in the main
1511 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1512 // stride and scale are constants (offset and limit often are).
1513 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 ) {
1514 // For positive stride, the pre-loop limit always uses a MAX function
1515 // and the main loop a MIN function. For negative stride these are
1516 // reversed.
1518 // Also for positive stride*scale the affine function is increasing, so the
1519 // pre-loop must check for underflow and the post-loop for overflow.
1520 // Negative stride*scale reverses this; pre-loop checks for overflow and
1521 // post-loop for underflow.
1523 Node *scale = _igvn.intcon(scale_con);
1524 set_ctrl(scale, C->root());
1526 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
1527 // The overflow limit: scale*I+offset < upper_limit
1528 // For main-loop compute
1529 // ( if (scale > 0) /* and stride > 0 */
1530 // I < (upper_limit-offset)/scale
1531 // else /* scale < 0 and stride < 0 */
1532 // I > (upper_limit-offset)/scale
1533 // )
1534 //
1535 // (upper_limit-offset) may overflow or underflow.
1536 // But it is fine since main loop will either have
1537 // less iterations or will be skipped in such case.
1538 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
1540 // The underflow limit: low_limit <= scale*I+offset.
1541 // For pre-loop compute
1542 // NOT(scale*I+offset >= low_limit)
1543 // scale*I+offset < low_limit
1544 // ( if (scale > 0) /* and stride > 0 */
1545 // I < (low_limit-offset)/scale
1546 // else /* scale < 0 and stride < 0 */
1547 // I > (low_limit-offset)/scale
1548 // )
1550 if (low_limit->get_int() == -max_jint) {
1551 if (!RangeLimitCheck) return;
1552 // We need this guard when scale*pre_limit+offset >= limit
1553 // due to underflow. So we need execute pre-loop until
1554 // scale*I+offset >= min_int. But (min_int-offset) will
1555 // underflow when offset > 0 and X will be > original_limit
1556 // when stride > 0. To avoid it we replace positive offset with 0.
1557 //
1558 // Also (min_int+1 == -max_int) is used instead of min_int here
1559 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1560 Node* shift = _igvn.intcon(31);
1561 set_ctrl(shift, C->root());
1562 Node* sign = new (C) RShiftINode(offset, shift);
1563 register_new_node(sign, pre_ctrl);
1564 offset = new (C) AndINode(offset, sign);
1565 register_new_node(offset, pre_ctrl);
1566 } else {
1567 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1568 // The only problem we have here when offset == min_int
1569 // since (0-min_int) == min_int. It may be fine for stride > 0
1570 // but for stride < 0 X will be < original_limit. To avoid it
1571 // max(pre_limit, original_limit) is used in do_range_check().
1572 }
1573 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1574 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
1576 } else { // stride_con*scale_con < 0
1577 // For negative stride*scale pre-loop checks for overflow and
1578 // post-loop for underflow.
1579 //
1580 // The overflow limit: scale*I+offset < upper_limit
1581 // For pre-loop compute
1582 // NOT(scale*I+offset < upper_limit)
1583 // scale*I+offset >= upper_limit
1584 // scale*I+offset+1 > upper_limit
1585 // ( if (scale < 0) /* and stride > 0 */
1586 // I < (upper_limit-(offset+1))/scale
1587 // else /* scale > 0 and stride < 0 */
1588 // I > (upper_limit-(offset+1))/scale
1589 // )
1590 //
1591 // (upper_limit-offset-1) may underflow or overflow.
1592 // To avoid it min(pre_limit, original_limit) is used
1593 // in do_range_check() for stride > 0 and max() for < 0.
1594 Node *one = _igvn.intcon(1);
1595 set_ctrl(one, C->root());
1597 Node *plus_one = new (C) AddINode(offset, one);
1598 register_new_node( plus_one, pre_ctrl );
1599 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
1600 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
1602 if (low_limit->get_int() == -max_jint) {
1603 if (!RangeLimitCheck) return;
1604 // We need this guard when scale*main_limit+offset >= limit
1605 // due to underflow. So we need execute main-loop while
1606 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
1607 // underflow when (offset+1) > 0 and X will be < main_limit
1608 // when scale < 0 (and stride > 0). To avoid it we replace
1609 // positive (offset+1) with 0.
1610 //
1611 // Also (min_int+1 == -max_int) is used instead of min_int here
1612 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
1613 Node* shift = _igvn.intcon(31);
1614 set_ctrl(shift, C->root());
1615 Node* sign = new (C) RShiftINode(plus_one, shift);
1616 register_new_node(sign, pre_ctrl);
1617 plus_one = new (C) AndINode(plus_one, sign);
1618 register_new_node(plus_one, pre_ctrl);
1619 } else {
1620 assert(low_limit->get_int() == 0, "wrong low limit for range check");
1621 // The only problem we have here when offset == max_int
1622 // since (max_int+1) == min_int and (0-min_int) == min_int.
1623 // But it is fine since main loop will either have
1624 // less iterations or will be skipped in such case.
1625 }
1626 // The underflow limit: low_limit <= scale*I+offset.
1627 // For main-loop compute
1628 // scale*I+offset+1 > low_limit
1629 // ( if (scale < 0) /* and stride > 0 */
1630 // I < (low_limit-(offset+1))/scale
1631 // else /* scale > 0 and stride < 0 */
1632 // I > (low_limit-(offset+1))/scale
1633 // )
1635 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
1636 }
1637 }
1640 //------------------------------is_scaled_iv---------------------------------
1641 // Return true if exp is a constant times an induction var
1642 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1643 if (exp == iv) {
1644 if (p_scale != NULL) {
1645 *p_scale = 1;
1646 }
1647 return true;
1648 }
1649 int opc = exp->Opcode();
1650 if (opc == Op_MulI) {
1651 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1652 if (p_scale != NULL) {
1653 *p_scale = exp->in(2)->get_int();
1654 }
1655 return true;
1656 }
1657 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1658 if (p_scale != NULL) {
1659 *p_scale = exp->in(1)->get_int();
1660 }
1661 return true;
1662 }
1663 } else if (opc == Op_LShiftI) {
1664 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1665 if (p_scale != NULL) {
1666 *p_scale = 1 << exp->in(2)->get_int();
1667 }
1668 return true;
1669 }
1670 }
1671 return false;
1672 }
1674 //-----------------------------is_scaled_iv_plus_offset------------------------------
1675 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1676 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1677 if (is_scaled_iv(exp, iv, p_scale)) {
1678 if (p_offset != NULL) {
1679 Node *zero = _igvn.intcon(0);
1680 set_ctrl(zero, C->root());
1681 *p_offset = zero;
1682 }
1683 return true;
1684 }
1685 int opc = exp->Opcode();
1686 if (opc == Op_AddI) {
1687 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1688 if (p_offset != NULL) {
1689 *p_offset = exp->in(2);
1690 }
1691 return true;
1692 }
1693 if (exp->in(2)->is_Con()) {
1694 Node* offset2 = NULL;
1695 if (depth < 2 &&
1696 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1697 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1698 if (p_offset != NULL) {
1699 Node *ctrl_off2 = get_ctrl(offset2);
1700 Node* offset = new (C) AddINode(offset2, exp->in(2));
1701 register_new_node(offset, ctrl_off2);
1702 *p_offset = offset;
1703 }
1704 return true;
1705 }
1706 }
1707 } else if (opc == Op_SubI) {
1708 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1709 if (p_offset != NULL) {
1710 Node *zero = _igvn.intcon(0);
1711 set_ctrl(zero, C->root());
1712 Node *ctrl_off = get_ctrl(exp->in(2));
1713 Node* offset = new (C) SubINode(zero, exp->in(2));
1714 register_new_node(offset, ctrl_off);
1715 *p_offset = offset;
1716 }
1717 return true;
1718 }
1719 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1720 if (p_offset != NULL) {
1721 *p_scale *= -1;
1722 *p_offset = exp->in(1);
1723 }
1724 return true;
1725 }
1726 }
1727 return false;
1728 }
1730 //------------------------------do_range_check---------------------------------
1731 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1732 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1733 #ifndef PRODUCT
1734 if (PrintOpto && VerifyLoopOptimizations) {
1735 tty->print("Range Check Elimination ");
1736 loop->dump_head();
1737 } else if (TraceLoopOpts) {
1738 tty->print("RangeCheck ");
1739 loop->dump_head();
1740 }
1741 #endif
1742 assert(RangeCheckElimination, "");
1743 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1744 assert(cl->is_main_loop(), "");
1746 // protect against stride not being a constant
1747 if (!cl->stride_is_con())
1748 return;
1750 // Find the trip counter; we are iteration splitting based on it
1751 Node *trip_counter = cl->phi();
1752 // Find the main loop limit; we will trim it's iterations
1753 // to not ever trip end tests
1754 Node *main_limit = cl->limit();
1756 // Need to find the main-loop zero-trip guard
1757 Node *ctrl = cl->in(LoopNode::EntryControl);
1758 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1759 Node *iffm = ctrl->in(0);
1760 assert(iffm->Opcode() == Op_If, "");
1761 Node *bolzm = iffm->in(1);
1762 assert(bolzm->Opcode() == Op_Bool, "");
1763 Node *cmpzm = bolzm->in(1);
1764 assert(cmpzm->is_Cmp(), "");
1765 Node *opqzm = cmpzm->in(2);
1766 // Can not optimize a loop if zero-trip Opaque1 node is optimized
1767 // away and then another round of loop opts attempted.
1768 if (opqzm->Opcode() != Op_Opaque1)
1769 return;
1770 assert(opqzm->in(1) == main_limit, "do not understand situation");
1772 // Find the pre-loop limit; we will expand it's iterations to
1773 // not ever trip low tests.
1774 Node *p_f = iffm->in(0);
1775 assert(p_f->Opcode() == Op_IfFalse, "");
1776 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1777 assert(pre_end->loopnode()->is_pre_loop(), "");
1778 Node *pre_opaq1 = pre_end->limit();
1779 // Occasionally it's possible for a pre-loop Opaque1 node to be
1780 // optimized away and then another round of loop opts attempted.
1781 // We can not optimize this particular loop in that case.
1782 if (pre_opaq1->Opcode() != Op_Opaque1)
1783 return;
1784 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1785 Node *pre_limit = pre_opaq->in(1);
1787 // Where do we put new limit calculations
1788 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1790 // Ensure the original loop limit is available from the
1791 // pre-loop Opaque1 node.
1792 Node *orig_limit = pre_opaq->original_loop_limit();
1793 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1794 return;
1796 // Must know if its a count-up or count-down loop
1798 int stride_con = cl->stride_con();
1799 Node *zero = _igvn.intcon(0);
1800 Node *one = _igvn.intcon(1);
1801 // Use symmetrical int range [-max_jint,max_jint]
1802 Node *mini = _igvn.intcon(-max_jint);
1803 set_ctrl(zero, C->root());
1804 set_ctrl(one, C->root());
1805 set_ctrl(mini, C->root());
1807 // Range checks that do not dominate the loop backedge (ie.
1808 // conditionally executed) can lengthen the pre loop limit beyond
1809 // the original loop limit. To prevent this, the pre limit is
1810 // (for stride > 0) MINed with the original loop limit (MAXed
1811 // stride < 0) when some range_check (rc) is conditionally
1812 // executed.
1813 bool conditional_rc = false;
1815 // Check loop body for tests of trip-counter plus loop-invariant vs
1816 // loop-invariant.
1817 for( uint i = 0; i < loop->_body.size(); i++ ) {
1818 Node *iff = loop->_body[i];
1819 if( iff->Opcode() == Op_If ) { // Test?
1821 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1822 // we need loop unswitching instead of iteration splitting.
1823 Node *exit = loop->is_loop_exit(iff);
1824 if( !exit ) continue;
1825 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1827 // Get boolean condition to test
1828 Node *i1 = iff->in(1);
1829 if( !i1->is_Bool() ) continue;
1830 BoolNode *bol = i1->as_Bool();
1831 BoolTest b_test = bol->_test;
1832 // Flip sense of test if exit condition is flipped
1833 if( flip )
1834 b_test = b_test.negate();
1836 // Get compare
1837 Node *cmp = bol->in(1);
1839 // Look for trip_counter + offset vs limit
1840 Node *rc_exp = cmp->in(1);
1841 Node *limit = cmp->in(2);
1842 jint scale_con= 1; // Assume trip counter not scaled
1844 Node *limit_c = get_ctrl(limit);
1845 if( loop->is_member(get_loop(limit_c) ) ) {
1846 // Compare might have operands swapped; commute them
1847 b_test = b_test.commute();
1848 rc_exp = cmp->in(2);
1849 limit = cmp->in(1);
1850 limit_c = get_ctrl(limit);
1851 if( loop->is_member(get_loop(limit_c) ) )
1852 continue; // Both inputs are loop varying; cannot RCE
1853 }
1854 // Here we know 'limit' is loop invariant
1856 // 'limit' maybe pinned below the zero trip test (probably from a
1857 // previous round of rce), in which case, it can't be used in the
1858 // zero trip test expression which must occur before the zero test's if.
1859 if( limit_c == ctrl ) {
1860 continue; // Don't rce this check but continue looking for other candidates.
1861 }
1863 // Check for scaled induction variable plus an offset
1864 Node *offset = NULL;
1866 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1867 continue;
1868 }
1870 Node *offset_c = get_ctrl(offset);
1871 if( loop->is_member( get_loop(offset_c) ) )
1872 continue; // Offset is not really loop invariant
1873 // Here we know 'offset' is loop invariant.
1875 // As above for the 'limit', the 'offset' maybe pinned below the
1876 // zero trip test.
1877 if( offset_c == ctrl ) {
1878 continue; // Don't rce this check but continue looking for other candidates.
1879 }
1880 #ifdef ASSERT
1881 if (TraceRangeLimitCheck) {
1882 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
1883 bol->dump(2);
1884 }
1885 #endif
1886 // At this point we have the expression as:
1887 // scale_con * trip_counter + offset :: limit
1888 // where scale_con, offset and limit are loop invariant. Trip_counter
1889 // monotonically increases by stride_con, a constant. Both (or either)
1890 // stride_con and scale_con can be negative which will flip about the
1891 // sense of the test.
1893 // Adjust pre and main loop limits to guard the correct iteration set
1894 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1895 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1896 // The underflow and overflow limits: 0 <= scale*I+offset < limit
1897 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
1898 if (!conditional_rc) {
1899 // (0-offset)/scale could be outside of loop iterations range.
1900 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1901 }
1902 } else {
1903 #ifndef PRODUCT
1904 if( PrintOpto )
1905 tty->print_cr("missed RCE opportunity");
1906 #endif
1907 continue; // In release mode, ignore it
1908 }
1909 } else { // Otherwise work on normal compares
1910 switch( b_test._test ) {
1911 case BoolTest::gt:
1912 // Fall into GE case
1913 case BoolTest::ge:
1914 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
1915 scale_con = -scale_con;
1916 offset = new (C) SubINode( zero, offset );
1917 register_new_node( offset, pre_ctrl );
1918 limit = new (C) SubINode( zero, limit );
1919 register_new_node( limit, pre_ctrl );
1920 // Fall into LE case
1921 case BoolTest::le:
1922 if (b_test._test != BoolTest::gt) {
1923 // Convert X <= Y to X < Y+1
1924 limit = new (C) AddINode( limit, one );
1925 register_new_node( limit, pre_ctrl );
1926 }
1927 // Fall into LT case
1928 case BoolTest::lt:
1929 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
1930 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
1931 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
1932 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
1933 if (!conditional_rc) {
1934 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
1935 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
1936 // still be outside of loop range.
1937 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
1938 }
1939 break;
1940 default:
1941 #ifndef PRODUCT
1942 if( PrintOpto )
1943 tty->print_cr("missed RCE opportunity");
1944 #endif
1945 continue; // Unhandled case
1946 }
1947 }
1949 // Kill the eliminated test
1950 C->set_major_progress();
1951 Node *kill_con = _igvn.intcon( 1-flip );
1952 set_ctrl(kill_con, C->root());
1953 _igvn.replace_input_of(iff, 1, kill_con);
1954 // Find surviving projection
1955 assert(iff->is_If(), "");
1956 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1957 // Find loads off the surviving projection; remove their control edge
1958 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1959 Node* cd = dp->fast_out(i); // Control-dependent node
1960 if( cd->is_Load() ) { // Loads can now float around in the loop
1961 // Allow the load to float around in the loop, or before it
1962 // but NOT before the pre-loop.
1963 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
1964 --i;
1965 --imax;
1966 }
1967 }
1969 } // End of is IF
1971 }
1973 // Update loop limits
1974 if (conditional_rc) {
1975 pre_limit = (stride_con > 0) ? (Node*)new (C) MinINode(pre_limit, orig_limit)
1976 : (Node*)new (C) MaxINode(pre_limit, orig_limit);
1977 register_new_node(pre_limit, pre_ctrl);
1978 }
1979 _igvn.hash_delete(pre_opaq);
1980 pre_opaq->set_req(1, pre_limit);
1982 // Note:: we are making the main loop limit no longer precise;
1983 // need to round up based on stride.
1984 cl->set_nonexact_trip_count();
1985 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
1986 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1987 // Hopefully, compiler will optimize for powers of 2.
1988 Node *ctrl = get_ctrl(main_limit);
1989 Node *stride = cl->stride();
1990 Node *init = cl->init_trip();
1991 Node *span = new (C) SubINode(main_limit,init);
1992 register_new_node(span,ctrl);
1993 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
1994 Node *add = new (C) AddINode(span,rndup);
1995 register_new_node(add,ctrl);
1996 Node *div = new (C) DivINode(0,add,stride);
1997 register_new_node(div,ctrl);
1998 Node *mul = new (C) MulINode(div,stride);
1999 register_new_node(mul,ctrl);
2000 Node *newlim = new (C) AddINode(mul,init);
2001 register_new_node(newlim,ctrl);
2002 main_limit = newlim;
2003 }
2005 Node *main_cle = cl->loopexit();
2006 Node *main_bol = main_cle->in(1);
2007 // Hacking loop bounds; need private copies of exit test
2008 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
2009 _igvn.hash_delete(main_cle);
2010 main_bol = main_bol->clone();// Clone a private BoolNode
2011 register_new_node( main_bol, main_cle->in(0) );
2012 main_cle->set_req(1,main_bol);
2013 }
2014 Node *main_cmp = main_bol->in(1);
2015 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
2016 _igvn.hash_delete(main_bol);
2017 main_cmp = main_cmp->clone();// Clone a private CmpNode
2018 register_new_node( main_cmp, main_cle->in(0) );
2019 main_bol->set_req(1,main_cmp);
2020 }
2021 // Hack the now-private loop bounds
2022 _igvn.replace_input_of(main_cmp, 2, main_limit);
2023 // The OpaqueNode is unshared by design
2024 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
2025 _igvn.replace_input_of(opqzm, 1, main_limit);
2026 }
2028 //------------------------------DCE_loop_body----------------------------------
2029 // Remove simplistic dead code from loop body
2030 void IdealLoopTree::DCE_loop_body() {
2031 for( uint i = 0; i < _body.size(); i++ )
2032 if( _body.at(i)->outcnt() == 0 )
2033 _body.map( i--, _body.pop() );
2034 }
2037 //------------------------------adjust_loop_exit_prob--------------------------
2038 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
2039 // Replace with a 1-in-10 exit guess.
2040 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
2041 Node *test = tail();
2042 while( test != _head ) {
2043 uint top = test->Opcode();
2044 if( top == Op_IfTrue || top == Op_IfFalse ) {
2045 int test_con = ((ProjNode*)test)->_con;
2046 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
2047 IfNode *iff = test->in(0)->as_If();
2048 if( iff->outcnt() == 2 ) { // Ignore dead tests
2049 Node *bol = iff->in(1);
2050 if( bol && bol->req() > 1 && bol->in(1) &&
2051 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
2052 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
2053 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
2054 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
2055 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
2056 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
2057 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
2058 return; // Allocation loops RARELY take backedge
2059 // Find the OTHER exit path from the IF
2060 Node* ex = iff->proj_out(1-test_con);
2061 float p = iff->_prob;
2062 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
2063 if( top == Op_IfTrue ) {
2064 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
2065 iff->_prob = PROB_STATIC_FREQUENT;
2066 }
2067 } else {
2068 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
2069 iff->_prob = PROB_STATIC_INFREQUENT;
2070 }
2071 }
2072 }
2073 }
2074 }
2075 test = phase->idom(test);
2076 }
2077 }
2080 //------------------------------policy_do_remove_empty_loop--------------------
2081 // Micro-benchmark spamming. Policy is to always remove empty loops.
2082 // The 'DO' part is to replace the trip counter with the value it will
2083 // have on the last iteration. This will break the loop.
2084 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
2085 // Minimum size must be empty loop
2086 if (_body.size() > EMPTY_LOOP_SIZE)
2087 return false;
2089 if (!_head->is_CountedLoop())
2090 return false; // Dead loop
2091 CountedLoopNode *cl = _head->as_CountedLoop();
2092 if (!cl->is_valid_counted_loop())
2093 return false; // Malformed loop
2094 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
2095 return false; // Infinite loop
2097 #ifdef ASSERT
2098 // Ensure only one phi which is the iv.
2099 Node* iv = NULL;
2100 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
2101 Node* n = cl->fast_out(i);
2102 if (n->Opcode() == Op_Phi) {
2103 assert(iv == NULL, "Too many phis" );
2104 iv = n;
2105 }
2106 }
2107 assert(iv == cl->phi(), "Wrong phi" );
2108 #endif
2110 // main and post loops have explicitly created zero trip guard
2111 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
2112 if (needs_guard) {
2113 // Skip guard if values not overlap.
2114 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
2115 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
2116 int stride_con = cl->stride_con();
2117 if (stride_con > 0) {
2118 needs_guard = (init_t->_hi >= limit_t->_lo);
2119 } else {
2120 needs_guard = (init_t->_lo <= limit_t->_hi);
2121 }
2122 }
2123 if (needs_guard) {
2124 // Check for an obvious zero trip guard.
2125 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
2126 if (inctrl->Opcode() == Op_IfTrue) {
2127 // The test should look like just the backedge of a CountedLoop
2128 Node* iff = inctrl->in(0);
2129 if (iff->is_If()) {
2130 Node* bol = iff->in(1);
2131 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
2132 Node* cmp = bol->in(1);
2133 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
2134 needs_guard = false;
2135 }
2136 }
2137 }
2138 }
2139 }
2141 #ifndef PRODUCT
2142 if (PrintOpto) {
2143 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
2144 this->dump_head();
2145 } else if (TraceLoopOpts) {
2146 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
2147 this->dump_head();
2148 }
2149 #endif
2151 if (needs_guard) {
2152 // Peel the loop to ensure there's a zero trip guard
2153 Node_List old_new;
2154 phase->do_peeling(this, old_new);
2155 }
2157 // Replace the phi at loop head with the final value of the last
2158 // iteration. Then the CountedLoopEnd will collapse (backedge never
2159 // taken) and all loop-invariant uses of the exit values will be correct.
2160 Node *phi = cl->phi();
2161 Node *exact_limit = phase->exact_limit(this);
2162 if (exact_limit != cl->limit()) {
2163 // We also need to replace the original limit to collapse loop exit.
2164 Node* cmp = cl->loopexit()->cmp_node();
2165 assert(cl->limit() == cmp->in(2), "sanity");
2166 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
2167 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
2168 }
2169 // Note: the final value after increment should not overflow since
2170 // counted loop has limit check predicate.
2171 Node *final = new (phase->C) SubINode( exact_limit, cl->stride() );
2172 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
2173 phase->_igvn.replace_node(phi,final);
2174 phase->C->set_major_progress();
2175 return true;
2176 }
2178 //------------------------------policy_do_one_iteration_loop-------------------
2179 // Convert one iteration loop into normal code.
2180 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
2181 if (!_head->as_Loop()->is_valid_counted_loop())
2182 return false; // Only for counted loop
2184 CountedLoopNode *cl = _head->as_CountedLoop();
2185 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
2186 return false;
2187 }
2189 #ifndef PRODUCT
2190 if(TraceLoopOpts) {
2191 tty->print("OneIteration ");
2192 this->dump_head();
2193 }
2194 #endif
2196 Node *init_n = cl->init_trip();
2197 #ifdef ASSERT
2198 // Loop boundaries should be constant since trip count is exact.
2199 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
2200 #endif
2201 // Replace the phi at loop head with the value of the init_trip.
2202 // Then the CountedLoopEnd will collapse (backedge will not be taken)
2203 // and all loop-invariant uses of the exit values will be correct.
2204 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
2205 phase->C->set_major_progress();
2206 return true;
2207 }
2209 //=============================================================================
2210 //------------------------------iteration_split_impl---------------------------
2211 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
2212 // Compute exact loop trip count if possible.
2213 compute_exact_trip_count(phase);
2215 // Convert one iteration loop into normal code.
2216 if (policy_do_one_iteration_loop(phase))
2217 return true;
2219 // Check and remove empty loops (spam micro-benchmarks)
2220 if (policy_do_remove_empty_loop(phase))
2221 return true; // Here we removed an empty loop
2223 bool should_peel = policy_peeling(phase); // Should we peel?
2225 bool should_unswitch = policy_unswitching(phase);
2227 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
2228 // This removes loop-invariant tests (usually null checks).
2229 if (!_head->is_CountedLoop()) { // Non-counted loop
2230 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
2231 // Partial peel succeeded so terminate this round of loop opts
2232 return false;
2233 }
2234 if (should_peel) { // Should we peel?
2235 #ifndef PRODUCT
2236 if (PrintOpto) tty->print_cr("should_peel");
2237 #endif
2238 phase->do_peeling(this,old_new);
2239 } else if (should_unswitch) {
2240 phase->do_unswitching(this, old_new);
2241 }
2242 return true;
2243 }
2244 CountedLoopNode *cl = _head->as_CountedLoop();
2246 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
2248 // Do nothing special to pre- and post- loops
2249 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
2251 // Compute loop trip count from profile data
2252 compute_profile_trip_cnt(phase);
2254 // Before attempting fancy unrolling, RCE or alignment, see if we want
2255 // to completely unroll this loop or do loop unswitching.
2256 if (cl->is_normal_loop()) {
2257 if (should_unswitch) {
2258 phase->do_unswitching(this, old_new);
2259 return true;
2260 }
2261 bool should_maximally_unroll = policy_maximally_unroll(phase);
2262 if (should_maximally_unroll) {
2263 // Here we did some unrolling and peeling. Eventually we will
2264 // completely unroll this loop and it will no longer be a loop.
2265 phase->do_maximally_unroll(this,old_new);
2266 return true;
2267 }
2268 }
2270 // Skip next optimizations if running low on nodes. Note that
2271 // policy_unswitching and policy_maximally_unroll have this check.
2272 uint nodes_left = MaxNodeLimit - (uint) phase->C->live_nodes();
2273 if ((2 * _body.size()) > nodes_left) {
2274 return true;
2275 }
2277 // Counted loops may be peeled, may need some iterations run up
2278 // front for RCE, and may want to align loop refs to a cache
2279 // line. Thus we clone a full loop up front whose trip count is
2280 // at least 1 (if peeling), but may be several more.
2282 // The main loop will start cache-line aligned with at least 1
2283 // iteration of the unrolled body (zero-trip test required) and
2284 // will have some range checks removed.
2286 // A post-loop will finish any odd iterations (leftover after
2287 // unrolling), plus any needed for RCE purposes.
2289 bool should_unroll = policy_unroll(phase);
2291 bool should_rce = policy_range_check(phase);
2293 bool should_align = policy_align(phase);
2295 // If not RCE'ing (iteration splitting) or Aligning, then we do not
2296 // need a pre-loop. We may still need to peel an initial iteration but
2297 // we will not be needing an unknown number of pre-iterations.
2298 //
2299 // Basically, if may_rce_align reports FALSE first time through,
2300 // we will not be able to later do RCE or Aligning on this loop.
2301 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
2303 // If we have any of these conditions (RCE, alignment, unrolling) met, then
2304 // we switch to the pre-/main-/post-loop model. This model also covers
2305 // peeling.
2306 if (should_rce || should_align || should_unroll) {
2307 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
2308 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
2310 // Adjust the pre- and main-loop limits to let the pre and post loops run
2311 // with full checks, but the main-loop with no checks. Remove said
2312 // checks from the main body.
2313 if (should_rce)
2314 phase->do_range_check(this,old_new);
2316 // Double loop body for unrolling. Adjust the minimum-trip test (will do
2317 // twice as many iterations as before) and the main body limit (only do
2318 // an even number of trips). If we are peeling, we might enable some RCE
2319 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
2320 // peeling.
2321 if (should_unroll && !should_peel)
2322 phase->do_unroll(this,old_new, true);
2324 // Adjust the pre-loop limits to align the main body
2325 // iterations.
2326 if (should_align)
2327 Unimplemented();
2329 } else { // Else we have an unchanged counted loop
2330 if (should_peel) // Might want to peel but do nothing else
2331 phase->do_peeling(this,old_new);
2332 }
2333 return true;
2334 }
2337 //=============================================================================
2338 //------------------------------iteration_split--------------------------------
2339 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
2340 // Recursively iteration split nested loops
2341 if (_child && !_child->iteration_split(phase, old_new))
2342 return false;
2344 // Clean out prior deadwood
2345 DCE_loop_body();
2348 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
2349 // Replace with a 1-in-10 exit guess.
2350 if (_parent /*not the root loop*/ &&
2351 !_irreducible &&
2352 // Also ignore the occasional dead backedge
2353 !tail()->is_top()) {
2354 adjust_loop_exit_prob(phase);
2355 }
2357 // Gate unrolling, RCE and peeling efforts.
2358 if (!_child && // If not an inner loop, do not split
2359 !_irreducible &&
2360 _allow_optimizations &&
2361 !tail()->is_top()) { // Also ignore the occasional dead backedge
2362 if (!_has_call) {
2363 if (!iteration_split_impl(phase, old_new)) {
2364 return false;
2365 }
2366 } else if (policy_unswitching(phase)) {
2367 phase->do_unswitching(this, old_new);
2368 }
2369 }
2371 // Minor offset re-organization to remove loop-fallout uses of
2372 // trip counter when there was no major reshaping.
2373 phase->reorg_offsets(this);
2375 if (_next && !_next->iteration_split(phase, old_new))
2376 return false;
2377 return true;
2378 }
2381 //=============================================================================
2382 // Process all the loops in the loop tree and replace any fill
2383 // patterns with an intrisc version.
2384 bool PhaseIdealLoop::do_intrinsify_fill() {
2385 bool changed = false;
2386 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2387 IdealLoopTree* lpt = iter.current();
2388 changed |= intrinsify_fill(lpt);
2389 }
2390 return changed;
2391 }
2394 // Examine an inner loop looking for a a single store of an invariant
2395 // value in a unit stride loop,
2396 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2397 Node*& shift, Node*& con) {
2398 const char* msg = NULL;
2399 Node* msg_node = NULL;
2401 store_value = NULL;
2402 con = NULL;
2403 shift = NULL;
2405 // Process the loop looking for stores. If there are multiple
2406 // stores or extra control flow give at this point.
2407 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2408 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2409 Node* n = lpt->_body.at(i);
2410 if (n->outcnt() == 0) continue; // Ignore dead
2411 if (n->is_Store()) {
2412 if (store != NULL) {
2413 msg = "multiple stores";
2414 break;
2415 }
2416 int opc = n->Opcode();
2417 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
2418 msg = "oop fills not handled";
2419 break;
2420 }
2421 Node* value = n->in(MemNode::ValueIn);
2422 if (!lpt->is_invariant(value)) {
2423 msg = "variant store value";
2424 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2425 msg = "not array address";
2426 }
2427 store = n;
2428 store_value = value;
2429 } else if (n->is_If() && n != head->loopexit()) {
2430 msg = "extra control flow";
2431 msg_node = n;
2432 }
2433 }
2435 if (store == NULL) {
2436 // No store in loop
2437 return false;
2438 }
2440 if (msg == NULL && head->stride_con() != 1) {
2441 // could handle negative strides too
2442 if (head->stride_con() < 0) {
2443 msg = "negative stride";
2444 } else {
2445 msg = "non-unit stride";
2446 }
2447 }
2449 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2450 msg = "can't handle store address";
2451 msg_node = store->in(MemNode::Address);
2452 }
2454 if (msg == NULL &&
2455 (!store->in(MemNode::Memory)->is_Phi() ||
2456 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2457 msg = "store memory isn't proper phi";
2458 msg_node = store->in(MemNode::Memory);
2459 }
2461 // Make sure there is an appropriate fill routine
2462 BasicType t = store->as_Mem()->memory_type();
2463 const char* fill_name;
2464 if (msg == NULL &&
2465 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2466 msg = "unsupported store";
2467 msg_node = store;
2468 }
2470 if (msg != NULL) {
2471 #ifndef PRODUCT
2472 if (TraceOptimizeFill) {
2473 tty->print_cr("not fill intrinsic candidate: %s", msg);
2474 if (msg_node != NULL) msg_node->dump();
2475 }
2476 #endif
2477 return false;
2478 }
2480 // Make sure the address expression can be handled. It should be
2481 // head->phi * elsize + con. head->phi might have a ConvI2L.
2482 Node* elements[4];
2483 Node* conv = NULL;
2484 bool found_index = false;
2485 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2486 for (int e = 0; e < count; e++) {
2487 Node* n = elements[e];
2488 if (n->is_Con() && con == NULL) {
2489 con = n;
2490 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2491 Node* value = n->in(1);
2492 #ifdef _LP64
2493 if (value->Opcode() == Op_ConvI2L) {
2494 conv = value;
2495 value = value->in(1);
2496 }
2497 #endif
2498 if (value != head->phi()) {
2499 msg = "unhandled shift in address";
2500 } else {
2501 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
2502 msg = "scale doesn't match";
2503 } else {
2504 found_index = true;
2505 shift = n;
2506 }
2507 }
2508 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2509 if (n->in(1) == head->phi()) {
2510 found_index = true;
2511 conv = n;
2512 } else {
2513 msg = "unhandled input to ConvI2L";
2514 }
2515 } else if (n == head->phi()) {
2516 // no shift, check below for allowed cases
2517 found_index = true;
2518 } else {
2519 msg = "unhandled node in address";
2520 msg_node = n;
2521 }
2522 }
2524 if (count == -1) {
2525 msg = "malformed address expression";
2526 msg_node = store;
2527 }
2529 if (!found_index) {
2530 msg = "missing use of index";
2531 }
2533 // byte sized items won't have a shift
2534 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2535 msg = "can't find shift";
2536 msg_node = store;
2537 }
2539 if (msg != NULL) {
2540 #ifndef PRODUCT
2541 if (TraceOptimizeFill) {
2542 tty->print_cr("not fill intrinsic: %s", msg);
2543 if (msg_node != NULL) msg_node->dump();
2544 }
2545 #endif
2546 return false;
2547 }
2549 // No make sure all the other nodes in the loop can be handled
2550 VectorSet ok(Thread::current()->resource_area());
2552 // store related values are ok
2553 ok.set(store->_idx);
2554 ok.set(store->in(MemNode::Memory)->_idx);
2556 CountedLoopEndNode* loop_exit = head->loopexit();
2557 guarantee(loop_exit != NULL, "no loop exit node");
2559 // Loop structure is ok
2560 ok.set(head->_idx);
2561 ok.set(loop_exit->_idx);
2562 ok.set(head->phi()->_idx);
2563 ok.set(head->incr()->_idx);
2564 ok.set(loop_exit->cmp_node()->_idx);
2565 ok.set(loop_exit->in(1)->_idx);
2567 // Address elements are ok
2568 if (con) ok.set(con->_idx);
2569 if (shift) ok.set(shift->_idx);
2570 if (conv) ok.set(conv->_idx);
2572 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2573 Node* n = lpt->_body.at(i);
2574 if (n->outcnt() == 0) continue; // Ignore dead
2575 if (ok.test(n->_idx)) continue;
2576 // Backedge projection is ok
2577 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
2578 if (!n->is_AddP()) {
2579 msg = "unhandled node";
2580 msg_node = n;
2581 break;
2582 }
2583 }
2585 // Make sure no unexpected values are used outside the loop
2586 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2587 Node* n = lpt->_body.at(i);
2588 // These values can be replaced with other nodes if they are used
2589 // outside the loop.
2590 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2591 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2592 Node* use = iter.get();
2593 if (!lpt->_body.contains(use)) {
2594 msg = "node is used outside loop";
2595 // lpt->_body.dump();
2596 msg_node = n;
2597 break;
2598 }
2599 }
2600 }
2602 #ifdef ASSERT
2603 if (TraceOptimizeFill) {
2604 if (msg != NULL) {
2605 tty->print_cr("no fill intrinsic: %s", msg);
2606 if (msg_node != NULL) msg_node->dump();
2607 } else {
2608 tty->print_cr("fill intrinsic for:");
2609 }
2610 store->dump();
2611 if (Verbose) {
2612 lpt->_body.dump();
2613 }
2614 }
2615 #endif
2617 return msg == NULL;
2618 }
2622 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2623 // Only for counted inner loops
2624 if (!lpt->is_counted() || !lpt->is_inner()) {
2625 return false;
2626 }
2628 // Must have constant stride
2629 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2630 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
2631 return false;
2632 }
2634 // Check that the body only contains a store of a loop invariant
2635 // value that is indexed by the loop phi.
2636 Node* store = NULL;
2637 Node* store_value = NULL;
2638 Node* shift = NULL;
2639 Node* offset = NULL;
2640 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2641 return false;
2642 }
2644 #ifndef PRODUCT
2645 if (TraceLoopOpts) {
2646 tty->print("ArrayFill ");
2647 lpt->dump_head();
2648 }
2649 #endif
2651 // Now replace the whole loop body by a call to a fill routine that
2652 // covers the same region as the loop.
2653 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2655 // Build an expression for the beginning of the copy region
2656 Node* index = head->init_trip();
2657 #ifdef _LP64
2658 index = new (C) ConvI2LNode(index);
2659 _igvn.register_new_node_with_optimizer(index);
2660 #endif
2661 if (shift != NULL) {
2662 // byte arrays don't require a shift but others do.
2663 index = new (C) LShiftXNode(index, shift->in(2));
2664 _igvn.register_new_node_with_optimizer(index);
2665 }
2666 index = new (C) AddPNode(base, base, index);
2667 _igvn.register_new_node_with_optimizer(index);
2668 Node* from = new (C) AddPNode(base, index, offset);
2669 _igvn.register_new_node_with_optimizer(from);
2670 // Compute the number of elements to copy
2671 Node* len = new (C) SubINode(head->limit(), head->init_trip());
2672 _igvn.register_new_node_with_optimizer(len);
2674 BasicType t = store->as_Mem()->memory_type();
2675 bool aligned = false;
2676 if (offset != NULL && head->init_trip()->is_Con()) {
2677 int element_size = type2aelembytes(t);
2678 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2679 }
2681 // Build a call to the fill routine
2682 const char* fill_name;
2683 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2684 assert(fill != NULL, "what?");
2686 // Convert float/double to int/long for fill routines
2687 if (t == T_FLOAT) {
2688 store_value = new (C) MoveF2INode(store_value);
2689 _igvn.register_new_node_with_optimizer(store_value);
2690 } else if (t == T_DOUBLE) {
2691 store_value = new (C) MoveD2LNode(store_value);
2692 _igvn.register_new_node_with_optimizer(store_value);
2693 }
2695 if (CCallingConventionRequiresIntsAsLongs &&
2696 // See StubRoutines::select_fill_function for types. FLOAT has been converted to INT.
2697 (t == T_FLOAT || t == T_INT || is_subword_type(t))) {
2698 store_value = new (C) ConvI2LNode(store_value);
2699 _igvn.register_new_node_with_optimizer(store_value);
2700 }
2702 Node* mem_phi = store->in(MemNode::Memory);
2703 Node* result_ctrl;
2704 Node* result_mem;
2705 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2706 CallLeafNode *call = new (C) CallLeafNoFPNode(call_type, fill,
2707 fill_name, TypeAryPtr::get_array_body_type(t));
2708 uint cnt = 0;
2709 call->init_req(TypeFunc::Parms + cnt++, from);
2710 call->init_req(TypeFunc::Parms + cnt++, store_value);
2711 if (CCallingConventionRequiresIntsAsLongs) {
2712 call->init_req(TypeFunc::Parms + cnt++, C->top());
2713 }
2714 #ifdef _LP64
2715 len = new (C) ConvI2LNode(len);
2716 _igvn.register_new_node_with_optimizer(len);
2717 #endif
2718 call->init_req(TypeFunc::Parms + cnt++, len);
2719 #ifdef _LP64
2720 call->init_req(TypeFunc::Parms + cnt++, 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) ProjNode(call,TypeFunc::Control);
2729 _igvn.register_new_node_with_optimizer(result_ctrl);
2730 result_mem = new (C) ProjNode(call,TypeFunc::Memory);
2731 _igvn.register_new_node_with_optimizer(result_mem);
2733 /* Disable following optimization until proper fix (add missing checks).
2735 // If this fill is tightly coupled to an allocation and overwrites
2736 // the whole body, allow it to take over the zeroing.
2737 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2738 if (alloc != NULL && alloc->is_AllocateArray()) {
2739 Node* length = alloc->as_AllocateArray()->Ideal_length();
2740 if (head->limit() == length &&
2741 head->init_trip() == _igvn.intcon(0)) {
2742 if (TraceOptimizeFill) {
2743 tty->print_cr("Eliminated zeroing in allocation");
2744 }
2745 alloc->maybe_set_complete(&_igvn);
2746 } else {
2747 #ifdef ASSERT
2748 if (TraceOptimizeFill) {
2749 tty->print_cr("filling array but bounds don't match");
2750 alloc->dump();
2751 head->init_trip()->dump();
2752 head->limit()->dump();
2753 length->dump();
2754 }
2755 #endif
2756 }
2757 }
2758 */
2760 // Redirect the old control and memory edges that are outside the loop.
2761 Node* exit = head->loopexit()->proj_out(0);
2762 // Sometimes the memory phi of the head is used as the outgoing
2763 // state of the loop. It's safe in this case to replace it with the
2764 // result_mem.
2765 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2766 _igvn.replace_node(exit, result_ctrl);
2767 _igvn.replace_node(store, result_mem);
2768 // Any uses the increment outside of the loop become the loop limit.
2769 _igvn.replace_node(head->incr(), head->limit());
2771 // Disconnect the head from the loop.
2772 for (uint i = 0; i < lpt->_body.size(); i++) {
2773 Node* n = lpt->_body.at(i);
2774 _igvn.replace_node(n, C->top());
2775 }
2777 return true;
2778 }