1 /* 2 * Copyright (c) 2000, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 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/castnode.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/convertnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opaquenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 42 //------------------------------is_loop_exit----------------------------------- 43 // Given an IfNode, return the loop-exiting projection or NULL if both 44 // arms remain in the loop. 45 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 46 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 47 PhaseIdealLoop *phase = _phase; 48 // Test is an IfNode, has 2 projections. If BOTH are in the loop 49 // we need loop unswitching instead of peeling. 50 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 51 return iff->raw_out(0); 52 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 53 return iff->raw_out(1); 54 return NULL; 55 } 56 57 58 //============================================================================= 59 60 61 //------------------------------record_for_igvn---------------------------- 62 // Put loop body on igvn work list 63 void IdealLoopTree::record_for_igvn() { 64 for( uint i = 0; i < _body.size(); i++ ) { 65 Node *n = _body.at(i); 66 _phase->_igvn._worklist.push(n); 67 } 68 } 69 70 //------------------------------compute_exact_trip_count----------------------- 71 // Compute loop exact trip count if possible. Do not recalculate trip count for 72 // split loops (pre-main-post) which have their limits and inits behind Opaque node. 73 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) { 74 if (!_head->as_Loop()->is_valid_counted_loop()) { 75 return; 76 } 77 CountedLoopNode* cl = _head->as_CountedLoop(); 78 // Trip count may become nonexact for iteration split loops since 79 // RCE modifies limits. Note, _trip_count value is not reset since 80 // it is used to limit unrolling of main loop. 81 cl->set_nonexact_trip_count(); 82 83 // Loop's test should be part of loop. 84 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 85 return; // Infinite loop 86 87 #ifdef ASSERT 88 BoolTest::mask bt = cl->loopexit()->test_trip(); 89 assert(bt == BoolTest::lt || bt == BoolTest::gt || 90 bt == BoolTest::ne, "canonical test is expected"); 91 #endif 92 93 Node* init_n = cl->init_trip(); 94 Node* limit_n = cl->limit(); 95 if (init_n != NULL && init_n->is_Con() && 96 limit_n != NULL && limit_n->is_Con()) { 97 // Use longs to avoid integer overflow. 98 int stride_con = cl->stride_con(); 99 jlong init_con = cl->init_trip()->get_int(); 100 jlong limit_con = cl->limit()->get_int(); 101 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 102 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 103 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { 104 // Set exact trip count. 105 cl->set_exact_trip_count((uint)trip_count); 106 } 107 } 108 } 109 110 //------------------------------compute_profile_trip_cnt---------------------------- 111 // Compute loop trip count from profile data as 112 // (backedge_count + loop_exit_count) / loop_exit_count 113 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 114 if (!_head->is_CountedLoop()) { 115 return; 116 } 117 CountedLoopNode* head = _head->as_CountedLoop(); 118 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 119 return; // Already computed 120 } 121 float trip_cnt = (float)max_jint; // default is big 122 123 Node* back = head->in(LoopNode::LoopBackControl); 124 while (back != head) { 125 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 126 back->in(0) && 127 back->in(0)->is_If() && 128 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 129 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 130 break; 131 } 132 back = phase->idom(back); 133 } 134 if (back != head) { 135 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 136 back->in(0), "if-projection exists"); 137 IfNode* back_if = back->in(0)->as_If(); 138 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 139 140 // Now compute a loop exit count 141 float loop_exit_cnt = 0.0f; 142 for( uint i = 0; i < _body.size(); i++ ) { 143 Node *n = _body[i]; 144 if( n->is_If() ) { 145 IfNode *iff = n->as_If(); 146 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 147 Node *exit = is_loop_exit(iff); 148 if( exit ) { 149 float exit_prob = iff->_prob; 150 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 151 if (exit_prob > PROB_MIN) { 152 float exit_cnt = iff->_fcnt * exit_prob; 153 loop_exit_cnt += exit_cnt; 154 } 155 } 156 } 157 } 158 } 159 if (loop_exit_cnt > 0.0f) { 160 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 161 } else { 162 // No exit count so use 163 trip_cnt = loop_back_cnt; 164 } 165 } 166 #ifndef PRODUCT 167 if (TraceProfileTripCount) { 168 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 169 } 170 #endif 171 head->set_profile_trip_cnt(trip_cnt); 172 } 173 174 //---------------------is_invariant_addition----------------------------- 175 // Return nonzero index of invariant operand for an Add or Sub 176 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 177 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 178 int op = n->Opcode(); 179 if (op == Op_AddI || op == Op_SubI) { 180 bool in1_invar = this->is_invariant(n->in(1)); 181 bool in2_invar = this->is_invariant(n->in(2)); 182 if (in1_invar && !in2_invar) return 1; 183 if (!in1_invar && in2_invar) return 2; 184 } 185 return 0; 186 } 187 188 //---------------------reassociate_add_sub----------------------------- 189 // Reassociate invariant add and subtract expressions: 190 // 191 // inv1 + (x + inv2) => ( inv1 + inv2) + x 192 // (x + inv2) + inv1 => ( inv1 + inv2) + x 193 // inv1 + (x - inv2) => ( inv1 - inv2) + x 194 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 195 // (x + inv2) - inv1 => (-inv1 + inv2) + x 196 // (x - inv2) + inv1 => ( inv1 - inv2) + x 197 // (x - inv2) - inv1 => (-inv1 - inv2) + x 198 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 199 // inv1 - (x - inv2) => ( inv1 + inv2) - x 200 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 201 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 202 // inv1 - (x + inv2) => ( inv1 - inv2) - x 203 // 204 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 205 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 206 if (is_invariant(n1)) return NULL; 207 int inv1_idx = is_invariant_addition(n1, phase); 208 if (!inv1_idx) return NULL; 209 // Don't mess with add of constant (igvn moves them to expression tree root.) 210 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 211 Node* inv1 = n1->in(inv1_idx); 212 Node* n2 = n1->in(3 - inv1_idx); 213 int inv2_idx = is_invariant_addition(n2, phase); 214 if (!inv2_idx) return NULL; 215 Node* x = n2->in(3 - inv2_idx); 216 Node* inv2 = n2->in(inv2_idx); 217 218 bool neg_x = n2->is_Sub() && inv2_idx == 1; 219 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 220 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 221 if (n1->is_Sub() && inv1_idx == 1) { 222 neg_x = !neg_x; 223 neg_inv2 = !neg_inv2; 224 } 225 Node* inv1_c = phase->get_ctrl(inv1); 226 Node* inv2_c = phase->get_ctrl(inv2); 227 Node* n_inv1; 228 if (neg_inv1) { 229 Node *zero = phase->_igvn.intcon(0); 230 phase->set_ctrl(zero, phase->C->root()); 231 n_inv1 = new SubINode(zero, inv1); 232 phase->register_new_node(n_inv1, inv1_c); 233 } else { 234 n_inv1 = inv1; 235 } 236 Node* inv; 237 if (neg_inv2) { 238 inv = new SubINode(n_inv1, inv2); 239 } else { 240 inv = new AddINode(n_inv1, inv2); 241 } 242 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 243 244 Node* addx; 245 if (neg_x) { 246 addx = new SubINode(inv, x); 247 } else { 248 addx = new AddINode(x, inv); 249 } 250 phase->register_new_node(addx, phase->get_ctrl(x)); 251 phase->_igvn.replace_node(n1, addx); 252 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 253 _body.yank(n1); 254 return addx; 255 } 256 257 //---------------------reassociate_invariants----------------------------- 258 // Reassociate invariant expressions: 259 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 260 for (int i = _body.size() - 1; i >= 0; i--) { 261 Node *n = _body.at(i); 262 for (int j = 0; j < 5; j++) { 263 Node* nn = reassociate_add_sub(n, phase); 264 if (nn == NULL) break; 265 n = nn; // again 266 }; 267 } 268 } 269 270 //------------------------------policy_peeling--------------------------------- 271 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 272 // make some loop-invariant test (usually a null-check) happen before the loop. 273 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 274 Node *test = ((IdealLoopTree*)this)->tail(); 275 int body_size = ((IdealLoopTree*)this)->_body.size(); 276 // Peeling does loop cloning which can result in O(N^2) node construction 277 if( body_size > 255 /* Prevent overflow for large body_size */ 278 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) { 279 return false; // too large to safely clone 280 } 281 while( test != _head ) { // Scan till run off top of loop 282 if( test->is_If() ) { // Test? 283 Node *ctrl = phase->get_ctrl(test->in(1)); 284 if (ctrl->is_top()) 285 return false; // Found dead test on live IF? No peeling! 286 // Standard IF only has one input value to check for loop invariance 287 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 288 // Condition is not a member of this loop? 289 if( !is_member(phase->get_loop(ctrl)) && 290 is_loop_exit(test) ) 291 return true; // Found reason to peel! 292 } 293 // Walk up dominators to loop _head looking for test which is 294 // executed on every path thru loop. 295 test = phase->idom(test); 296 } 297 return false; 298 } 299 300 //------------------------------peeled_dom_test_elim--------------------------- 301 // If we got the effect of peeling, either by actually peeling or by making 302 // a pre-loop which must execute at least once, we can remove all 303 // loop-invariant dominated tests in the main body. 304 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 305 bool progress = true; 306 while( progress ) { 307 progress = false; // Reset for next iteration 308 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 309 Node *test = prev->in(0); 310 while( test != loop->_head ) { // Scan till run off top of loop 311 312 int p_op = prev->Opcode(); 313 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 314 test->is_If() && // Test? 315 !test->in(1)->is_Con() && // And not already obvious? 316 // Condition is not a member of this loop? 317 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 318 // Walk loop body looking for instances of this test 319 for( uint i = 0; i < loop->_body.size(); i++ ) { 320 Node *n = loop->_body.at(i); 321 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 322 // IfNode was dominated by version in peeled loop body 323 progress = true; 324 dominated_by( old_new[prev->_idx], n ); 325 } 326 } 327 } 328 prev = test; 329 test = idom(test); 330 } // End of scan tests in loop 331 332 } // End of while( progress ) 333 } 334 335 //------------------------------do_peeling------------------------------------- 336 // Peel the first iteration of the given loop. 337 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 338 // The pre-loop illegally has 2 control users (old & new loops). 339 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 340 // Do this by making the old-loop fall-in edges act as if they came 341 // around the loopback from the prior iteration (follow the old-loop 342 // backedges) and then map to the new peeled iteration. This leaves 343 // the pre-loop with only 1 user (the new peeled iteration), but the 344 // peeled-loop backedge has 2 users. 345 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 346 // extra backedge user. 347 // 348 // orig 349 // 350 // stmt1 351 // | 352 // v 353 // loop predicate 354 // | 355 // v 356 // loop<----+ 357 // | | 358 // stmt2 | 359 // | | 360 // v | 361 // if ^ 362 // / \ | 363 // / \ | 364 // v v | 365 // false true | 366 // / \ | 367 // / ----+ 368 // | 369 // v 370 // exit 371 // 372 // 373 // after clone loop 374 // 375 // stmt1 376 // | 377 // v 378 // loop predicate 379 // / \ 380 // clone / \ orig 381 // / \ 382 // / \ 383 // v v 384 // +---->loop clone loop<----+ 385 // | | | | 386 // | stmt2 clone stmt2 | 387 // | | | | 388 // | v v | 389 // ^ if clone If ^ 390 // | / \ / \ | 391 // | / \ / \ | 392 // | v v v v | 393 // | true false false true | 394 // | / \ / \ | 395 // +---- \ / ----+ 396 // \ / 397 // 1v v2 398 // region 399 // | 400 // v 401 // exit 402 // 403 // 404 // after peel and predicate move 405 // 406 // stmt1 407 // / 408 // / 409 // clone / orig 410 // / 411 // / +----------+ 412 // / | | 413 // / loop predicate | 414 // / | | 415 // v v | 416 // TOP-->loop clone loop<----+ | 417 // | | | | 418 // stmt2 clone stmt2 | | 419 // | | | ^ 420 // v v | | 421 // if clone If ^ | 422 // / \ / \ | | 423 // / \ / \ | | 424 // v v v v | | 425 // true false false true | | 426 // | \ / \ | | 427 // | \ / ----+ ^ 428 // | \ / | 429 // | 1v v2 | 430 // v region | 431 // | | | 432 // | v | 433 // | exit | 434 // | | 435 // +--------------->-----------------+ 436 // 437 // 438 // final graph 439 // 440 // stmt1 441 // | 442 // v 443 // stmt2 clone 444 // | 445 // v 446 // if clone 447 // / | 448 // / | 449 // v v 450 // false true 451 // | | 452 // | v 453 // | loop predicate 454 // | | 455 // | v 456 // | loop<----+ 457 // | | | 458 // | stmt2 | 459 // | | | 460 // | v | 461 // v if ^ 462 // | / \ | 463 // | / \ | 464 // | v v | 465 // | false true | 466 // | | \ | 467 // v v --+ 468 // region 469 // | 470 // v 471 // exit 472 // 473 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 474 475 C->set_major_progress(); 476 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 477 // 'pre' loop from the main and the 'pre' can no longer have it's 478 // iterations adjusted. Therefore, we need to declare this loop as 479 // no longer a 'main' loop; it will need new pre and post loops before 480 // we can do further RCE. 481 #ifndef PRODUCT 482 if (TraceLoopOpts) { 483 tty->print("Peel "); 484 loop->dump_head(); 485 } 486 #endif 487 Node* head = loop->_head; 488 bool counted_loop = head->is_CountedLoop(); 489 if (counted_loop) { 490 CountedLoopNode *cl = head->as_CountedLoop(); 491 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 492 cl->set_trip_count(cl->trip_count() - 1); 493 if (cl->is_main_loop()) { 494 cl->set_normal_loop(); 495 #ifndef PRODUCT 496 if (PrintOpto && VerifyLoopOptimizations) { 497 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 498 loop->dump_head(); 499 } 500 #endif 501 } 502 } 503 Node* entry = head->in(LoopNode::EntryControl); 504 505 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 506 // The pre-loop illegally has 2 control users (old & new loops). 507 clone_loop( loop, old_new, dom_depth(head) ); 508 509 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 510 // Do this by making the old-loop fall-in edges act as if they came 511 // around the loopback from the prior iteration (follow the old-loop 512 // backedges) and then map to the new peeled iteration. This leaves 513 // the pre-loop with only 1 user (the new peeled iteration), but the 514 // peeled-loop backedge has 2 users. 515 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 516 _igvn.hash_delete(head); 517 head->set_req(LoopNode::EntryControl, new_entry); 518 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 519 Node* old = head->fast_out(j); 520 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 521 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 522 if (!new_exit_value ) // Backedge value is ALSO loop invariant? 523 // Then loop body backedge value remains the same. 524 new_exit_value = old->in(LoopNode::LoopBackControl); 525 _igvn.hash_delete(old); 526 old->set_req(LoopNode::EntryControl, new_exit_value); 527 } 528 } 529 530 531 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 532 // extra backedge user. 533 Node* new_head = old_new[head->_idx]; 534 _igvn.hash_delete(new_head); 535 new_head->set_req(LoopNode::LoopBackControl, C->top()); 536 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 537 Node* use = new_head->fast_out(j2); 538 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 539 _igvn.hash_delete(use); 540 use->set_req(LoopNode::LoopBackControl, C->top()); 541 } 542 } 543 544 545 // Step 4: Correct dom-depth info. Set to loop-head depth. 546 int dd = dom_depth(head); 547 set_idom(head, head->in(1), dd); 548 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 549 Node *old = loop->_body.at(j3); 550 Node *nnn = old_new[old->_idx]; 551 if (!has_ctrl(nnn)) 552 set_idom(nnn, idom(nnn), dd-1); 553 } 554 555 // Now force out all loop-invariant dominating tests. The optimizer 556 // finds some, but we _know_ they are all useless. 557 peeled_dom_test_elim(loop,old_new); 558 559 loop->record_for_igvn(); 560 } 561 562 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 563 564 //------------------------------policy_maximally_unroll------------------------ 565 // Calculate exact loop trip count and return true if loop can be maximally 566 // unrolled. 567 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 568 CountedLoopNode *cl = _head->as_CountedLoop(); 569 assert(cl->is_normal_loop(), ""); 570 if (!cl->is_valid_counted_loop()) 571 return false; // Malformed counted loop 572 573 if (!cl->has_exact_trip_count()) { 574 // Trip count is not exact. 575 return false; 576 } 577 578 uint trip_count = cl->trip_count(); 579 // Note, max_juint is used to indicate unknown trip count. 580 assert(trip_count > 1, "one iteration loop should be optimized out already"); 581 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 582 583 // Real policy: if we maximally unroll, does it get too big? 584 // Allow the unrolled mess to get larger than standard loop 585 // size. After all, it will no longer be a loop. 586 uint body_size = _body.size(); 587 uint unroll_limit = (uint)LoopUnrollLimit * 4; 588 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 589 if (trip_count > unroll_limit || body_size > unroll_limit) { 590 return false; 591 } 592 593 // Fully unroll a loop with few iterations regardless next 594 // conditions since following loop optimizations will split 595 // such loop anyway (pre-main-post). 596 if (trip_count <= 3) 597 return true; 598 599 // Take into account that after unroll conjoined heads and tails will fold, 600 // otherwise policy_unroll() may allow more unrolling than max unrolling. 601 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; 602 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; 603 if (body_size != tst_body_size) // Check for int overflow 604 return false; 605 if (new_body_size > unroll_limit || 606 // Unrolling can result in a large amount of node construction 607 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) { 608 return false; 609 } 610 611 // Do not unroll a loop with String intrinsics code. 612 // String intrinsics are large and have loops. 613 for (uint k = 0; k < _body.size(); k++) { 614 Node* n = _body.at(k); 615 switch (n->Opcode()) { 616 case Op_StrComp: 617 case Op_StrEquals: 618 case Op_StrIndexOf: 619 case Op_EncodeISOArray: 620 case Op_AryEq: { 621 return false; 622 } 623 #if INCLUDE_RTM_OPT 624 case Op_FastLock: 625 case Op_FastUnlock: { 626 // Don't unroll RTM locking code because it is large. 627 if (UseRTMLocking) { 628 return false; 629 } 630 } 631 #endif 632 } // switch 633 } 634 635 return true; // Do maximally unroll 636 } 637 638 639 //------------------------------policy_unroll---------------------------------- 640 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 641 // the loop is a CountedLoop and the body is small enough. 642 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { 643 644 CountedLoopNode *cl = _head->as_CountedLoop(); 645 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 646 647 if (!cl->is_valid_counted_loop()) 648 return false; // Malformed counted loop 649 650 // Protect against over-unrolling. 651 // After split at least one iteration will be executed in pre-loop. 652 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; 653 654 int future_unroll_ct = cl->unrolled_count() * 2; 655 if (future_unroll_ct > LoopMaxUnroll) return false; 656 657 // Check for initial stride being a small enough constant 658 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; 659 660 // Don't unroll if the next round of unrolling would push us 661 // over the expected trip count of the loop. One is subtracted 662 // from the expected trip count because the pre-loop normally 663 // executes 1 iteration. 664 if (UnrollLimitForProfileCheck > 0 && 665 cl->profile_trip_cnt() != COUNT_UNKNOWN && 666 future_unroll_ct > UnrollLimitForProfileCheck && 667 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 668 return false; 669 } 670 671 // When unroll count is greater than LoopUnrollMin, don't unroll if: 672 // the residual iterations are more than 10% of the trip count 673 // and rounds of "unroll,optimize" are not making significant progress 674 // Progress defined as current size less than 20% larger than previous size. 675 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 676 future_unroll_ct > LoopUnrollMin && 677 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 678 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 679 return false; 680 } 681 682 Node *init_n = cl->init_trip(); 683 Node *limit_n = cl->limit(); 684 int stride_con = cl->stride_con(); 685 // Non-constant bounds. 686 // Protect against over-unrolling when init or/and limit are not constant 687 // (so that trip_count's init value is maxint) but iv range is known. 688 if (init_n == NULL || !init_n->is_Con() || 689 limit_n == NULL || !limit_n->is_Con()) { 690 Node* phi = cl->phi(); 691 if (phi != NULL) { 692 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 693 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 694 int next_stride = stride_con * 2; // stride after this unroll 695 if (next_stride > 0) { 696 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 697 iv_type->_lo + next_stride > iv_type->_hi) { 698 return false; // over-unrolling 699 } 700 } else if (next_stride < 0) { 701 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 702 iv_type->_hi + next_stride < iv_type->_lo) { 703 return false; // over-unrolling 704 } 705 } 706 } 707 } 708 709 // After unroll limit will be adjusted: new_limit = limit-stride. 710 // Bailout if adjustment overflow. 711 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 712 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || 713 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) 714 return false; // overflow 715 716 // Adjust body_size to determine if we unroll or not 717 uint body_size = _body.size(); 718 // Key test to unroll loop in CRC32 java code 719 int xors_in_loop = 0; 720 // Also count ModL, DivL and MulL which expand mightly 721 for (uint k = 0; k < _body.size(); k++) { 722 Node* n = _body.at(k); 723 switch (n->Opcode()) { 724 case Op_XorI: xors_in_loop++; break; // CRC32 java code 725 case Op_ModL: body_size += 30; break; 726 case Op_DivL: body_size += 30; break; 727 case Op_MulL: body_size += 10; break; 728 case Op_StrComp: 729 case Op_StrEquals: 730 case Op_StrIndexOf: 731 case Op_EncodeISOArray: 732 case Op_AryEq: { 733 // Do not unroll a loop with String intrinsics code. 734 // String intrinsics are large and have loops. 735 return false; 736 } 737 #if INCLUDE_RTM_OPT 738 case Op_FastLock: 739 case Op_FastUnlock: { 740 // Don't unroll RTM locking code because it is large. 741 if (UseRTMLocking) { 742 return false; 743 } 744 } 745 #endif 746 } // switch 747 } 748 749 // Check for being too big 750 if (body_size > (uint)LoopUnrollLimit) { 751 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 752 // Normal case: loop too big 753 return false; 754 } 755 756 // Unroll once! (Each trip will soon do double iterations) 757 return true; 758 } 759 760 //------------------------------policy_align----------------------------------- 761 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 762 // expression that does the alignment. Note that only one array base can be 763 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 764 // if we vectorize short memory ops into longer memory ops, we may want to 765 // increase alignment. 766 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 767 return false; 768 } 769 770 //------------------------------policy_range_check----------------------------- 771 // Return TRUE or FALSE if the loop should be range-check-eliminated. 772 // Actually we do iteration-splitting, a more powerful form of RCE. 773 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 774 if (!RangeCheckElimination) return false; 775 776 CountedLoopNode *cl = _head->as_CountedLoop(); 777 // If we unrolled with no intention of doing RCE and we later 778 // changed our minds, we got no pre-loop. Either we need to 779 // make a new pre-loop, or we gotta disallow RCE. 780 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 781 Node *trip_counter = cl->phi(); 782 783 // Check loop body for tests of trip-counter plus loop-invariant vs 784 // loop-invariant. 785 for (uint i = 0; i < _body.size(); i++) { 786 Node *iff = _body[i]; 787 if (iff->Opcode() == Op_If) { // Test? 788 789 // Comparing trip+off vs limit 790 Node *bol = iff->in(1); 791 if (bol->req() != 2) continue; // dead constant test 792 if (!bol->is_Bool()) { 793 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 794 continue; 795 } 796 if (bol->as_Bool()->_test._test == BoolTest::ne) 797 continue; // not RC 798 799 Node *cmp = bol->in(1); 800 Node *rc_exp = cmp->in(1); 801 Node *limit = cmp->in(2); 802 803 Node *limit_c = phase->get_ctrl(limit); 804 if( limit_c == phase->C->top() ) 805 return false; // Found dead test on live IF? No RCE! 806 if( is_member(phase->get_loop(limit_c) ) ) { 807 // Compare might have operands swapped; commute them 808 rc_exp = cmp->in(2); 809 limit = cmp->in(1); 810 limit_c = phase->get_ctrl(limit); 811 if( is_member(phase->get_loop(limit_c) ) ) 812 continue; // Both inputs are loop varying; cannot RCE 813 } 814 815 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 816 continue; 817 } 818 // Yeah! Found a test like 'trip+off vs limit' 819 // Test is an IfNode, has 2 projections. If BOTH are in the loop 820 // we need loop unswitching instead of iteration splitting. 821 if( is_loop_exit(iff) ) 822 return true; // Found reason to split iterations 823 } // End of is IF 824 } 825 826 return false; 827 } 828 829 //------------------------------policy_peel_only------------------------------- 830 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 831 // for unrolling loops with NO array accesses. 832 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 833 834 for( uint i = 0; i < _body.size(); i++ ) 835 if( _body[i]->is_Mem() ) 836 return false; 837 838 // No memory accesses at all! 839 return true; 840 } 841 842 //------------------------------clone_up_backedge_goo-------------------------- 843 // If Node n lives in the back_ctrl block and cannot float, we clone a private 844 // version of n in preheader_ctrl block and return that, otherwise return n. 845 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { 846 if( get_ctrl(n) != back_ctrl ) return n; 847 848 // Only visit once 849 if (visited.test_set(n->_idx)) { 850 Node *x = clones.find(n->_idx); 851 if (x != NULL) 852 return x; 853 return n; 854 } 855 856 Node *x = NULL; // If required, a clone of 'n' 857 // Check for 'n' being pinned in the backedge. 858 if( n->in(0) && n->in(0) == back_ctrl ) { 859 assert(clones.find(n->_idx) == NULL, "dead loop"); 860 x = n->clone(); // Clone a copy of 'n' to preheader 861 clones.push(x, n->_idx); 862 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 863 } 864 865 // Recursive fixup any other input edges into x. 866 // If there are no changes we can just return 'n', otherwise 867 // we need to clone a private copy and change it. 868 for( uint i = 1; i < n->req(); i++ ) { 869 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); 870 if( g != n->in(i) ) { 871 if( !x ) { 872 assert(clones.find(n->_idx) == NULL, "dead loop"); 873 x = n->clone(); 874 clones.push(x, n->_idx); 875 } 876 x->set_req(i, g); 877 } 878 } 879 if( x ) { // x can legally float to pre-header location 880 register_new_node( x, preheader_ctrl ); 881 return x; 882 } else { // raise n to cover LCA of uses 883 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 884 } 885 return n; 886 } 887 888 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { 889 Node* castii = new CastIINode(incr, TypeInt::INT, true); 890 castii->set_req(0, ctrl); 891 register_new_node(castii, ctrl); 892 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { 893 Node* n = incr->fast_out(i); 894 if (n->is_Phi() && n->in(0) == loop) { 895 int nrep = n->replace_edge(incr, castii); 896 return true; 897 } 898 } 899 return false; 900 } 901 902 //------------------------------insert_pre_post_loops-------------------------- 903 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 904 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 905 // alignment. Useful to unroll loops that do no array accesses. 906 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 907 908 #ifndef PRODUCT 909 if (TraceLoopOpts) { 910 if (peel_only) 911 tty->print("PeelMainPost "); 912 else 913 tty->print("PreMainPost "); 914 loop->dump_head(); 915 } 916 #endif 917 C->set_major_progress(); 918 919 // Find common pieces of the loop being guarded with pre & post loops 920 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 921 assert( main_head->is_normal_loop(), "" ); 922 CountedLoopEndNode *main_end = main_head->loopexit(); 923 guarantee(main_end != NULL, "no loop exit node"); 924 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 925 uint dd_main_head = dom_depth(main_head); 926 uint max = main_head->outcnt(); 927 928 Node *pre_header= main_head->in(LoopNode::EntryControl); 929 Node *init = main_head->init_trip(); 930 Node *incr = main_end ->incr(); 931 Node *limit = main_end ->limit(); 932 Node *stride = main_end ->stride(); 933 Node *cmp = main_end ->cmp_node(); 934 BoolTest::mask b_test = main_end->test_trip(); 935 936 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 937 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 938 if( bol->outcnt() != 1 ) { 939 bol = bol->clone(); 940 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 941 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol); 942 } 943 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 944 if( cmp->outcnt() != 1 ) { 945 cmp = cmp->clone(); 946 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 947 _igvn.replace_input_of(bol, 1, cmp); 948 } 949 950 //------------------------------ 951 // Step A: Create Post-Loop. 952 Node* main_exit = main_end->proj_out(false); 953 assert( main_exit->Opcode() == Op_IfFalse, "" ); 954 int dd_main_exit = dom_depth(main_exit); 955 956 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 957 // loop pre-header illegally has 2 control users (old & new loops). 958 clone_loop( loop, old_new, dd_main_exit ); 959 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 960 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 961 post_head->set_post_loop(main_head); 962 963 // Reduce the post-loop trip count. 964 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 965 post_end->_prob = PROB_FAIR; 966 967 // Build the main-loop normal exit. 968 IfFalseNode *new_main_exit = new IfFalseNode(main_end); 969 _igvn.register_new_node_with_optimizer( new_main_exit ); 970 set_idom(new_main_exit, main_end, dd_main_exit ); 971 set_loop(new_main_exit, loop->_parent); 972 973 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 974 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 975 // (the main-loop trip-counter exit value) because we will be changing 976 // the exit value (via unrolling) so we cannot constant-fold away the zero 977 // trip guard until all unrolling is done. 978 Node *zer_opaq = new Opaque1Node(C, incr); 979 Node *zer_cmp = new CmpINode( zer_opaq, limit ); 980 Node *zer_bol = new BoolNode( zer_cmp, b_test ); 981 register_new_node( zer_opaq, new_main_exit ); 982 register_new_node( zer_cmp , new_main_exit ); 983 register_new_node( zer_bol , new_main_exit ); 984 985 // Build the IfNode 986 IfNode *zer_iff = new IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 987 _igvn.register_new_node_with_optimizer( zer_iff ); 988 set_idom(zer_iff, new_main_exit, dd_main_exit); 989 set_loop(zer_iff, loop->_parent); 990 991 // Plug in the false-path, taken if we need to skip post-loop 992 _igvn.replace_input_of(main_exit, 0, zer_iff); 993 set_idom(main_exit, zer_iff, dd_main_exit); 994 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 995 // Make the true-path, must enter the post loop 996 Node *zer_taken = new IfTrueNode( zer_iff ); 997 _igvn.register_new_node_with_optimizer( zer_taken ); 998 set_idom(zer_taken, zer_iff, dd_main_exit); 999 set_loop(zer_taken, loop->_parent); 1000 // Plug in the true path 1001 _igvn.hash_delete( post_head ); 1002 post_head->set_req(LoopNode::EntryControl, zer_taken); 1003 set_idom(post_head, zer_taken, dd_main_exit); 1004 1005 Arena *a = Thread::current()->resource_area(); 1006 VectorSet visited(a); 1007 Node_Stack clones(a, main_head->back_control()->outcnt()); 1008 // Step A3: Make the fall-in values to the post-loop come from the 1009 // fall-out values of the main-loop. 1010 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 1011 Node* main_phi = main_head->fast_out(i); 1012 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 1013 Node *post_phi = old_new[main_phi->_idx]; 1014 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 1015 post_head->init_control(), 1016 main_phi->in(LoopNode::LoopBackControl), 1017 visited, clones); 1018 _igvn.hash_delete(post_phi); 1019 post_phi->set_req( LoopNode::EntryControl, fallmain ); 1020 } 1021 } 1022 1023 // Update local caches for next stanza 1024 main_exit = new_main_exit; 1025 1026 1027 //------------------------------ 1028 // Step B: Create Pre-Loop. 1029 1030 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 1031 // loop pre-header illegally has 2 control users (old & new loops). 1032 clone_loop( loop, old_new, dd_main_head ); 1033 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1034 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1035 pre_head->set_pre_loop(main_head); 1036 Node *pre_incr = old_new[incr->_idx]; 1037 1038 // Reduce the pre-loop trip count. 1039 pre_end->_prob = PROB_FAIR; 1040 1041 // Find the pre-loop normal exit. 1042 Node* pre_exit = pre_end->proj_out(false); 1043 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 1044 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end); 1045 _igvn.register_new_node_with_optimizer( new_pre_exit ); 1046 set_idom(new_pre_exit, pre_end, dd_main_head); 1047 set_loop(new_pre_exit, loop->_parent); 1048 1049 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1050 // pre-loop, the main-loop may not execute at all. Later in life this 1051 // zero-trip guard will become the minimum-trip guard when we unroll 1052 // the main-loop. 1053 Node *min_opaq = new Opaque1Node(C, limit); 1054 Node *min_cmp = new CmpINode( pre_incr, min_opaq ); 1055 Node *min_bol = new BoolNode( min_cmp, b_test ); 1056 register_new_node( min_opaq, new_pre_exit ); 1057 register_new_node( min_cmp , new_pre_exit ); 1058 register_new_node( min_bol , new_pre_exit ); 1059 1060 // Build the IfNode (assume the main-loop is executed always). 1061 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1062 _igvn.register_new_node_with_optimizer( min_iff ); 1063 set_idom(min_iff, new_pre_exit, dd_main_head); 1064 set_loop(min_iff, loop->_parent); 1065 1066 // Plug in the false-path, taken if we need to skip main-loop 1067 _igvn.hash_delete( pre_exit ); 1068 pre_exit->set_req(0, min_iff); 1069 set_idom(pre_exit, min_iff, dd_main_head); 1070 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1071 // Make the true-path, must enter the main loop 1072 Node *min_taken = new IfTrueNode( min_iff ); 1073 _igvn.register_new_node_with_optimizer( min_taken ); 1074 set_idom(min_taken, min_iff, dd_main_head); 1075 set_loop(min_taken, loop->_parent); 1076 // Plug in the true path 1077 _igvn.hash_delete( main_head ); 1078 main_head->set_req(LoopNode::EntryControl, min_taken); 1079 set_idom(main_head, min_taken, dd_main_head); 1080 1081 visited.Clear(); 1082 clones.clear(); 1083 // Step B3: Make the fall-in values to the main-loop come from the 1084 // fall-out values of the pre-loop. 1085 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1086 Node* main_phi = main_head->fast_out(i2); 1087 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1088 Node *pre_phi = old_new[main_phi->_idx]; 1089 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1090 main_head->init_control(), 1091 pre_phi->in(LoopNode::LoopBackControl), 1092 visited, clones); 1093 _igvn.hash_delete(main_phi); 1094 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1095 } 1096 } 1097 1098 // Nodes inside the loop may be control dependent on a predicate 1099 // that was moved before the preloop. If the back branch of the main 1100 // or post loops becomes dead, those nodes won't be dependent on the 1101 // test that guards that loop nest anymore which could lead to an 1102 // incorrect array access because it executes independently of the 1103 // test that was guarding the loop nest. We add a special CastII on 1104 // the if branch that enters the loop, between the input induction 1105 // variable value and the induction variable Phi to preserve correct 1106 // dependencies. 1107 1108 // CastII for the post loop: 1109 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); 1110 assert(inserted, "no castII inserted"); 1111 1112 // CastII for the main loop: 1113 inserted = cast_incr_before_loop(pre_incr, min_taken, main_head); 1114 assert(inserted, "no castII inserted"); 1115 1116 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1117 // RCE and alignment may change this later. 1118 Node *cmp_end = pre_end->cmp_node(); 1119 assert( cmp_end->in(2) == limit, "" ); 1120 Node *pre_limit = new AddINode( init, stride ); 1121 1122 // Save the original loop limit in this Opaque1 node for 1123 // use by range check elimination. 1124 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit); 1125 1126 register_new_node( pre_limit, pre_head->in(0) ); 1127 register_new_node( pre_opaq , pre_head->in(0) ); 1128 1129 // Since no other users of pre-loop compare, I can hack limit directly 1130 assert( cmp_end->outcnt() == 1, "no other users" ); 1131 _igvn.hash_delete(cmp_end); 1132 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1133 1134 // Special case for not-equal loop bounds: 1135 // Change pre loop test, main loop test, and the 1136 // main loop guard test to use lt or gt depending on stride 1137 // direction: 1138 // positive stride use < 1139 // negative stride use > 1140 // 1141 // not-equal test is kept for post loop to handle case 1142 // when init > limit when stride > 0 (and reverse). 1143 1144 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1145 1146 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1147 // Modify pre loop end condition 1148 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1149 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test); 1150 register_new_node( new_bol0, pre_head->in(0) ); 1151 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0); 1152 // Modify main loop guard condition 1153 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1154 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test); 1155 register_new_node( new_bol1, new_pre_exit ); 1156 _igvn.hash_delete(min_iff); 1157 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1158 // Modify main loop end condition 1159 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1160 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test); 1161 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1162 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2); 1163 } 1164 1165 // Flag main loop 1166 main_head->set_main_loop(); 1167 if( peel_only ) main_head->set_main_no_pre_loop(); 1168 1169 // Subtract a trip count for the pre-loop. 1170 main_head->set_trip_count(main_head->trip_count() - 1); 1171 1172 // It's difficult to be precise about the trip-counts 1173 // for the pre/post loops. They are usually very short, 1174 // so guess that 4 trips is a reasonable value. 1175 post_head->set_profile_trip_cnt(4.0); 1176 pre_head->set_profile_trip_cnt(4.0); 1177 1178 // Now force out all loop-invariant dominating tests. The optimizer 1179 // finds some, but we _know_ they are all useless. 1180 peeled_dom_test_elim(loop,old_new); 1181 loop->record_for_igvn(); 1182 } 1183 1184 //------------------------------is_invariant----------------------------- 1185 // Return true if n is invariant 1186 bool IdealLoopTree::is_invariant(Node* n) const { 1187 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1188 if (n_c->is_top()) return false; 1189 return !is_member(_phase->get_loop(n_c)); 1190 } 1191 1192 1193 //------------------------------do_unroll-------------------------------------- 1194 // Unroll the loop body one step - make each trip do 2 iterations. 1195 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1196 assert(LoopUnrollLimit, ""); 1197 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1198 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1199 assert(loop_end, ""); 1200 #ifndef PRODUCT 1201 if (PrintOpto && VerifyLoopOptimizations) { 1202 tty->print("Unrolling "); 1203 loop->dump_head(); 1204 } else if (TraceLoopOpts) { 1205 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1206 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1207 } else { 1208 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1209 } 1210 loop->dump_head(); 1211 } 1212 #endif 1213 1214 // Remember loop node count before unrolling to detect 1215 // if rounds of unroll,optimize are making progress 1216 loop_head->set_node_count_before_unroll(loop->_body.size()); 1217 1218 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1219 Node *limit = loop_head->limit(); 1220 Node *init = loop_head->init_trip(); 1221 Node *stride = loop_head->stride(); 1222 1223 Node *opaq = NULL; 1224 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1225 // Search for zero-trip guard. 1226 assert( loop_head->is_main_loop(), "" ); 1227 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1228 Node *iff = ctrl->in(0); 1229 assert( iff->Opcode() == Op_If, "" ); 1230 Node *bol = iff->in(1); 1231 assert( bol->Opcode() == Op_Bool, "" ); 1232 Node *cmp = bol->in(1); 1233 assert( cmp->Opcode() == Op_CmpI, "" ); 1234 opaq = cmp->in(2); 1235 // Occasionally it's possible for a zero-trip guard Opaque1 node to be 1236 // optimized away and then another round of loop opts attempted. 1237 // We can not optimize this particular loop in that case. 1238 if (opaq->Opcode() != Op_Opaque1) 1239 return; // Cannot find zero-trip guard! Bail out! 1240 // Zero-trip test uses an 'opaque' node which is not shared. 1241 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1242 } 1243 1244 C->set_major_progress(); 1245 1246 Node* new_limit = NULL; 1247 if (UnrollLimitCheck) { 1248 int stride_con = stride->get_int(); 1249 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1250 uint old_trip_count = loop_head->trip_count(); 1251 // Verify that unroll policy result is still valid. 1252 assert(old_trip_count > 1 && 1253 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1254 1255 // Adjust loop limit to keep valid iterations number after unroll. 1256 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1257 // which may overflow. 1258 if (!adjust_min_trip) { 1259 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1260 "odd trip count for maximally unroll"); 1261 // Don't need to adjust limit for maximally unroll since trip count is even. 1262 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1263 // Loop's limit is constant. Loop's init could be constant when pre-loop 1264 // become peeled iteration. 1265 jlong init_con = init->get_int(); 1266 // We can keep old loop limit if iterations count stays the same: 1267 // old_trip_count == new_trip_count * 2 1268 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1269 // so we also don't need to adjust zero trip test. 1270 jlong limit_con = limit->get_int(); 1271 // (stride_con*2) not overflow since stride_con <= 8. 1272 int new_stride_con = stride_con * 2; 1273 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1274 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1275 // New trip count should satisfy next conditions. 1276 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1277 uint new_trip_count = (uint)trip_count; 1278 adjust_min_trip = (old_trip_count != new_trip_count*2); 1279 } 1280 1281 if (adjust_min_trip) { 1282 // Step 2: Adjust the trip limit if it is called for. 1283 // The adjustment amount is -stride. Need to make sure if the 1284 // adjustment underflows or overflows, then the main loop is skipped. 1285 Node* cmp = loop_end->cmp_node(); 1286 assert(cmp->in(2) == limit, "sanity"); 1287 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1288 1289 // Verify that policy_unroll result is still valid. 1290 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1291 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1292 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); 1293 1294 if (limit->is_Con()) { 1295 // The check in policy_unroll and the assert above guarantee 1296 // no underflow if limit is constant. 1297 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1298 set_ctrl(new_limit, C->root()); 1299 } else { 1300 // Limit is not constant. 1301 if (loop_head->unrolled_count() == 1) { // only for first unroll 1302 // Separate limit by Opaque node in case it is an incremented 1303 // variable from previous loop to avoid using pre-incremented 1304 // value which could increase register pressure. 1305 // Otherwise reorg_offsets() optimization will create a separate 1306 // Opaque node for each use of trip-counter and as result 1307 // zero trip guard limit will be different from loop limit. 1308 assert(has_ctrl(opaq), "should have it"); 1309 Node* opaq_ctrl = get_ctrl(opaq); 1310 limit = new Opaque2Node( C, limit ); 1311 register_new_node( limit, opaq_ctrl ); 1312 } 1313 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) || 1314 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) { 1315 // No underflow. 1316 new_limit = new SubINode(limit, stride); 1317 } else { 1318 // (limit - stride) may underflow. 1319 // Clamp the adjustment value with MININT or MAXINT: 1320 // 1321 // new_limit = limit-stride 1322 // if (stride > 0) 1323 // new_limit = (limit < new_limit) ? MININT : new_limit; 1324 // else 1325 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1326 // 1327 BoolTest::mask bt = loop_end->test_trip(); 1328 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1329 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1330 set_ctrl(adj_max, C->root()); 1331 Node* old_limit = NULL; 1332 Node* adj_limit = NULL; 1333 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1334 if (loop_head->unrolled_count() > 1 && 1335 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1336 limit->in(CMoveNode::IfTrue) == adj_max && 1337 bol->as_Bool()->_test._test == bt && 1338 bol->in(1)->Opcode() == Op_CmpI && 1339 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1340 // Loop was unrolled before. 1341 // Optimize the limit to avoid nested CMove: 1342 // use original limit as old limit. 1343 old_limit = bol->in(1)->in(1); 1344 // Adjust previous adjusted limit. 1345 adj_limit = limit->in(CMoveNode::IfFalse); 1346 adj_limit = new SubINode(adj_limit, stride); 1347 } else { 1348 old_limit = limit; 1349 adj_limit = new SubINode(limit, stride); 1350 } 1351 assert(old_limit != NULL && adj_limit != NULL, ""); 1352 register_new_node( adj_limit, ctrl ); // adjust amount 1353 Node* adj_cmp = new CmpINode(old_limit, adj_limit); 1354 register_new_node( adj_cmp, ctrl ); 1355 Node* adj_bool = new BoolNode(adj_cmp, bt); 1356 register_new_node( adj_bool, ctrl ); 1357 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1358 } 1359 register_new_node(new_limit, ctrl); 1360 } 1361 assert(new_limit != NULL, ""); 1362 // Replace in loop test. 1363 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1364 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1365 // Don't need to create new test since only one user. 1366 _igvn.hash_delete(cmp); 1367 cmp->set_req(2, new_limit); 1368 } else { 1369 // Create new test since it is shared. 1370 Node* ctrl2 = loop_end->in(0); 1371 Node* cmp2 = cmp->clone(); 1372 cmp2->set_req(2, new_limit); 1373 register_new_node(cmp2, ctrl2); 1374 Node* bol2 = loop_end->in(1)->clone(); 1375 bol2->set_req(1, cmp2); 1376 register_new_node(bol2, ctrl2); 1377 _igvn.replace_input_of(loop_end, 1, bol2); 1378 } 1379 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1380 // Make it a 1-trip test (means at least 2 trips). 1381 1382 // Guard test uses an 'opaque' node which is not shared. Hence I 1383 // can edit it's inputs directly. Hammer in the new limit for the 1384 // minimum-trip guard. 1385 assert(opaq->outcnt() == 1, ""); 1386 _igvn.replace_input_of(opaq, 1, new_limit); 1387 } 1388 1389 // Adjust max trip count. The trip count is intentionally rounded 1390 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1391 // the main, unrolled, part of the loop will never execute as it is protected 1392 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1393 // and later determined that part of the unrolled loop was dead. 1394 loop_head->set_trip_count(old_trip_count / 2); 1395 1396 // Double the count of original iterations in the unrolled loop body. 1397 loop_head->double_unrolled_count(); 1398 1399 } else { // LoopLimitCheck 1400 1401 // Adjust max trip count. The trip count is intentionally rounded 1402 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1403 // the main, unrolled, part of the loop will never execute as it is protected 1404 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1405 // and later determined that part of the unrolled loop was dead. 1406 loop_head->set_trip_count(loop_head->trip_count() / 2); 1407 1408 // Double the count of original iterations in the unrolled loop body. 1409 loop_head->double_unrolled_count(); 1410 1411 // ----------- 1412 // Step 2: Cut back the trip counter for an unroll amount of 2. 1413 // Loop will normally trip (limit - init)/stride_con. Since it's a 1414 // CountedLoop this is exact (stride divides limit-init exactly). 1415 // We are going to double the loop body, so we want to knock off any 1416 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 1417 Node *span = new SubINode( limit, init ); 1418 register_new_node( span, ctrl ); 1419 Node *trip = new DivINode( 0, span, stride ); 1420 register_new_node( trip, ctrl ); 1421 Node *mtwo = _igvn.intcon(-2); 1422 set_ctrl(mtwo, C->root()); 1423 Node *rond = new AndINode( trip, mtwo ); 1424 register_new_node( rond, ctrl ); 1425 Node *spn2 = new MulINode( rond, stride ); 1426 register_new_node( spn2, ctrl ); 1427 new_limit = new AddINode( spn2, init ); 1428 register_new_node( new_limit, ctrl ); 1429 1430 // Hammer in the new limit 1431 Node *ctrl2 = loop_end->in(0); 1432 Node *cmp2 = new CmpINode( loop_head->incr(), new_limit ); 1433 register_new_node( cmp2, ctrl2 ); 1434 Node *bol2 = new BoolNode( cmp2, loop_end->test_trip() ); 1435 register_new_node( bol2, ctrl2 ); 1436 _igvn.replace_input_of(loop_end, CountedLoopEndNode::TestValue, bol2); 1437 1438 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1439 // Make it a 1-trip test (means at least 2 trips). 1440 if( adjust_min_trip ) { 1441 assert( new_limit != NULL, "" ); 1442 // Guard test uses an 'opaque' node which is not shared. Hence I 1443 // can edit it's inputs directly. Hammer in the new limit for the 1444 // minimum-trip guard. 1445 assert( opaq->outcnt() == 1, "" ); 1446 _igvn.hash_delete(opaq); 1447 opaq->set_req(1, new_limit); 1448 } 1449 } // LoopLimitCheck 1450 1451 // --------- 1452 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1453 // represents the odd iterations; since the loop trips an even number of 1454 // times its backedge is never taken. Kill the backedge. 1455 uint dd = dom_depth(loop_head); 1456 clone_loop( loop, old_new, dd ); 1457 1458 // Make backedges of the clone equal to backedges of the original. 1459 // Make the fall-in from the original come from the fall-out of the clone. 1460 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1461 Node* phi = loop_head->fast_out(j); 1462 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1463 Node *newphi = old_new[phi->_idx]; 1464 _igvn.hash_delete( phi ); 1465 _igvn.hash_delete( newphi ); 1466 1467 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1468 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1469 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1470 } 1471 } 1472 Node *clone_head = old_new[loop_head->_idx]; 1473 _igvn.hash_delete( clone_head ); 1474 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1475 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1476 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1477 loop->_head = clone_head; // New loop header 1478 1479 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1480 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1481 1482 // Kill the clone's backedge 1483 Node *newcle = old_new[loop_end->_idx]; 1484 _igvn.hash_delete( newcle ); 1485 Node *one = _igvn.intcon(1); 1486 set_ctrl(one, C->root()); 1487 newcle->set_req(1, one); 1488 // Force clone into same loop body 1489 uint max = loop->_body.size(); 1490 for( uint k = 0; k < max; k++ ) { 1491 Node *old = loop->_body.at(k); 1492 Node *nnn = old_new[old->_idx]; 1493 loop->_body.push(nnn); 1494 if (!has_ctrl(old)) 1495 set_loop(nnn, loop); 1496 } 1497 1498 loop->record_for_igvn(); 1499 } 1500 1501 //------------------------------do_maximally_unroll---------------------------- 1502 1503 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1504 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1505 assert(cl->has_exact_trip_count(), "trip count is not exact"); 1506 assert(cl->trip_count() > 0, ""); 1507 #ifndef PRODUCT 1508 if (TraceLoopOpts) { 1509 tty->print("MaxUnroll %d ", cl->trip_count()); 1510 loop->dump_head(); 1511 } 1512 #endif 1513 1514 // If loop is tripping an odd number of times, peel odd iteration 1515 if ((cl->trip_count() & 1) == 1) { 1516 do_peeling(loop, old_new); 1517 } 1518 1519 // Now its tripping an even number of times remaining. Double loop body. 1520 // Do not adjust pre-guards; they are not needed and do not exist. 1521 if (cl->trip_count() > 0) { 1522 assert((cl->trip_count() & 1) == 0, "missed peeling"); 1523 do_unroll(loop, old_new, false); 1524 } 1525 } 1526 1527 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) { 1528 if (SuperWordReductions == false) return; 1529 1530 for (uint i = 0; i < loop->_body.size(); i++) { 1531 Node *n = loop->_body[i]; 1532 if (n != NULL && n->is_Phi()) { 1533 PhiNode *phi = n->as_Phi(); 1534 // for definitions which are loop inclusive and not tripcounts 1535 if (phi != NULL && phi->is_tripcount() == false) { 1536 uint len = n->len(); 1537 Node *defNode = n->in(len - 1); 1538 1539 if (defNode != NULL) { 1540 Node *n_ctrl = get_ctrl(defNode); 1541 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) { 1542 uint flags = defNode->flags(); 1543 1544 // This is a loop carried depedence, set if needed 1545 if ((flags & Node::Flag_is_loop_carried_dep) != Node::Flag_is_loop_carried_dep) 1546 defNode->add_flag(Node::Flag_is_loop_carried_dep); 1547 1548 // now test it to see if it fits the standard pattern for a reduction operator 1549 switch (defNode->Opcode()) { 1550 // the communitative operators 1551 case Op_AddI: 1552 case Op_AddL: 1553 case Op_AddF: 1554 case Op_AddD: 1555 case Op_MulI: 1556 case Op_MulL: 1557 case Op_MulF: 1558 case Op_MulD: 1559 if ((flags & Node::Flag_has_reduction) != Node::Flag_has_reduction) 1560 { 1561 // To be a reduction, the arithmetic node must source the phi and provide a def to it 1562 int end = defNode->req(); 1563 for (int j = 1; j < end; j++) { 1564 Node* in = defNode->in(j); 1565 if (in == n){ 1566 defNode->add_flag(Node::Flag_has_reduction); 1567 break; 1568 } 1569 } 1570 } 1571 break; 1572 1573 // the non communitative operators 1574 case Op_DivI: 1575 case Op_DivL: 1576 case Op_DivF: 1577 case Op_DivD: 1578 case Op_SubI: 1579 case Op_SubL: 1580 case Op_SubF: 1581 case Op_SubD: 1582 // TODO: these need ordering constraints 1583 break; 1584 1585 default: 1586 break; 1587 } 1588 } 1589 } 1590 } 1591 } 1592 } 1593 } 1594 1595 //------------------------------dominates_backedge--------------------------------- 1596 // Returns true if ctrl is executed on every complete iteration 1597 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1598 assert(ctrl->is_CFG(), "must be control"); 1599 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1600 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1601 } 1602 1603 //------------------------------adjust_limit----------------------------------- 1604 // Helper function for add_constraint(). 1605 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { 1606 // Compute "I :: (limit-offset)/scale" 1607 Node *con = new SubINode(rc_limit, offset); 1608 register_new_node(con, pre_ctrl); 1609 Node *X = new DivINode(0, con, scale); 1610 register_new_node(X, pre_ctrl); 1611 1612 // Adjust loop limit 1613 loop_limit = (stride_con > 0) 1614 ? (Node*)(new MinINode(loop_limit, X)) 1615 : (Node*)(new MaxINode(loop_limit, X)); 1616 register_new_node(loop_limit, pre_ctrl); 1617 return loop_limit; 1618 } 1619 1620 //------------------------------add_constraint--------------------------------- 1621 // Constrain the main loop iterations so the conditions: 1622 // low_limit <= scale_con * I + offset < upper_limit 1623 // always holds true. That is, either increase the number of iterations in 1624 // the pre-loop or the post-loop until the condition holds true in the main 1625 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1626 // stride and scale are constants (offset and limit often are). 1627 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 ) { 1628 // For positive stride, the pre-loop limit always uses a MAX function 1629 // and the main loop a MIN function. For negative stride these are 1630 // reversed. 1631 1632 // Also for positive stride*scale the affine function is increasing, so the 1633 // pre-loop must check for underflow and the post-loop for overflow. 1634 // Negative stride*scale reverses this; pre-loop checks for overflow and 1635 // post-loop for underflow. 1636 1637 Node *scale = _igvn.intcon(scale_con); 1638 set_ctrl(scale, C->root()); 1639 1640 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 1641 // The overflow limit: scale*I+offset < upper_limit 1642 // For main-loop compute 1643 // ( if (scale > 0) /* and stride > 0 */ 1644 // I < (upper_limit-offset)/scale 1645 // else /* scale < 0 and stride < 0 */ 1646 // I > (upper_limit-offset)/scale 1647 // ) 1648 // 1649 // (upper_limit-offset) may overflow or underflow. 1650 // But it is fine since main loop will either have 1651 // less iterations or will be skipped in such case. 1652 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); 1653 1654 // The underflow limit: low_limit <= scale*I+offset. 1655 // For pre-loop compute 1656 // NOT(scale*I+offset >= low_limit) 1657 // scale*I+offset < low_limit 1658 // ( if (scale > 0) /* and stride > 0 */ 1659 // I < (low_limit-offset)/scale 1660 // else /* scale < 0 and stride < 0 */ 1661 // I > (low_limit-offset)/scale 1662 // ) 1663 1664 if (low_limit->get_int() == -max_jint) { 1665 if (!RangeLimitCheck) return; 1666 // We need this guard when scale*pre_limit+offset >= limit 1667 // due to underflow. So we need execute pre-loop until 1668 // scale*I+offset >= min_int. But (min_int-offset) will 1669 // underflow when offset > 0 and X will be > original_limit 1670 // when stride > 0. To avoid it we replace positive offset with 0. 1671 // 1672 // Also (min_int+1 == -max_int) is used instead of min_int here 1673 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1674 Node* shift = _igvn.intcon(31); 1675 set_ctrl(shift, C->root()); 1676 Node* sign = new RShiftINode(offset, shift); 1677 register_new_node(sign, pre_ctrl); 1678 offset = new AndINode(offset, sign); 1679 register_new_node(offset, pre_ctrl); 1680 } else { 1681 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1682 // The only problem we have here when offset == min_int 1683 // since (0-min_int) == min_int. It may be fine for stride > 0 1684 // but for stride < 0 X will be < original_limit. To avoid it 1685 // max(pre_limit, original_limit) is used in do_range_check(). 1686 } 1687 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1688 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); 1689 1690 } else { // stride_con*scale_con < 0 1691 // For negative stride*scale pre-loop checks for overflow and 1692 // post-loop for underflow. 1693 // 1694 // The overflow limit: scale*I+offset < upper_limit 1695 // For pre-loop compute 1696 // NOT(scale*I+offset < upper_limit) 1697 // scale*I+offset >= upper_limit 1698 // scale*I+offset+1 > upper_limit 1699 // ( if (scale < 0) /* and stride > 0 */ 1700 // I < (upper_limit-(offset+1))/scale 1701 // else /* scale > 0 and stride < 0 */ 1702 // I > (upper_limit-(offset+1))/scale 1703 // ) 1704 // 1705 // (upper_limit-offset-1) may underflow or overflow. 1706 // To avoid it min(pre_limit, original_limit) is used 1707 // in do_range_check() for stride > 0 and max() for < 0. 1708 Node *one = _igvn.intcon(1); 1709 set_ctrl(one, C->root()); 1710 1711 Node *plus_one = new AddINode(offset, one); 1712 register_new_node( plus_one, pre_ctrl ); 1713 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1714 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); 1715 1716 if (low_limit->get_int() == -max_jint) { 1717 if (!RangeLimitCheck) return; 1718 // We need this guard when scale*main_limit+offset >= limit 1719 // due to underflow. So we need execute main-loop while 1720 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 1721 // underflow when (offset+1) > 0 and X will be < main_limit 1722 // when scale < 0 (and stride > 0). To avoid it we replace 1723 // positive (offset+1) with 0. 1724 // 1725 // Also (min_int+1 == -max_int) is used instead of min_int here 1726 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1727 Node* shift = _igvn.intcon(31); 1728 set_ctrl(shift, C->root()); 1729 Node* sign = new RShiftINode(plus_one, shift); 1730 register_new_node(sign, pre_ctrl); 1731 plus_one = new AndINode(plus_one, sign); 1732 register_new_node(plus_one, pre_ctrl); 1733 } else { 1734 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1735 // The only problem we have here when offset == max_int 1736 // since (max_int+1) == min_int and (0-min_int) == min_int. 1737 // But it is fine since main loop will either have 1738 // less iterations or will be skipped in such case. 1739 } 1740 // The underflow limit: low_limit <= scale*I+offset. 1741 // For main-loop compute 1742 // scale*I+offset+1 > low_limit 1743 // ( if (scale < 0) /* and stride > 0 */ 1744 // I < (low_limit-(offset+1))/scale 1745 // else /* scale > 0 and stride < 0 */ 1746 // I > (low_limit-(offset+1))/scale 1747 // ) 1748 1749 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); 1750 } 1751 } 1752 1753 1754 //------------------------------is_scaled_iv--------------------------------- 1755 // Return true if exp is a constant times an induction var 1756 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1757 if (exp == iv) { 1758 if (p_scale != NULL) { 1759 *p_scale = 1; 1760 } 1761 return true; 1762 } 1763 int opc = exp->Opcode(); 1764 if (opc == Op_MulI) { 1765 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1766 if (p_scale != NULL) { 1767 *p_scale = exp->in(2)->get_int(); 1768 } 1769 return true; 1770 } 1771 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1772 if (p_scale != NULL) { 1773 *p_scale = exp->in(1)->get_int(); 1774 } 1775 return true; 1776 } 1777 } else if (opc == Op_LShiftI) { 1778 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1779 if (p_scale != NULL) { 1780 *p_scale = 1 << exp->in(2)->get_int(); 1781 } 1782 return true; 1783 } 1784 } 1785 return false; 1786 } 1787 1788 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1789 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1790 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1791 if (is_scaled_iv(exp, iv, p_scale)) { 1792 if (p_offset != NULL) { 1793 Node *zero = _igvn.intcon(0); 1794 set_ctrl(zero, C->root()); 1795 *p_offset = zero; 1796 } 1797 return true; 1798 } 1799 int opc = exp->Opcode(); 1800 if (opc == Op_AddI) { 1801 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1802 if (p_offset != NULL) { 1803 *p_offset = exp->in(2); 1804 } 1805 return true; 1806 } 1807 if (exp->in(2)->is_Con()) { 1808 Node* offset2 = NULL; 1809 if (depth < 2 && 1810 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1811 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1812 if (p_offset != NULL) { 1813 Node *ctrl_off2 = get_ctrl(offset2); 1814 Node* offset = new AddINode(offset2, exp->in(2)); 1815 register_new_node(offset, ctrl_off2); 1816 *p_offset = offset; 1817 } 1818 return true; 1819 } 1820 } 1821 } else if (opc == Op_SubI) { 1822 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1823 if (p_offset != NULL) { 1824 Node *zero = _igvn.intcon(0); 1825 set_ctrl(zero, C->root()); 1826 Node *ctrl_off = get_ctrl(exp->in(2)); 1827 Node* offset = new SubINode(zero, exp->in(2)); 1828 register_new_node(offset, ctrl_off); 1829 *p_offset = offset; 1830 } 1831 return true; 1832 } 1833 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1834 if (p_offset != NULL) { 1835 *p_scale *= -1; 1836 *p_offset = exp->in(1); 1837 } 1838 return true; 1839 } 1840 } 1841 return false; 1842 } 1843 1844 //------------------------------do_range_check--------------------------------- 1845 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1846 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1847 #ifndef PRODUCT 1848 if (PrintOpto && VerifyLoopOptimizations) { 1849 tty->print("Range Check Elimination "); 1850 loop->dump_head(); 1851 } else if (TraceLoopOpts) { 1852 tty->print("RangeCheck "); 1853 loop->dump_head(); 1854 } 1855 #endif 1856 assert(RangeCheckElimination, ""); 1857 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1858 assert(cl->is_main_loop(), ""); 1859 1860 // protect against stride not being a constant 1861 if (!cl->stride_is_con()) 1862 return; 1863 1864 // Find the trip counter; we are iteration splitting based on it 1865 Node *trip_counter = cl->phi(); 1866 // Find the main loop limit; we will trim it's iterations 1867 // to not ever trip end tests 1868 Node *main_limit = cl->limit(); 1869 1870 // Need to find the main-loop zero-trip guard 1871 Node *ctrl = cl->in(LoopNode::EntryControl); 1872 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 1873 Node *iffm = ctrl->in(0); 1874 assert(iffm->Opcode() == Op_If, ""); 1875 Node *bolzm = iffm->in(1); 1876 assert(bolzm->Opcode() == Op_Bool, ""); 1877 Node *cmpzm = bolzm->in(1); 1878 assert(cmpzm->is_Cmp(), ""); 1879 Node *opqzm = cmpzm->in(2); 1880 // Can not optimize a loop if zero-trip Opaque1 node is optimized 1881 // away and then another round of loop opts attempted. 1882 if (opqzm->Opcode() != Op_Opaque1) 1883 return; 1884 assert(opqzm->in(1) == main_limit, "do not understand situation"); 1885 1886 // Find the pre-loop limit; we will expand it's iterations to 1887 // not ever trip low tests. 1888 Node *p_f = iffm->in(0); 1889 assert(p_f->Opcode() == Op_IfFalse, ""); 1890 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 1891 assert(pre_end->loopnode()->is_pre_loop(), ""); 1892 Node *pre_opaq1 = pre_end->limit(); 1893 // Occasionally it's possible for a pre-loop Opaque1 node to be 1894 // optimized away and then another round of loop opts attempted. 1895 // We can not optimize this particular loop in that case. 1896 if (pre_opaq1->Opcode() != Op_Opaque1) 1897 return; 1898 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 1899 Node *pre_limit = pre_opaq->in(1); 1900 1901 // Where do we put new limit calculations 1902 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 1903 1904 // Ensure the original loop limit is available from the 1905 // pre-loop Opaque1 node. 1906 Node *orig_limit = pre_opaq->original_loop_limit(); 1907 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 1908 return; 1909 1910 // Must know if its a count-up or count-down loop 1911 1912 int stride_con = cl->stride_con(); 1913 Node *zero = _igvn.intcon(0); 1914 Node *one = _igvn.intcon(1); 1915 // Use symmetrical int range [-max_jint,max_jint] 1916 Node *mini = _igvn.intcon(-max_jint); 1917 set_ctrl(zero, C->root()); 1918 set_ctrl(one, C->root()); 1919 set_ctrl(mini, C->root()); 1920 1921 // Range checks that do not dominate the loop backedge (ie. 1922 // conditionally executed) can lengthen the pre loop limit beyond 1923 // the original loop limit. To prevent this, the pre limit is 1924 // (for stride > 0) MINed with the original loop limit (MAXed 1925 // stride < 0) when some range_check (rc) is conditionally 1926 // executed. 1927 bool conditional_rc = false; 1928 1929 // Check loop body for tests of trip-counter plus loop-invariant vs 1930 // loop-invariant. 1931 for( uint i = 0; i < loop->_body.size(); i++ ) { 1932 Node *iff = loop->_body[i]; 1933 if( iff->Opcode() == Op_If ) { // Test? 1934 1935 // Test is an IfNode, has 2 projections. If BOTH are in the loop 1936 // we need loop unswitching instead of iteration splitting. 1937 Node *exit = loop->is_loop_exit(iff); 1938 if( !exit ) continue; 1939 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 1940 1941 // Get boolean condition to test 1942 Node *i1 = iff->in(1); 1943 if( !i1->is_Bool() ) continue; 1944 BoolNode *bol = i1->as_Bool(); 1945 BoolTest b_test = bol->_test; 1946 // Flip sense of test if exit condition is flipped 1947 if( flip ) 1948 b_test = b_test.negate(); 1949 1950 // Get compare 1951 Node *cmp = bol->in(1); 1952 1953 // Look for trip_counter + offset vs limit 1954 Node *rc_exp = cmp->in(1); 1955 Node *limit = cmp->in(2); 1956 jint scale_con= 1; // Assume trip counter not scaled 1957 1958 Node *limit_c = get_ctrl(limit); 1959 if( loop->is_member(get_loop(limit_c) ) ) { 1960 // Compare might have operands swapped; commute them 1961 b_test = b_test.commute(); 1962 rc_exp = cmp->in(2); 1963 limit = cmp->in(1); 1964 limit_c = get_ctrl(limit); 1965 if( loop->is_member(get_loop(limit_c) ) ) 1966 continue; // Both inputs are loop varying; cannot RCE 1967 } 1968 // Here we know 'limit' is loop invariant 1969 1970 // 'limit' maybe pinned below the zero trip test (probably from a 1971 // previous round of rce), in which case, it can't be used in the 1972 // zero trip test expression which must occur before the zero test's if. 1973 if( limit_c == ctrl ) { 1974 continue; // Don't rce this check but continue looking for other candidates. 1975 } 1976 1977 // Check for scaled induction variable plus an offset 1978 Node *offset = NULL; 1979 1980 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 1981 continue; 1982 } 1983 1984 Node *offset_c = get_ctrl(offset); 1985 if( loop->is_member( get_loop(offset_c) ) ) 1986 continue; // Offset is not really loop invariant 1987 // Here we know 'offset' is loop invariant. 1988 1989 // As above for the 'limit', the 'offset' maybe pinned below the 1990 // zero trip test. 1991 if( offset_c == ctrl ) { 1992 continue; // Don't rce this check but continue looking for other candidates. 1993 } 1994 #ifdef ASSERT 1995 if (TraceRangeLimitCheck) { 1996 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 1997 bol->dump(2); 1998 } 1999 #endif 2000 // At this point we have the expression as: 2001 // scale_con * trip_counter + offset :: limit 2002 // where scale_con, offset and limit are loop invariant. Trip_counter 2003 // monotonically increases by stride_con, a constant. Both (or either) 2004 // stride_con and scale_con can be negative which will flip about the 2005 // sense of the test. 2006 2007 // Adjust pre and main loop limits to guard the correct iteration set 2008 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 2009 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 2010 // The underflow and overflow limits: 0 <= scale*I+offset < limit 2011 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); 2012 if (!conditional_rc) { 2013 // (0-offset)/scale could be outside of loop iterations range. 2014 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2015 } 2016 } else { 2017 #ifndef PRODUCT 2018 if( PrintOpto ) 2019 tty->print_cr("missed RCE opportunity"); 2020 #endif 2021 continue; // In release mode, ignore it 2022 } 2023 } else { // Otherwise work on normal compares 2024 switch( b_test._test ) { 2025 case BoolTest::gt: 2026 // Fall into GE case 2027 case BoolTest::ge: 2028 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 2029 scale_con = -scale_con; 2030 offset = new SubINode( zero, offset ); 2031 register_new_node( offset, pre_ctrl ); 2032 limit = new SubINode( zero, limit ); 2033 register_new_node( limit, pre_ctrl ); 2034 // Fall into LE case 2035 case BoolTest::le: 2036 if (b_test._test != BoolTest::gt) { 2037 // Convert X <= Y to X < Y+1 2038 limit = new AddINode( limit, one ); 2039 register_new_node( limit, pre_ctrl ); 2040 } 2041 // Fall into LT case 2042 case BoolTest::lt: 2043 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 2044 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 2045 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 2046 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); 2047 if (!conditional_rc) { 2048 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 2049 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 2050 // still be outside of loop range. 2051 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2052 } 2053 break; 2054 default: 2055 #ifndef PRODUCT 2056 if( PrintOpto ) 2057 tty->print_cr("missed RCE opportunity"); 2058 #endif 2059 continue; // Unhandled case 2060 } 2061 } 2062 2063 // Kill the eliminated test 2064 C->set_major_progress(); 2065 Node *kill_con = _igvn.intcon( 1-flip ); 2066 set_ctrl(kill_con, C->root()); 2067 _igvn.replace_input_of(iff, 1, kill_con); 2068 // Find surviving projection 2069 assert(iff->is_If(), ""); 2070 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 2071 // Find loads off the surviving projection; remove their control edge 2072 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2073 Node* cd = dp->fast_out(i); // Control-dependent node 2074 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop 2075 // Allow the load to float around in the loop, or before it 2076 // but NOT before the pre-loop. 2077 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 2078 --i; 2079 --imax; 2080 } 2081 } 2082 2083 } // End of is IF 2084 2085 } 2086 2087 // Update loop limits 2088 if (conditional_rc) { 2089 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit) 2090 : (Node*)new MaxINode(pre_limit, orig_limit); 2091 register_new_node(pre_limit, pre_ctrl); 2092 } 2093 _igvn.replace_input_of(pre_opaq, 1, pre_limit); 2094 2095 // Note:: we are making the main loop limit no longer precise; 2096 // need to round up based on stride. 2097 cl->set_nonexact_trip_count(); 2098 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case 2099 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 2100 // Hopefully, compiler will optimize for powers of 2. 2101 Node *ctrl = get_ctrl(main_limit); 2102 Node *stride = cl->stride(); 2103 Node *init = cl->init_trip()->uncast(); 2104 Node *span = new SubINode(main_limit,init); 2105 register_new_node(span,ctrl); 2106 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 2107 Node *add = new AddINode(span,rndup); 2108 register_new_node(add,ctrl); 2109 Node *div = new DivINode(0,add,stride); 2110 register_new_node(div,ctrl); 2111 Node *mul = new MulINode(div,stride); 2112 register_new_node(mul,ctrl); 2113 Node *newlim = new AddINode(mul,init); 2114 register_new_node(newlim,ctrl); 2115 main_limit = newlim; 2116 } 2117 2118 Node *main_cle = cl->loopexit(); 2119 Node *main_bol = main_cle->in(1); 2120 // Hacking loop bounds; need private copies of exit test 2121 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 2122 main_bol = main_bol->clone();// Clone a private BoolNode 2123 register_new_node( main_bol, main_cle->in(0) ); 2124 _igvn.replace_input_of(main_cle, 1, main_bol); 2125 } 2126 Node *main_cmp = main_bol->in(1); 2127 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 2128 main_cmp = main_cmp->clone();// Clone a private CmpNode 2129 register_new_node( main_cmp, main_cle->in(0) ); 2130 _igvn.replace_input_of(main_bol, 1, main_cmp); 2131 } 2132 // Hack the now-private loop bounds 2133 _igvn.replace_input_of(main_cmp, 2, main_limit); 2134 // The OpaqueNode is unshared by design 2135 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 2136 _igvn.replace_input_of(opqzm, 1, main_limit); 2137 } 2138 2139 //------------------------------DCE_loop_body---------------------------------- 2140 // Remove simplistic dead code from loop body 2141 void IdealLoopTree::DCE_loop_body() { 2142 for( uint i = 0; i < _body.size(); i++ ) 2143 if( _body.at(i)->outcnt() == 0 ) 2144 _body.map( i--, _body.pop() ); 2145 } 2146 2147 2148 //------------------------------adjust_loop_exit_prob-------------------------- 2149 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2150 // Replace with a 1-in-10 exit guess. 2151 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 2152 Node *test = tail(); 2153 while( test != _head ) { 2154 uint top = test->Opcode(); 2155 if( top == Op_IfTrue || top == Op_IfFalse ) { 2156 int test_con = ((ProjNode*)test)->_con; 2157 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2158 IfNode *iff = test->in(0)->as_If(); 2159 if( iff->outcnt() == 2 ) { // Ignore dead tests 2160 Node *bol = iff->in(1); 2161 if( bol && bol->req() > 1 && bol->in(1) && 2162 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 2163 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 2164 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 2165 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 2166 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 2167 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 2168 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 2169 return; // Allocation loops RARELY take backedge 2170 // Find the OTHER exit path from the IF 2171 Node* ex = iff->proj_out(1-test_con); 2172 float p = iff->_prob; 2173 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 2174 if( top == Op_IfTrue ) { 2175 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2176 iff->_prob = PROB_STATIC_FREQUENT; 2177 } 2178 } else { 2179 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2180 iff->_prob = PROB_STATIC_INFREQUENT; 2181 } 2182 } 2183 } 2184 } 2185 } 2186 test = phase->idom(test); 2187 } 2188 } 2189 2190 2191 //------------------------------policy_do_remove_empty_loop-------------------- 2192 // Micro-benchmark spamming. Policy is to always remove empty loops. 2193 // The 'DO' part is to replace the trip counter with the value it will 2194 // have on the last iteration. This will break the loop. 2195 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 2196 // Minimum size must be empty loop 2197 if (_body.size() > EMPTY_LOOP_SIZE) 2198 return false; 2199 2200 if (!_head->is_CountedLoop()) 2201 return false; // Dead loop 2202 CountedLoopNode *cl = _head->as_CountedLoop(); 2203 if (!cl->is_valid_counted_loop()) 2204 return false; // Malformed loop 2205 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 2206 return false; // Infinite loop 2207 2208 #ifdef ASSERT 2209 // Ensure only one phi which is the iv. 2210 Node* iv = NULL; 2211 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 2212 Node* n = cl->fast_out(i); 2213 if (n->Opcode() == Op_Phi) { 2214 assert(iv == NULL, "Too many phis" ); 2215 iv = n; 2216 } 2217 } 2218 assert(iv == cl->phi(), "Wrong phi" ); 2219 #endif 2220 2221 // main and post loops have explicitly created zero trip guard 2222 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 2223 if (needs_guard) { 2224 // Skip guard if values not overlap. 2225 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 2226 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 2227 int stride_con = cl->stride_con(); 2228 if (stride_con > 0) { 2229 needs_guard = (init_t->_hi >= limit_t->_lo); 2230 } else { 2231 needs_guard = (init_t->_lo <= limit_t->_hi); 2232 } 2233 } 2234 if (needs_guard) { 2235 // Check for an obvious zero trip guard. 2236 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 2237 if (inctrl->Opcode() == Op_IfTrue) { 2238 // The test should look like just the backedge of a CountedLoop 2239 Node* iff = inctrl->in(0); 2240 if (iff->is_If()) { 2241 Node* bol = iff->in(1); 2242 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { 2243 Node* cmp = bol->in(1); 2244 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { 2245 needs_guard = false; 2246 } 2247 } 2248 } 2249 } 2250 } 2251 2252 #ifndef PRODUCT 2253 if (PrintOpto) { 2254 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 2255 this->dump_head(); 2256 } else if (TraceLoopOpts) { 2257 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 2258 this->dump_head(); 2259 } 2260 #endif 2261 2262 if (needs_guard) { 2263 // Peel the loop to ensure there's a zero trip guard 2264 Node_List old_new; 2265 phase->do_peeling(this, old_new); 2266 } 2267 2268 // Replace the phi at loop head with the final value of the last 2269 // iteration. Then the CountedLoopEnd will collapse (backedge never 2270 // taken) and all loop-invariant uses of the exit values will be correct. 2271 Node *phi = cl->phi(); 2272 Node *exact_limit = phase->exact_limit(this); 2273 if (exact_limit != cl->limit()) { 2274 // We also need to replace the original limit to collapse loop exit. 2275 Node* cmp = cl->loopexit()->cmp_node(); 2276 assert(cl->limit() == cmp->in(2), "sanity"); 2277 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 2278 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 2279 } 2280 // Note: the final value after increment should not overflow since 2281 // counted loop has limit check predicate. 2282 Node *final = new SubINode( exact_limit, cl->stride() ); 2283 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 2284 phase->_igvn.replace_node(phi,final); 2285 phase->C->set_major_progress(); 2286 return true; 2287 } 2288 2289 //------------------------------policy_do_one_iteration_loop------------------- 2290 // Convert one iteration loop into normal code. 2291 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 2292 if (!_head->as_Loop()->is_valid_counted_loop()) 2293 return false; // Only for counted loop 2294 2295 CountedLoopNode *cl = _head->as_CountedLoop(); 2296 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 2297 return false; 2298 } 2299 2300 #ifndef PRODUCT 2301 if(TraceLoopOpts) { 2302 tty->print("OneIteration "); 2303 this->dump_head(); 2304 } 2305 #endif 2306 2307 Node *init_n = cl->init_trip(); 2308 #ifdef ASSERT 2309 // Loop boundaries should be constant since trip count is exact. 2310 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 2311 #endif 2312 // Replace the phi at loop head with the value of the init_trip. 2313 // Then the CountedLoopEnd will collapse (backedge will not be taken) 2314 // and all loop-invariant uses of the exit values will be correct. 2315 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 2316 phase->C->set_major_progress(); 2317 return true; 2318 } 2319 2320 //============================================================================= 2321 //------------------------------iteration_split_impl--------------------------- 2322 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 2323 // Compute exact loop trip count if possible. 2324 compute_exact_trip_count(phase); 2325 2326 // Convert one iteration loop into normal code. 2327 if (policy_do_one_iteration_loop(phase)) 2328 return true; 2329 2330 // Check and remove empty loops (spam micro-benchmarks) 2331 if (policy_do_remove_empty_loop(phase)) 2332 return true; // Here we removed an empty loop 2333 2334 bool should_peel = policy_peeling(phase); // Should we peel? 2335 2336 bool should_unswitch = policy_unswitching(phase); 2337 2338 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 2339 // This removes loop-invariant tests (usually null checks). 2340 if (!_head->is_CountedLoop()) { // Non-counted loop 2341 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 2342 // Partial peel succeeded so terminate this round of loop opts 2343 return false; 2344 } 2345 if (should_peel) { // Should we peel? 2346 #ifndef PRODUCT 2347 if (PrintOpto) tty->print_cr("should_peel"); 2348 #endif 2349 phase->do_peeling(this,old_new); 2350 } else if (should_unswitch) { 2351 phase->do_unswitching(this, old_new); 2352 } 2353 return true; 2354 } 2355 CountedLoopNode *cl = _head->as_CountedLoop(); 2356 2357 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 2358 2359 // Do nothing special to pre- and post- loops 2360 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 2361 2362 // Compute loop trip count from profile data 2363 compute_profile_trip_cnt(phase); 2364 2365 // Before attempting fancy unrolling, RCE or alignment, see if we want 2366 // to completely unroll this loop or do loop unswitching. 2367 if (cl->is_normal_loop()) { 2368 if (should_unswitch) { 2369 phase->do_unswitching(this, old_new); 2370 return true; 2371 } 2372 bool should_maximally_unroll = policy_maximally_unroll(phase); 2373 if (should_maximally_unroll) { 2374 // Here we did some unrolling and peeling. Eventually we will 2375 // completely unroll this loop and it will no longer be a loop. 2376 phase->do_maximally_unroll(this,old_new); 2377 return true; 2378 } 2379 } 2380 2381 // Skip next optimizations if running low on nodes. Note that 2382 // policy_unswitching and policy_maximally_unroll have this check. 2383 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes(); 2384 if ((int)(2 * _body.size()) > nodes_left) { 2385 return true; 2386 } 2387 2388 // Counted loops may be peeled, may need some iterations run up 2389 // front for RCE, and may want to align loop refs to a cache 2390 // line. Thus we clone a full loop up front whose trip count is 2391 // at least 1 (if peeling), but may be several more. 2392 2393 // The main loop will start cache-line aligned with at least 1 2394 // iteration of the unrolled body (zero-trip test required) and 2395 // will have some range checks removed. 2396 2397 // A post-loop will finish any odd iterations (leftover after 2398 // unrolling), plus any needed for RCE purposes. 2399 2400 bool should_unroll = policy_unroll(phase); 2401 2402 bool should_rce = policy_range_check(phase); 2403 2404 bool should_align = policy_align(phase); 2405 2406 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2407 // need a pre-loop. We may still need to peel an initial iteration but 2408 // we will not be needing an unknown number of pre-iterations. 2409 // 2410 // Basically, if may_rce_align reports FALSE first time through, 2411 // we will not be able to later do RCE or Aligning on this loop. 2412 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2413 2414 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2415 // we switch to the pre-/main-/post-loop model. This model also covers 2416 // peeling. 2417 if (should_rce || should_align || should_unroll) { 2418 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2419 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2420 2421 // Adjust the pre- and main-loop limits to let the pre and post loops run 2422 // with full checks, but the main-loop with no checks. Remove said 2423 // checks from the main body. 2424 if (should_rce) 2425 phase->do_range_check(this,old_new); 2426 2427 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2428 // twice as many iterations as before) and the main body limit (only do 2429 // an even number of trips). If we are peeling, we might enable some RCE 2430 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2431 // peeling. 2432 if (should_unroll && !should_peel) { 2433 phase->mark_reductions(this); 2434 phase->do_unroll(this, old_new, true); 2435 } 2436 2437 // Adjust the pre-loop limits to align the main body 2438 // iterations. 2439 if (should_align) 2440 Unimplemented(); 2441 2442 } else { // Else we have an unchanged counted loop 2443 if (should_peel) // Might want to peel but do nothing else 2444 phase->do_peeling(this,old_new); 2445 } 2446 return true; 2447 } 2448 2449 2450 //============================================================================= 2451 //------------------------------iteration_split-------------------------------- 2452 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2453 // Recursively iteration split nested loops 2454 if (_child && !_child->iteration_split(phase, old_new)) 2455 return false; 2456 2457 // Clean out prior deadwood 2458 DCE_loop_body(); 2459 2460 2461 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2462 // Replace with a 1-in-10 exit guess. 2463 if (_parent /*not the root loop*/ && 2464 !_irreducible && 2465 // Also ignore the occasional dead backedge 2466 !tail()->is_top()) { 2467 adjust_loop_exit_prob(phase); 2468 } 2469 2470 // Gate unrolling, RCE and peeling efforts. 2471 if (!_child && // If not an inner loop, do not split 2472 !_irreducible && 2473 _allow_optimizations && 2474 !tail()->is_top()) { // Also ignore the occasional dead backedge 2475 if (!_has_call) { 2476 if (!iteration_split_impl(phase, old_new)) { 2477 return false; 2478 } 2479 } else if (policy_unswitching(phase)) { 2480 phase->do_unswitching(this, old_new); 2481 } 2482 } 2483 2484 // Minor offset re-organization to remove loop-fallout uses of 2485 // trip counter when there was no major reshaping. 2486 phase->reorg_offsets(this); 2487 2488 if (_next && !_next->iteration_split(phase, old_new)) 2489 return false; 2490 return true; 2491 } 2492 2493 2494 //============================================================================= 2495 // Process all the loops in the loop tree and replace any fill 2496 // patterns with an intrisc version. 2497 bool PhaseIdealLoop::do_intrinsify_fill() { 2498 bool changed = false; 2499 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2500 IdealLoopTree* lpt = iter.current(); 2501 changed |= intrinsify_fill(lpt); 2502 } 2503 return changed; 2504 } 2505 2506 2507 // Examine an inner loop looking for a a single store of an invariant 2508 // value in a unit stride loop, 2509 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2510 Node*& shift, Node*& con) { 2511 const char* msg = NULL; 2512 Node* msg_node = NULL; 2513 2514 store_value = NULL; 2515 con = NULL; 2516 shift = NULL; 2517 2518 // Process the loop looking for stores. If there are multiple 2519 // stores or extra control flow give at this point. 2520 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2521 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2522 Node* n = lpt->_body.at(i); 2523 if (n->outcnt() == 0) continue; // Ignore dead 2524 if (n->is_Store()) { 2525 if (store != NULL) { 2526 msg = "multiple stores"; 2527 break; 2528 } 2529 int opc = n->Opcode(); 2530 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 2531 msg = "oop fills not handled"; 2532 break; 2533 } 2534 Node* value = n->in(MemNode::ValueIn); 2535 if (!lpt->is_invariant(value)) { 2536 msg = "variant store value"; 2537 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2538 msg = "not array address"; 2539 } 2540 store = n; 2541 store_value = value; 2542 } else if (n->is_If() && n != head->loopexit()) { 2543 msg = "extra control flow"; 2544 msg_node = n; 2545 } 2546 } 2547 2548 if (store == NULL) { 2549 // No store in loop 2550 return false; 2551 } 2552 2553 if (msg == NULL && head->stride_con() != 1) { 2554 // could handle negative strides too 2555 if (head->stride_con() < 0) { 2556 msg = "negative stride"; 2557 } else { 2558 msg = "non-unit stride"; 2559 } 2560 } 2561 2562 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2563 msg = "can't handle store address"; 2564 msg_node = store->in(MemNode::Address); 2565 } 2566 2567 if (msg == NULL && 2568 (!store->in(MemNode::Memory)->is_Phi() || 2569 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2570 msg = "store memory isn't proper phi"; 2571 msg_node = store->in(MemNode::Memory); 2572 } 2573 2574 // Make sure there is an appropriate fill routine 2575 BasicType t = store->as_Mem()->memory_type(); 2576 const char* fill_name; 2577 if (msg == NULL && 2578 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2579 msg = "unsupported store"; 2580 msg_node = store; 2581 } 2582 2583 if (msg != NULL) { 2584 #ifndef PRODUCT 2585 if (TraceOptimizeFill) { 2586 tty->print_cr("not fill intrinsic candidate: %s", msg); 2587 if (msg_node != NULL) msg_node->dump(); 2588 } 2589 #endif 2590 return false; 2591 } 2592 2593 // Make sure the address expression can be handled. It should be 2594 // head->phi * elsize + con. head->phi might have a ConvI2L. 2595 Node* elements[4]; 2596 Node* conv = NULL; 2597 bool found_index = false; 2598 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2599 for (int e = 0; e < count; e++) { 2600 Node* n = elements[e]; 2601 if (n->is_Con() && con == NULL) { 2602 con = n; 2603 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2604 Node* value = n->in(1); 2605 #ifdef _LP64 2606 if (value->Opcode() == Op_ConvI2L) { 2607 conv = value; 2608 value = value->in(1); 2609 } 2610 #endif 2611 if (value != head->phi()) { 2612 msg = "unhandled shift in address"; 2613 } else { 2614 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 2615 msg = "scale doesn't match"; 2616 } else { 2617 found_index = true; 2618 shift = n; 2619 } 2620 } 2621 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2622 if (n->in(1) == head->phi()) { 2623 found_index = true; 2624 conv = n; 2625 } else { 2626 msg = "unhandled input to ConvI2L"; 2627 } 2628 } else if (n == head->phi()) { 2629 // no shift, check below for allowed cases 2630 found_index = true; 2631 } else { 2632 msg = "unhandled node in address"; 2633 msg_node = n; 2634 } 2635 } 2636 2637 if (count == -1) { 2638 msg = "malformed address expression"; 2639 msg_node = store; 2640 } 2641 2642 if (!found_index) { 2643 msg = "missing use of index"; 2644 } 2645 2646 // byte sized items won't have a shift 2647 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2648 msg = "can't find shift"; 2649 msg_node = store; 2650 } 2651 2652 if (msg != NULL) { 2653 #ifndef PRODUCT 2654 if (TraceOptimizeFill) { 2655 tty->print_cr("not fill intrinsic: %s", msg); 2656 if (msg_node != NULL) msg_node->dump(); 2657 } 2658 #endif 2659 return false; 2660 } 2661 2662 // No make sure all the other nodes in the loop can be handled 2663 VectorSet ok(Thread::current()->resource_area()); 2664 2665 // store related values are ok 2666 ok.set(store->_idx); 2667 ok.set(store->in(MemNode::Memory)->_idx); 2668 2669 CountedLoopEndNode* loop_exit = head->loopexit(); 2670 guarantee(loop_exit != NULL, "no loop exit node"); 2671 2672 // Loop structure is ok 2673 ok.set(head->_idx); 2674 ok.set(loop_exit->_idx); 2675 ok.set(head->phi()->_idx); 2676 ok.set(head->incr()->_idx); 2677 ok.set(loop_exit->cmp_node()->_idx); 2678 ok.set(loop_exit->in(1)->_idx); 2679 2680 // Address elements are ok 2681 if (con) ok.set(con->_idx); 2682 if (shift) ok.set(shift->_idx); 2683 if (conv) ok.set(conv->_idx); 2684 2685 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2686 Node* n = lpt->_body.at(i); 2687 if (n->outcnt() == 0) continue; // Ignore dead 2688 if (ok.test(n->_idx)) continue; 2689 // Backedge projection is ok 2690 if (n->is_IfTrue() && n->in(0) == loop_exit) continue; 2691 if (!n->is_AddP()) { 2692 msg = "unhandled node"; 2693 msg_node = n; 2694 break; 2695 } 2696 } 2697 2698 // Make sure no unexpected values are used outside the loop 2699 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2700 Node* n = lpt->_body.at(i); 2701 // These values can be replaced with other nodes if they are used 2702 // outside the loop. 2703 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2704 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2705 Node* use = iter.get(); 2706 if (!lpt->_body.contains(use)) { 2707 msg = "node is used outside loop"; 2708 // lpt->_body.dump(); 2709 msg_node = n; 2710 break; 2711 } 2712 } 2713 } 2714 2715 #ifdef ASSERT 2716 if (TraceOptimizeFill) { 2717 if (msg != NULL) { 2718 tty->print_cr("no fill intrinsic: %s", msg); 2719 if (msg_node != NULL) msg_node->dump(); 2720 } else { 2721 tty->print_cr("fill intrinsic for:"); 2722 } 2723 store->dump(); 2724 if (Verbose) { 2725 lpt->_body.dump(); 2726 } 2727 } 2728 #endif 2729 2730 return msg == NULL; 2731 } 2732 2733 2734 2735 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2736 // Only for counted inner loops 2737 if (!lpt->is_counted() || !lpt->is_inner()) { 2738 return false; 2739 } 2740 2741 // Must have constant stride 2742 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2743 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 2744 return false; 2745 } 2746 2747 // Check that the body only contains a store of a loop invariant 2748 // value that is indexed by the loop phi. 2749 Node* store = NULL; 2750 Node* store_value = NULL; 2751 Node* shift = NULL; 2752 Node* offset = NULL; 2753 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2754 return false; 2755 } 2756 2757 #ifndef PRODUCT 2758 if (TraceLoopOpts) { 2759 tty->print("ArrayFill "); 2760 lpt->dump_head(); 2761 } 2762 #endif 2763 2764 // Now replace the whole loop body by a call to a fill routine that 2765 // covers the same region as the loop. 2766 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2767 2768 // Build an expression for the beginning of the copy region 2769 Node* index = head->init_trip(); 2770 #ifdef _LP64 2771 index = new ConvI2LNode(index); 2772 _igvn.register_new_node_with_optimizer(index); 2773 #endif 2774 if (shift != NULL) { 2775 // byte arrays don't require a shift but others do. 2776 index = new LShiftXNode(index, shift->in(2)); 2777 _igvn.register_new_node_with_optimizer(index); 2778 } 2779 index = new AddPNode(base, base, index); 2780 _igvn.register_new_node_with_optimizer(index); 2781 Node* from = new AddPNode(base, index, offset); 2782 _igvn.register_new_node_with_optimizer(from); 2783 // Compute the number of elements to copy 2784 Node* len = new SubINode(head->limit(), head->init_trip()); 2785 _igvn.register_new_node_with_optimizer(len); 2786 2787 BasicType t = store->as_Mem()->memory_type(); 2788 bool aligned = false; 2789 if (offset != NULL && head->init_trip()->is_Con()) { 2790 int element_size = type2aelembytes(t); 2791 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2792 } 2793 2794 // Build a call to the fill routine 2795 const char* fill_name; 2796 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2797 assert(fill != NULL, "what?"); 2798 2799 // Convert float/double to int/long for fill routines 2800 if (t == T_FLOAT) { 2801 store_value = new MoveF2INode(store_value); 2802 _igvn.register_new_node_with_optimizer(store_value); 2803 } else if (t == T_DOUBLE) { 2804 store_value = new MoveD2LNode(store_value); 2805 _igvn.register_new_node_with_optimizer(store_value); 2806 } 2807 2808 if (CCallingConventionRequiresIntsAsLongs && 2809 // See StubRoutines::select_fill_function for types. FLOAT has been converted to INT. 2810 (t == T_FLOAT || t == T_INT || is_subword_type(t))) { 2811 store_value = new ConvI2LNode(store_value); 2812 _igvn.register_new_node_with_optimizer(store_value); 2813 } 2814 2815 Node* mem_phi = store->in(MemNode::Memory); 2816 Node* result_ctrl; 2817 Node* result_mem; 2818 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2819 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill, 2820 fill_name, TypeAryPtr::get_array_body_type(t)); 2821 uint cnt = 0; 2822 call->init_req(TypeFunc::Parms + cnt++, from); 2823 call->init_req(TypeFunc::Parms + cnt++, store_value); 2824 if (CCallingConventionRequiresIntsAsLongs) { 2825 call->init_req(TypeFunc::Parms + cnt++, C->top()); 2826 } 2827 #ifdef _LP64 2828 len = new ConvI2LNode(len); 2829 _igvn.register_new_node_with_optimizer(len); 2830 #endif 2831 call->init_req(TypeFunc::Parms + cnt++, len); 2832 #ifdef _LP64 2833 call->init_req(TypeFunc::Parms + cnt++, C->top()); 2834 #endif 2835 call->init_req(TypeFunc::Control, head->init_control()); 2836 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. 2837 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); 2838 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr)); 2839 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr)); 2840 _igvn.register_new_node_with_optimizer(call); 2841 result_ctrl = new ProjNode(call,TypeFunc::Control); 2842 _igvn.register_new_node_with_optimizer(result_ctrl); 2843 result_mem = new ProjNode(call,TypeFunc::Memory); 2844 _igvn.register_new_node_with_optimizer(result_mem); 2845 2846 /* Disable following optimization until proper fix (add missing checks). 2847 2848 // If this fill is tightly coupled to an allocation and overwrites 2849 // the whole body, allow it to take over the zeroing. 2850 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 2851 if (alloc != NULL && alloc->is_AllocateArray()) { 2852 Node* length = alloc->as_AllocateArray()->Ideal_length(); 2853 if (head->limit() == length && 2854 head->init_trip() == _igvn.intcon(0)) { 2855 if (TraceOptimizeFill) { 2856 tty->print_cr("Eliminated zeroing in allocation"); 2857 } 2858 alloc->maybe_set_complete(&_igvn); 2859 } else { 2860 #ifdef ASSERT 2861 if (TraceOptimizeFill) { 2862 tty->print_cr("filling array but bounds don't match"); 2863 alloc->dump(); 2864 head->init_trip()->dump(); 2865 head->limit()->dump(); 2866 length->dump(); 2867 } 2868 #endif 2869 } 2870 } 2871 */ 2872 2873 // Redirect the old control and memory edges that are outside the loop. 2874 Node* exit = head->loopexit()->proj_out(0); 2875 // Sometimes the memory phi of the head is used as the outgoing 2876 // state of the loop. It's safe in this case to replace it with the 2877 // result_mem. 2878 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 2879 _igvn.replace_node(exit, result_ctrl); 2880 _igvn.replace_node(store, result_mem); 2881 // Any uses the increment outside of the loop become the loop limit. 2882 _igvn.replace_node(head->incr(), head->limit()); 2883 2884 // Disconnect the head from the loop. 2885 for (uint i = 0; i < lpt->_body.size(); i++) { 2886 Node* n = lpt->_body.at(i); 2887 _igvn.replace_node(n, C->top()); 2888 } 2889 2890 return true; 2891 }