Mercurial > hg > truffle
annotate src/share/vm/opto/loopTransform.cpp @ 3902:11a4af030e4b
7071709: JSR 292: switchpoint invalidation should be pushed not pulled
Reviewed-by: never
| author | twisti |
|---|---|
| date | Fri, 02 Sep 2011 04:28:59 -0700 |
| parents | 6987871cfb9b |
| children | da6a29fb0da5 |
| 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 || | |
| 3782 | 86 bt == BoolTest::ne, "canonical test is expected"); |
| 2465 | 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. | |
|
3845
c96c3eb1efae
7068051: SIGSEGV in PhaseIdealLoop::build_loop_late_post
kvn
parents:
3788
diff
changeset
|
512 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; |
| 2445 | 513 _igvn.hash_delete(head); |
|
3845
c96c3eb1efae
7068051: SIGSEGV in PhaseIdealLoop::build_loop_late_post
kvn
parents:
3788
diff
changeset
|
514 head->set_req(LoopNode::EntryControl, new_entry); |
| 2445 | 515 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { |
| 516 Node* old = head->fast_out(j); | |
| 517 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { | |
|
3845
c96c3eb1efae
7068051: SIGSEGV in PhaseIdealLoop::build_loop_late_post
kvn
parents:
3788
diff
changeset
|
518 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; |
| 2445 | 519 if (!new_exit_value ) // Backedge value is ALSO loop invariant? |
| 0 | 520 // Then loop body backedge value remains the same. |
| 521 new_exit_value = old->in(LoopNode::LoopBackControl); | |
| 522 _igvn.hash_delete(old); | |
| 523 old->set_req(LoopNode::EntryControl, new_exit_value); | |
| 524 } | |
| 525 } | |
| 526 | |
| 527 | |
| 528 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the | |
| 529 // extra backedge user. | |
| 2445 | 530 Node* new_head = old_new[head->_idx]; |
| 531 _igvn.hash_delete(new_head); | |
| 532 new_head->set_req(LoopNode::LoopBackControl, C->top()); | |
| 533 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { | |
| 534 Node* use = new_head->fast_out(j2); | |
| 535 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { | |
| 0 | 536 _igvn.hash_delete(use); |
| 537 use->set_req(LoopNode::LoopBackControl, C->top()); | |
| 538 } | |
| 539 } | |
| 540 | |
| 541 | |
| 542 // Step 4: Correct dom-depth info. Set to loop-head depth. | |
| 2445 | 543 int dd = dom_depth(head); |
| 544 set_idom(head, head->in(1), dd); | |
| 0 | 545 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { |
| 546 Node *old = loop->_body.at(j3); | |
| 547 Node *nnn = old_new[old->_idx]; | |
| 548 if (!has_ctrl(nnn)) | |
| 549 set_idom(nnn, idom(nnn), dd-1); | |
| 550 // While we're at it, remove any SafePoints from the peeled code | |
| 2445 | 551 if (old->Opcode() == Op_SafePoint) { |
| 0 | 552 Node *nnn = old_new[old->_idx]; |
| 553 lazy_replace(nnn,nnn->in(TypeFunc::Control)); | |
| 554 } | |
| 555 } | |
| 556 | |
| 557 // Now force out all loop-invariant dominating tests. The optimizer | |
| 558 // finds some, but we _know_ they are all useless. | |
| 559 peeled_dom_test_elim(loop,old_new); | |
| 560 | |
| 561 loop->record_for_igvn(); | |
| 562 } | |
| 563 | |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
564 #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
|
565 |
| 0 | 566 //------------------------------policy_maximally_unroll------------------------ |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
567 // 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
|
568 // unrolled. |
| 0 | 569 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { |
| 570 CountedLoopNode *cl = _head->as_CountedLoop(); | |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
571 assert(cl->is_normal_loop(), ""); |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
572 if (!cl->is_valid_counted_loop()) |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
573 return false; // Malformed counted loop |
| 0 | 574 |
| 2465 | 575 if (!cl->has_exact_trip_count()) { |
| 576 // Trip count is not exact. | |
| 0 | 577 return false; |
| 578 } | |
| 579 | |
| 2465 | 580 uint trip_count = cl->trip_count(); |
| 581 // Note, max_juint is used to indicate unknown trip count. | |
| 582 assert(trip_count > 1, "one iteration loop should be optimized out already"); | |
| 583 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); | |
| 0 | 584 |
| 585 // Real policy: if we maximally unroll, does it get too big? | |
| 586 // Allow the unrolled mess to get larger than standard loop | |
| 587 // size. After all, it will no longer be a loop. | |
| 588 uint body_size = _body.size(); | |
| 589 uint unroll_limit = (uint)LoopUnrollLimit * 4; | |
| 590 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); | |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
591 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
|
592 return false; |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
593 } |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
594 |
| 3345 | 595 // Fully unroll a loop with few iterations regardless next |
| 596 // conditions since following loop optimizations will split | |
| 597 // such loop anyway (pre-main-post). | |
| 598 if (trip_count <= 3) | |
| 599 return true; | |
| 600 | |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
601 // 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
|
602 // 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
|
603 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
|
604 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
|
605 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
|
606 return false; |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
607 if (new_body_size > unroll_limit || |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
608 // 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
|
609 new_body_size >= MaxNodeLimit - phase->C->unique()) { |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
610 return false; |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
611 } |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
612 |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
613 // 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
|
614 // String intrinsics are large and have loops. |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
615 for (uint k = 0; k < _body.size(); k++) { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
616 Node* n = _body.at(k); |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
617 switch (n->Opcode()) { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
618 case Op_StrComp: |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
619 case Op_StrEquals: |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
620 case Op_StrIndexOf: |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
621 case Op_AryEq: { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
622 return false; |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
623 } |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
624 } // switch |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
625 } |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
626 |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
627 return true; // Do maximally unroll |
| 0 | 628 } |
| 629 | |
| 630 | |
| 3333 | 631 #define MAX_UNROLL 16 // maximum number of unrolls for main loop |
| 632 | |
| 0 | 633 //------------------------------policy_unroll---------------------------------- |
| 634 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if | |
| 635 // the loop is a CountedLoop and the body is small enough. | |
| 636 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { | |
| 637 | |
| 638 CountedLoopNode *cl = _head->as_CountedLoop(); | |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
639 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); |
| 0 | 640 |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
641 if (!cl->is_valid_counted_loop()) |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
642 return false; // Malformed counted loop |
| 0 | 643 |
| 3345 | 644 // Protect against over-unrolling. |
| 645 // After split at least one iteration will be executed in pre-loop. | |
| 646 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; | |
| 0 | 647 |
| 3333 | 648 int future_unroll_ct = cl->unrolled_count() * 2; |
| 649 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
|
650 |
| 3333 | 651 // Check for initial stride being a small enough constant |
| 652 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; | |
| 0 | 653 |
| 654 // Don't unroll if the next round of unrolling would push us | |
| 655 // over the expected trip count of the loop. One is subtracted | |
| 656 // from the expected trip count because the pre-loop normally | |
| 657 // executes 1 iteration. | |
| 658 if (UnrollLimitForProfileCheck > 0 && | |
| 659 cl->profile_trip_cnt() != COUNT_UNKNOWN && | |
| 660 future_unroll_ct > UnrollLimitForProfileCheck && | |
| 661 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { | |
| 662 return false; | |
| 663 } | |
| 664 | |
| 665 // When unroll count is greater than LoopUnrollMin, don't unroll if: | |
| 666 // the residual iterations are more than 10% of the trip count | |
| 667 // and rounds of "unroll,optimize" are not making significant progress | |
| 668 // Progress defined as current size less than 20% larger than previous size. | |
| 669 if (UseSuperWord && cl->node_count_before_unroll() > 0 && | |
| 670 future_unroll_ct > LoopUnrollMin && | |
| 671 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && | |
| 672 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { | |
| 673 return false; | |
| 674 } | |
| 675 | |
| 676 Node *init_n = cl->init_trip(); | |
| 677 Node *limit_n = cl->limit(); | |
| 3345 | 678 int stride_con = cl->stride_con(); |
| 0 | 679 // Non-constant bounds. |
| 680 // Protect against over-unrolling when init or/and limit are not constant | |
| 681 // (so that trip_count's init value is maxint) but iv range is known. | |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
682 if (init_n == NULL || !init_n->is_Con() || |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
683 limit_n == NULL || !limit_n->is_Con()) { |
| 0 | 684 Node* phi = cl->phi(); |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
685 if (phi != NULL) { |
| 0 | 686 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); |
| 687 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); | |
| 3345 | 688 int next_stride = stride_con * 2; // stride after this unroll |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
689 if (next_stride > 0) { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
690 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
|
691 iv_type->_lo + next_stride > iv_type->_hi) { |
| 0 | 692 return false; // over-unrolling |
| 693 } | |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
694 } else if (next_stride < 0) { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
695 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
|
696 iv_type->_hi + next_stride < iv_type->_lo) { |
| 0 | 697 return false; // over-unrolling |
| 698 } | |
| 699 } | |
| 700 } | |
| 701 } | |
| 702 | |
| 3345 | 703 // After unroll limit will be adjusted: new_limit = limit-stride. |
| 704 // Bailout if adjustment overflow. | |
| 705 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); | |
| 706 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || | |
| 707 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) | |
| 708 return false; // overflow | |
| 709 | |
| 0 | 710 // Adjust body_size to determine if we unroll or not |
| 711 uint body_size = _body.size(); | |
| 712 // 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
|
713 for (uint k = 0; k < _body.size(); k++) { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
714 Node* n = _body.at(k); |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
715 switch (n->Opcode()) { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
716 case Op_ModL: body_size += 30; break; |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
717 case Op_DivL: body_size += 30; break; |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
718 case Op_MulL: body_size += 10; break; |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
719 case Op_StrComp: |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
720 case Op_StrEquals: |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
721 case Op_StrIndexOf: |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
722 case Op_AryEq: { |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
723 // 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
|
724 // String intrinsics are large and have loops. |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
725 return false; |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
726 } |
|
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
727 } // switch |
| 0 | 728 } |
| 729 | |
| 730 // Check for being too big | |
|
2412
f9424955eb18
7029152: Ideal nodes for String intrinsics miss memory edge optimization
kvn
parents:
2403
diff
changeset
|
731 if (body_size > (uint)LoopUnrollLimit) { |
| 3345 | 732 // Normal case: loop too big |
| 0 | 733 return false; |
| 734 } | |
| 735 | |
| 736 // Unroll once! (Each trip will soon do double iterations) | |
| 737 return true; | |
| 738 } | |
| 739 | |
| 740 //------------------------------policy_align----------------------------------- | |
| 741 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the | |
| 742 // expression that does the alignment. Note that only one array base can be | |
| 605 | 743 // aligned in a loop (unless the VM guarantees mutual alignment). Note that |
| 0 | 744 // if we vectorize short memory ops into longer memory ops, we may want to |
| 745 // increase alignment. | |
| 746 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { | |
| 747 return false; | |
| 748 } | |
| 749 | |
| 750 //------------------------------policy_range_check----------------------------- | |
| 751 // Return TRUE or FALSE if the loop should be range-check-eliminated. | |
| 752 // Actually we do iteration-splitting, a more powerful form of RCE. | |
| 753 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { | |
| 3345 | 754 if (!RangeCheckElimination) return false; |
| 0 | 755 |
| 756 CountedLoopNode *cl = _head->as_CountedLoop(); | |
| 757 // If we unrolled with no intention of doing RCE and we later | |
| 758 // changed our minds, we got no pre-loop. Either we need to | |
| 759 // make a new pre-loop, or we gotta disallow RCE. | |
| 3345 | 760 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. |
| 0 | 761 Node *trip_counter = cl->phi(); |
| 762 | |
| 763 // Check loop body for tests of trip-counter plus loop-invariant vs | |
| 764 // loop-invariant. | |
| 3345 | 765 for (uint i = 0; i < _body.size(); i++) { |
| 0 | 766 Node *iff = _body[i]; |
| 3345 | 767 if (iff->Opcode() == Op_If) { // Test? |
| 0 | 768 |
| 769 // Comparing trip+off vs limit | |
| 770 Node *bol = iff->in(1); | |
| 3345 | 771 if (bol->req() != 2) continue; // dead constant test |
| 1172 | 772 if (!bol->is_Bool()) { |
| 773 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); | |
| 774 continue; | |
| 775 } | |
| 3345 | 776 if (bol->as_Bool()->_test._test == BoolTest::ne) |
| 777 continue; // not RC | |
| 778 | |
| 0 | 779 Node *cmp = bol->in(1); |
| 780 | |
| 781 Node *rc_exp = cmp->in(1); | |
| 782 Node *limit = cmp->in(2); | |
| 783 | |
| 784 Node *limit_c = phase->get_ctrl(limit); | |
| 785 if( limit_c == phase->C->top() ) | |
| 786 return false; // Found dead test on live IF? No RCE! | |
| 787 if( is_member(phase->get_loop(limit_c) ) ) { | |
| 788 // Compare might have operands swapped; commute them | |
| 789 rc_exp = cmp->in(2); | |
| 790 limit = cmp->in(1); | |
| 791 limit_c = phase->get_ctrl(limit); | |
| 792 if( is_member(phase->get_loop(limit_c) ) ) | |
| 793 continue; // Both inputs are loop varying; cannot RCE | |
| 794 } | |
| 795 | |
| 796 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { | |
| 797 continue; | |
| 798 } | |
| 799 // Yeah! Found a test like 'trip+off vs limit' | |
| 800 // Test is an IfNode, has 2 projections. If BOTH are in the loop | |
| 801 // we need loop unswitching instead of iteration splitting. | |
| 802 if( is_loop_exit(iff) ) | |
| 803 return true; // Found reason to split iterations | |
| 804 } // End of is IF | |
| 805 } | |
| 806 | |
| 807 return false; | |
| 808 } | |
| 809 | |
| 810 //------------------------------policy_peel_only------------------------------- | |
| 811 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful | |
| 812 // for unrolling loops with NO array accesses. | |
| 813 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { | |
| 814 | |
| 815 for( uint i = 0; i < _body.size(); i++ ) | |
| 816 if( _body[i]->is_Mem() ) | |
| 817 return false; | |
| 818 | |
| 819 // No memory accesses at all! | |
| 820 return true; | |
| 821 } | |
| 822 | |
| 823 //------------------------------clone_up_backedge_goo-------------------------- | |
| 824 // If Node n lives in the back_ctrl block and cannot float, we clone a private | |
| 825 // version of n in preheader_ctrl block and return that, otherwise return n. | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
826 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { |
| 0 | 827 if( get_ctrl(n) != back_ctrl ) return n; |
| 828 | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
829 // Only visit once |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
830 if (visited.test_set(n->_idx)) { |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
831 Node *x = clones.find(n->_idx); |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
832 if (x != NULL) |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
833 return x; |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
834 return n; |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
835 } |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
836 |
| 0 | 837 Node *x = NULL; // If required, a clone of 'n' |
| 838 // Check for 'n' being pinned in the backedge. | |
| 839 if( n->in(0) && n->in(0) == back_ctrl ) { | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
840 assert(clones.find(n->_idx) == NULL, "dead loop"); |
| 0 | 841 x = n->clone(); // Clone a copy of 'n' to preheader |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
842 clones.push(x, n->_idx); |
| 0 | 843 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader |
| 844 } | |
| 845 | |
| 846 // Recursive fixup any other input edges into x. | |
| 847 // If there are no changes we can just return 'n', otherwise | |
| 848 // we need to clone a private copy and change it. | |
| 849 for( uint i = 1; i < n->req(); i++ ) { | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
850 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); |
| 0 | 851 if( g != n->in(i) ) { |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
852 if( !x ) { |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
853 assert(clones.find(n->_idx) == NULL, "dead loop"); |
| 0 | 854 x = n->clone(); |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
855 clones.push(x, n->_idx); |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
856 } |
| 0 | 857 x->set_req(i, g); |
| 858 } | |
| 859 } | |
| 860 if( x ) { // x can legally float to pre-header location | |
| 861 register_new_node( x, preheader_ctrl ); | |
| 862 return x; | |
| 863 } else { // raise n to cover LCA of uses | |
| 864 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); | |
| 865 } | |
| 866 return n; | |
| 867 } | |
| 868 | |
| 869 //------------------------------insert_pre_post_loops-------------------------- | |
| 870 // Insert pre and post loops. If peel_only is set, the pre-loop can not have | |
| 871 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no | |
| 872 // alignment. Useful to unroll loops that do no array accesses. | |
| 873 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { | |
| 874 | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
875 #ifndef PRODUCT |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
876 if (TraceLoopOpts) { |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
877 if (peel_only) |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
878 tty->print("PeelMainPost "); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
879 else |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
880 tty->print("PreMainPost "); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
881 loop->dump_head(); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
882 } |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
883 #endif |
| 0 | 884 C->set_major_progress(); |
| 885 | |
| 886 // Find common pieces of the loop being guarded with pre & post loops | |
| 887 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); | |
| 888 assert( main_head->is_normal_loop(), "" ); | |
| 889 CountedLoopEndNode *main_end = main_head->loopexit(); | |
| 890 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); | |
| 891 uint dd_main_head = dom_depth(main_head); | |
| 892 uint max = main_head->outcnt(); | |
| 893 | |
| 894 Node *pre_header= main_head->in(LoopNode::EntryControl); | |
| 895 Node *init = main_head->init_trip(); | |
| 896 Node *incr = main_end ->incr(); | |
| 897 Node *limit = main_end ->limit(); | |
| 898 Node *stride = main_end ->stride(); | |
| 899 Node *cmp = main_end ->cmp_node(); | |
| 900 BoolTest::mask b_test = main_end->test_trip(); | |
| 901 | |
| 902 // Need only 1 user of 'bol' because I will be hacking the loop bounds. | |
| 903 Node *bol = main_end->in(CountedLoopEndNode::TestValue); | |
| 904 if( bol->outcnt() != 1 ) { | |
| 905 bol = bol->clone(); | |
| 906 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); | |
| 907 _igvn.hash_delete(main_end); | |
| 908 main_end->set_req(CountedLoopEndNode::TestValue, bol); | |
| 909 } | |
| 910 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. | |
| 911 if( cmp->outcnt() != 1 ) { | |
| 912 cmp = cmp->clone(); | |
| 913 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); | |
| 914 _igvn.hash_delete(bol); | |
| 915 bol->set_req(1, cmp); | |
| 916 } | |
| 917 | |
| 918 //------------------------------ | |
| 919 // Step A: Create Post-Loop. | |
| 920 Node* main_exit = main_end->proj_out(false); | |
| 921 assert( main_exit->Opcode() == Op_IfFalse, "" ); | |
| 922 int dd_main_exit = dom_depth(main_exit); | |
| 923 | |
| 924 // Step A1: Clone the loop body. The clone becomes the post-loop. The main | |
| 925 // loop pre-header illegally has 2 control users (old & new loops). | |
| 926 clone_loop( loop, old_new, dd_main_exit ); | |
| 927 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); | |
| 928 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); | |
| 929 post_head->set_post_loop(main_head); | |
| 930 | |
|
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
931 // 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
|
932 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
|
933 post_end->_prob = PROB_FAIR; |
|
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
934 |
| 0 | 935 // Build the main-loop normal exit. |
| 936 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end); | |
| 937 _igvn.register_new_node_with_optimizer( new_main_exit ); | |
| 938 set_idom(new_main_exit, main_end, dd_main_exit ); | |
| 939 set_loop(new_main_exit, loop->_parent); | |
| 940 | |
| 941 // Step A2: Build a zero-trip guard for the post-loop. After leaving the | |
| 942 // main-loop, the post-loop may not execute at all. We 'opaque' the incr | |
| 943 // (the main-loop trip-counter exit value) because we will be changing | |
| 944 // the exit value (via unrolling) so we cannot constant-fold away the zero | |
| 945 // trip guard until all unrolling is done. | |
|
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
113
diff
changeset
|
946 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr); |
| 0 | 947 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit ); |
| 948 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test ); | |
| 949 register_new_node( zer_opaq, new_main_exit ); | |
| 950 register_new_node( zer_cmp , new_main_exit ); | |
| 951 register_new_node( zer_bol , new_main_exit ); | |
| 952 | |
| 953 // Build the IfNode | |
| 954 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); | |
| 955 _igvn.register_new_node_with_optimizer( zer_iff ); | |
| 956 set_idom(zer_iff, new_main_exit, dd_main_exit); | |
| 957 set_loop(zer_iff, loop->_parent); | |
| 958 | |
| 959 // Plug in the false-path, taken if we need to skip post-loop | |
| 960 _igvn.hash_delete( main_exit ); | |
| 961 main_exit->set_req(0, zer_iff); | |
| 962 _igvn._worklist.push(main_exit); | |
| 963 set_idom(main_exit, zer_iff, dd_main_exit); | |
| 964 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); | |
| 965 // Make the true-path, must enter the post loop | |
| 966 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff ); | |
| 967 _igvn.register_new_node_with_optimizer( zer_taken ); | |
| 968 set_idom(zer_taken, zer_iff, dd_main_exit); | |
| 969 set_loop(zer_taken, loop->_parent); | |
| 970 // Plug in the true path | |
| 971 _igvn.hash_delete( post_head ); | |
| 972 post_head->set_req(LoopNode::EntryControl, zer_taken); | |
| 973 set_idom(post_head, zer_taken, dd_main_exit); | |
| 974 | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
975 Arena *a = Thread::current()->resource_area(); |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
976 VectorSet visited(a); |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
977 Node_Stack clones(a, main_head->back_control()->outcnt()); |
| 0 | 978 // Step A3: Make the fall-in values to the post-loop come from the |
| 979 // fall-out values of the main-loop. | |
| 980 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { | |
| 981 Node* main_phi = main_head->fast_out(i); | |
| 982 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { | |
| 983 Node *post_phi = old_new[main_phi->_idx]; | |
| 984 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), | |
| 985 post_head->init_control(), | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
986 main_phi->in(LoopNode::LoopBackControl), |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
987 visited, clones); |
| 0 | 988 _igvn.hash_delete(post_phi); |
| 989 post_phi->set_req( LoopNode::EntryControl, fallmain ); | |
| 990 } | |
| 991 } | |
| 992 | |
| 993 // Update local caches for next stanza | |
| 994 main_exit = new_main_exit; | |
| 995 | |
| 996 | |
| 997 //------------------------------ | |
| 998 // Step B: Create Pre-Loop. | |
| 999 | |
| 1000 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main | |
| 1001 // loop pre-header illegally has 2 control users (old & new loops). | |
| 1002 clone_loop( loop, old_new, dd_main_head ); | |
| 1003 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); | |
| 1004 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); | |
| 1005 pre_head->set_pre_loop(main_head); | |
| 1006 Node *pre_incr = old_new[incr->_idx]; | |
| 1007 | |
|
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
1008 // 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
|
1009 pre_end->_prob = PROB_FAIR; |
|
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
1010 |
| 0 | 1011 // Find the pre-loop normal exit. |
| 1012 Node* pre_exit = pre_end->proj_out(false); | |
| 1013 assert( pre_exit->Opcode() == Op_IfFalse, "" ); | |
| 1014 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end); | |
| 1015 _igvn.register_new_node_with_optimizer( new_pre_exit ); | |
| 1016 set_idom(new_pre_exit, pre_end, dd_main_head); | |
| 1017 set_loop(new_pre_exit, loop->_parent); | |
| 1018 | |
| 1019 // Step B2: Build a zero-trip guard for the main-loop. After leaving the | |
| 1020 // pre-loop, the main-loop may not execute at all. Later in life this | |
| 1021 // zero-trip guard will become the minimum-trip guard when we unroll | |
| 1022 // the main-loop. | |
|
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
113
diff
changeset
|
1023 Node *min_opaq = new (C, 2) Opaque1Node(C, limit); |
| 0 | 1024 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq ); |
| 1025 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test ); | |
| 1026 register_new_node( min_opaq, new_pre_exit ); | |
| 1027 register_new_node( min_cmp , new_pre_exit ); | |
| 1028 register_new_node( min_bol , new_pre_exit ); | |
| 1029 | |
|
400
cc80376deb0c
6667595: Set probability FAIR for pre-, post- loops and ALWAYS for main loop
kvn
parents:
367
diff
changeset
|
1030 // 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
|
1031 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); |
| 0 | 1032 _igvn.register_new_node_with_optimizer( min_iff ); |
| 1033 set_idom(min_iff, new_pre_exit, dd_main_head); | |
| 1034 set_loop(min_iff, loop->_parent); | |
| 1035 | |
| 1036 // Plug in the false-path, taken if we need to skip main-loop | |
| 1037 _igvn.hash_delete( pre_exit ); | |
| 1038 pre_exit->set_req(0, min_iff); | |
| 1039 set_idom(pre_exit, min_iff, dd_main_head); | |
| 1040 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); | |
| 1041 // Make the true-path, must enter the main loop | |
| 1042 Node *min_taken = new (C, 1) IfTrueNode( min_iff ); | |
| 1043 _igvn.register_new_node_with_optimizer( min_taken ); | |
| 1044 set_idom(min_taken, min_iff, dd_main_head); | |
| 1045 set_loop(min_taken, loop->_parent); | |
| 1046 // Plug in the true path | |
| 1047 _igvn.hash_delete( main_head ); | |
| 1048 main_head->set_req(LoopNode::EntryControl, min_taken); | |
| 1049 set_idom(main_head, min_taken, dd_main_head); | |
| 1050 | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
1051 visited.Clear(); |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
1052 clones.clear(); |
| 0 | 1053 // Step B3: Make the fall-in values to the main-loop come from the |
| 1054 // fall-out values of the pre-loop. | |
| 1055 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { | |
| 1056 Node* main_phi = main_head->fast_out(i2); | |
| 1057 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { | |
| 1058 Node *pre_phi = old_new[main_phi->_idx]; | |
| 1059 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), | |
| 1060 main_head->init_control(), | |
|
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
1061 pre_phi->in(LoopNode::LoopBackControl), |
|
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3782
diff
changeset
|
1062 visited, clones); |
| 0 | 1063 _igvn.hash_delete(main_phi); |
| 1064 main_phi->set_req( LoopNode::EntryControl, fallpre ); | |
| 1065 } | |
| 1066 } | |
| 1067 | |
| 1068 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). | |
| 1069 // RCE and alignment may change this later. | |
| 1070 Node *cmp_end = pre_end->cmp_node(); | |
| 1071 assert( cmp_end->in(2) == limit, "" ); | |
| 1072 Node *pre_limit = new (C, 3) AddINode( init, stride ); | |
| 1073 | |
| 1074 // Save the original loop limit in this Opaque1 node for | |
| 1075 // use by range check elimination. | |
|
216
8d191a7697e2
6715633: when matching a memory node the adr_type should not change
kvn
parents:
113
diff
changeset
|
1076 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit); |
| 0 | 1077 |
| 1078 register_new_node( pre_limit, pre_head->in(0) ); | |
| 1079 register_new_node( pre_opaq , pre_head->in(0) ); | |
| 1080 | |
| 1081 // Since no other users of pre-loop compare, I can hack limit directly | |
| 1082 assert( cmp_end->outcnt() == 1, "no other users" ); | |
| 1083 _igvn.hash_delete(cmp_end); | |
| 1084 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); | |
| 1085 | |
| 1086 // Special case for not-equal loop bounds: | |
| 1087 // Change pre loop test, main loop test, and the | |
| 1088 // main loop guard test to use lt or gt depending on stride | |
| 1089 // direction: | |
| 1090 // positive stride use < | |
| 1091 // negative stride use > | |
| 3782 | 1092 // |
| 1093 // not-equal test is kept for post loop to handle case | |
| 1094 // when init > limit when stride > 0 (and reverse). | |
| 0 | 1095 |
| 1096 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { | |
| 1097 | |
| 1098 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; | |
| 1099 // Modify pre loop end condition | |
| 1100 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); | |
| 1101 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test); | |
| 1102 register_new_node( new_bol0, pre_head->in(0) ); | |
| 1103 _igvn.hash_delete(pre_end); | |
| 1104 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); | |
| 1105 // Modify main loop guard condition | |
| 1106 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); | |
| 1107 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test); | |
| 1108 register_new_node( new_bol1, new_pre_exit ); | |
| 1109 _igvn.hash_delete(min_iff); | |
| 1110 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); | |
| 1111 // Modify main loop end condition | |
| 1112 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); | |
| 1113 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test); | |
| 1114 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); | |
| 1115 _igvn.hash_delete(main_end); | |
| 1116 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); | |
| 1117 } | |
| 1118 | |
| 1119 // Flag main loop | |
| 1120 main_head->set_main_loop(); | |
| 1121 if( peel_only ) main_head->set_main_no_pre_loop(); | |
| 1122 | |
| 3345 | 1123 // Subtract a trip count for the pre-loop. |
| 1124 main_head->set_trip_count(main_head->trip_count() - 1); | |
| 1125 | |
| 0 | 1126 // It's difficult to be precise about the trip-counts |
| 1127 // for the pre/post loops. They are usually very short, | |
| 1128 // so guess that 4 trips is a reasonable value. | |
| 1129 post_head->set_profile_trip_cnt(4.0); | |
| 1130 pre_head->set_profile_trip_cnt(4.0); | |
| 1131 | |
| 1132 // Now force out all loop-invariant dominating tests. The optimizer | |
| 1133 // finds some, but we _know_ they are all useless. | |
| 1134 peeled_dom_test_elim(loop,old_new); | |
| 1135 } | |
| 1136 | |
| 1137 //------------------------------is_invariant----------------------------- | |
| 1138 // Return true if n is invariant | |
| 1139 bool IdealLoopTree::is_invariant(Node* n) const { | |
| 1172 | 1140 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; |
| 0 | 1141 if (n_c->is_top()) return false; |
| 1142 return !is_member(_phase->get_loop(n_c)); | |
| 1143 } | |
| 1144 | |
| 1145 | |
| 1146 //------------------------------do_unroll-------------------------------------- | |
| 1147 // Unroll the loop body one step - make each trip do 2 iterations. | |
| 1148 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
|
1149 assert(LoopUnrollLimit, ""); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1150 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1151 CountedLoopEndNode *loop_end = loop_head->loopexit(); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1152 assert(loop_end, ""); |
| 0 | 1153 #ifndef PRODUCT |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1154 if (PrintOpto && VerifyLoopOptimizations) { |
| 0 | 1155 tty->print("Unrolling "); |
| 1156 loop->dump_head(); | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1157 } else if (TraceLoopOpts) { |
| 2465 | 1158 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { |
| 3345 | 1159 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1160 } else { |
| 3345 | 1161 tty->print("Unroll %d ", loop_head->unrolled_count()*2); |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
1162 } |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1163 loop->dump_head(); |
| 0 | 1164 } |
| 1165 #endif | |
| 1166 | |
| 1167 // Remember loop node count before unrolling to detect | |
| 1168 // if rounds of unroll,optimize are making progress | |
| 1169 loop_head->set_node_count_before_unroll(loop->_body.size()); | |
| 1170 | |
| 1171 Node *ctrl = loop_head->in(LoopNode::EntryControl); | |
| 1172 Node *limit = loop_head->limit(); | |
| 1173 Node *init = loop_head->init_trip(); | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1174 Node *stride = loop_head->stride(); |
| 0 | 1175 |
| 1176 Node *opaq = NULL; | |
| 3345 | 1177 if (adjust_min_trip) { // If not maximally unrolling, need adjustment |
| 1178 // Search for zero-trip guard. | |
| 0 | 1179 assert( loop_head->is_main_loop(), "" ); |
| 1180 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); | |
| 1181 Node *iff = ctrl->in(0); | |
| 1182 assert( iff->Opcode() == Op_If, "" ); | |
| 1183 Node *bol = iff->in(1); | |
| 1184 assert( bol->Opcode() == Op_Bool, "" ); | |
| 1185 Node *cmp = bol->in(1); | |
| 1186 assert( cmp->Opcode() == Op_CmpI, "" ); | |
| 1187 opaq = cmp->in(2); | |
| 3345 | 1188 // Occasionally it's possible for a zero-trip guard Opaque1 node to be |
| 0 | 1189 // optimized away and then another round of loop opts attempted. |
| 1190 // We can not optimize this particular loop in that case. | |
| 3345 | 1191 if (opaq->Opcode() != Op_Opaque1) |
| 1192 return; // Cannot find zero-trip guard! Bail out! | |
| 1193 // Zero-trip test uses an 'opaque' node which is not shared. | |
| 1194 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); | |
| 0 | 1195 } |
| 1196 | |
| 1197 C->set_major_progress(); | |
| 1198 | |
| 3345 | 1199 Node* new_limit = NULL; |
| 1200 if (UnrollLimitCheck) { | |
| 1201 int stride_con = stride->get_int(); | |
| 1202 int stride_p = (stride_con > 0) ? stride_con : -stride_con; | |
| 1203 uint old_trip_count = loop_head->trip_count(); | |
| 1204 // Verify that unroll policy result is still valid. | |
| 1205 assert(old_trip_count > 1 && | |
| 1206 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); | |
| 0 | 1207 |
| 3345 | 1208 // Adjust loop limit to keep valid iterations number after unroll. |
| 1209 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride | |
| 1210 // which may overflow. | |
| 1211 if (!adjust_min_trip) { | |
| 1212 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, | |
| 1213 "odd trip count for maximally unroll"); | |
| 1214 // Don't need to adjust limit for maximally unroll since trip count is even. | |
| 1215 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { | |
| 1216 // Loop's limit is constant. Loop's init could be constant when pre-loop | |
| 1217 // become peeled iteration. | |
| 1218 long init_con = init->get_int(); | |
| 1219 // We can keep old loop limit if iterations count stays the same: | |
| 1220 // old_trip_count == new_trip_count * 2 | |
| 1221 // Note: since old_trip_count >= 2 then new_trip_count >= 1 | |
| 1222 // so we also don't need to adjust zero trip test. | |
| 1223 long limit_con = limit->get_int(); | |
| 1224 // (stride_con*2) not overflow since stride_con <= 8. | |
| 1225 int new_stride_con = stride_con * 2; | |
| 1226 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); | |
| 1227 long trip_count = (limit_con - init_con + stride_m)/new_stride_con; | |
| 1228 // New trip count should satisfy next conditions. | |
| 1229 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); | |
| 1230 uint new_trip_count = (uint)trip_count; | |
| 1231 adjust_min_trip = (old_trip_count != new_trip_count*2); | |
| 1232 } | |
| 1233 | |
| 1234 if (adjust_min_trip) { | |
| 1235 // Step 2: Adjust the trip limit if it is called for. | |
| 1236 // The adjustment amount is -stride. Need to make sure if the | |
| 1237 // adjustment underflows or overflows, then the main loop is skipped. | |
| 1238 Node* cmp = loop_end->cmp_node(); | |
| 1239 assert(cmp->in(2) == limit, "sanity"); | |
| 1240 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); | |
| 1241 | |
| 1242 // Verify that policy_unroll result is still valid. | |
| 1243 const TypeInt* limit_type = _igvn.type(limit)->is_int(); | |
| 1244 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || | |
| 1245 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); | |
| 0 | 1246 |
| 3345 | 1247 if (limit->is_Con()) { |
| 1248 // The check in policy_unroll and the assert above guarantee | |
| 1249 // no underflow if limit is constant. | |
| 1250 new_limit = _igvn.intcon(limit->get_int() - stride_con); | |
| 1251 set_ctrl(new_limit, C->root()); | |
| 1252 } else { | |
|
3348
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1253 // Limit is not constant. |
| 3367 | 1254 if (loop_head->unrolled_count() == 1) { // only for first unroll |
|
3348
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1255 // Separate limit by Opaque node in case it is an incremented |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1256 // variable from previous loop to avoid using pre-incremented |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1257 // value which could increase register pressure. |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1258 // Otherwise reorg_offsets() optimization will create a separate |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1259 // Opaque node for each use of trip-counter and as result |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1260 // zero trip guard limit will be different from loop limit. |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1261 assert(has_ctrl(opaq), "should have it"); |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1262 Node* opaq_ctrl = get_ctrl(opaq); |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1263 limit = new (C, 2) Opaque2Node( C, limit ); |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1264 register_new_node( limit, opaq_ctrl ); |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1265 } |
| 3345 | 1266 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) || |
| 1267 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) { | |
| 1268 // No underflow. | |
| 1269 new_limit = new (C, 3) SubINode(limit, stride); | |
| 1270 } else { | |
| 1271 // (limit - stride) may underflow. | |
| 1272 // Clamp the adjustment value with MININT or MAXINT: | |
| 1273 // | |
| 1274 // new_limit = limit-stride | |
| 1275 // if (stride > 0) | |
| 1276 // new_limit = (limit < new_limit) ? MININT : new_limit; | |
| 1277 // else | |
| 1278 // new_limit = (limit > new_limit) ? MAXINT : new_limit; | |
| 1279 // | |
| 1280 BoolTest::mask bt = loop_end->test_trip(); | |
| 1281 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); | |
| 1282 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); | |
| 1283 set_ctrl(adj_max, C->root()); | |
| 1284 Node* old_limit = NULL; | |
| 1285 Node* adj_limit = NULL; | |
| 1286 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; | |
| 1287 if (loop_head->unrolled_count() > 1 && | |
| 1288 limit->is_CMove() && limit->Opcode() == Op_CMoveI && | |
| 1289 limit->in(CMoveNode::IfTrue) == adj_max && | |
| 1290 bol->as_Bool()->_test._test == bt && | |
| 1291 bol->in(1)->Opcode() == Op_CmpI && | |
| 1292 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { | |
| 1293 // Loop was unrolled before. | |
| 1294 // Optimize the limit to avoid nested CMove: | |
| 1295 // use original limit as old limit. | |
| 1296 old_limit = bol->in(1)->in(1); | |
| 1297 // Adjust previous adjusted limit. | |
| 1298 adj_limit = limit->in(CMoveNode::IfFalse); | |
| 1299 adj_limit = new (C, 3) SubINode(adj_limit, stride); | |
| 1300 } else { | |
| 1301 old_limit = limit; | |
| 1302 adj_limit = new (C, 3) SubINode(limit, stride); | |
| 1303 } | |
| 1304 assert(old_limit != NULL && adj_limit != NULL, ""); | |
| 1305 register_new_node( adj_limit, ctrl ); // adjust amount | |
| 1306 Node* adj_cmp = new (C, 3) CmpINode(old_limit, adj_limit); | |
| 1307 register_new_node( adj_cmp, ctrl ); | |
| 1308 Node* adj_bool = new (C, 2) BoolNode(adj_cmp, bt); | |
| 1309 register_new_node( adj_bool, ctrl ); | |
| 1310 new_limit = new (C, 4) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); | |
| 1311 } | |
| 1312 register_new_node(new_limit, ctrl); | |
| 1313 } | |
| 1314 assert(new_limit != NULL, ""); | |
|
3348
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1315 // Replace in loop test. |
|
3398
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1316 assert(loop_end->in(1)->in(1) == cmp, "sanity"); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1317 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1318 // Don't need to create new test since only one user. |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1319 _igvn.hash_delete(cmp); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1320 cmp->set_req(2, new_limit); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1321 } else { |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1322 // Create new test since it is shared. |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1323 Node* ctrl2 = loop_end->in(0); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1324 Node* cmp2 = cmp->clone(); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1325 cmp2->set_req(2, new_limit); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1326 register_new_node(cmp2, ctrl2); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1327 Node* bol2 = loop_end->in(1)->clone(); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1328 bol2->set_req(1, cmp2); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1329 register_new_node(bol2, ctrl2); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1330 _igvn.hash_delete(loop_end); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1331 loop_end->set_req(1, bol2); |
|
789d04408ca3
7045693: java/util/EnumSet/EnumSetBash.java still failing intermittently
kvn
parents:
3383
diff
changeset
|
1332 } |
|
3348
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1333 // Step 3: Find the min-trip test guaranteed before a 'main' loop. |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1334 // Make it a 1-trip test (means at least 2 trips). |
| 3345 | 1335 |
|
3348
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1336 // Guard test uses an 'opaque' node which is not shared. Hence I |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1337 // can edit it's inputs directly. Hammer in the new limit for the |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1338 // minimum-trip guard. |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1339 assert(opaq->outcnt() == 1, ""); |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1340 _igvn.hash_delete(opaq); |
|
f879eafd5835
7042327: assert(opaq->outcnt() == 1 && opaq->in(1) == limit)
kvn
parents:
3345
diff
changeset
|
1341 opaq->set_req(1, new_limit); |
| 3345 | 1342 } |
| 1343 | |
| 1344 // Adjust max trip count. The trip count is intentionally rounded | |
| 1345 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, | |
| 1346 // the main, unrolled, part of the loop will never execute as it is protected | |
| 1347 // by the min-trip test. See bug 4834191 for a case where we over-unrolled | |
| 1348 // and later determined that part of the unrolled loop was dead. | |
| 1349 loop_head->set_trip_count(old_trip_count / 2); | |
| 1350 | |
| 1351 // Double the count of original iterations in the unrolled loop body. | |
| 1352 loop_head->double_unrolled_count(); | |
| 1353 | |
| 1354 } else { // LoopLimitCheck | |
| 1355 | |
| 1356 // Adjust max trip count. The trip count is intentionally rounded | |
| 1357 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, | |
| 1358 // the main, unrolled, part of the loop will never execute as it is protected | |
| 1359 // by the min-trip test. See bug 4834191 for a case where we over-unrolled | |
| 1360 // and later determined that part of the unrolled loop was dead. | |
| 1361 loop_head->set_trip_count(loop_head->trip_count() / 2); | |
| 1362 | |
| 1363 // Double the count of original iterations in the unrolled loop body. | |
| 1364 loop_head->double_unrolled_count(); | |
| 0 | 1365 |
| 3345 | 1366 // ----------- |
| 1367 // Step 2: Cut back the trip counter for an unroll amount of 2. | |
| 1368 // Loop will normally trip (limit - init)/stride_con. Since it's a | |
| 1369 // CountedLoop this is exact (stride divides limit-init exactly). | |
| 1370 // We are going to double the loop body, so we want to knock off any | |
| 1371 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. | |
| 1372 Node *span = new (C, 3) SubINode( limit, init ); | |
| 1373 register_new_node( span, ctrl ); | |
| 1374 Node *trip = new (C, 3) DivINode( 0, span, stride ); | |
| 1375 register_new_node( trip, ctrl ); | |
| 1376 Node *mtwo = _igvn.intcon(-2); | |
| 1377 set_ctrl(mtwo, C->root()); | |
| 1378 Node *rond = new (C, 3) AndINode( trip, mtwo ); | |
| 1379 register_new_node( rond, ctrl ); | |
| 1380 Node *spn2 = new (C, 3) MulINode( rond, stride ); | |
| 1381 register_new_node( spn2, ctrl ); | |
| 1382 new_limit = new (C, 3) AddINode( spn2, init ); | |
| 1383 register_new_node( new_limit, ctrl ); | |
| 1384 | |
| 1385 // Hammer in the new limit | |
| 1386 Node *ctrl2 = loop_end->in(0); | |
| 1387 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), new_limit ); | |
| 1388 register_new_node( cmp2, ctrl2 ); | |
| 1389 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() ); | |
| 1390 register_new_node( bol2, ctrl2 ); | |
| 1391 _igvn.hash_delete(loop_end); | |
| 1392 loop_end->set_req(CountedLoopEndNode::TestValue, bol2); | |
| 1393 | |
| 1394 // Step 3: Find the min-trip test guaranteed before a 'main' loop. | |
| 1395 // Make it a 1-trip test (means at least 2 trips). | |
| 1396 if( adjust_min_trip ) { | |
| 1397 assert( new_limit != NULL, "" ); | |
| 1398 // Guard test uses an 'opaque' node which is not shared. Hence I | |
| 1399 // can edit it's inputs directly. Hammer in the new limit for the | |
| 1400 // minimum-trip guard. | |
| 1401 assert( opaq->outcnt() == 1, "" ); | |
| 1402 _igvn.hash_delete(opaq); | |
| 1403 opaq->set_req(1, new_limit); | |
| 1404 } | |
| 1405 } // LoopLimitCheck | |
| 0 | 1406 |
| 1407 // --------- | |
| 1408 // Step 4: Clone the loop body. Move it inside the loop. This loop body | |
| 1409 // represents the odd iterations; since the loop trips an even number of | |
| 1410 // times its backedge is never taken. Kill the backedge. | |
| 1411 uint dd = dom_depth(loop_head); | |
| 1412 clone_loop( loop, old_new, dd ); | |
| 1413 | |
| 1414 // Make backedges of the clone equal to backedges of the original. | |
| 1415 // Make the fall-in from the original come from the fall-out of the clone. | |
| 1416 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { | |
| 1417 Node* phi = loop_head->fast_out(j); | |
| 1418 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { | |
| 1419 Node *newphi = old_new[phi->_idx]; | |
| 1420 _igvn.hash_delete( phi ); | |
| 1421 _igvn.hash_delete( newphi ); | |
| 1422 | |
| 1423 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); | |
| 1424 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); | |
| 1425 phi ->set_req(LoopNode::LoopBackControl, C->top()); | |
| 1426 } | |
| 1427 } | |
| 1428 Node *clone_head = old_new[loop_head->_idx]; | |
| 1429 _igvn.hash_delete( clone_head ); | |
| 1430 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); | |
| 1431 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); | |
| 1432 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); | |
| 1433 loop->_head = clone_head; // New loop header | |
| 1434 | |
| 1435 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); | |
| 1436 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); | |
| 1437 | |
| 1438 // Kill the clone's backedge | |
| 1439 Node *newcle = old_new[loop_end->_idx]; | |
| 1440 _igvn.hash_delete( newcle ); | |
| 1441 Node *one = _igvn.intcon(1); | |
| 1442 set_ctrl(one, C->root()); | |
| 1443 newcle->set_req(1, one); | |
| 1444 // Force clone into same loop body | |
| 1445 uint max = loop->_body.size(); | |
| 1446 for( uint k = 0; k < max; k++ ) { | |
| 1447 Node *old = loop->_body.at(k); | |
| 1448 Node *nnn = old_new[old->_idx]; | |
| 1449 loop->_body.push(nnn); | |
| 1450 if (!has_ctrl(old)) | |
| 1451 set_loop(nnn, loop); | |
| 1452 } | |
|
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
235
diff
changeset
|
1453 |
|
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
235
diff
changeset
|
1454 loop->record_for_igvn(); |
| 0 | 1455 } |
| 1456 | |
| 1457 //------------------------------do_maximally_unroll---------------------------- | |
| 1458 | |
| 1459 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { | |
| 1460 CountedLoopNode *cl = loop->_head->as_CountedLoop(); | |
| 3345 | 1461 assert(cl->has_exact_trip_count(), "trip count is not exact"); |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1462 assert(cl->trip_count() > 0, ""); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1463 #ifndef PRODUCT |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1464 if (TraceLoopOpts) { |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1465 tty->print("MaxUnroll %d ", cl->trip_count()); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1466 loop->dump_head(); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1467 } |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1468 #endif |
| 0 | 1469 |
| 1470 // 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
|
1471 if ((cl->trip_count() & 1) == 1) { |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1472 do_peeling(loop, old_new); |
| 0 | 1473 } |
| 1474 | |
| 1475 // Now its tripping an even number of times remaining. Double loop body. | |
| 1476 // 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
|
1477 if (cl->trip_count() > 0) { |
| 3345 | 1478 assert((cl->trip_count() & 1) == 0, "missed peeling"); |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1479 do_unroll(loop, old_new, false); |
| 0 | 1480 } |
| 1481 } | |
| 1482 | |
| 1483 //------------------------------dominates_backedge--------------------------------- | |
| 1484 // Returns true if ctrl is executed on every complete iteration | |
| 1485 bool IdealLoopTree::dominates_backedge(Node* ctrl) { | |
| 1486 assert(ctrl->is_CFG(), "must be control"); | |
| 1487 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); | |
| 1488 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; | |
| 1489 } | |
| 1490 | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1491 //------------------------------adjust_limit----------------------------------- |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1492 // Helper function for add_constraint(). |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1493 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1494 // Compute "I :: (limit-offset)/scale" |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1495 Node *con = new (C, 3) SubINode(rc_limit, offset); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1496 register_new_node(con, pre_ctrl); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1497 Node *X = new (C, 3) DivINode(0, con, scale); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1498 register_new_node(X, pre_ctrl); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1499 |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1500 // Adjust loop limit |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1501 loop_limit = (stride_con > 0) |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1502 ? (Node*)(new (C, 3) MinINode(loop_limit, X)) |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1503 : (Node*)(new (C, 3) MaxINode(loop_limit, X)); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1504 register_new_node(loop_limit, pre_ctrl); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1505 return loop_limit; |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1506 } |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1507 |
| 0 | 1508 //------------------------------add_constraint--------------------------------- |
| 3345 | 1509 // Constrain the main loop iterations so the conditions: |
| 1510 // low_limit <= scale_con * I + offset < upper_limit | |
| 0 | 1511 // always holds true. That is, either increase the number of iterations in |
| 1512 // the pre-loop or the post-loop until the condition holds true in the main | |
| 1513 // loop. Stride, scale, offset and limit are all loop invariant. Further, | |
| 1514 // stride and scale are constants (offset and limit often are). | |
| 3345 | 1515 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { |
| 0 | 1516 // For positive stride, the pre-loop limit always uses a MAX function |
| 1517 // and the main loop a MIN function. For negative stride these are | |
| 1518 // reversed. | |
| 1519 | |
| 1520 // Also for positive stride*scale the affine function is increasing, so the | |
| 1521 // pre-loop must check for underflow and the post-loop for overflow. | |
| 1522 // Negative stride*scale reverses this; pre-loop checks for overflow and | |
| 1523 // post-loop for underflow. | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1524 |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1525 Node *scale = _igvn.intcon(scale_con); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1526 set_ctrl(scale, C->root()); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1527 |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1528 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow |
| 3345 | 1529 // The overflow limit: scale*I+offset < upper_limit |
| 1530 // For main-loop compute | |
| 1531 // ( if (scale > 0) /* and stride > 0 */ | |
| 1532 // I < (upper_limit-offset)/scale | |
| 1533 // else /* scale < 0 and stride < 0 */ | |
| 1534 // I > (upper_limit-offset)/scale | |
| 1535 // ) | |
| 1536 // | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1537 // (upper_limit-offset) may overflow or underflow. |
| 3345 | 1538 // But it is fine since main loop will either have |
| 1539 // less iterations or will be skipped in such case. | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1540 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); |
| 0 | 1541 |
| 3345 | 1542 // The underflow limit: low_limit <= scale*I+offset. |
| 1543 // For pre-loop compute | |
| 1544 // NOT(scale*I+offset >= low_limit) | |
| 1545 // scale*I+offset < low_limit | |
| 1546 // ( if (scale > 0) /* and stride > 0 */ | |
| 1547 // I < (low_limit-offset)/scale | |
| 1548 // else /* scale < 0 and stride < 0 */ | |
| 1549 // I > (low_limit-offset)/scale | |
| 1550 // ) | |
| 1551 | |
| 1552 if (low_limit->get_int() == -max_jint) { | |
| 1553 if (!RangeLimitCheck) return; | |
| 1554 // We need this guard when scale*pre_limit+offset >= limit | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1555 // due to underflow. So we need execute pre-loop until |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1556 // scale*I+offset >= min_int. But (min_int-offset) will |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1557 // underflow when offset > 0 and X will be > original_limit |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1558 // when stride > 0. To avoid it we replace positive offset with 0. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1559 // |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1560 // Also (min_int+1 == -max_int) is used instead of min_int here |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1561 // to avoid problem with scale == -1 (min_int/(-1) == min_int). |
| 3345 | 1562 Node* shift = _igvn.intcon(31); |
| 1563 set_ctrl(shift, C->root()); | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1564 Node* sign = new (C, 3) RShiftINode(offset, shift); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1565 register_new_node(sign, pre_ctrl); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1566 offset = new (C, 3) AndINode(offset, sign); |
| 3345 | 1567 register_new_node(offset, pre_ctrl); |
| 1568 } else { | |
| 1569 assert(low_limit->get_int() == 0, "wrong low limit for range check"); | |
| 1570 // The only problem we have here when offset == min_int | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1571 // since (0-min_int) == min_int. It may be fine for stride > 0 |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1572 // but for stride < 0 X will be < original_limit. To avoid it |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1573 // max(pre_limit, original_limit) is used in do_range_check(). |
| 3345 | 1574 } |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1575 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1576 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); |
| 3345 | 1577 |
| 1578 } else { // stride_con*scale_con < 0 | |
| 1579 // For negative stride*scale pre-loop checks for overflow and | |
| 1580 // post-loop for underflow. | |
| 1581 // | |
| 1582 // The overflow limit: scale*I+offset < upper_limit | |
| 1583 // For pre-loop compute | |
| 1584 // NOT(scale*I+offset < upper_limit) | |
| 1585 // scale*I+offset >= upper_limit | |
| 1586 // scale*I+offset+1 > upper_limit | |
| 1587 // ( if (scale < 0) /* and stride > 0 */ | |
| 1588 // I < (upper_limit-(offset+1))/scale | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1589 // else /* scale > 0 and stride < 0 */ |
| 3345 | 1590 // I > (upper_limit-(offset+1))/scale |
| 1591 // ) | |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1592 // |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1593 // (upper_limit-offset-1) may underflow or overflow. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1594 // To avoid it min(pre_limit, original_limit) is used |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1595 // in do_range_check() for stride > 0 and max() for < 0. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1596 Node *one = _igvn.intcon(1); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1597 set_ctrl(one, C->root()); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1598 |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1599 Node *plus_one = new (C, 3) AddINode(offset, one); |
| 3345 | 1600 register_new_node( plus_one, pre_ctrl ); |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1601 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1602 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); |
| 3345 | 1603 |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1604 if (low_limit->get_int() == -max_jint) { |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1605 if (!RangeLimitCheck) return; |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1606 // We need this guard when scale*main_limit+offset >= limit |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1607 // due to underflow. So we need execute main-loop while |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1608 // scale*I+offset+1 > min_int. But (min_int-offset-1) will |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1609 // underflow when (offset+1) > 0 and X will be < main_limit |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1610 // when scale < 0 (and stride > 0). To avoid it we replace |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1611 // positive (offset+1) with 0. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1612 // |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1613 // Also (min_int+1 == -max_int) is used instead of min_int here |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1614 // to avoid problem with scale == -1 (min_int/(-1) == min_int). |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1615 Node* shift = _igvn.intcon(31); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1616 set_ctrl(shift, C->root()); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1617 Node* sign = new (C, 3) RShiftINode(plus_one, shift); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1618 register_new_node(sign, pre_ctrl); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1619 plus_one = new (C, 3) AndINode(plus_one, sign); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1620 register_new_node(plus_one, pre_ctrl); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1621 } else { |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1622 assert(low_limit->get_int() == 0, "wrong low limit for range check"); |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1623 // The only problem we have here when offset == max_int |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1624 // since (max_int+1) == min_int and (0-min_int) == min_int. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1625 // But it is fine since main loop will either have |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1626 // less iterations or will be skipped in such case. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1627 } |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1628 // The underflow limit: low_limit <= scale*I+offset. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1629 // For main-loop compute |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1630 // scale*I+offset+1 > low_limit |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1631 // ( if (scale < 0) /* and stride > 0 */ |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1632 // I < (low_limit-(offset+1))/scale |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1633 // else /* scale > 0 and stride < 0 */ |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1634 // I > (low_limit-(offset+1))/scale |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1635 // ) |
| 3345 | 1636 |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1637 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); |
| 0 | 1638 } |
| 1639 } | |
| 1640 | |
| 1641 | |
| 1642 //------------------------------is_scaled_iv--------------------------------- | |
| 1643 // Return true if exp is a constant times an induction var | |
| 1644 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { | |
| 1645 if (exp == iv) { | |
| 1646 if (p_scale != NULL) { | |
| 1647 *p_scale = 1; | |
| 1648 } | |
| 1649 return true; | |
| 1650 } | |
| 1651 int opc = exp->Opcode(); | |
| 1652 if (opc == Op_MulI) { | |
| 1653 if (exp->in(1) == iv && exp->in(2)->is_Con()) { | |
| 1654 if (p_scale != NULL) { | |
| 1655 *p_scale = exp->in(2)->get_int(); | |
| 1656 } | |
| 1657 return true; | |
| 1658 } | |
| 1659 if (exp->in(2) == iv && exp->in(1)->is_Con()) { | |
| 1660 if (p_scale != NULL) { | |
| 1661 *p_scale = exp->in(1)->get_int(); | |
| 1662 } | |
| 1663 return true; | |
| 1664 } | |
| 1665 } else if (opc == Op_LShiftI) { | |
| 1666 if (exp->in(1) == iv && exp->in(2)->is_Con()) { | |
| 1667 if (p_scale != NULL) { | |
| 1668 *p_scale = 1 << exp->in(2)->get_int(); | |
| 1669 } | |
| 1670 return true; | |
| 1671 } | |
| 1672 } | |
| 1673 return false; | |
| 1674 } | |
| 1675 | |
| 1676 //-----------------------------is_scaled_iv_plus_offset------------------------------ | |
| 1677 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) | |
| 1678 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { | |
| 1679 if (is_scaled_iv(exp, iv, p_scale)) { | |
| 1680 if (p_offset != NULL) { | |
| 1681 Node *zero = _igvn.intcon(0); | |
| 1682 set_ctrl(zero, C->root()); | |
| 1683 *p_offset = zero; | |
| 1684 } | |
| 1685 return true; | |
| 1686 } | |
| 1687 int opc = exp->Opcode(); | |
| 1688 if (opc == Op_AddI) { | |
| 1689 if (is_scaled_iv(exp->in(1), iv, p_scale)) { | |
| 1690 if (p_offset != NULL) { | |
| 1691 *p_offset = exp->in(2); | |
| 1692 } | |
| 1693 return true; | |
| 1694 } | |
| 1695 if (exp->in(2)->is_Con()) { | |
| 1696 Node* offset2 = NULL; | |
| 1697 if (depth < 2 && | |
| 1698 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, | |
| 1699 p_offset != NULL ? &offset2 : NULL, depth+1)) { | |
| 1700 if (p_offset != NULL) { | |
| 1701 Node *ctrl_off2 = get_ctrl(offset2); | |
| 1702 Node* offset = new (C, 3) AddINode(offset2, exp->in(2)); | |
| 1703 register_new_node(offset, ctrl_off2); | |
| 1704 *p_offset = offset; | |
| 1705 } | |
| 1706 return true; | |
| 1707 } | |
| 1708 } | |
| 1709 } else if (opc == Op_SubI) { | |
| 1710 if (is_scaled_iv(exp->in(1), iv, p_scale)) { | |
| 1711 if (p_offset != NULL) { | |
| 1712 Node *zero = _igvn.intcon(0); | |
| 1713 set_ctrl(zero, C->root()); | |
| 1714 Node *ctrl_off = get_ctrl(exp->in(2)); | |
| 1715 Node* offset = new (C, 3) SubINode(zero, exp->in(2)); | |
| 1716 register_new_node(offset, ctrl_off); | |
| 1717 *p_offset = offset; | |
| 1718 } | |
| 1719 return true; | |
| 1720 } | |
| 1721 if (is_scaled_iv(exp->in(2), iv, p_scale)) { | |
| 1722 if (p_offset != NULL) { | |
| 1723 *p_scale *= -1; | |
| 1724 *p_offset = exp->in(1); | |
| 1725 } | |
| 1726 return true; | |
| 1727 } | |
| 1728 } | |
| 1729 return false; | |
| 1730 } | |
| 1731 | |
| 1732 //------------------------------do_range_check--------------------------------- | |
| 1733 // Eliminate range-checks and other trip-counter vs loop-invariant tests. | |
| 1734 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { | |
| 1735 #ifndef PRODUCT | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1736 if (PrintOpto && VerifyLoopOptimizations) { |
| 0 | 1737 tty->print("Range Check Elimination "); |
| 1738 loop->dump_head(); | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1739 } else if (TraceLoopOpts) { |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1740 tty->print("RangeCheck "); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1741 loop->dump_head(); |
| 0 | 1742 } |
| 1743 #endif | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1744 assert(RangeCheckElimination, ""); |
| 0 | 1745 CountedLoopNode *cl = loop->_head->as_CountedLoop(); |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1746 assert(cl->is_main_loop(), ""); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1747 |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1748 // protect against stride not being a constant |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1749 if (!cl->stride_is_con()) |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1750 return; |
| 0 | 1751 |
| 1752 // Find the trip counter; we are iteration splitting based on it | |
| 1753 Node *trip_counter = cl->phi(); | |
| 1754 // Find the main loop limit; we will trim it's iterations | |
| 1755 // to not ever trip end tests | |
| 1756 Node *main_limit = cl->limit(); | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1757 |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1758 // Need to find the main-loop zero-trip guard |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1759 Node *ctrl = cl->in(LoopNode::EntryControl); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1760 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1761 Node *iffm = ctrl->in(0); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1762 assert(iffm->Opcode() == Op_If, ""); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1763 Node *bolzm = iffm->in(1); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1764 assert(bolzm->Opcode() == Op_Bool, ""); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1765 Node *cmpzm = bolzm->in(1); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1766 assert(cmpzm->is_Cmp(), ""); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1767 Node *opqzm = cmpzm->in(2); |
| 3345 | 1768 // Can not optimize a loop if zero-trip Opaque1 node is optimized |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1769 // away and then another round of loop opts attempted. |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1770 if (opqzm->Opcode() != Op_Opaque1) |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1771 return; |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1772 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
|
1773 |
| 0 | 1774 // Find the pre-loop limit; we will expand it's iterations to |
| 1775 // not ever trip low tests. | |
| 1776 Node *p_f = iffm->in(0); | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1777 assert(p_f->Opcode() == Op_IfFalse, ""); |
| 0 | 1778 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
|
1779 assert(pre_end->loopnode()->is_pre_loop(), ""); |
| 0 | 1780 Node *pre_opaq1 = pre_end->limit(); |
| 1781 // Occasionally it's possible for a pre-loop Opaque1 node to be | |
| 1782 // optimized away and then another round of loop opts attempted. | |
| 1783 // 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
|
1784 if (pre_opaq1->Opcode() != Op_Opaque1) |
| 0 | 1785 return; |
| 1786 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; | |
| 1787 Node *pre_limit = pre_opaq->in(1); | |
| 1788 | |
| 1789 // Where do we put new limit calculations | |
| 1790 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); | |
| 1791 | |
| 1792 // Ensure the original loop limit is available from the | |
| 1793 // pre-loop Opaque1 node. | |
| 1794 Node *orig_limit = pre_opaq->original_loop_limit(); | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
1795 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) |
| 0 | 1796 return; |
| 1797 | |
| 1798 // Must know if its a count-up or count-down loop | |
| 1799 | |
| 1800 int stride_con = cl->stride_con(); | |
| 1801 Node *zero = _igvn.intcon(0); | |
| 1802 Node *one = _igvn.intcon(1); | |
| 3345 | 1803 // Use symmetrical int range [-max_jint,max_jint] |
| 1804 Node *mini = _igvn.intcon(-max_jint); | |
| 0 | 1805 set_ctrl(zero, C->root()); |
| 1806 set_ctrl(one, C->root()); | |
| 3345 | 1807 set_ctrl(mini, C->root()); |
| 0 | 1808 |
| 1809 // Range checks that do not dominate the loop backedge (ie. | |
| 1810 // conditionally executed) can lengthen the pre loop limit beyond | |
| 1811 // the original loop limit. To prevent this, the pre limit is | |
| 1812 // (for stride > 0) MINed with the original loop limit (MAXed | |
| 1813 // stride < 0) when some range_check (rc) is conditionally | |
| 1814 // executed. | |
| 1815 bool conditional_rc = false; | |
| 1816 | |
| 1817 // Check loop body for tests of trip-counter plus loop-invariant vs | |
| 1818 // loop-invariant. | |
| 1819 for( uint i = 0; i < loop->_body.size(); i++ ) { | |
| 1820 Node *iff = loop->_body[i]; | |
| 1821 if( iff->Opcode() == Op_If ) { // Test? | |
| 1822 | |
| 1823 // Test is an IfNode, has 2 projections. If BOTH are in the loop | |
| 1824 // we need loop unswitching instead of iteration splitting. | |
| 1825 Node *exit = loop->is_loop_exit(iff); | |
| 1826 if( !exit ) continue; | |
| 1827 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; | |
| 1828 | |
| 1829 // Get boolean condition to test | |
| 1830 Node *i1 = iff->in(1); | |
| 1831 if( !i1->is_Bool() ) continue; | |
| 1832 BoolNode *bol = i1->as_Bool(); | |
| 1833 BoolTest b_test = bol->_test; | |
| 1834 // Flip sense of test if exit condition is flipped | |
| 1835 if( flip ) | |
| 1836 b_test = b_test.negate(); | |
| 1837 | |
| 1838 // Get compare | |
| 1839 Node *cmp = bol->in(1); | |
| 1840 | |
| 1841 // Look for trip_counter + offset vs limit | |
| 1842 Node *rc_exp = cmp->in(1); | |
| 1843 Node *limit = cmp->in(2); | |
| 1844 jint scale_con= 1; // Assume trip counter not scaled | |
| 1845 | |
| 1846 Node *limit_c = get_ctrl(limit); | |
| 1847 if( loop->is_member(get_loop(limit_c) ) ) { | |
| 1848 // Compare might have operands swapped; commute them | |
| 1849 b_test = b_test.commute(); | |
| 1850 rc_exp = cmp->in(2); | |
| 1851 limit = cmp->in(1); | |
| 1852 limit_c = get_ctrl(limit); | |
| 1853 if( loop->is_member(get_loop(limit_c) ) ) | |
| 1854 continue; // Both inputs are loop varying; cannot RCE | |
| 1855 } | |
| 1856 // Here we know 'limit' is loop invariant | |
| 1857 | |
| 1858 // 'limit' maybe pinned below the zero trip test (probably from a | |
| 1859 // previous round of rce), in which case, it can't be used in the | |
| 1860 // zero trip test expression which must occur before the zero test's if. | |
| 1861 if( limit_c == ctrl ) { | |
| 1862 continue; // Don't rce this check but continue looking for other candidates. | |
| 1863 } | |
| 1864 | |
| 1865 // Check for scaled induction variable plus an offset | |
| 1866 Node *offset = NULL; | |
| 1867 | |
| 1868 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { | |
| 1869 continue; | |
| 1870 } | |
| 1871 | |
| 1872 Node *offset_c = get_ctrl(offset); | |
| 1873 if( loop->is_member( get_loop(offset_c) ) ) | |
| 1874 continue; // Offset is not really loop invariant | |
| 1875 // Here we know 'offset' is loop invariant. | |
| 1876 | |
| 1877 // As above for the 'limit', the 'offset' maybe pinned below the | |
| 1878 // zero trip test. | |
| 1879 if( offset_c == ctrl ) { | |
| 1880 continue; // Don't rce this check but continue looking for other candidates. | |
| 1881 } | |
| 3345 | 1882 #ifdef ASSERT |
| 1883 if (TraceRangeLimitCheck) { | |
| 1884 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); | |
| 1885 bol->dump(2); | |
| 1886 } | |
| 1887 #endif | |
| 0 | 1888 // At this point we have the expression as: |
| 1889 // scale_con * trip_counter + offset :: limit | |
| 1890 // where scale_con, offset and limit are loop invariant. Trip_counter | |
| 1891 // monotonically increases by stride_con, a constant. Both (or either) | |
| 1892 // stride_con and scale_con can be negative which will flip about the | |
| 1893 // sense of the test. | |
| 1894 | |
| 1895 // Adjust pre and main loop limits to guard the correct iteration set | |
| 1896 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests | |
| 1897 if( b_test._test == BoolTest::lt ) { // Range checks always use lt | |
| 3345 | 1898 // The underflow and overflow limits: 0 <= scale*I+offset < limit |
| 1899 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); | |
| 0 | 1900 if (!conditional_rc) { |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1901 // (0-offset)/scale could be outside of loop iterations range. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1902 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; |
| 0 | 1903 } |
| 1904 } else { | |
| 1905 #ifndef PRODUCT | |
| 1906 if( PrintOpto ) | |
| 1907 tty->print_cr("missed RCE opportunity"); | |
| 1908 #endif | |
| 1909 continue; // In release mode, ignore it | |
| 1910 } | |
| 1911 } else { // Otherwise work on normal compares | |
| 1912 switch( b_test._test ) { | |
| 3345 | 1913 case BoolTest::gt: |
| 1914 // Fall into GE case | |
| 1915 case BoolTest::ge: | |
| 1916 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit | |
| 0 | 1917 scale_con = -scale_con; |
| 1918 offset = new (C, 3) SubINode( zero, offset ); | |
| 1919 register_new_node( offset, pre_ctrl ); | |
| 1920 limit = new (C, 3) SubINode( zero, limit ); | |
| 1921 register_new_node( limit, pre_ctrl ); | |
| 1922 // Fall into LE case | |
| 3345 | 1923 case BoolTest::le: |
| 1924 if (b_test._test != BoolTest::gt) { | |
| 1925 // Convert X <= Y to X < Y+1 | |
| 1926 limit = new (C, 3) AddINode( limit, one ); | |
| 1927 register_new_node( limit, pre_ctrl ); | |
| 1928 } | |
| 0 | 1929 // Fall into LT case |
| 1930 case BoolTest::lt: | |
| 3345 | 1931 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1932 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1933 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. |
| 3345 | 1934 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); |
| 0 | 1935 if (!conditional_rc) { |
|
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1936 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1937 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1938 // still be outside of loop range. |
|
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3367
diff
changeset
|
1939 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; |
| 0 | 1940 } |
| 1941 break; | |
| 1942 default: | |
| 1943 #ifndef PRODUCT | |
| 1944 if( PrintOpto ) | |
| 1945 tty->print_cr("missed RCE opportunity"); | |
| 1946 #endif | |
| 1947 continue; // Unhandled case | |
| 1948 } | |
| 1949 } | |
| 1950 | |
| 1951 // Kill the eliminated test | |
| 1952 C->set_major_progress(); | |
| 1953 Node *kill_con = _igvn.intcon( 1-flip ); | |
| 1954 set_ctrl(kill_con, C->root()); | |
| 1955 _igvn.hash_delete(iff); | |
| 1956 iff->set_req(1, kill_con); | |
| 1957 _igvn._worklist.push(iff); | |
| 1958 // Find surviving projection | |
| 1959 assert(iff->is_If(), ""); | |
| 1960 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); | |
| 1961 // Find loads off the surviving projection; remove their control edge | |
| 1962 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { | |
| 1963 Node* cd = dp->fast_out(i); // Control-dependent node | |
| 1964 if( cd->is_Load() ) { // Loads can now float around in the loop | |
| 1965 _igvn.hash_delete(cd); | |
| 1966 // Allow the load to float around in the loop, or before it | |
| 1967 // but NOT before the pre-loop. | |
| 1968 cd->set_req(0, ctrl); // ctrl, not NULL | |
| 1969 _igvn._worklist.push(cd); | |
| 1970 --i; | |
| 1971 --imax; | |
| 1972 } | |
| 1973 } | |
| 1974 | |
| 1975 } // End of is IF | |
| 1976 | |
| 1977 } | |
| 1978 | |
| 1979 // Update loop limits | |
| 1980 if (conditional_rc) { | |
| 1981 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit) | |
| 1982 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit); | |
| 1983 register_new_node(pre_limit, pre_ctrl); | |
| 1984 } | |
| 1985 _igvn.hash_delete(pre_opaq); | |
| 1986 pre_opaq->set_req(1, pre_limit); | |
| 1987 | |
| 1988 // Note:: we are making the main loop limit no longer precise; | |
| 1989 // need to round up based on stride. | |
| 3345 | 1990 cl->set_nonexact_trip_count(); |
| 1991 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case | |
| 0 | 1992 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init |
| 1993 // Hopefully, compiler will optimize for powers of 2. | |
| 1994 Node *ctrl = get_ctrl(main_limit); | |
| 1995 Node *stride = cl->stride(); | |
| 1996 Node *init = cl->init_trip(); | |
| 1997 Node *span = new (C, 3) SubINode(main_limit,init); | |
| 1998 register_new_node(span,ctrl); | |
| 1999 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); | |
| 2000 Node *add = new (C, 3) AddINode(span,rndup); | |
| 2001 register_new_node(add,ctrl); | |
| 2002 Node *div = new (C, 3) DivINode(0,add,stride); | |
| 2003 register_new_node(div,ctrl); | |
| 2004 Node *mul = new (C, 3) MulINode(div,stride); | |
| 2005 register_new_node(mul,ctrl); | |
| 2006 Node *newlim = new (C, 3) AddINode(mul,init); | |
| 2007 register_new_node(newlim,ctrl); | |
| 2008 main_limit = newlim; | |
| 2009 } | |
| 2010 | |
| 2011 Node *main_cle = cl->loopexit(); | |
| 2012 Node *main_bol = main_cle->in(1); | |
| 2013 // Hacking loop bounds; need private copies of exit test | |
| 2014 if( main_bol->outcnt() > 1 ) {// BoolNode shared? | |
| 2015 _igvn.hash_delete(main_cle); | |
| 2016 main_bol = main_bol->clone();// Clone a private BoolNode | |
| 2017 register_new_node( main_bol, main_cle->in(0) ); | |
| 2018 main_cle->set_req(1,main_bol); | |
| 2019 } | |
| 2020 Node *main_cmp = main_bol->in(1); | |
| 2021 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? | |
| 2022 _igvn.hash_delete(main_bol); | |
| 2023 main_cmp = main_cmp->clone();// Clone a private CmpNode | |
| 2024 register_new_node( main_cmp, main_cle->in(0) ); | |
| 2025 main_bol->set_req(1,main_cmp); | |
| 2026 } | |
| 2027 // Hack the now-private loop bounds | |
| 2028 _igvn.hash_delete(main_cmp); | |
| 2029 main_cmp->set_req(2, main_limit); | |
| 2030 _igvn._worklist.push(main_cmp); | |
| 2031 // The OpaqueNode is unshared by design | |
| 2032 _igvn.hash_delete(opqzm); | |
| 2033 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); | |
| 2034 opqzm->set_req(1,main_limit); | |
| 2035 _igvn._worklist.push(opqzm); | |
| 2036 } | |
| 2037 | |
| 2038 //------------------------------DCE_loop_body---------------------------------- | |
| 2039 // Remove simplistic dead code from loop body | |
| 2040 void IdealLoopTree::DCE_loop_body() { | |
| 2041 for( uint i = 0; i < _body.size(); i++ ) | |
| 2042 if( _body.at(i)->outcnt() == 0 ) | |
| 2043 _body.map( i--, _body.pop() ); | |
| 2044 } | |
| 2045 | |
| 2046 | |
| 2047 //------------------------------adjust_loop_exit_prob-------------------------- | |
| 2048 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. | |
| 2049 // Replace with a 1-in-10 exit guess. | |
| 2050 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { | |
| 2051 Node *test = tail(); | |
| 2052 while( test != _head ) { | |
| 2053 uint top = test->Opcode(); | |
| 2054 if( top == Op_IfTrue || top == Op_IfFalse ) { | |
| 2055 int test_con = ((ProjNode*)test)->_con; | |
| 2056 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); | |
| 2057 IfNode *iff = test->in(0)->as_If(); | |
| 2058 if( iff->outcnt() == 2 ) { // Ignore dead tests | |
| 2059 Node *bol = iff->in(1); | |
| 2060 if( bol && bol->req() > 1 && bol->in(1) && | |
| 2061 ((bol->in(1)->Opcode() == Op_StorePConditional ) || | |
|
420
a1980da045cc
6462850: generate biased locking code in C2 ideal graph
kvn
parents:
401
diff
changeset
|
2062 (bol->in(1)->Opcode() == Op_StoreIConditional ) || |
| 0 | 2063 (bol->in(1)->Opcode() == Op_StoreLConditional ) || |
| 2064 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || | |
| 2065 (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
|
2066 (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
|
2067 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) |
| 0 | 2068 return; // Allocation loops RARELY take backedge |
| 2069 // Find the OTHER exit path from the IF | |
| 2070 Node* ex = iff->proj_out(1-test_con); | |
| 2071 float p = iff->_prob; | |
| 2072 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { | |
| 2073 if( top == Op_IfTrue ) { | |
| 2074 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { | |
| 2075 iff->_prob = PROB_STATIC_FREQUENT; | |
| 2076 } | |
| 2077 } else { | |
| 2078 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { | |
| 2079 iff->_prob = PROB_STATIC_INFREQUENT; | |
| 2080 } | |
| 2081 } | |
| 2082 } | |
| 2083 } | |
| 2084 } | |
| 2085 test = phase->idom(test); | |
| 2086 } | |
| 2087 } | |
| 2088 | |
| 2089 | |
| 2090 //------------------------------policy_do_remove_empty_loop-------------------- | |
| 2091 // Micro-benchmark spamming. Policy is to always remove empty loops. | |
| 2092 // The 'DO' part is to replace the trip counter with the value it will | |
| 2093 // have on the last iteration. This will break the loop. | |
| 2094 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { | |
| 2095 // Minimum size must be empty loop | |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
2096 if (_body.size() > EMPTY_LOOP_SIZE) |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2097 return false; |
| 0 | 2098 |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2099 if (!_head->is_CountedLoop()) |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2100 return false; // Dead loop |
| 0 | 2101 CountedLoopNode *cl = _head->as_CountedLoop(); |
|
3850
6987871cfb9b
7077439: Possible reference through NULL in loopPredicate.cpp:726
kvn
parents:
3845
diff
changeset
|
2102 if (!cl->is_valid_counted_loop()) |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2103 return false; // Malformed loop |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2104 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) |
| 0 | 2105 return false; // Infinite loop |
| 2403 | 2106 |
| 0 | 2107 #ifdef ASSERT |
| 2108 // Ensure only one phi which is the iv. | |
| 2109 Node* iv = NULL; | |
| 2110 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { | |
| 2111 Node* n = cl->fast_out(i); | |
| 2112 if (n->Opcode() == Op_Phi) { | |
| 2113 assert(iv == NULL, "Too many phis" ); | |
| 2114 iv = n; | |
| 2115 } | |
| 2116 } | |
| 2117 assert(iv == cl->phi(), "Wrong phi" ); | |
| 2118 #endif | |
| 2403 | 2119 |
| 2120 // main and post loops have explicitly created zero trip guard | |
| 2121 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); | |
| 2122 if (needs_guard) { | |
| 2465 | 2123 // Skip guard if values not overlap. |
| 2124 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); | |
| 2125 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); | |
| 2126 int stride_con = cl->stride_con(); | |
| 2127 if (stride_con > 0) { | |
| 2128 needs_guard = (init_t->_hi >= limit_t->_lo); | |
| 2129 } else { | |
| 2130 needs_guard = (init_t->_lo <= limit_t->_hi); | |
| 2131 } | |
| 2132 } | |
| 2133 if (needs_guard) { | |
| 2403 | 2134 // Check for an obvious zero trip guard. |
| 2445 | 2135 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); |
| 2403 | 2136 if (inctrl->Opcode() == Op_IfTrue) { |
| 2137 // The test should look like just the backedge of a CountedLoop | |
| 2138 Node* iff = inctrl->in(0); | |
| 2139 if (iff->is_If()) { | |
| 2140 Node* bol = iff->in(1); | |
| 2141 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { | |
| 2142 Node* cmp = bol->in(1); | |
| 2143 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { | |
| 2144 needs_guard = false; | |
| 2145 } | |
| 2146 } | |
| 2147 } | |
| 2148 } | |
| 2149 } | |
| 2150 | |
| 2151 #ifndef PRODUCT | |
| 2152 if (PrintOpto) { | |
| 2153 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); | |
| 2154 this->dump_head(); | |
| 2155 } else if (TraceLoopOpts) { | |
| 2156 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); | |
| 2157 this->dump_head(); | |
| 2158 } | |
| 2159 #endif | |
| 2160 | |
| 2161 if (needs_guard) { | |
| 2162 // Peel the loop to ensure there's a zero trip guard | |
| 2163 Node_List old_new; | |
| 2164 phase->do_peeling(this, old_new); | |
| 2165 } | |
| 2166 | |
| 0 | 2167 // Replace the phi at loop head with the final value of the last |
| 2168 // iteration. Then the CountedLoopEnd will collapse (backedge never | |
| 2169 // taken) and all loop-invariant uses of the exit values will be correct. | |
| 2170 Node *phi = cl->phi(); | |
| 3345 | 2171 Node *exact_limit = phase->exact_limit(this); |
| 2172 if (exact_limit != cl->limit()) { | |
| 2173 // We also need to replace the original limit to collapse loop exit. | |
| 2174 Node* cmp = cl->loopexit()->cmp_node(); | |
| 2175 assert(cl->limit() == cmp->in(2), "sanity"); | |
| 2176 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist | |
| 2177 phase->_igvn.hash_delete(cmp); | |
| 2178 cmp->set_req(2, exact_limit); | |
| 2179 phase->_igvn._worklist.push(cmp); // put cmp on worklist | |
| 2180 } | |
| 2181 // Note: the final value after increment should not overflow since | |
| 2182 // counted loop has limit check predicate. | |
| 2183 Node *final = new (phase->C, 3) SubINode( exact_limit, cl->stride() ); | |
| 0 | 2184 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
|
2185 phase->_igvn.replace_node(phi,final); |
| 0 | 2186 phase->C->set_major_progress(); |
| 2187 return true; | |
| 2188 } | |
| 2189 | |
| 2465 | 2190 //------------------------------policy_do_one_iteration_loop------------------- |
| 2191 // Convert one iteration loop into normal code. | |
| 2192 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { | |
| 2193 if (!_head->as_Loop()->is_valid_counted_loop()) | |
| 2194 return false; // Only for counted loop | |
| 2195 | |
| 2196 CountedLoopNode *cl = _head->as_CountedLoop(); | |
| 2197 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { | |
| 2198 return false; | |
| 2199 } | |
| 2200 | |
| 2201 #ifndef PRODUCT | |
| 2202 if(TraceLoopOpts) { | |
| 2203 tty->print("OneIteration "); | |
| 2204 this->dump_head(); | |
| 2205 } | |
| 2206 #endif | |
| 2207 | |
| 2208 Node *init_n = cl->init_trip(); | |
| 2209 #ifdef ASSERT | |
| 2210 // Loop boundaries should be constant since trip count is exact. | |
| 2211 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); | |
| 2212 #endif | |
| 2213 // Replace the phi at loop head with the value of the init_trip. | |
| 2214 // Then the CountedLoopEnd will collapse (backedge will not be taken) | |
| 2215 // and all loop-invariant uses of the exit values will be correct. | |
| 2216 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); | |
| 2217 phase->C->set_major_progress(); | |
| 2218 return true; | |
| 2219 } | |
| 0 | 2220 |
| 2221 //============================================================================= | |
| 2222 //------------------------------iteration_split_impl--------------------------- | |
|
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2223 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { |
| 2465 | 2224 // Compute exact loop trip count if possible. |
| 2225 compute_exact_trip_count(phase); | |
| 2226 | |
| 2227 // Convert one iteration loop into normal code. | |
| 2228 if (policy_do_one_iteration_loop(phase)) | |
| 2229 return true; | |
| 2230 | |
| 0 | 2231 // Check and remove empty loops (spam micro-benchmarks) |
| 2465 | 2232 if (policy_do_remove_empty_loop(phase)) |
| 1172 | 2233 return true; // Here we removed an empty loop |
| 0 | 2234 |
| 2235 bool should_peel = policy_peeling(phase); // Should we peel? | |
| 2236 | |
| 2237 bool should_unswitch = policy_unswitching(phase); | |
| 2238 | |
| 2239 // Non-counted loops may be peeled; exactly 1 iteration is peeled. | |
| 2240 // This removes loop-invariant tests (usually null checks). | |
| 2465 | 2241 if (!_head->is_CountedLoop()) { // Non-counted loop |
| 0 | 2242 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
|
2243 // 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
|
2244 return false; |
| 0 | 2245 } |
| 2465 | 2246 if (should_peel) { // Should we peel? |
| 0 | 2247 #ifndef PRODUCT |
| 2248 if (PrintOpto) tty->print_cr("should_peel"); | |
| 2249 #endif | |
| 2250 phase->do_peeling(this,old_new); | |
| 2465 | 2251 } else if (should_unswitch) { |
| 0 | 2252 phase->do_unswitching(this, old_new); |
| 2253 } | |
|
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2254 return true; |
| 0 | 2255 } |
| 2256 CountedLoopNode *cl = _head->as_CountedLoop(); | |
| 2257 | |
|
3850
6987871cfb9b
7077439: Possible reference through NULL in loopPredicate.cpp:726
kvn
parents:
3845
diff
changeset
|
2258 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops |
| 0 | 2259 |
| 2260 // Do nothing special to pre- and post- loops | |
| 2465 | 2261 if (cl->is_pre_loop() || cl->is_post_loop()) return true; |
| 0 | 2262 |
| 2263 // Compute loop trip count from profile data | |
| 2264 compute_profile_trip_cnt(phase); | |
| 2265 | |
| 2266 // Before attempting fancy unrolling, RCE or alignment, see if we want | |
| 2267 // to completely unroll this loop or do loop unswitching. | |
| 2465 | 2268 if (cl->is_normal_loop()) { |
|
789
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
2269 if (should_unswitch) { |
|
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
2270 phase->do_unswitching(this, old_new); |
|
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
2271 return true; |
|
273b2358ef1a
6837146: Should perform unswitch before maximally unroll in loop transformation
cfang
parents:
605
diff
changeset
|
2272 } |
| 0 | 2273 bool should_maximally_unroll = policy_maximally_unroll(phase); |
| 2465 | 2274 if (should_maximally_unroll) { |
| 0 | 2275 // Here we did some unrolling and peeling. Eventually we will |
| 2276 // completely unroll this loop and it will no longer be a loop. | |
| 2277 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
|
2278 return true; |
| 0 | 2279 } |
| 2280 } | |
| 2281 | |
|
2453
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
2282 // 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
|
2283 // 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
|
2284 uint nodes_left = MaxNodeLimit - phase->C->unique(); |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
2285 if ((2 * _body.size()) > nodes_left) { |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
2286 return true; |
|
d7a3fed1c1c9
7004547: regular loop unroll should not unroll more than max unrolling
kvn
parents:
2448
diff
changeset
|
2287 } |
| 0 | 2288 |
| 2289 // Counted loops may be peeled, may need some iterations run up | |
| 2290 // front for RCE, and may want to align loop refs to a cache | |
| 2291 // line. Thus we clone a full loop up front whose trip count is | |
| 2292 // at least 1 (if peeling), but may be several more. | |
| 2293 | |
| 2294 // The main loop will start cache-line aligned with at least 1 | |
| 2295 // iteration of the unrolled body (zero-trip test required) and | |
| 2296 // will have some range checks removed. | |
| 2297 | |
| 2298 // A post-loop will finish any odd iterations (leftover after | |
| 2299 // unrolling), plus any needed for RCE purposes. | |
| 2300 | |
| 2301 bool should_unroll = policy_unroll(phase); | |
| 2302 | |
| 2303 bool should_rce = policy_range_check(phase); | |
| 2304 | |
| 2305 bool should_align = policy_align(phase); | |
| 2306 | |
| 2307 // If not RCE'ing (iteration splitting) or Aligning, then we do not | |
| 2308 // need a pre-loop. We may still need to peel an initial iteration but | |
| 2309 // we will not be needing an unknown number of pre-iterations. | |
| 2310 // | |
| 2311 // Basically, if may_rce_align reports FALSE first time through, | |
| 2312 // we will not be able to later do RCE or Aligning on this loop. | |
| 2313 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; | |
| 2314 | |
| 2315 // If we have any of these conditions (RCE, alignment, unrolling) met, then | |
| 2316 // we switch to the pre-/main-/post-loop model. This model also covers | |
| 2317 // peeling. | |
| 2465 | 2318 if (should_rce || should_align || should_unroll) { |
| 2319 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops | |
| 0 | 2320 phase->insert_pre_post_loops(this,old_new, !may_rce_align); |
| 2321 | |
| 2322 // Adjust the pre- and main-loop limits to let the pre and post loops run | |
| 2323 // with full checks, but the main-loop with no checks. Remove said | |
| 2324 // checks from the main body. | |
| 2465 | 2325 if (should_rce) |
| 0 | 2326 phase->do_range_check(this,old_new); |
| 2327 | |
| 2328 // Double loop body for unrolling. Adjust the minimum-trip test (will do | |
| 2329 // twice as many iterations as before) and the main body limit (only do | |
| 2330 // an even number of trips). If we are peeling, we might enable some RCE | |
| 2331 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if | |
| 2332 // peeling. | |
| 2465 | 2333 if (should_unroll && !should_peel) |
| 2334 phase->do_unroll(this,old_new, true); | |
| 0 | 2335 |
| 2336 // Adjust the pre-loop limits to align the main body | |
| 2337 // iterations. | |
| 2465 | 2338 if (should_align) |
| 0 | 2339 Unimplemented(); |
| 2340 | |
| 2341 } else { // Else we have an unchanged counted loop | |
| 2465 | 2342 if (should_peel) // Might want to peel but do nothing else |
| 0 | 2343 phase->do_peeling(this,old_new); |
| 2344 } | |
|
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2345 return true; |
| 0 | 2346 } |
| 2347 | |
| 2348 | |
| 2349 //============================================================================= | |
| 2350 //------------------------------iteration_split-------------------------------- | |
|
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2351 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { |
| 0 | 2352 // Recursively iteration split nested loops |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2353 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
|
2354 return false; |
| 0 | 2355 |
| 2356 // Clean out prior deadwood | |
| 2357 DCE_loop_body(); | |
| 2358 | |
| 2359 | |
| 2360 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. | |
| 2361 // Replace with a 1-in-10 exit guess. | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2362 if (_parent /*not the root loop*/ && |
| 0 | 2363 !_irreducible && |
| 2364 // Also ignore the occasional dead backedge | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2365 !tail()->is_top()) { |
| 0 | 2366 adjust_loop_exit_prob(phase); |
| 2367 } | |
| 2368 | |
| 2369 // Gate unrolling, RCE and peeling efforts. | |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2370 if (!_child && // If not an inner loop, do not split |
| 0 | 2371 !_irreducible && |
|
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
0
diff
changeset
|
2372 _allow_optimizations && |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2373 !tail()->is_top()) { // Also ignore the occasional dead backedge |
| 0 | 2374 if (!_has_call) { |
|
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2375 if (!iteration_split_impl(phase, old_new)) { |
| 1172 | 2376 return false; |
| 2377 } | |
| 0 | 2378 } else if (policy_unswitching(phase)) { |
| 2379 phase->do_unswitching(this, old_new); | |
| 2380 } | |
| 2381 } | |
| 2382 | |
| 2383 // 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
|
2384 // trip counter when there was no major reshaping. |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2385 phase->reorg_offsets(this); |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2386 |
|
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
1972
diff
changeset
|
2387 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
|
2388 return false; |
|
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
400
diff
changeset
|
2389 return true; |
| 0 | 2390 } |
| 1172 | 2391 |
| 2392 | |
| 2445 | 2393 //============================================================================= |
| 1763 | 2394 // Process all the loops in the loop tree and replace any fill |
| 2395 // patterns with an intrisc version. | |
| 2396 bool PhaseIdealLoop::do_intrinsify_fill() { | |
| 2397 bool changed = false; | |
| 2398 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
| 2399 IdealLoopTree* lpt = iter.current(); | |
| 2400 changed |= intrinsify_fill(lpt); | |
| 2401 } | |
| 2402 return changed; | |
| 2403 } | |
| 2404 | |
| 2405 | |
| 2406 // Examine an inner loop looking for a a single store of an invariant | |
| 2407 // value in a unit stride loop, | |
| 2408 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, | |
| 2409 Node*& shift, Node*& con) { | |
| 2410 const char* msg = NULL; | |
| 2411 Node* msg_node = NULL; | |
| 2412 | |
| 2413 store_value = NULL; | |
| 2414 con = NULL; | |
| 2415 shift = NULL; | |
| 2416 | |
| 2417 // Process the loop looking for stores. If there are multiple | |
| 2418 // stores or extra control flow give at this point. | |
| 2419 CountedLoopNode* head = lpt->_head->as_CountedLoop(); | |
| 2420 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { | |
| 2421 Node* n = lpt->_body.at(i); | |
| 2422 if (n->outcnt() == 0) continue; // Ignore dead | |
| 2423 if (n->is_Store()) { | |
| 2424 if (store != NULL) { | |
| 2425 msg = "multiple stores"; | |
| 2426 break; | |
| 2427 } | |
| 2428 int opc = n->Opcode(); | |
| 2429 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) { | |
| 2430 msg = "oop fills not handled"; | |
| 2431 break; | |
| 2432 } | |
| 2433 Node* value = n->in(MemNode::ValueIn); | |
| 2434 if (!lpt->is_invariant(value)) { | |
| 2435 msg = "variant store value"; | |
|
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2436 } 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
|
2437 msg = "not array address"; |
| 1763 | 2438 } |
| 2439 store = n; | |
| 2440 store_value = value; | |
| 2441 } else if (n->is_If() && n != head->loopexit()) { | |
| 2442 msg = "extra control flow"; | |
| 2443 msg_node = n; | |
| 2444 } | |
| 2445 } | |
| 2446 | |
| 2447 if (store == NULL) { | |
| 2448 // No store in loop | |
| 2449 return false; | |
| 2450 } | |
| 2451 | |
| 2452 if (msg == NULL && head->stride_con() != 1) { | |
| 2453 // could handle negative strides too | |
| 2454 if (head->stride_con() < 0) { | |
| 2455 msg = "negative stride"; | |
| 2456 } else { | |
| 2457 msg = "non-unit stride"; | |
| 2458 } | |
| 2459 } | |
| 2460 | |
| 2461 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { | |
| 2462 msg = "can't handle store address"; | |
| 2463 msg_node = store->in(MemNode::Address); | |
| 2464 } | |
| 2465 | |
|
1813
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2466 if (msg == NULL && |
|
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2467 (!store->in(MemNode::Memory)->is_Phi() || |
|
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2468 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { |
|
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2469 msg = "store memory isn't proper phi"; |
|
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2470 msg_node = store->in(MemNode::Memory); |
|
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2471 } |
|
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2472 |
| 1763 | 2473 // Make sure there is an appropriate fill routine |
| 2474 BasicType t = store->as_Mem()->memory_type(); | |
| 2475 const char* fill_name; | |
| 2476 if (msg == NULL && | |
| 2477 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { | |
| 2478 msg = "unsupported store"; | |
| 2479 msg_node = store; | |
| 2480 } | |
| 2481 | |
| 2482 if (msg != NULL) { | |
| 2483 #ifndef PRODUCT | |
| 2484 if (TraceOptimizeFill) { | |
| 2485 tty->print_cr("not fill intrinsic candidate: %s", msg); | |
| 2486 if (msg_node != NULL) msg_node->dump(); | |
| 2487 } | |
| 2488 #endif | |
| 2489 return false; | |
| 2490 } | |
| 2491 | |
| 2492 // Make sure the address expression can be handled. It should be | |
| 2493 // head->phi * elsize + con. head->phi might have a ConvI2L. | |
| 2494 Node* elements[4]; | |
| 2495 Node* conv = NULL; | |
|
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2496 bool found_index = false; |
| 1763 | 2497 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); |
| 2498 for (int e = 0; e < count; e++) { | |
| 2499 Node* n = elements[e]; | |
| 2500 if (n->is_Con() && con == NULL) { | |
| 2501 con = n; | |
| 2502 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { | |
| 2503 Node* value = n->in(1); | |
| 2504 #ifdef _LP64 | |
| 2505 if (value->Opcode() == Op_ConvI2L) { | |
| 2506 conv = value; | |
| 2507 value = value->in(1); | |
| 2508 } | |
| 2509 #endif | |
| 2510 if (value != head->phi()) { | |
| 2511 msg = "unhandled shift in address"; | |
| 2512 } else { | |
|
2448
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2513 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
|
2514 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
|
2515 } else { |
|
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2516 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
|
2517 shift = n; |
|
8b2317d732ec
7026957: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int()) failed
never
parents:
2445
diff
changeset
|
2518 } |
| 1763 | 2519 } |
| 2520 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { | |
| 2521 if (n->in(1) == head->phi()) { | |
|
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2522 found_index = true; |
| 1763 | 2523 conv = n; |
| 2524 } else { | |
| 2525 msg = "unhandled input to ConvI2L"; | |
| 2526 } | |
| 2527 } else if (n == head->phi()) { | |
| 2528 // no shift, check below for allowed cases | |
|
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2529 found_index = true; |
| 1763 | 2530 } else { |
| 2531 msg = "unhandled node in address"; | |
| 2532 msg_node = n; | |
| 2533 } | |
| 2534 } | |
| 2535 | |
| 2536 if (count == -1) { | |
| 2537 msg = "malformed address expression"; | |
| 2538 msg_node = store; | |
| 2539 } | |
| 2540 | |
|
1785
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2541 if (!found_index) { |
|
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2542 msg = "missing use of index"; |
|
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2543 } |
|
5e4f03302987
6982533: Crash in ~StubRoutines::jbyte_fill with AggressiveOpts enabled
never
parents:
1763
diff
changeset
|
2544 |
| 1763 | 2545 // byte sized items won't have a shift |
| 2546 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { | |
| 2547 msg = "can't find shift"; | |
| 2548 msg_node = store; | |
| 2549 } | |
| 2550 | |
| 2551 if (msg != NULL) { | |
| 2552 #ifndef PRODUCT | |
| 2553 if (TraceOptimizeFill) { | |
| 2554 tty->print_cr("not fill intrinsic: %s", msg); | |
| 2555 if (msg_node != NULL) msg_node->dump(); | |
| 2556 } | |
| 2557 #endif | |
| 2558 return false; | |
| 2559 } | |
| 2560 | |
| 2561 // No make sure all the other nodes in the loop can be handled | |
| 2562 VectorSet ok(Thread::current()->resource_area()); | |
| 2563 | |
| 2564 // store related values are ok | |
| 2565 ok.set(store->_idx); | |
| 2566 ok.set(store->in(MemNode::Memory)->_idx); | |
| 2567 | |
| 2568 // Loop structure is ok | |
| 2569 ok.set(head->_idx); | |
| 2570 ok.set(head->loopexit()->_idx); | |
| 2571 ok.set(head->phi()->_idx); | |
| 2572 ok.set(head->incr()->_idx); | |
| 2573 ok.set(head->loopexit()->cmp_node()->_idx); | |
| 2574 ok.set(head->loopexit()->in(1)->_idx); | |
| 2575 | |
| 2576 // Address elements are ok | |
| 2577 if (con) ok.set(con->_idx); | |
| 2578 if (shift) ok.set(shift->_idx); | |
| 2579 if (conv) ok.set(conv->_idx); | |
| 2580 | |
| 2581 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { | |
| 2582 Node* n = lpt->_body.at(i); | |
| 2583 if (n->outcnt() == 0) continue; // Ignore dead | |
| 2584 if (ok.test(n->_idx)) continue; | |
| 2585 // Backedge projection is ok | |
| 2586 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue; | |
| 2587 if (!n->is_AddP()) { | |
| 2588 msg = "unhandled node"; | |
| 2589 msg_node = n; | |
| 2590 break; | |
| 2591 } | |
| 2592 } | |
| 2593 | |
| 2594 // Make sure no unexpected values are used outside the loop | |
| 2595 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { | |
| 2596 Node* n = lpt->_body.at(i); | |
| 2597 // These values can be replaced with other nodes if they are used | |
| 2598 // outside the loop. | |
|
1813
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2599 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue; |
| 1763 | 2600 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { |
| 2601 Node* use = iter.get(); | |
| 2602 if (!lpt->_body.contains(use)) { | |
| 2603 msg = "node is used outside loop"; | |
| 2604 // lpt->_body.dump(); | |
| 2605 msg_node = n; | |
| 2606 break; | |
| 2607 } | |
| 2608 } | |
| 2609 } | |
| 2610 | |
| 2611 #ifdef ASSERT | |
| 2612 if (TraceOptimizeFill) { | |
| 2613 if (msg != NULL) { | |
| 2614 tty->print_cr("no fill intrinsic: %s", msg); | |
| 2615 if (msg_node != NULL) msg_node->dump(); | |
| 2616 } else { | |
| 2617 tty->print_cr("fill intrinsic for:"); | |
| 2618 } | |
| 2619 store->dump(); | |
| 2620 if (Verbose) { | |
| 2621 lpt->_body.dump(); | |
| 2622 } | |
| 2623 } | |
| 2624 #endif | |
| 2625 | |
| 2626 return msg == NULL; | |
| 2627 } | |
| 2628 | |
| 2629 | |
| 2630 | |
| 2631 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { | |
| 2632 // Only for counted inner loops | |
| 2633 if (!lpt->is_counted() || !lpt->is_inner()) { | |
| 2634 return false; | |
| 2635 } | |
| 2636 | |
| 2637 // Must have constant stride | |
| 2638 CountedLoopNode* head = lpt->_head->as_CountedLoop(); | |
|
3850
6987871cfb9b
7077439: Possible reference through NULL in loopPredicate.cpp:726
kvn
parents:
3845
diff
changeset
|
2639 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { |
| 1763 | 2640 return false; |
| 2641 } | |
| 2642 | |
| 2643 // Check that the body only contains a store of a loop invariant | |
| 2644 // value that is indexed by the loop phi. | |
| 2645 Node* store = NULL; | |
| 2646 Node* store_value = NULL; | |
| 2647 Node* shift = NULL; | |
| 2648 Node* offset = NULL; | |
| 2649 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { | |
| 2650 return false; | |
| 2651 } | |
| 2652 | |
| 2445 | 2653 #ifndef PRODUCT |
| 2654 if (TraceLoopOpts) { | |
| 2655 tty->print("ArrayFill "); | |
| 2656 lpt->dump_head(); | |
| 2657 } | |
| 2658 #endif | |
| 2659 | |
| 1763 | 2660 // Now replace the whole loop body by a call to a fill routine that |
| 2661 // covers the same region as the loop. | |
| 2662 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); | |
| 2663 | |
| 2664 // Build an expression for the beginning of the copy region | |
| 2665 Node* index = head->init_trip(); | |
| 2666 #ifdef _LP64 | |
| 2667 index = new (C, 2) ConvI2LNode(index); | |
| 2668 _igvn.register_new_node_with_optimizer(index); | |
| 2669 #endif | |
| 2670 if (shift != NULL) { | |
| 2671 // byte arrays don't require a shift but others do. | |
| 2672 index = new (C, 3) LShiftXNode(index, shift->in(2)); | |
| 2673 _igvn.register_new_node_with_optimizer(index); | |
| 2674 } | |
| 2675 index = new (C, 4) AddPNode(base, base, index); | |
| 2676 _igvn.register_new_node_with_optimizer(index); | |
| 2677 Node* from = new (C, 4) AddPNode(base, index, offset); | |
| 2678 _igvn.register_new_node_with_optimizer(from); | |
| 2679 // Compute the number of elements to copy | |
| 2680 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip()); | |
| 2681 _igvn.register_new_node_with_optimizer(len); | |
| 2682 | |
| 2683 BasicType t = store->as_Mem()->memory_type(); | |
| 2684 bool aligned = false; | |
| 2685 if (offset != NULL && head->init_trip()->is_Con()) { | |
| 2686 int element_size = type2aelembytes(t); | |
| 2687 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; | |
| 2688 } | |
| 2689 | |
| 2690 // Build a call to the fill routine | |
| 2691 const char* fill_name; | |
| 2692 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); | |
| 2693 assert(fill != NULL, "what?"); | |
| 2694 | |
| 2695 // Convert float/double to int/long for fill routines | |
| 2696 if (t == T_FLOAT) { | |
| 2697 store_value = new (C, 2) MoveF2INode(store_value); | |
| 2698 _igvn.register_new_node_with_optimizer(store_value); | |
| 2699 } else if (t == T_DOUBLE) { | |
| 2700 store_value = new (C, 2) MoveD2LNode(store_value); | |
| 2701 _igvn.register_new_node_with_optimizer(store_value); | |
| 2702 } | |
| 2703 | |
| 2704 Node* mem_phi = store->in(MemNode::Memory); | |
| 2705 Node* result_ctrl; | |
| 2706 Node* result_mem; | |
| 2707 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); | |
| 2708 int size = call_type->domain()->cnt(); | |
| 2709 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill, | |
| 2710 fill_name, TypeAryPtr::get_array_body_type(t)); | |
| 2711 call->init_req(TypeFunc::Parms+0, from); | |
| 2712 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
|
2713 #ifdef _LP64 |
|
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2714 len = new (C, 2) ConvI2LNode(len); |
|
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2715 _igvn.register_new_node_with_optimizer(len); |
|
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2716 #endif |
| 1763 | 2717 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
|
2718 #ifdef _LP64 |
|
75588558f1bf
6980792: Crash "exception happened outside interpreter, nmethods and vtable stubs (1)"
never
parents:
1813
diff
changeset
|
2719 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
|
2720 #endif |
| 1763 | 2721 call->init_req( TypeFunc::Control, head->init_control()); |
| 2722 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o | |
| 2723 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) ); | |
| 2724 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) ); | |
| 2725 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) ); | |
| 2726 _igvn.register_new_node_with_optimizer(call); | |
| 2727 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); | |
| 2728 _igvn.register_new_node_with_optimizer(result_ctrl); | |
| 2729 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory); | |
| 2730 _igvn.register_new_node_with_optimizer(result_mem); | |
| 2731 | |
| 2732 // If this fill is tightly coupled to an allocation and overwrites | |
| 2733 // the whole body, allow it to take over the zeroing. | |
| 2734 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); | |
| 2735 if (alloc != NULL && alloc->is_AllocateArray()) { | |
| 2736 Node* length = alloc->as_AllocateArray()->Ideal_length(); | |
| 2737 if (head->limit() == length && | |
| 2738 head->init_trip() == _igvn.intcon(0)) { | |
| 2739 if (TraceOptimizeFill) { | |
| 2740 tty->print_cr("Eliminated zeroing in allocation"); | |
| 2741 } | |
| 2742 alloc->maybe_set_complete(&_igvn); | |
| 2743 } else { | |
| 2744 #ifdef ASSERT | |
| 2745 if (TraceOptimizeFill) { | |
| 2746 tty->print_cr("filling array but bounds don't match"); | |
| 2747 alloc->dump(); | |
| 2748 head->init_trip()->dump(); | |
| 2749 head->limit()->dump(); | |
| 2750 length->dump(); | |
| 2751 } | |
| 2752 #endif | |
| 2753 } | |
| 2754 } | |
| 2755 | |
| 2756 // Redirect the old control and memory edges that are outside the loop. | |
| 2757 Node* exit = head->loopexit()->proj_out(0); | |
|
1813
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2758 // 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
|
2759 // 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
|
2760 // result_mem. |
|
c77e8f982901
6984979: OptimizeFill misses some cases with an odd memory graph
never
parents:
1785
diff
changeset
|
2761 _igvn.replace_node(store->in(MemNode::Memory), result_mem); |
| 1763 | 2762 _igvn.replace_node(exit, result_ctrl); |
| 2763 _igvn.replace_node(store, result_mem); | |
| 2764 // Any uses the increment outside of the loop become the loop limit. | |
| 2765 _igvn.replace_node(head->incr(), head->limit()); | |
| 2766 | |
| 2767 // Disconnect the head from the loop. | |
| 2768 for (uint i = 0; i < lpt->_body.size(); i++) { | |
| 2769 Node* n = lpt->_body.at(i); | |
| 2770 _igvn.replace_node(n, C->top()); | |
| 2771 } | |
| 2772 | |
| 2773 return true; | |
| 2774 } |