Mercurial > hg > truffle
annotate src/share/vm/opto/loopTransform.cpp @ 3333:ae93231c7a1f
7039652: Performance regression after 7004547 changes
Summary: Use unrolled_count() to limit unrolling and use the stride check only for initial stride value.
Reviewed-by: never
author | kvn |
---|---|
date | Thu, 28 Apr 2011 16:40:23 -0700 |
parents | 6c97c830fb6f |
children | bad7ecd0b6ed |
rev | line source |
---|---|
0 | 1 /* |
2426
1d1603768966
7010070: Update all 2010 Oracle-changed OpenJDK files to have the proper copyright dates - second pass
trims
parents:
2417
diff
changeset
|
2 * Copyright (c) 2000, 2011, Oracle and/or its affiliates. All rights reserved. |
0 | 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1303
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1303
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1303
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 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" | |
0 | 37 |
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 } | |
52 | |
53 | |
54 //============================================================================= | |
55 | |
56 | |
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 } | |
65 | |
2465 | 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(); | |
78 | |
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 | |
82 | |
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 | |
88 | |
89 Node* init_n = cl->init_trip(); | |
90 Node* limit_n = cl->limit(); | |
91 if (init_n != NULL && init_n->is_Con() && | |
92 limit_n != NULL && limit_n->is_Con()) { | |
93 // Use longs to avoid integer overflow. | |
94 int stride_con = cl->stride_con(); | |
95 long init_con = cl->init_trip()->get_int(); | |
96 long limit_con = cl->limit()->get_int(); | |
97 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); | |
98 long trip_count = (limit_con - init_con + stride_m)/stride_con; | |
99 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { | |
100 // Set exact trip count. | |
101 cl->set_exact_trip_count((uint)trip_count); | |
102 } | |
103 } | |
104 } | |
105 | |
0 | 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 | |
118 | |
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; | |
135 | |
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 } | |
169 | |
170 //---------------------is_invariant_addition----------------------------- | |
171 // Return nonzero index of invariant operand for an Add or Sub | |
605 | 172 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. |
0 | 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 } | |
183 | |
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); | |
213 | |
214 bool neg_x = n2->is_Sub() && inv2_idx == 1; | |
215 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; | |
216 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; | |
217 if (n1->is_Sub() && inv1_idx == 1) { | |
218 neg_x = !neg_x; | |
219 neg_inv2 = !neg_inv2; | |
220 } | |
221 Node* inv1_c = phase->get_ctrl(inv1); | |
222 Node* inv2_c = phase->get_ctrl(inv2); | |
223 Node* n_inv1; | |
224 if (neg_inv1) { | |
225 Node *zero = phase->_igvn.intcon(0); | |
226 phase->set_ctrl(zero, phase->C->root()); | |
227 n_inv1 = new (phase->C, 3) SubINode(zero, inv1); | |
228 phase->register_new_node(n_inv1, inv1_c); | |
229 } else { | |
230 n_inv1 = inv1; | |
231 } | |
232 Node* inv; | |
233 if (neg_inv2) { | |
234 inv = new (phase->C, 3) SubINode(n_inv1, inv2); | |
235 } else { | |
236 inv = new (phase->C, 3) AddINode(n_inv1, inv2); | |
237 } | |
238 phase->register_new_node(inv, phase->get_early_ctrl(inv)); | |
239 | |
240 Node* addx; | |
241 if (neg_x) { | |
242 addx = new (phase->C, 3) SubINode(inv, x); | |
243 } else { | |
244 addx = new (phase->C, 3) AddINode(x, inv); | |
245 } | |
246 phase->register_new_node(addx, phase->get_ctrl(x)); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
247 phase->_igvn.replace_node(n1, addx); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
248 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
249 _body.yank(n1); |
0 | 250 return addx; |
251 } | |
252 | |
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 } | |
265 | |
266 //------------------------------policy_peeling--------------------------------- | |
267 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can | |
268 // make some loop-invariant test (usually a null-check) happen before the loop. | |
269 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { | |
270 Node *test = ((IdealLoopTree*)this)->tail(); | |
271 int body_size = ((IdealLoopTree*)this)->_body.size(); | |
272 int uniq = phase->C->unique(); | |
273 // Peeling does loop cloning which can result in O(N^2) node construction | |
274 if( body_size > 255 /* Prevent overflow for large body_size */ | |
275 || (body_size * body_size + uniq > MaxNodeLimit) ) { | |
276 return false; // too large to safely clone | |
277 } | |
278 while( test != _head ) { // Scan till run off top of loop | |
279 if( test->is_If() ) { // Test? | |
280 Node *ctrl = phase->get_ctrl(test->in(1)); | |
281 if (ctrl->is_top()) | |
282 return false; // Found dead test on live IF? No peeling! | |
283 // Standard IF only has one input value to check for loop invariance | |
284 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); | |
285 // Condition is not a member of this loop? | |
286 if( !is_member(phase->get_loop(ctrl)) && | |
287 is_loop_exit(test) ) | |
288 return true; // Found reason to peel! | |
289 } | |
290 // Walk up dominators to loop _head looking for test which is | |
291 // executed on every path thru loop. | |
292 test = phase->idom(test); | |
293 } | |
294 return false; | |
295 } | |
296 | |
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 | |
308 | |
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 | |
328 | |
329 } // End of while( progress ) | |
330 } | |
331 | |
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. | |
2445 | 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 // | |
0 | 470 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { |
471 | |
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. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
478 #ifndef PRODUCT |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
479 if (TraceLoopOpts) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
480 tty->print("Peel "); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
481 loop->dump_head(); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
482 } |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
483 #endif |
2445 | 484 Node* head = loop->_head; |
485 bool counted_loop = head->is_CountedLoop(); | |
486 if (counted_loop) { | |
487 CountedLoopNode *cl = head->as_CountedLoop(); | |
0 | 488 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); |
489 cl->set_trip_count(cl->trip_count() - 1); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
490 if (cl->is_main_loop()) { |
0 | 491 cl->set_normal_loop(); |
492 #ifndef PRODUCT | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
493 if (PrintOpto && VerifyLoopOptimizations) { |
0 | 494 tty->print("Peeling a 'main' loop; resetting to 'normal' "); |
495 loop->dump_head(); | |
496 } | |
497 #endif | |
498 } | |
499 } | |
2445 | 500 Node* entry = head->in(LoopNode::EntryControl); |
0 | 501 |
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). | |
2445 | 504 clone_loop( loop, old_new, dom_depth(head) ); |
0 | 505 |
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. | |
2445 | 512 Node* new_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx]; |
513 new_exit_value = move_loop_predicates(entry, new_exit_value); | |
514 _igvn.hash_delete(head); | |
515 head->set_req(LoopNode::EntryControl, new_exit_value); | |
516 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { | |
517 Node* old = head->fast_out(j); | |
518 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { | |
519 new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; | |
520 if (!new_exit_value ) // Backedge value is ALSO loop invariant? | |
0 | 521 // Then loop body backedge value remains the same. |
522 new_exit_value = old->in(LoopNode::LoopBackControl); | |
523 _igvn.hash_delete(old); | |
524 old->set_req(LoopNode::EntryControl, new_exit_value); | |
525 } | |
526 } | |
527 | |
528 | |
529 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the | |
530 // extra backedge user. | |
2445 | 531 Node* new_head = old_new[head->_idx]; |
532 _igvn.hash_delete(new_head); | |
533 new_head->set_req(LoopNode::LoopBackControl, C->top()); | |
534 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { | |
535 Node* use = new_head->fast_out(j2); | |
536 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { | |
0 | 537 _igvn.hash_delete(use); |
538 use->set_req(LoopNode::LoopBackControl, C->top()); | |
539 } | |
540 } | |
541 | |
542 | |
543 // Step 4: Correct dom-depth info. Set to loop-head depth. | |
2445 | 544 int dd = dom_depth(head); |
545 set_idom(head, head->in(1), dd); | |
0 | 546 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { |
547 Node *old = loop->_body.at(j3); | |
548 Node *nnn = old_new[old->_idx]; | |
549 if (!has_ctrl(nnn)) | |
550 set_idom(nnn, idom(nnn), dd-1); | |
551 // While we're at it, remove any SafePoints from the peeled code | |
2445 | 552 if (old->Opcode() == Op_SafePoint) { |
0 | 553 Node *nnn = old_new[old->_idx]; |
554 lazy_replace(nnn,nnn->in(TypeFunc::Control)); | |
555 } | |
556 } | |
557 | |
558 // Now force out all loop-invariant dominating tests. The optimizer | |
559 // finds some, but we _know_ they are all useless. | |
560 peeled_dom_test_elim(loop,old_new); | |
561 | |
562 loop->record_for_igvn(); | |
563 } | |
564 | |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
565 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
566 |
0 | 567 //------------------------------policy_maximally_unroll------------------------ |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
568 // Calculate exact loop trip count and return true if loop can be maximally |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
569 // unrolled. |
0 | 570 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { |
571 CountedLoopNode *cl = _head->as_CountedLoop(); | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
572 assert(cl->is_normal_loop(), ""); |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
573 if (!cl->is_valid_counted_loop()) |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
574 return false; // Malformed counted loop |
0 | 575 |
2465 | 576 if (!cl->has_exact_trip_count()) { |
577 // Trip count is not exact. | |
0 | 578 return false; |
579 } | |
580 | |
2465 | 581 uint trip_count = cl->trip_count(); |
582 // Note, max_juint is used to indicate unknown trip count. | |
583 assert(trip_count > 1, "one iteration loop should be optimized out already"); | |
584 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); | |
0 | 585 |
586 // Real policy: if we maximally unroll, does it get too big? | |
587 // Allow the unrolled mess to get larger than standard loop | |
588 // size. After all, it will no longer be a loop. | |
589 uint body_size = _body.size(); | |
590 uint unroll_limit = (uint)LoopUnrollLimit * 4; | |
591 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
592 if (trip_count > unroll_limit || body_size > unroll_limit) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
593 return false; |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
594 } |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
595 |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
596 // Take into account that after unroll conjoined heads and tails will fold, |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
597 // otherwise policy_unroll() may allow more unrolling than max unrolling. |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
598 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
599 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
600 if (body_size != tst_body_size) // Check for int overflow |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
601 return false; |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
602 if (new_body_size > unroll_limit || |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
603 // Unrolling can result in a large amount of node construction |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
604 new_body_size >= MaxNodeLimit - phase->C->unique()) { |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
605 return false; |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
606 } |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
607 |
2417 | 608 // Currently we don't have policy to optimize one iteration loops. |
609 // Maximally unrolling transformation is used for that: | |
610 // it is peeled and the original loop become non reachable (dead). | |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
611 // Also fully unroll a loop with few iterations regardless next |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
612 // conditions since following loop optimizations will split |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
613 // such loop anyway (pre-main-post). |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
614 if (trip_count <= 3) |
2417 | 615 return true; |
616 | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
617 // Do not unroll a loop with String intrinsics code. |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
618 // String intrinsics are large and have loops. |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
619 for (uint k = 0; k < _body.size(); k++) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
620 Node* n = _body.at(k); |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
621 switch (n->Opcode()) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
622 case Op_StrComp: |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
623 case Op_StrEquals: |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
624 case Op_StrIndexOf: |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
625 case Op_AryEq: { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
626 return false; |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
627 } |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
628 } // switch |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
629 } |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
630 |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
631 return true; // Do maximally unroll |
0 | 632 } |
633 | |
634 | |
3333 | 635 #define MAX_UNROLL 16 // maximum number of unrolls for main loop |
636 | |
0 | 637 //------------------------------policy_unroll---------------------------------- |
638 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if | |
639 // the loop is a CountedLoop and the body is small enough. | |
640 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { | |
641 | |
642 CountedLoopNode *cl = _head->as_CountedLoop(); | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
643 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); |
0 | 644 |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
645 if (!cl->is_valid_counted_loop()) |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
646 return false; // Malformed counted loop |
0 | 647 |
648 // protect against over-unrolling | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
649 if (cl->trip_count() <= 1) return false; |
0 | 650 |
3333 | 651 int future_unroll_ct = cl->unrolled_count() * 2; |
652 if (future_unroll_ct > MAX_UNROLL) return false; | |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
653 |
3333 | 654 // Check for initial stride being a small enough constant |
655 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; | |
0 | 656 |
657 // Don't unroll if the next round of unrolling would push us | |
658 // over the expected trip count of the loop. One is subtracted | |
659 // from the expected trip count because the pre-loop normally | |
660 // executes 1 iteration. | |
661 if (UnrollLimitForProfileCheck > 0 && | |
662 cl->profile_trip_cnt() != COUNT_UNKNOWN && | |
663 future_unroll_ct > UnrollLimitForProfileCheck && | |
664 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { | |
665 return false; | |
666 } | |
667 | |
668 // When unroll count is greater than LoopUnrollMin, don't unroll if: | |
669 // the residual iterations are more than 10% of the trip count | |
670 // and rounds of "unroll,optimize" are not making significant progress | |
671 // Progress defined as current size less than 20% larger than previous size. | |
672 if (UseSuperWord && cl->node_count_before_unroll() > 0 && | |
673 future_unroll_ct > LoopUnrollMin && | |
674 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && | |
675 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { | |
676 return false; | |
677 } | |
678 | |
679 Node *init_n = cl->init_trip(); | |
680 Node *limit_n = cl->limit(); | |
681 // Non-constant bounds. | |
682 // Protect against over-unrolling when init or/and limit are not constant | |
683 // (so that trip_count's init value is maxint) but iv range is known. | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
684 if (init_n == NULL || !init_n->is_Con() || |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
685 limit_n == NULL || !limit_n->is_Con()) { |
0 | 686 Node* phi = cl->phi(); |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
687 if (phi != NULL) { |
0 | 688 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); |
689 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); | |
690 int next_stride = cl->stride_con() * 2; // stride after this unroll | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
691 if (next_stride > 0) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
692 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
693 iv_type->_lo + next_stride > iv_type->_hi) { |
0 | 694 return false; // over-unrolling |
695 } | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
696 } else if (next_stride < 0) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
697 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
698 iv_type->_hi + next_stride < iv_type->_lo) { |
0 | 699 return false; // over-unrolling |
700 } | |
701 } | |
702 } | |
703 } | |
704 | |
705 // Adjust body_size to determine if we unroll or not | |
706 uint body_size = _body.size(); | |
707 // Key test to unroll CaffeineMark's Logic test | |
708 int xors_in_loop = 0; | |
709 // Also count ModL, DivL and MulL which expand mightly | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
710 for (uint k = 0; k < _body.size(); k++) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
711 Node* n = _body.at(k); |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
712 switch (n->Opcode()) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
713 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
714 case Op_ModL: body_size += 30; break; |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
715 case Op_DivL: body_size += 30; break; |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
716 case Op_MulL: body_size += 10; break; |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
717 case Op_StrComp: |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
718 case Op_StrEquals: |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
719 case Op_StrIndexOf: |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
720 case Op_AryEq: { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
721 // Do not unroll a loop with String intrinsics code. |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
722 // String intrinsics are large and have loops. |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
723 return false; |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
724 } |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
725 } // switch |
0 | 726 } |
727 | |
728 // Check for being too big | |
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
729 if (body_size > (uint)LoopUnrollLimit) { |
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
730 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; |
0 | 731 // Normal case: loop too big |
732 return false; | |
733 } | |
734 | |
735 // Unroll once! (Each trip will soon do double iterations) | |
736 return true; | |
737 } | |
738 | |
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 | |
605 | 742 // aligned in a loop (unless the VM guarantees mutual alignment). Note that |
0 | 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 } | |
748 | |
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; | |
754 | |
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(); | |
761 | |
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? | |
767 | |
768 // Comparing trip+off vs limit | |
769 Node *bol = iff->in(1); | |
770 if( bol->req() != 2 ) continue; // dead constant test | |
1172 | 771 if (!bol->is_Bool()) { |
772 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); | |
773 continue; | |
774 } | |
0 | 775 Node *cmp = bol->in(1); |
776 | |
777 Node *rc_exp = cmp->in(1); | |
778 Node *limit = cmp->in(2); | |
779 | |
780 Node *limit_c = phase->get_ctrl(limit); | |
781 if( limit_c == phase->C->top() ) | |
782 return false; // Found dead test on live IF? No RCE! | |
783 if( is_member(phase->get_loop(limit_c) ) ) { | |
784 // Compare might have operands swapped; commute them | |
785 rc_exp = cmp->in(2); | |
786 limit = cmp->in(1); | |
787 limit_c = phase->get_ctrl(limit); | |
788 if( is_member(phase->get_loop(limit_c) ) ) | |
789 continue; // Both inputs are loop varying; cannot RCE | |
790 } | |
791 | |
792 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { | |
793 continue; | |
794 } | |
795 // Yeah! Found a test like 'trip+off vs limit' | |
796 // Test is an IfNode, has 2 projections. If BOTH are in the loop | |
797 // we need loop unswitching instead of iteration splitting. | |
798 if( is_loop_exit(iff) ) | |
799 return true; // Found reason to split iterations | |
800 } // End of is IF | |
801 } | |
802 | |
803 return false; | |
804 } | |
805 | |
806 //------------------------------policy_peel_only------------------------------- | |
807 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful | |
808 // for unrolling loops with NO array accesses. | |
809 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { | |
810 | |
811 for( uint i = 0; i < _body.size(); i++ ) | |
812 if( _body[i]->is_Mem() ) | |
813 return false; | |
814 | |
815 // No memory accesses at all! | |
816 return true; | |
817 } | |
818 | |
819 //------------------------------clone_up_backedge_goo-------------------------- | |
820 // If Node n lives in the back_ctrl block and cannot float, we clone a private | |
821 // version of n in preheader_ctrl block and return that, otherwise return n. | |
822 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) { | |
823 if( get_ctrl(n) != back_ctrl ) return n; | |
824 | |
825 Node *x = NULL; // If required, a clone of 'n' | |
826 // Check for 'n' being pinned in the backedge. | |
827 if( n->in(0) && n->in(0) == back_ctrl ) { | |
828 x = n->clone(); // Clone a copy of 'n' to preheader | |
829 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader | |
830 } | |
831 | |
832 // Recursive fixup any other input edges into x. | |
833 // If there are no changes we can just return 'n', otherwise | |
834 // we need to clone a private copy and change it. | |
835 for( uint i = 1; i < n->req(); i++ ) { | |
836 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) ); | |
837 if( g != n->in(i) ) { | |
838 if( !x ) | |
839 x = n->clone(); | |
840 x->set_req(i, g); | |
841 } | |
842 } | |
843 if( x ) { // x can legally float to pre-header location | |
844 register_new_node( x, preheader_ctrl ); | |
845 return x; | |
846 } else { // raise n to cover LCA of uses | |
847 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); | |
848 } | |
849 return n; | |
850 } | |
851 | |
852 //------------------------------insert_pre_post_loops-------------------------- | |
853 // Insert pre and post loops. If peel_only is set, the pre-loop can not have | |
854 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no | |
855 // alignment. Useful to unroll loops that do no array accesses. | |
856 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { | |
857 | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
858 #ifndef PRODUCT |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
859 if (TraceLoopOpts) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
860 if (peel_only) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
861 tty->print("PeelMainPost "); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
862 else |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
863 tty->print("PreMainPost "); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
864 loop->dump_head(); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
865 } |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
866 #endif |
0 | 867 C->set_major_progress(); |
868 | |
869 // Find common pieces of the loop being guarded with pre & post loops | |
870 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); | |
871 assert( main_head->is_normal_loop(), "" ); | |
872 CountedLoopEndNode *main_end = main_head->loopexit(); | |
873 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); | |
874 uint dd_main_head = dom_depth(main_head); | |
875 uint max = main_head->outcnt(); | |
876 | |
877 Node *pre_header= main_head->in(LoopNode::EntryControl); | |
878 Node *init = main_head->init_trip(); | |
879 Node *incr = main_end ->incr(); | |
880 Node *limit = main_end ->limit(); | |
881 Node *stride = main_end ->stride(); | |
882 Node *cmp = main_end ->cmp_node(); | |
883 BoolTest::mask b_test = main_end->test_trip(); | |
884 | |
885 // Need only 1 user of 'bol' because I will be hacking the loop bounds. | |
886 Node *bol = main_end->in(CountedLoopEndNode::TestValue); | |
887 if( bol->outcnt() != 1 ) { | |
888 bol = bol->clone(); | |
889 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); | |
890 _igvn.hash_delete(main_end); | |
891 main_end->set_req(CountedLoopEndNode::TestValue, bol); | |
892 } | |
893 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. | |
894 if( cmp->outcnt() != 1 ) { | |
895 cmp = cmp->clone(); | |
896 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); | |
897 _igvn.hash_delete(bol); | |
898 bol->set_req(1, cmp); | |
899 } | |
900 | |
901 //------------------------------ | |
902 // Step A: Create Post-Loop. | |
903 Node* main_exit = main_end->proj_out(false); | |
904 assert( main_exit->Opcode() == Op_IfFalse, "" ); | |
905 int dd_main_exit = dom_depth(main_exit); | |
906 | |
907 // Step A1: Clone the loop body. The clone becomes the post-loop. The main | |
908 // loop pre-header illegally has 2 control users (old & new loops). | |
909 clone_loop( loop, old_new, dd_main_exit ); | |
910 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); | |
911 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); | |
912 post_head->set_post_loop(main_head); | |
913 | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
914 // Reduce the post-loop trip count. |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
915 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
916 post_end->_prob = PROB_FAIR; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
917 |
0 | 918 // Build the main-loop normal exit. |
919 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end); | |
920 _igvn.register_new_node_with_optimizer( new_main_exit ); | |
921 set_idom(new_main_exit, main_end, dd_main_exit ); | |
922 set_loop(new_main_exit, loop->_parent); | |
923 | |
924 // Step A2: Build a zero-trip guard for the post-loop. After leaving the | |
925 // main-loop, the post-loop may not execute at all. We 'opaque' the incr | |
926 // (the main-loop trip-counter exit value) because we will be changing | |
927 // the exit value (via unrolling) so we cannot constant-fold away the zero | |
928 // trip guard until all unrolling is done. | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
113
diff
changeset
|
929 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr); |
0 | 930 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit ); |
931 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test ); | |
932 register_new_node( zer_opaq, new_main_exit ); | |
933 register_new_node( zer_cmp , new_main_exit ); | |
934 register_new_node( zer_bol , new_main_exit ); | |
935 | |
936 // Build the IfNode | |
937 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); | |
938 _igvn.register_new_node_with_optimizer( zer_iff ); | |
939 set_idom(zer_iff, new_main_exit, dd_main_exit); | |
940 set_loop(zer_iff, loop->_parent); | |
941 | |
942 // Plug in the false-path, taken if we need to skip post-loop | |
943 _igvn.hash_delete( main_exit ); | |
944 main_exit->set_req(0, zer_iff); | |
945 _igvn._worklist.push(main_exit); | |
946 set_idom(main_exit, zer_iff, dd_main_exit); | |
947 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); | |
948 // Make the true-path, must enter the post loop | |
949 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff ); | |
950 _igvn.register_new_node_with_optimizer( zer_taken ); | |
951 set_idom(zer_taken, zer_iff, dd_main_exit); | |
952 set_loop(zer_taken, loop->_parent); | |
953 // Plug in the true path | |
954 _igvn.hash_delete( post_head ); | |
955 post_head->set_req(LoopNode::EntryControl, zer_taken); | |
956 set_idom(post_head, zer_taken, dd_main_exit); | |
957 | |
958 // Step A3: Make the fall-in values to the post-loop come from the | |
959 // fall-out values of the main-loop. | |
960 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { | |
961 Node* main_phi = main_head->fast_out(i); | |
962 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { | |
963 Node *post_phi = old_new[main_phi->_idx]; | |
964 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), | |
965 post_head->init_control(), | |
966 main_phi->in(LoopNode::LoopBackControl)); | |
967 _igvn.hash_delete(post_phi); | |
968 post_phi->set_req( LoopNode::EntryControl, fallmain ); | |
969 } | |
970 } | |
971 | |
972 // Update local caches for next stanza | |
973 main_exit = new_main_exit; | |
974 | |
975 | |
976 //------------------------------ | |
977 // Step B: Create Pre-Loop. | |
978 | |
979 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main | |
980 // loop pre-header illegally has 2 control users (old & new loops). | |
981 clone_loop( loop, old_new, dd_main_head ); | |
982 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); | |
983 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); | |
984 pre_head->set_pre_loop(main_head); | |
985 Node *pre_incr = old_new[incr->_idx]; | |
986 | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
987 // Reduce the pre-loop trip count. |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
988 pre_end->_prob = PROB_FAIR; |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
989 |
0 | 990 // Find the pre-loop normal exit. |
991 Node* pre_exit = pre_end->proj_out(false); | |
992 assert( pre_exit->Opcode() == Op_IfFalse, "" ); | |
993 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end); | |
994 _igvn.register_new_node_with_optimizer( new_pre_exit ); | |
995 set_idom(new_pre_exit, pre_end, dd_main_head); | |
996 set_loop(new_pre_exit, loop->_parent); | |
997 | |
998 // Step B2: Build a zero-trip guard for the main-loop. After leaving the | |
999 // pre-loop, the main-loop may not execute at all. Later in life this | |
1000 // zero-trip guard will become the minimum-trip guard when we unroll | |
1001 // the main-loop. | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
113
diff
changeset
|
1002 Node *min_opaq = new (C, 2) Opaque1Node(C, limit); |
0 | 1003 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq ); |
1004 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test ); | |
1005 register_new_node( min_opaq, new_pre_exit ); | |
1006 register_new_node( min_cmp , new_pre_exit ); | |
1007 register_new_node( min_bol , new_pre_exit ); | |
1008 | |
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
1009 // Build the IfNode (assume the main-loop is executed always). |
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
1010 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); |
0 | 1011 _igvn.register_new_node_with_optimizer( min_iff ); |
1012 set_idom(min_iff, new_pre_exit, dd_main_head); | |
1013 set_loop(min_iff, loop->_parent); | |
1014 | |
1015 // Plug in the false-path, taken if we need to skip main-loop | |
1016 _igvn.hash_delete( pre_exit ); | |
1017 pre_exit->set_req(0, min_iff); | |
1018 set_idom(pre_exit, min_iff, dd_main_head); | |
1019 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); | |
1020 // Make the true-path, must enter the main loop | |
1021 Node *min_taken = new (C, 1) IfTrueNode( min_iff ); | |
1022 _igvn.register_new_node_with_optimizer( min_taken ); | |
1023 set_idom(min_taken, min_iff, dd_main_head); | |
1024 set_loop(min_taken, loop->_parent); | |
1025 // Plug in the true path | |
1026 _igvn.hash_delete( main_head ); | |
1027 main_head->set_req(LoopNode::EntryControl, min_taken); | |
1028 set_idom(main_head, min_taken, dd_main_head); | |
1029 | |
1030 // Step B3: Make the fall-in values to the main-loop come from the | |
1031 // fall-out values of the pre-loop. | |
1032 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { | |
1033 Node* main_phi = main_head->fast_out(i2); | |
1034 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { | |
1035 Node *pre_phi = old_new[main_phi->_idx]; | |
1036 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), | |
1037 main_head->init_control(), | |
1038 pre_phi->in(LoopNode::LoopBackControl)); | |
1039 _igvn.hash_delete(main_phi); | |
1040 main_phi->set_req( LoopNode::EntryControl, fallpre ); | |
1041 } | |
1042 } | |
1043 | |
1044 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). | |
1045 // RCE and alignment may change this later. | |
1046 Node *cmp_end = pre_end->cmp_node(); | |
1047 assert( cmp_end->in(2) == limit, "" ); | |
1048 Node *pre_limit = new (C, 3) AddINode( init, stride ); | |
1049 | |
1050 // Save the original loop limit in this Opaque1 node for | |
1051 // use by range check elimination. | |
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
113
diff
changeset
|
1052 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit); |
0 | 1053 |
1054 register_new_node( pre_limit, pre_head->in(0) ); | |
1055 register_new_node( pre_opaq , pre_head->in(0) ); | |
1056 | |
1057 // Since no other users of pre-loop compare, I can hack limit directly | |
1058 assert( cmp_end->outcnt() == 1, "no other users" ); | |
1059 _igvn.hash_delete(cmp_end); | |
1060 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); | |
1061 | |
1062 // Special case for not-equal loop bounds: | |
1063 // Change pre loop test, main loop test, and the | |
1064 // main loop guard test to use lt or gt depending on stride | |
1065 // direction: | |
1066 // positive stride use < | |
1067 // negative stride use > | |
1068 | |
1069 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { | |
1070 | |
1071 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; | |
1072 // Modify pre loop end condition | |
1073 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); | |
1074 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test); | |
1075 register_new_node( new_bol0, pre_head->in(0) ); | |
1076 _igvn.hash_delete(pre_end); | |
1077 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); | |
1078 // Modify main loop guard condition | |
1079 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); | |
1080 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test); | |
1081 register_new_node( new_bol1, new_pre_exit ); | |
1082 _igvn.hash_delete(min_iff); | |
1083 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); | |
1084 // Modify main loop end condition | |
1085 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); | |
1086 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test); | |
1087 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); | |
1088 _igvn.hash_delete(main_end); | |
1089 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); | |
1090 } | |
1091 | |
1092 // Flag main loop | |
1093 main_head->set_main_loop(); | |
1094 if( peel_only ) main_head->set_main_no_pre_loop(); | |
1095 | |
1096 // It's difficult to be precise about the trip-counts | |
1097 // for the pre/post loops. They are usually very short, | |
1098 // so guess that 4 trips is a reasonable value. | |
1099 post_head->set_profile_trip_cnt(4.0); | |
1100 pre_head->set_profile_trip_cnt(4.0); | |
1101 | |
1102 // Now force out all loop-invariant dominating tests. The optimizer | |
1103 // finds some, but we _know_ they are all useless. | |
1104 peeled_dom_test_elim(loop,old_new); | |
1105 } | |
1106 | |
1107 //------------------------------is_invariant----------------------------- | |
1108 // Return true if n is invariant | |
1109 bool IdealLoopTree::is_invariant(Node* n) const { | |
1172 | 1110 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; |
0 | 1111 if (n_c->is_top()) return false; |
1112 return !is_member(_phase->get_loop(n_c)); | |
1113 } | |
1114 | |
1115 | |
1116 //------------------------------do_unroll-------------------------------------- | |
1117 // Unroll the loop body one step - make each trip do 2 iterations. | |
1118 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1119 assert(LoopUnrollLimit, ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1120 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1121 CountedLoopEndNode *loop_end = loop_head->loopexit(); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1122 assert(loop_end, ""); |
0 | 1123 #ifndef PRODUCT |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1124 if (PrintOpto && VerifyLoopOptimizations) { |
0 | 1125 tty->print("Unrolling "); |
1126 loop->dump_head(); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1127 } else if (TraceLoopOpts) { |
2465 | 1128 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1129 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1130 } else { |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1131 tty->print("Unroll %d ", loop_head->unrolled_count()*2); |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1132 } |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1133 loop->dump_head(); |
0 | 1134 } |
1135 #endif | |
1136 | |
1137 // Remember loop node count before unrolling to detect | |
1138 // if rounds of unroll,optimize are making progress | |
1139 loop_head->set_node_count_before_unroll(loop->_body.size()); | |
1140 | |
1141 Node *ctrl = loop_head->in(LoopNode::EntryControl); | |
1142 Node *limit = loop_head->limit(); | |
1143 Node *init = loop_head->init_trip(); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1144 Node *stride = loop_head->stride(); |
0 | 1145 |
1146 Node *opaq = NULL; | |
1147 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment | |
1148 assert( loop_head->is_main_loop(), "" ); | |
1149 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); | |
1150 Node *iff = ctrl->in(0); | |
1151 assert( iff->Opcode() == Op_If, "" ); | |
1152 Node *bol = iff->in(1); | |
1153 assert( bol->Opcode() == Op_Bool, "" ); | |
1154 Node *cmp = bol->in(1); | |
1155 assert( cmp->Opcode() == Op_CmpI, "" ); | |
1156 opaq = cmp->in(2); | |
1157 // Occasionally it's possible for a pre-loop Opaque1 node to be | |
1158 // optimized away and then another round of loop opts attempted. | |
1159 // We can not optimize this particular loop in that case. | |
1160 if( opaq->Opcode() != Op_Opaque1 ) | |
1161 return; // Cannot find pre-loop! Bail out! | |
1162 } | |
1163 | |
1164 C->set_major_progress(); | |
1165 | |
1166 // Adjust max trip count. The trip count is intentionally rounded | |
1167 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, | |
1168 // the main, unrolled, part of the loop will never execute as it is protected | |
1169 // by the min-trip test. See bug 4834191 for a case where we over-unrolled | |
1170 // and later determined that part of the unrolled loop was dead. | |
1171 loop_head->set_trip_count(loop_head->trip_count() / 2); | |
1172 | |
1173 // Double the count of original iterations in the unrolled loop body. | |
1174 loop_head->double_unrolled_count(); | |
1175 | |
1176 // ----------- | |
1177 // Step 2: Cut back the trip counter for an unroll amount of 2. | |
1178 // Loop will normally trip (limit - init)/stride_con. Since it's a | |
1179 // CountedLoop this is exact (stride divides limit-init exactly). | |
1180 // We are going to double the loop body, so we want to knock off any | |
1181 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. | |
1182 Node *span = new (C, 3) SubINode( limit, init ); | |
1183 register_new_node( span, ctrl ); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1184 Node *trip = new (C, 3) DivINode( 0, span, stride ); |
0 | 1185 register_new_node( trip, ctrl ); |
1186 Node *mtwo = _igvn.intcon(-2); | |
1187 set_ctrl(mtwo, C->root()); | |
1188 Node *rond = new (C, 3) AndINode( trip, mtwo ); | |
1189 register_new_node( rond, ctrl ); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1190 Node *spn2 = new (C, 3) MulINode( rond, stride ); |
0 | 1191 register_new_node( spn2, ctrl ); |
1192 Node *lim2 = new (C, 3) AddINode( spn2, init ); | |
1193 register_new_node( lim2, ctrl ); | |
1194 | |
1195 // Hammer in the new limit | |
1196 Node *ctrl2 = loop_end->in(0); | |
1197 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 ); | |
1198 register_new_node( cmp2, ctrl2 ); | |
1199 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() ); | |
1200 register_new_node( bol2, ctrl2 ); | |
1201 _igvn.hash_delete(loop_end); | |
1202 loop_end->set_req(CountedLoopEndNode::TestValue, bol2); | |
1203 | |
1204 // Step 3: Find the min-trip test guaranteed before a 'main' loop. | |
1205 // Make it a 1-trip test (means at least 2 trips). | |
1206 if( adjust_min_trip ) { | |
1207 // Guard test uses an 'opaque' node which is not shared. Hence I | |
1208 // can edit it's inputs directly. Hammer in the new limit for the | |
1209 // minimum-trip guard. | |
1210 assert( opaq->outcnt() == 1, "" ); | |
1211 _igvn.hash_delete(opaq); | |
1212 opaq->set_req(1, lim2); | |
1213 } | |
1214 | |
1215 // --------- | |
1216 // Step 4: Clone the loop body. Move it inside the loop. This loop body | |
1217 // represents the odd iterations; since the loop trips an even number of | |
1218 // times its backedge is never taken. Kill the backedge. | |
1219 uint dd = dom_depth(loop_head); | |
1220 clone_loop( loop, old_new, dd ); | |
1221 | |
1222 // Make backedges of the clone equal to backedges of the original. | |
1223 // Make the fall-in from the original come from the fall-out of the clone. | |
1224 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { | |
1225 Node* phi = loop_head->fast_out(j); | |
1226 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { | |
1227 Node *newphi = old_new[phi->_idx]; | |
1228 _igvn.hash_delete( phi ); | |
1229 _igvn.hash_delete( newphi ); | |
1230 | |
1231 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); | |
1232 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); | |
1233 phi ->set_req(LoopNode::LoopBackControl, C->top()); | |
1234 } | |
1235 } | |
1236 Node *clone_head = old_new[loop_head->_idx]; | |
1237 _igvn.hash_delete( clone_head ); | |
1238 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); | |
1239 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); | |
1240 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); | |
1241 loop->_head = clone_head; // New loop header | |
1242 | |
1243 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); | |
1244 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); | |
1245 | |
1246 // Kill the clone's backedge | |
1247 Node *newcle = old_new[loop_end->_idx]; | |
1248 _igvn.hash_delete( newcle ); | |
1249 Node *one = _igvn.intcon(1); | |
1250 set_ctrl(one, C->root()); | |
1251 newcle->set_req(1, one); | |
1252 // Force clone into same loop body | |
1253 uint max = loop->_body.size(); | |
1254 for( uint k = 0; k < max; k++ ) { | |
1255 Node *old = loop->_body.at(k); | |
1256 Node *nnn = old_new[old->_idx]; | |
1257 loop->_body.push(nnn); | |
1258 if (!has_ctrl(old)) | |
1259 set_loop(nnn, loop); | |
1260 } | |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
235
diff
changeset
|
1261 |
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
235
diff
changeset
|
1262 loop->record_for_igvn(); |
0 | 1263 } |
1264 | |
1265 //------------------------------do_maximally_unroll---------------------------- | |
1266 | |
1267 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { | |
1268 CountedLoopNode *cl = loop->_head->as_CountedLoop(); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1269 assert(cl->trip_count() > 0, ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1270 #ifndef PRODUCT |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1271 if (TraceLoopOpts) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1272 tty->print("MaxUnroll %d ", cl->trip_count()); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1273 loop->dump_head(); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1274 } |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1275 #endif |
0 | 1276 |
1277 // If loop is tripping an odd number of times, peel odd iteration | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1278 if ((cl->trip_count() & 1) == 1) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1279 do_peeling(loop, old_new); |
0 | 1280 } |
1281 | |
1282 // Now its tripping an even number of times remaining. Double loop body. | |
1283 // Do not adjust pre-guards; they are not needed and do not exist. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1284 if (cl->trip_count() > 0) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1285 do_unroll(loop, old_new, false); |
0 | 1286 } |
1287 } | |
1288 | |
1289 //------------------------------dominates_backedge--------------------------------- | |
1290 // Returns true if ctrl is executed on every complete iteration | |
1291 bool IdealLoopTree::dominates_backedge(Node* ctrl) { | |
1292 assert(ctrl->is_CFG(), "must be control"); | |
1293 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); | |
1294 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; | |
1295 } | |
1296 | |
1297 //------------------------------add_constraint--------------------------------- | |
1298 // Constrain the main loop iterations so the condition: | |
1299 // scale_con * I + offset < limit | |
1300 // always holds true. That is, either increase the number of iterations in | |
1301 // the pre-loop or the post-loop until the condition holds true in the main | |
1302 // loop. Stride, scale, offset and limit are all loop invariant. Further, | |
1303 // stride and scale are constants (offset and limit often are). | |
1304 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { | |
1305 | |
1306 // Compute "I :: (limit-offset)/scale_con" | |
1307 Node *con = new (C, 3) SubINode( limit, offset ); | |
1308 register_new_node( con, pre_ctrl ); | |
1309 Node *scale = _igvn.intcon(scale_con); | |
1310 set_ctrl(scale, C->root()); | |
1311 Node *X = new (C, 3) DivINode( 0, con, scale ); | |
1312 register_new_node( X, pre_ctrl ); | |
1313 | |
1314 // For positive stride, the pre-loop limit always uses a MAX function | |
1315 // and the main loop a MIN function. For negative stride these are | |
1316 // reversed. | |
1317 | |
1318 // Also for positive stride*scale the affine function is increasing, so the | |
1319 // pre-loop must check for underflow and the post-loop for overflow. | |
1320 // Negative stride*scale reverses this; pre-loop checks for overflow and | |
1321 // post-loop for underflow. | |
1322 if( stride_con*scale_con > 0 ) { | |
1323 // Compute I < (limit-offset)/scale_con | |
1324 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X) | |
1325 *main_limit = (stride_con > 0) | |
1326 ? (Node*)(new (C, 3) MinINode( *main_limit, X )) | |
1327 : (Node*)(new (C, 3) MaxINode( *main_limit, X )); | |
1328 register_new_node( *main_limit, pre_ctrl ); | |
1329 | |
1330 } else { | |
1331 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I | |
1332 // Add the negation of the main-loop constraint to the pre-loop. | |
1333 // See footnote [++] below for a derivation of the limit expression. | |
1334 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1); | |
1335 set_ctrl(incr, C->root()); | |
1336 Node *adj = new (C, 3) AddINode( X, incr ); | |
1337 register_new_node( adj, pre_ctrl ); | |
1338 *pre_limit = (scale_con > 0) | |
1339 ? (Node*)new (C, 3) MinINode( *pre_limit, adj ) | |
1340 : (Node*)new (C, 3) MaxINode( *pre_limit, adj ); | |
1341 register_new_node( *pre_limit, pre_ctrl ); | |
1342 | |
1343 // [++] Here's the algebra that justifies the pre-loop limit expression: | |
1344 // | |
1345 // NOT( scale_con * I + offset < limit ) | |
1346 // == | |
1347 // scale_con * I + offset >= limit | |
1348 // == | |
1349 // SGN(scale_con) * I >= (limit-offset)/|scale_con| | |
1350 // == | |
1351 // (limit-offset)/|scale_con| <= I * SGN(scale_con) | |
1352 // == | |
1353 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con) | |
1354 // == | |
1355 // ( if (scale_con > 0) /*common case*/ | |
1356 // (limit-offset)/scale_con - 1 < I | |
1357 // else | |
1358 // (limit-offset)/scale_con + 1 > I | |
1359 // ) | |
1360 // ( if (scale_con > 0) /*common case*/ | |
1361 // (limit-offset)/scale_con + SGN(-scale_con) < I | |
1362 // else | |
1363 // (limit-offset)/scale_con + SGN(-scale_con) > I | |
1364 } | |
1365 } | |
1366 | |
1367 | |
1368 //------------------------------is_scaled_iv--------------------------------- | |
1369 // Return true if exp is a constant times an induction var | |
1370 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { | |
1371 if (exp == iv) { | |
1372 if (p_scale != NULL) { | |
1373 *p_scale = 1; | |
1374 } | |
1375 return true; | |
1376 } | |
1377 int opc = exp->Opcode(); | |
1378 if (opc == Op_MulI) { | |
1379 if (exp->in(1) == iv && exp->in(2)->is_Con()) { | |
1380 if (p_scale != NULL) { | |
1381 *p_scale = exp->in(2)->get_int(); | |
1382 } | |
1383 return true; | |
1384 } | |
1385 if (exp->in(2) == iv && exp->in(1)->is_Con()) { | |
1386 if (p_scale != NULL) { | |
1387 *p_scale = exp->in(1)->get_int(); | |
1388 } | |
1389 return true; | |
1390 } | |
1391 } else if (opc == Op_LShiftI) { | |
1392 if (exp->in(1) == iv && exp->in(2)->is_Con()) { | |
1393 if (p_scale != NULL) { | |
1394 *p_scale = 1 << exp->in(2)->get_int(); | |
1395 } | |
1396 return true; | |
1397 } | |
1398 } | |
1399 return false; | |
1400 } | |
1401 | |
1402 //-----------------------------is_scaled_iv_plus_offset------------------------------ | |
1403 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) | |
1404 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { | |
1405 if (is_scaled_iv(exp, iv, p_scale)) { | |
1406 if (p_offset != NULL) { | |
1407 Node *zero = _igvn.intcon(0); | |
1408 set_ctrl(zero, C->root()); | |
1409 *p_offset = zero; | |
1410 } | |
1411 return true; | |
1412 } | |
1413 int opc = exp->Opcode(); | |
1414 if (opc == Op_AddI) { | |
1415 if (is_scaled_iv(exp->in(1), iv, p_scale)) { | |
1416 if (p_offset != NULL) { | |
1417 *p_offset = exp->in(2); | |
1418 } | |
1419 return true; | |
1420 } | |
1421 if (exp->in(2)->is_Con()) { | |
1422 Node* offset2 = NULL; | |
1423 if (depth < 2 && | |
1424 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, | |
1425 p_offset != NULL ? &offset2 : NULL, depth+1)) { | |
1426 if (p_offset != NULL) { | |
1427 Node *ctrl_off2 = get_ctrl(offset2); | |
1428 Node* offset = new (C, 3) AddINode(offset2, exp->in(2)); | |
1429 register_new_node(offset, ctrl_off2); | |
1430 *p_offset = offset; | |
1431 } | |
1432 return true; | |
1433 } | |
1434 } | |
1435 } else if (opc == Op_SubI) { | |
1436 if (is_scaled_iv(exp->in(1), iv, p_scale)) { | |
1437 if (p_offset != NULL) { | |
1438 Node *zero = _igvn.intcon(0); | |
1439 set_ctrl(zero, C->root()); | |
1440 Node *ctrl_off = get_ctrl(exp->in(2)); | |
1441 Node* offset = new (C, 3) SubINode(zero, exp->in(2)); | |
1442 register_new_node(offset, ctrl_off); | |
1443 *p_offset = offset; | |
1444 } | |
1445 return true; | |
1446 } | |
1447 if (is_scaled_iv(exp->in(2), iv, p_scale)) { | |
1448 if (p_offset != NULL) { | |
1449 *p_scale *= -1; | |
1450 *p_offset = exp->in(1); | |
1451 } | |
1452 return true; | |
1453 } | |
1454 } | |
1455 return false; | |
1456 } | |
1457 | |
1458 //------------------------------do_range_check--------------------------------- | |
1459 // Eliminate range-checks and other trip-counter vs loop-invariant tests. | |
1460 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { | |
1461 #ifndef PRODUCT | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1462 if (PrintOpto && VerifyLoopOptimizations) { |
0 | 1463 tty->print("Range Check Elimination "); |
1464 loop->dump_head(); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1465 } else if (TraceLoopOpts) { |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1466 tty->print("RangeCheck "); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1467 loop->dump_head(); |
0 | 1468 } |
1469 #endif | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1470 assert(RangeCheckElimination, ""); |
0 | 1471 CountedLoopNode *cl = loop->_head->as_CountedLoop(); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1472 assert(cl->is_main_loop(), ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1473 |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1474 // protect against stride not being a constant |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1475 if (!cl->stride_is_con()) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1476 return; |
0 | 1477 |
1478 // Find the trip counter; we are iteration splitting based on it | |
1479 Node *trip_counter = cl->phi(); | |
1480 // Find the main loop limit; we will trim it's iterations | |
1481 // to not ever trip end tests | |
1482 Node *main_limit = cl->limit(); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1483 |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1484 // Need to find the main-loop zero-trip guard |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1485 Node *ctrl = cl->in(LoopNode::EntryControl); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1486 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1487 Node *iffm = ctrl->in(0); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1488 assert(iffm->Opcode() == Op_If, ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1489 Node *bolzm = iffm->in(1); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1490 assert(bolzm->Opcode() == Op_Bool, ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1491 Node *cmpzm = bolzm->in(1); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1492 assert(cmpzm->is_Cmp(), ""); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1493 Node *opqzm = cmpzm->in(2); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1494 // Can not optimize a loop if pre-loop Opaque1 node is optimized |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1495 // away and then another round of loop opts attempted. |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1496 if (opqzm->Opcode() != Op_Opaque1) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1497 return; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1498 assert(opqzm->in(1) == main_limit, "do not understand situation"); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1499 |
0 | 1500 // Find the pre-loop limit; we will expand it's iterations to |
1501 // not ever trip low tests. | |
1502 Node *p_f = iffm->in(0); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1503 assert(p_f->Opcode() == Op_IfFalse, ""); |
0 | 1504 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1505 assert(pre_end->loopnode()->is_pre_loop(), ""); |
0 | 1506 Node *pre_opaq1 = pre_end->limit(); |
1507 // Occasionally it's possible for a pre-loop Opaque1 node to be | |
1508 // optimized away and then another round of loop opts attempted. | |
1509 // We can not optimize this particular loop in that case. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1510 if (pre_opaq1->Opcode() != Op_Opaque1) |
0 | 1511 return; |
1512 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; | |
1513 Node *pre_limit = pre_opaq->in(1); | |
1514 | |
1515 // Where do we put new limit calculations | |
1516 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); | |
1517 | |
1518 // Ensure the original loop limit is available from the | |
1519 // pre-loop Opaque1 node. | |
1520 Node *orig_limit = pre_opaq->original_loop_limit(); | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1521 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) |
0 | 1522 return; |
1523 | |
1524 // Must know if its a count-up or count-down loop | |
1525 | |
1526 int stride_con = cl->stride_con(); | |
1527 Node *zero = _igvn.intcon(0); | |
1528 Node *one = _igvn.intcon(1); | |
1529 set_ctrl(zero, C->root()); | |
1530 set_ctrl(one, C->root()); | |
1531 | |
1532 // Range checks that do not dominate the loop backedge (ie. | |
1533 // conditionally executed) can lengthen the pre loop limit beyond | |
1534 // the original loop limit. To prevent this, the pre limit is | |
1535 // (for stride > 0) MINed with the original loop limit (MAXed | |
1536 // stride < 0) when some range_check (rc) is conditionally | |
1537 // executed. | |
1538 bool conditional_rc = false; | |
1539 | |
1540 // Check loop body for tests of trip-counter plus loop-invariant vs | |
1541 // loop-invariant. | |
1542 for( uint i = 0; i < loop->_body.size(); i++ ) { | |
1543 Node *iff = loop->_body[i]; | |
1544 if( iff->Opcode() == Op_If ) { // Test? | |
1545 | |
1546 // Test is an IfNode, has 2 projections. If BOTH are in the loop | |
1547 // we need loop unswitching instead of iteration splitting. | |
1548 Node *exit = loop->is_loop_exit(iff); | |
1549 if( !exit ) continue; | |
1550 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; | |
1551 | |
1552 // Get boolean condition to test | |
1553 Node *i1 = iff->in(1); | |
1554 if( !i1->is_Bool() ) continue; | |
1555 BoolNode *bol = i1->as_Bool(); | |
1556 BoolTest b_test = bol->_test; | |
1557 // Flip sense of test if exit condition is flipped | |
1558 if( flip ) | |
1559 b_test = b_test.negate(); | |
1560 | |
1561 // Get compare | |
1562 Node *cmp = bol->in(1); | |
1563 | |
1564 // Look for trip_counter + offset vs limit | |
1565 Node *rc_exp = cmp->in(1); | |
1566 Node *limit = cmp->in(2); | |
1567 jint scale_con= 1; // Assume trip counter not scaled | |
1568 | |
1569 Node *limit_c = get_ctrl(limit); | |
1570 if( loop->is_member(get_loop(limit_c) ) ) { | |
1571 // Compare might have operands swapped; commute them | |
1572 b_test = b_test.commute(); | |
1573 rc_exp = cmp->in(2); | |
1574 limit = cmp->in(1); | |
1575 limit_c = get_ctrl(limit); | |
1576 if( loop->is_member(get_loop(limit_c) ) ) | |
1577 continue; // Both inputs are loop varying; cannot RCE | |
1578 } | |
1579 // Here we know 'limit' is loop invariant | |
1580 | |
1581 // 'limit' maybe pinned below the zero trip test (probably from a | |
1582 // previous round of rce), in which case, it can't be used in the | |
1583 // zero trip test expression which must occur before the zero test's if. | |
1584 if( limit_c == ctrl ) { | |
1585 continue; // Don't rce this check but continue looking for other candidates. | |
1586 } | |
1587 | |
1588 // Check for scaled induction variable plus an offset | |
1589 Node *offset = NULL; | |
1590 | |
1591 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { | |
1592 continue; | |
1593 } | |
1594 | |
1595 Node *offset_c = get_ctrl(offset); | |
1596 if( loop->is_member( get_loop(offset_c) ) ) | |
1597 continue; // Offset is not really loop invariant | |
1598 // Here we know 'offset' is loop invariant. | |
1599 | |
1600 // As above for the 'limit', the 'offset' maybe pinned below the | |
1601 // zero trip test. | |
1602 if( offset_c == ctrl ) { | |
1603 continue; // Don't rce this check but continue looking for other candidates. | |
1604 } | |
1605 | |
1606 // At this point we have the expression as: | |
1607 // scale_con * trip_counter + offset :: limit | |
1608 // where scale_con, offset and limit are loop invariant. Trip_counter | |
1609 // monotonically increases by stride_con, a constant. Both (or either) | |
1610 // stride_con and scale_con can be negative which will flip about the | |
1611 // sense of the test. | |
1612 | |
1613 // Adjust pre and main loop limits to guard the correct iteration set | |
1614 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests | |
1615 if( b_test._test == BoolTest::lt ) { // Range checks always use lt | |
1616 // The overflow limit: scale*I+offset < limit | |
1617 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); | |
1618 // The underflow limit: 0 <= scale*I+offset. | |
1619 // Some math yields: -scale*I-(offset+1) < 0 | |
1620 Node *plus_one = new (C, 3) AddINode( offset, one ); | |
1621 register_new_node( plus_one, pre_ctrl ); | |
1622 Node *neg_offset = new (C, 3) SubINode( zero, plus_one ); | |
1623 register_new_node( neg_offset, pre_ctrl ); | |
1624 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit ); | |
1625 if (!conditional_rc) { | |
1626 conditional_rc = !loop->dominates_backedge(iff); | |
1627 } | |
1628 } else { | |
1629 #ifndef PRODUCT | |
1630 if( PrintOpto ) | |
1631 tty->print_cr("missed RCE opportunity"); | |
1632 #endif | |
1633 continue; // In release mode, ignore it | |
1634 } | |
1635 } else { // Otherwise work on normal compares | |
1636 switch( b_test._test ) { | |
1637 case BoolTest::ge: // Convert X >= Y to -X <= -Y | |
1638 scale_con = -scale_con; | |
1639 offset = new (C, 3) SubINode( zero, offset ); | |
1640 register_new_node( offset, pre_ctrl ); | |
1641 limit = new (C, 3) SubINode( zero, limit ); | |
1642 register_new_node( limit, pre_ctrl ); | |
1643 // Fall into LE case | |
1644 case BoolTest::le: // Convert X <= Y to X < Y+1 | |
1645 limit = new (C, 3) AddINode( limit, one ); | |
1646 register_new_node( limit, pre_ctrl ); | |
1647 // Fall into LT case | |
1648 case BoolTest::lt: | |
1649 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); | |
1650 if (!conditional_rc) { | |
1651 conditional_rc = !loop->dominates_backedge(iff); | |
1652 } | |
1653 break; | |
1654 default: | |
1655 #ifndef PRODUCT | |
1656 if( PrintOpto ) | |
1657 tty->print_cr("missed RCE opportunity"); | |
1658 #endif | |
1659 continue; // Unhandled case | |
1660 } | |
1661 } | |
1662 | |
1663 // Kill the eliminated test | |
1664 C->set_major_progress(); | |
1665 Node *kill_con = _igvn.intcon( 1-flip ); | |
1666 set_ctrl(kill_con, C->root()); | |
1667 _igvn.hash_delete(iff); | |
1668 iff->set_req(1, kill_con); | |
1669 _igvn._worklist.push(iff); | |
1670 // Find surviving projection | |
1671 assert(iff->is_If(), ""); | |
1672 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); | |
1673 // Find loads off the surviving projection; remove their control edge | |
1674 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { | |
1675 Node* cd = dp->fast_out(i); // Control-dependent node | |
1676 if( cd->is_Load() ) { // Loads can now float around in the loop | |
1677 _igvn.hash_delete(cd); | |
1678 // Allow the load to float around in the loop, or before it | |
1679 // but NOT before the pre-loop. | |
1680 cd->set_req(0, ctrl); // ctrl, not NULL | |
1681 _igvn._worklist.push(cd); | |
1682 --i; | |
1683 --imax; | |
1684 } | |
1685 } | |
1686 | |
1687 } // End of is IF | |
1688 | |
1689 } | |
1690 | |
1691 // Update loop limits | |
1692 if (conditional_rc) { | |
1693 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit) | |
1694 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit); | |
1695 register_new_node(pre_limit, pre_ctrl); | |
1696 } | |
1697 _igvn.hash_delete(pre_opaq); | |
1698 pre_opaq->set_req(1, pre_limit); | |
1699 | |
1700 // Note:: we are making the main loop limit no longer precise; | |
1701 // need to round up based on stride. | |
1702 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case | |
1703 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init | |
1704 // Hopefully, compiler will optimize for powers of 2. | |
1705 Node *ctrl = get_ctrl(main_limit); | |
1706 Node *stride = cl->stride(); | |
1707 Node *init = cl->init_trip(); | |
1708 Node *span = new (C, 3) SubINode(main_limit,init); | |
1709 register_new_node(span,ctrl); | |
1710 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); | |
1711 Node *add = new (C, 3) AddINode(span,rndup); | |
1712 register_new_node(add,ctrl); | |
1713 Node *div = new (C, 3) DivINode(0,add,stride); | |
1714 register_new_node(div,ctrl); | |
1715 Node *mul = new (C, 3) MulINode(div,stride); | |
1716 register_new_node(mul,ctrl); | |
1717 Node *newlim = new (C, 3) AddINode(mul,init); | |
1718 register_new_node(newlim,ctrl); | |
1719 main_limit = newlim; | |
1720 } | |
1721 | |
1722 Node *main_cle = cl->loopexit(); | |
1723 Node *main_bol = main_cle->in(1); | |
1724 // Hacking loop bounds; need private copies of exit test | |
1725 if( main_bol->outcnt() > 1 ) {// BoolNode shared? | |
1726 _igvn.hash_delete(main_cle); | |
1727 main_bol = main_bol->clone();// Clone a private BoolNode | |
1728 register_new_node( main_bol, main_cle->in(0) ); | |
1729 main_cle->set_req(1,main_bol); | |
1730 } | |
1731 Node *main_cmp = main_bol->in(1); | |
1732 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? | |
1733 _igvn.hash_delete(main_bol); | |
1734 main_cmp = main_cmp->clone();// Clone a private CmpNode | |
1735 register_new_node( main_cmp, main_cle->in(0) ); | |
1736 main_bol->set_req(1,main_cmp); | |
1737 } | |
1738 // Hack the now-private loop bounds | |
1739 _igvn.hash_delete(main_cmp); | |
1740 main_cmp->set_req(2, main_limit); | |
1741 _igvn._worklist.push(main_cmp); | |
1742 // The OpaqueNode is unshared by design | |
1743 _igvn.hash_delete(opqzm); | |
1744 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); | |
1745 opqzm->set_req(1,main_limit); | |
1746 _igvn._worklist.push(opqzm); | |
1747 } | |
1748 | |
1749 //------------------------------DCE_loop_body---------------------------------- | |
1750 // Remove simplistic dead code from loop body | |
1751 void IdealLoopTree::DCE_loop_body() { | |
1752 for( uint i = 0; i < _body.size(); i++ ) | |
1753 if( _body.at(i)->outcnt() == 0 ) | |
1754 _body.map( i--, _body.pop() ); | |
1755 } | |
1756 | |
1757 | |
1758 //------------------------------adjust_loop_exit_prob-------------------------- | |
1759 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. | |
1760 // Replace with a 1-in-10 exit guess. | |
1761 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { | |
1762 Node *test = tail(); | |
1763 while( test != _head ) { | |
1764 uint top = test->Opcode(); | |
1765 if( top == Op_IfTrue || top == Op_IfFalse ) { | |
1766 int test_con = ((ProjNode*)test)->_con; | |
1767 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); | |
1768 IfNode *iff = test->in(0)->as_If(); | |
1769 if( iff->outcnt() == 2 ) { // Ignore dead tests | |
1770 Node *bol = iff->in(1); | |
1771 if( bol && bol->req() > 1 && bol->in(1) && | |
1772 ((bol->in(1)->Opcode() == Op_StorePConditional ) || | |
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
401
diff
changeset
|
1773 (bol->in(1)->Opcode() == Op_StoreIConditional ) || |
0 | 1774 (bol->in(1)->Opcode() == Op_StoreLConditional ) || |
1775 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || | |
1776 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
39
diff
changeset
|
1777 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
39
diff
changeset
|
1778 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) |
0 | 1779 return; // Allocation loops RARELY take backedge |
1780 // Find the OTHER exit path from the IF | |
1781 Node* ex = iff->proj_out(1-test_con); | |
1782 float p = iff->_prob; | |
1783 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { | |
1784 if( top == Op_IfTrue ) { | |
1785 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { | |
1786 iff->_prob = PROB_STATIC_FREQUENT; | |
1787 } | |
1788 } else { | |
1789 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { | |
1790 iff->_prob = PROB_STATIC_INFREQUENT; | |
1791 } | |
1792 } | |
1793 } | |
1794 } | |
1795 } | |
1796 test = phase->idom(test); | |
1797 } | |
1798 } | |
1799 | |
1800 | |
1801 //------------------------------policy_do_remove_empty_loop-------------------- | |
1802 // Micro-benchmark spamming. Policy is to always remove empty loops. | |
1803 // The 'DO' part is to replace the trip counter with the value it will | |
1804 // have on the last iteration. This will break the loop. | |
1805 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { | |
1806 // Minimum size must be empty loop | |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1807 if (_body.size() > EMPTY_LOOP_SIZE) |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1808 return false; |
0 | 1809 |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1810 if (!_head->is_CountedLoop()) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1811 return false; // Dead loop |
0 | 1812 CountedLoopNode *cl = _head->as_CountedLoop(); |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1813 if (!cl->loopexit()) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1814 return false; // Malformed loop |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1815 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) |
0 | 1816 return false; // Infinite loop |
2403 | 1817 |
0 | 1818 #ifdef ASSERT |
1819 // Ensure only one phi which is the iv. | |
1820 Node* iv = NULL; | |
1821 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { | |
1822 Node* n = cl->fast_out(i); | |
1823 if (n->Opcode() == Op_Phi) { | |
1824 assert(iv == NULL, "Too many phis" ); | |
1825 iv = n; | |
1826 } | |
1827 } | |
1828 assert(iv == cl->phi(), "Wrong phi" ); | |
1829 #endif | |
2403 | 1830 |
1831 // main and post loops have explicitly created zero trip guard | |
1832 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); | |
1833 if (needs_guard) { | |
2465 | 1834 // Skip guard if values not overlap. |
1835 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); | |
1836 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); | |
1837 int stride_con = cl->stride_con(); | |
1838 if (stride_con > 0) { | |
1839 needs_guard = (init_t->_hi >= limit_t->_lo); | |
1840 } else { | |
1841 needs_guard = (init_t->_lo <= limit_t->_hi); | |
1842 } | |
1843 } | |
1844 if (needs_guard) { | |
2403 | 1845 // Check for an obvious zero trip guard. |
2445 | 1846 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); |
2403 | 1847 if (inctrl->Opcode() == Op_IfTrue) { |
1848 // The test should look like just the backedge of a CountedLoop | |
1849 Node* iff = inctrl->in(0); | |
1850 if (iff->is_If()) { | |
1851 Node* bol = iff->in(1); | |
1852 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { | |
1853 Node* cmp = bol->in(1); | |
1854 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { | |
1855 needs_guard = false; | |
1856 } | |
1857 } | |
1858 } | |
1859 } | |
1860 } | |
1861 | |
1862 #ifndef PRODUCT | |
1863 if (PrintOpto) { | |
1864 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); | |
1865 this->dump_head(); | |
1866 } else if (TraceLoopOpts) { | |
1867 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); | |
1868 this->dump_head(); | |
1869 } | |
1870 #endif | |
1871 | |
1872 if (needs_guard) { | |
1873 // Peel the loop to ensure there's a zero trip guard | |
1874 Node_List old_new; | |
1875 phase->do_peeling(this, old_new); | |
1876 } | |
1877 | |
0 | 1878 // Replace the phi at loop head with the final value of the last |
1879 // iteration. Then the CountedLoopEnd will collapse (backedge never | |
1880 // taken) and all loop-invariant uses of the exit values will be correct. | |
1881 Node *phi = cl->phi(); | |
1882 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() ); | |
1883 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
1884 phase->_igvn.replace_node(phi,final); |
0 | 1885 phase->C->set_major_progress(); |
1886 return true; | |
1887 } | |
1888 | |
2465 | 1889 //------------------------------policy_do_one_iteration_loop------------------- |
1890 // Convert one iteration loop into normal code. | |
1891 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { | |
1892 if (!_head->as_Loop()->is_valid_counted_loop()) | |
1893 return false; // Only for counted loop | |
1894 | |
1895 CountedLoopNode *cl = _head->as_CountedLoop(); | |
1896 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { | |
1897 return false; | |
1898 } | |
1899 | |
1900 #ifndef PRODUCT | |
1901 if(TraceLoopOpts) { | |
1902 tty->print("OneIteration "); | |
1903 this->dump_head(); | |
1904 } | |
1905 #endif | |
1906 | |
1907 Node *init_n = cl->init_trip(); | |
1908 #ifdef ASSERT | |
1909 // Loop boundaries should be constant since trip count is exact. | |
1910 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); | |
1911 #endif | |
1912 // Replace the phi at loop head with the value of the init_trip. | |
1913 // Then the CountedLoopEnd will collapse (backedge will not be taken) | |
1914 // and all loop-invariant uses of the exit values will be correct. | |
1915 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); | |
1916 phase->C->set_major_progress(); | |
1917 return true; | |
1918 } | |
0 | 1919 |
1920 //============================================================================= | |
1921 //------------------------------iteration_split_impl--------------------------- | |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
1922 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { |
2465 | 1923 // Compute exact loop trip count if possible. |
1924 compute_exact_trip_count(phase); | |
1925 | |
1926 // Convert one iteration loop into normal code. | |
1927 if (policy_do_one_iteration_loop(phase)) | |
1928 return true; | |
1929 | |
0 | 1930 // Check and remove empty loops (spam micro-benchmarks) |
2465 | 1931 if (policy_do_remove_empty_loop(phase)) |
1172 | 1932 return true; // Here we removed an empty loop |
0 | 1933 |
1934 bool should_peel = policy_peeling(phase); // Should we peel? | |
1935 | |
1936 bool should_unswitch = policy_unswitching(phase); | |
1937 | |
1938 // Non-counted loops may be peeled; exactly 1 iteration is peeled. | |
1939 // This removes loop-invariant tests (usually null checks). | |
2465 | 1940 if (!_head->is_CountedLoop()) { // Non-counted loop |
0 | 1941 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
1942 // Partial peel succeeded so terminate this round of loop opts |
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
1943 return false; |
0 | 1944 } |
2465 | 1945 if (should_peel) { // Should we peel? |
0 | 1946 #ifndef PRODUCT |
1947 if (PrintOpto) tty->print_cr("should_peel"); | |
1948 #endif | |
1949 phase->do_peeling(this,old_new); | |
2465 | 1950 } else if (should_unswitch) { |
0 | 1951 phase->do_unswitching(this, old_new); |
1952 } | |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
1953 return true; |
0 | 1954 } |
1955 CountedLoopNode *cl = _head->as_CountedLoop(); | |
1956 | |
2465 | 1957 if (!cl->loopexit()) return true; // Ignore various kinds of broken loops |
0 | 1958 |
1959 // Do nothing special to pre- and post- loops | |
2465 | 1960 if (cl->is_pre_loop() || cl->is_post_loop()) return true; |
0 | 1961 |
1962 // Compute loop trip count from profile data | |
1963 compute_profile_trip_cnt(phase); | |
1964 | |
1965 // Before attempting fancy unrolling, RCE or alignment, see if we want | |
1966 // to completely unroll this loop or do loop unswitching. | |
2465 | 1967 if (cl->is_normal_loop()) { |
789
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
1968 if (should_unswitch) { |
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
1969 phase->do_unswitching(this, old_new); |
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
1970 return true; |
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
1971 } |
0 | 1972 bool should_maximally_unroll = policy_maximally_unroll(phase); |
2465 | 1973 if (should_maximally_unroll) { |
0 | 1974 // Here we did some unrolling and peeling. Eventually we will |
1975 // completely unroll this loop and it will no longer be a loop. | |
1976 phase->do_maximally_unroll(this,old_new); | |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
1977 return true; |
0 | 1978 } |
1979 } | |
1980 | |
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1981 // Skip next optimizations if running low on nodes. Note that |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1982 // policy_unswitching and policy_maximally_unroll have this check. |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1983 uint nodes_left = MaxNodeLimit - phase->C->unique(); |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1984 if ((2 * _body.size()) > nodes_left) { |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1985 return true; |
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1986 } |
0 | 1987 |
1988 // Counted loops may be peeled, may need some iterations run up | |
1989 // front for RCE, and may want to align loop refs to a cache | |
1990 // line. Thus we clone a full loop up front whose trip count is | |
1991 // at least 1 (if peeling), but may be several more. | |
1992 | |
1993 // The main loop will start cache-line aligned with at least 1 | |
1994 // iteration of the unrolled body (zero-trip test required) and | |
1995 // will have some range checks removed. | |
1996 | |
1997 // A post-loop will finish any odd iterations (leftover after | |
1998 // unrolling), plus any needed for RCE purposes. | |
1999 | |
2000 bool should_unroll = policy_unroll(phase); | |
2001 | |
2002 bool should_rce = policy_range_check(phase); | |
2003 | |
2004 bool should_align = policy_align(phase); | |
2005 | |
2006 // If not RCE'ing (iteration splitting) or Aligning, then we do not | |
2007 // need a pre-loop. We may still need to peel an initial iteration but | |
2008 // we will not be needing an unknown number of pre-iterations. | |
2009 // | |
2010 // Basically, if may_rce_align reports FALSE first time through, | |
2011 // we will not be able to later do RCE or Aligning on this loop. | |
2012 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; | |
2013 | |
2014 // If we have any of these conditions (RCE, alignment, unrolling) met, then | |
2015 // we switch to the pre-/main-/post-loop model. This model also covers | |
2016 // peeling. | |
2465 | 2017 if (should_rce || should_align || should_unroll) { |
2018 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops | |
0 | 2019 phase->insert_pre_post_loops(this,old_new, !may_rce_align); |
2020 | |
2021 // Adjust the pre- and main-loop limits to let the pre and post loops run | |
2022 // with full checks, but the main-loop with no checks. Remove said | |
2023 // checks from the main body. | |
2465 | 2024 if (should_rce) |
0 | 2025 phase->do_range_check(this,old_new); |
2026 | |
2027 // Double loop body for unrolling. Adjust the minimum-trip test (will do | |
2028 // twice as many iterations as before) and the main body limit (only do | |
2029 // an even number of trips). If we are peeling, we might enable some RCE | |
2030 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if | |
2031 // peeling. | |
2465 | 2032 if (should_unroll && !should_peel) |
2033 phase->do_unroll(this,old_new, true); | |
0 | 2034 |
2035 // Adjust the pre-loop limits to align the main body | |
2036 // iterations. | |
2465 | 2037 if (should_align) |
0 | 2038 Unimplemented(); |
2039 | |
2040 } else { // Else we have an unchanged counted loop | |
2465 | 2041 if (should_peel) // Might want to peel but do nothing else |
0 | 2042 phase->do_peeling(this,old_new); |
2043 } | |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2044 return true; |
0 | 2045 } |
2046 | |
2047 | |
2048 //============================================================================= | |
2049 //------------------------------iteration_split-------------------------------- | |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2050 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { |
0 | 2051 // Recursively iteration split nested loops |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2052 if (_child && !_child->iteration_split(phase, old_new)) |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2053 return false; |
0 | 2054 |
2055 // Clean out prior deadwood | |
2056 DCE_loop_body(); | |
2057 | |
2058 | |
2059 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. | |
2060 // Replace with a 1-in-10 exit guess. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2061 if (_parent /*not the root loop*/ && |
0 | 2062 !_irreducible && |
2063 // Also ignore the occasional dead backedge | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2064 !tail()->is_top()) { |
0 | 2065 adjust_loop_exit_prob(phase); |
2066 } | |
2067 | |
2068 // Gate unrolling, RCE and peeling efforts. | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2069 if (!_child && // If not an inner loop, do not split |
0 | 2070 !_irreducible && |
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
0
diff
changeset
|
2071 _allow_optimizations && |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2072 !tail()->is_top()) { // Also ignore the occasional dead backedge |
0 | 2073 if (!_has_call) { |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2074 if (!iteration_split_impl(phase, old_new)) { |
1172 | 2075 return false; |
2076 } | |
0 | 2077 } else if (policy_unswitching(phase)) { |
2078 phase->do_unswitching(this, old_new); | |
2079 } | |
2080 } | |
2081 | |
2082 // Minor offset re-organization to remove loop-fallout uses of | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2083 // trip counter when there was no major reshaping. |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2084 phase->reorg_offsets(this); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2085 |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2086 if (_next && !_next->iteration_split(phase, old_new)) |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2087 return false; |
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2088 return true; |
0 | 2089 } |
1172 | 2090 |
2091 | |
2445 | 2092 //============================================================================= |
1763 | 2093 // Process all the loops in the loop tree and replace any fill |
2094 // patterns with an intrisc version. | |
2095 bool PhaseIdealLoop::do_intrinsify_fill() { | |
2096 bool changed = false; | |
2097 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
2098 IdealLoopTree* lpt = iter.current(); | |
2099 changed |= intrinsify_fill(lpt); | |
2100 } | |
2101 return changed; | |
2102 } | |
2103 | |
2104 | |
2105 // Examine an inner loop looking for a a single store of an invariant | |
2106 // value in a unit stride loop, | |
2107 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, | |
2108 Node*& shift, Node*& con) { | |
2109 const char* msg = NULL; | |
2110 Node* msg_node = NULL; | |
2111 | |
2112 store_value = NULL; | |
2113 con = NULL; | |
2114 shift = NULL; | |
2115 | |
2116 // Process the loop looking for stores. If there are multiple | |
2117 // stores or extra control flow give at this point. | |
2118 CountedLoopNode* head = lpt->_head->as_CountedLoop(); | |
2119 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { | |
2120 Node* n = lpt->_body.at(i); | |
2121 if (n->outcnt() == 0) continue; // Ignore dead | |
2122 if (n->is_Store()) { | |
2123 if (store != NULL) { | |
2124 msg = "multiple stores"; | |
2125 break; | |
2126 } | |
2127 int opc = n->Opcode(); | |
2128 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) { | |
2129 msg = "oop fills not handled"; | |
2130 break; | |
2131 } | |
2132 Node* value = n->in(MemNode::ValueIn); | |
2133 if (!lpt->is_invariant(value)) { | |
2134 msg = "variant store value"; | |
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2135 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { |
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2136 msg = "not array address"; |
1763 | 2137 } |
2138 store = n; | |
2139 store_value = value; | |
2140 } else if (n->is_If() && n != head->loopexit()) { | |
2141 msg = "extra control flow"; | |
2142 msg_node = n; | |
2143 } | |
2144 } | |
2145 | |
2146 if (store == NULL) { | |
2147 // No store in loop | |
2148 return false; | |
2149 } | |
2150 | |
2151 if (msg == NULL && head->stride_con() != 1) { | |
2152 // could handle negative strides too | |
2153 if (head->stride_con() < 0) { | |
2154 msg = "negative stride"; | |
2155 } else { | |
2156 msg = "non-unit stride"; | |
2157 } | |
2158 } | |
2159 | |
2160 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { | |
2161 msg = "can't handle store address"; | |
2162 msg_node = store->in(MemNode::Address); | |
2163 } | |
2164 | |
1813
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2165 if (msg == NULL && |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2166 (!store->in(MemNode::Memory)->is_Phi() || |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2167 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2168 msg = "store memory isn't proper phi"; |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2169 msg_node = store->in(MemNode::Memory); |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2170 } |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2171 |
1763 | 2172 // Make sure there is an appropriate fill routine |
2173 BasicType t = store->as_Mem()->memory_type(); | |
2174 const char* fill_name; | |
2175 if (msg == NULL && | |
2176 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { | |
2177 msg = "unsupported store"; | |
2178 msg_node = store; | |
2179 } | |
2180 | |
2181 if (msg != NULL) { | |
2182 #ifndef PRODUCT | |
2183 if (TraceOptimizeFill) { | |
2184 tty->print_cr("not fill intrinsic candidate: %s", msg); | |
2185 if (msg_node != NULL) msg_node->dump(); | |
2186 } | |
2187 #endif | |
2188 return false; | |
2189 } | |
2190 | |
2191 // Make sure the address expression can be handled. It should be | |
2192 // head->phi * elsize + con. head->phi might have a ConvI2L. | |
2193 Node* elements[4]; | |
2194 Node* conv = NULL; | |
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2195 bool found_index = false; |
1763 | 2196 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); |
2197 for (int e = 0; e < count; e++) { | |
2198 Node* n = elements[e]; | |
2199 if (n->is_Con() && con == NULL) { | |
2200 con = n; | |
2201 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { | |
2202 Node* value = n->in(1); | |
2203 #ifdef _LP64 | |
2204 if (value->Opcode() == Op_ConvI2L) { | |
2205 conv = value; | |
2206 value = value->in(1); | |
2207 } | |
2208 #endif | |
2209 if (value != head->phi()) { | |
2210 msg = "unhandled shift in address"; | |
2211 } else { | |
2448
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2212 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { |
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2213 msg = "scale doesn't match"; |
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2214 } else { |
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2215 found_index = true; |
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2216 shift = n; |
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2217 } |
1763 | 2218 } |
2219 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { | |
2220 if (n->in(1) == head->phi()) { | |
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2221 found_index = true; |
1763 | 2222 conv = n; |
2223 } else { | |
2224 msg = "unhandled input to ConvI2L"; | |
2225 } | |
2226 } else if (n == head->phi()) { | |
2227 // no shift, check below for allowed cases | |
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2228 found_index = true; |
1763 | 2229 } else { |
2230 msg = "unhandled node in address"; | |
2231 msg_node = n; | |
2232 } | |
2233 } | |
2234 | |
2235 if (count == -1) { | |
2236 msg = "malformed address expression"; | |
2237 msg_node = store; | |
2238 } | |
2239 | |
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2240 if (!found_index) { |
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2241 msg = "missing use of index"; |
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2242 } |
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2243 |
1763 | 2244 // byte sized items won't have a shift |
2245 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { | |
2246 msg = "can't find shift"; | |
2247 msg_node = store; | |
2248 } | |
2249 | |
2250 if (msg != NULL) { | |
2251 #ifndef PRODUCT | |
2252 if (TraceOptimizeFill) { | |
2253 tty->print_cr("not fill intrinsic: %s", msg); | |
2254 if (msg_node != NULL) msg_node->dump(); | |
2255 } | |
2256 #endif | |
2257 return false; | |
2258 } | |
2259 | |
2260 // No make sure all the other nodes in the loop can be handled | |
2261 VectorSet ok(Thread::current()->resource_area()); | |
2262 | |
2263 // store related values are ok | |
2264 ok.set(store->_idx); | |
2265 ok.set(store->in(MemNode::Memory)->_idx); | |
2266 | |
2267 // Loop structure is ok | |
2268 ok.set(head->_idx); | |
2269 ok.set(head->loopexit()->_idx); | |
2270 ok.set(head->phi()->_idx); | |
2271 ok.set(head->incr()->_idx); | |
2272 ok.set(head->loopexit()->cmp_node()->_idx); | |
2273 ok.set(head->loopexit()->in(1)->_idx); | |
2274 | |
2275 // Address elements are ok | |
2276 if (con) ok.set(con->_idx); | |
2277 if (shift) ok.set(shift->_idx); | |
2278 if (conv) ok.set(conv->_idx); | |
2279 | |
2280 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { | |
2281 Node* n = lpt->_body.at(i); | |
2282 if (n->outcnt() == 0) continue; // Ignore dead | |
2283 if (ok.test(n->_idx)) continue; | |
2284 // Backedge projection is ok | |
2285 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue; | |
2286 if (!n->is_AddP()) { | |
2287 msg = "unhandled node"; | |
2288 msg_node = n; | |
2289 break; | |
2290 } | |
2291 } | |
2292 | |
2293 // Make sure no unexpected values are used outside the loop | |
2294 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { | |
2295 Node* n = lpt->_body.at(i); | |
2296 // These values can be replaced with other nodes if they are used | |
2297 // outside the loop. | |
1813
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2298 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue; |
1763 | 2299 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { |
2300 Node* use = iter.get(); | |
2301 if (!lpt->_body.contains(use)) { | |
2302 msg = "node is used outside loop"; | |
2303 // lpt->_body.dump(); | |
2304 msg_node = n; | |
2305 break; | |
2306 } | |
2307 } | |
2308 } | |
2309 | |
2310 #ifdef ASSERT | |
2311 if (TraceOptimizeFill) { | |
2312 if (msg != NULL) { | |
2313 tty->print_cr("no fill intrinsic: %s", msg); | |
2314 if (msg_node != NULL) msg_node->dump(); | |
2315 } else { | |
2316 tty->print_cr("fill intrinsic for:"); | |
2317 } | |
2318 store->dump(); | |
2319 if (Verbose) { | |
2320 lpt->_body.dump(); | |
2321 } | |
2322 } | |
2323 #endif | |
2324 | |
2325 return msg == NULL; | |
2326 } | |
2327 | |
2328 | |
2329 | |
2330 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { | |
2331 // Only for counted inner loops | |
2332 if (!lpt->is_counted() || !lpt->is_inner()) { | |
2333 return false; | |
2334 } | |
2335 | |
2336 // Must have constant stride | |
2337 CountedLoopNode* head = lpt->_head->as_CountedLoop(); | |
2338 if (!head->stride_is_con() || !head->is_normal_loop()) { | |
2339 return false; | |
2340 } | |
2341 | |
2342 // Check that the body only contains a store of a loop invariant | |
2343 // value that is indexed by the loop phi. | |
2344 Node* store = NULL; | |
2345 Node* store_value = NULL; | |
2346 Node* shift = NULL; | |
2347 Node* offset = NULL; | |
2348 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { | |
2349 return false; | |
2350 } | |
2351 | |
2445 | 2352 #ifndef PRODUCT |
2353 if (TraceLoopOpts) { | |
2354 tty->print("ArrayFill "); | |
2355 lpt->dump_head(); | |
2356 } | |
2357 #endif | |
2358 | |
1763 | 2359 // Now replace the whole loop body by a call to a fill routine that |
2360 // covers the same region as the loop. | |
2361 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); | |
2362 | |
2363 // Build an expression for the beginning of the copy region | |
2364 Node* index = head->init_trip(); | |
2365 #ifdef _LP64 | |
2366 index = new (C, 2) ConvI2LNode(index); | |
2367 _igvn.register_new_node_with_optimizer(index); | |
2368 #endif | |
2369 if (shift != NULL) { | |
2370 // byte arrays don't require a shift but others do. | |
2371 index = new (C, 3) LShiftXNode(index, shift->in(2)); | |
2372 _igvn.register_new_node_with_optimizer(index); | |
2373 } | |
2374 index = new (C, 4) AddPNode(base, base, index); | |
2375 _igvn.register_new_node_with_optimizer(index); | |
2376 Node* from = new (C, 4) AddPNode(base, index, offset); | |
2377 _igvn.register_new_node_with_optimizer(from); | |
2378 // Compute the number of elements to copy | |
2379 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip()); | |
2380 _igvn.register_new_node_with_optimizer(len); | |
2381 | |
2382 BasicType t = store->as_Mem()->memory_type(); | |
2383 bool aligned = false; | |
2384 if (offset != NULL && head->init_trip()->is_Con()) { | |
2385 int element_size = type2aelembytes(t); | |
2386 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; | |
2387 } | |
2388 | |
2389 // Build a call to the fill routine | |
2390 const char* fill_name; | |
2391 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); | |
2392 assert(fill != NULL, "what?"); | |
2393 | |
2394 // Convert float/double to int/long for fill routines | |
2395 if (t == T_FLOAT) { | |
2396 store_value = new (C, 2) MoveF2INode(store_value); | |
2397 _igvn.register_new_node_with_optimizer(store_value); | |
2398 } else if (t == T_DOUBLE) { | |
2399 store_value = new (C, 2) MoveD2LNode(store_value); | |
2400 _igvn.register_new_node_with_optimizer(store_value); | |
2401 } | |
2402 | |
2403 Node* mem_phi = store->in(MemNode::Memory); | |
2404 Node* result_ctrl; | |
2405 Node* result_mem; | |
2406 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); | |
2407 int size = call_type->domain()->cnt(); | |
2408 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill, | |
2409 fill_name, TypeAryPtr::get_array_body_type(t)); | |
2410 call->init_req(TypeFunc::Parms+0, from); | |
2411 call->init_req(TypeFunc::Parms+1, store_value); | |
1844
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2412 #ifdef _LP64 |
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2413 len = new (C, 2) ConvI2LNode(len); |
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2414 _igvn.register_new_node_with_optimizer(len); |
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2415 #endif |
1763 | 2416 call->init_req(TypeFunc::Parms+2, len); |
1844
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2417 #ifdef _LP64 |
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2418 call->init_req(TypeFunc::Parms+3, C->top()); |
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2419 #endif |
1763 | 2420 call->init_req( TypeFunc::Control, head->init_control()); |
2421 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o | |
2422 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) ); | |
2423 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) ); | |
2424 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) ); | |
2425 _igvn.register_new_node_with_optimizer(call); | |
2426 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); | |
2427 _igvn.register_new_node_with_optimizer(result_ctrl); | |
2428 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory); | |
2429 _igvn.register_new_node_with_optimizer(result_mem); | |
2430 | |
2431 // If this fill is tightly coupled to an allocation and overwrites | |
2432 // the whole body, allow it to take over the zeroing. | |
2433 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); | |
2434 if (alloc != NULL && alloc->is_AllocateArray()) { | |
2435 Node* length = alloc->as_AllocateArray()->Ideal_length(); | |
2436 if (head->limit() == length && | |
2437 head->init_trip() == _igvn.intcon(0)) { | |
2438 if (TraceOptimizeFill) { | |
2439 tty->print_cr("Eliminated zeroing in allocation"); | |
2440 } | |
2441 alloc->maybe_set_complete(&_igvn); | |
2442 } else { | |
2443 #ifdef ASSERT | |
2444 if (TraceOptimizeFill) { | |
2445 tty->print_cr("filling array but bounds don't match"); | |
2446 alloc->dump(); | |
2447 head->init_trip()->dump(); | |
2448 head->limit()->dump(); | |
2449 length->dump(); | |
2450 } | |
2451 #endif | |
2452 } | |
2453 } | |
2454 | |
2455 // Redirect the old control and memory edges that are outside the loop. | |
2456 Node* exit = head->loopexit()->proj_out(0); | |
1813
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2457 // Sometimes the memory phi of the head is used as the outgoing |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2458 // state of the loop. It's safe in this case to replace it with the |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2459 // result_mem. |
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2460 _igvn.replace_node(store->in(MemNode::Memory), result_mem); |
1763 | 2461 _igvn.replace_node(exit, result_ctrl); |
2462 _igvn.replace_node(store, result_mem); | |
2463 // Any uses the increment outside of the loop become the loop limit. | |
2464 _igvn.replace_node(head->incr(), head->limit()); | |
2465 | |
2466 // Disconnect the head from the loop. | |
2467 for (uint i = 0; i < lpt->_body.size(); i++) { | |
2468 Node* n = lpt->_body.at(i); | |
2469 _igvn.replace_node(n, C->top()); | |
2470 } | |
2471 | |
2472 return true; | |
2473 } |