1 /* 2 * Copyright (c) 2000, 2015, 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 #include "opto/superword.hpp" 42 #include "opto/vectornode.hpp" 43 44 //------------------------------is_loop_exit----------------------------------- 45 // Given an IfNode, return the loop-exiting projection or NULL if both 46 // arms remain in the loop. 47 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 48 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 49 PhaseIdealLoop *phase = _phase; 50 // Test is an IfNode, has 2 projections. If BOTH are in the loop 51 // we need loop unswitching instead of peeling. 52 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 53 return iff->raw_out(0); 54 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 55 return iff->raw_out(1); 56 return NULL; 57 } 58 59 60 //============================================================================= 61 62 63 //------------------------------record_for_igvn---------------------------- 64 // Put loop body on igvn work list 65 void IdealLoopTree::record_for_igvn() { 66 for( uint i = 0; i < _body.size(); i++ ) { 67 Node *n = _body.at(i); 68 _phase->_igvn._worklist.push(n); 69 } 70 } 71 72 //------------------------------compute_exact_trip_count----------------------- 73 // Compute loop exact trip count if possible. Do not recalculate trip count for 74 // split loops (pre-main-post) which have their limits and inits behind Opaque node. 75 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) { 76 if (!_head->as_Loop()->is_valid_counted_loop()) { 77 return; 78 } 79 CountedLoopNode* cl = _head->as_CountedLoop(); 80 // Trip count may become nonexact for iteration split loops since 81 // RCE modifies limits. Note, _trip_count value is not reset since 82 // it is used to limit unrolling of main loop. 83 cl->set_nonexact_trip_count(); 84 85 // Loop's test should be part of loop. 86 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 87 return; // Infinite loop 88 89 #ifdef ASSERT 90 BoolTest::mask bt = cl->loopexit()->test_trip(); 91 assert(bt == BoolTest::lt || bt == BoolTest::gt || 92 bt == BoolTest::ne, "canonical test is expected"); 93 #endif 94 95 Node* init_n = cl->init_trip(); 96 Node* limit_n = cl->limit(); 97 if (init_n != NULL && init_n->is_Con() && 98 limit_n != NULL && limit_n->is_Con()) { 99 // Use longs to avoid integer overflow. 100 int stride_con = cl->stride_con(); 101 jlong init_con = cl->init_trip()->get_int(); 102 jlong limit_con = cl->limit()->get_int(); 103 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 104 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 105 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { 106 // Set exact trip count. 107 cl->set_exact_trip_count((uint)trip_count); 108 } 109 } 110 } 111 112 //------------------------------compute_profile_trip_cnt---------------------------- 113 // Compute loop trip count from profile data as 114 // (backedge_count + loop_exit_count) / loop_exit_count 115 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 116 if (!_head->is_CountedLoop()) { 117 return; 118 } 119 CountedLoopNode* head = _head->as_CountedLoop(); 120 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 121 return; // Already computed 122 } 123 float trip_cnt = (float)max_jint; // default is big 124 125 Node* back = head->in(LoopNode::LoopBackControl); 126 while (back != head) { 127 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 128 back->in(0) && 129 back->in(0)->is_If() && 130 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 131 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 132 break; 133 } 134 back = phase->idom(back); 135 } 136 if (back != head) { 137 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 138 back->in(0), "if-projection exists"); 139 IfNode* back_if = back->in(0)->as_If(); 140 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 141 142 // Now compute a loop exit count 143 float loop_exit_cnt = 0.0f; 144 for( uint i = 0; i < _body.size(); i++ ) { 145 Node *n = _body[i]; 146 if( n->is_If() ) { 147 IfNode *iff = n->as_If(); 148 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 149 Node *exit = is_loop_exit(iff); 150 if( exit ) { 151 float exit_prob = iff->_prob; 152 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 153 if (exit_prob > PROB_MIN) { 154 float exit_cnt = iff->_fcnt * exit_prob; 155 loop_exit_cnt += exit_cnt; 156 } 157 } 158 } 159 } 160 } 161 if (loop_exit_cnt > 0.0f) { 162 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 163 } else { 164 // No exit count so use 165 trip_cnt = loop_back_cnt; 166 } 167 } 168 #ifndef PRODUCT 169 if (TraceProfileTripCount) { 170 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 171 } 172 #endif 173 head->set_profile_trip_cnt(trip_cnt); 174 } 175 176 //---------------------is_invariant_addition----------------------------- 177 // Return nonzero index of invariant operand for an Add or Sub 178 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 179 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 180 int op = n->Opcode(); 181 if (op == Op_AddI || op == Op_SubI) { 182 bool in1_invar = this->is_invariant(n->in(1)); 183 bool in2_invar = this->is_invariant(n->in(2)); 184 if (in1_invar && !in2_invar) return 1; 185 if (!in1_invar && in2_invar) return 2; 186 } 187 return 0; 188 } 189 190 //---------------------reassociate_add_sub----------------------------- 191 // Reassociate invariant add and subtract expressions: 192 // 193 // inv1 + (x + inv2) => ( inv1 + inv2) + x 194 // (x + inv2) + inv1 => ( inv1 + inv2) + x 195 // inv1 + (x - inv2) => ( inv1 - inv2) + x 196 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 197 // (x + inv2) - inv1 => (-inv1 + inv2) + x 198 // (x - inv2) + inv1 => ( inv1 - inv2) + x 199 // (x - inv2) - inv1 => (-inv1 - inv2) + x 200 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 201 // inv1 - (x - inv2) => ( inv1 + inv2) - x 202 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 203 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 204 // inv1 - (x + inv2) => ( inv1 - inv2) - x 205 // 206 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 207 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 208 if (is_invariant(n1)) return NULL; 209 int inv1_idx = is_invariant_addition(n1, phase); 210 if (!inv1_idx) return NULL; 211 // Don't mess with add of constant (igvn moves them to expression tree root.) 212 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 213 Node* inv1 = n1->in(inv1_idx); 214 Node* n2 = n1->in(3 - inv1_idx); 215 int inv2_idx = is_invariant_addition(n2, phase); 216 if (!inv2_idx) return NULL; 217 Node* x = n2->in(3 - inv2_idx); 218 Node* inv2 = n2->in(inv2_idx); 219 220 bool neg_x = n2->is_Sub() && inv2_idx == 1; 221 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 222 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 223 if (n1->is_Sub() && inv1_idx == 1) { 224 neg_x = !neg_x; 225 neg_inv2 = !neg_inv2; 226 } 227 Node* inv1_c = phase->get_ctrl(inv1); 228 Node* inv2_c = phase->get_ctrl(inv2); 229 Node* n_inv1; 230 if (neg_inv1) { 231 Node *zero = phase->_igvn.intcon(0); 232 phase->set_ctrl(zero, phase->C->root()); 233 n_inv1 = new SubINode(zero, inv1); 234 phase->register_new_node(n_inv1, inv1_c); 235 } else { 236 n_inv1 = inv1; 237 } 238 Node* inv; 239 if (neg_inv2) { 240 inv = new SubINode(n_inv1, inv2); 241 } else { 242 inv = new AddINode(n_inv1, inv2); 243 } 244 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 245 246 Node* addx; 247 if (neg_x) { 248 addx = new SubINode(inv, x); 249 } else { 250 addx = new AddINode(x, inv); 251 } 252 phase->register_new_node(addx, phase->get_ctrl(x)); 253 phase->_igvn.replace_node(n1, addx); 254 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 255 _body.yank(n1); 256 return addx; 257 } 258 259 //---------------------reassociate_invariants----------------------------- 260 // Reassociate invariant expressions: 261 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 262 for (int i = _body.size() - 1; i >= 0; i--) { 263 Node *n = _body.at(i); 264 for (int j = 0; j < 5; j++) { 265 Node* nn = reassociate_add_sub(n, phase); 266 if (nn == NULL) break; 267 n = nn; // again 268 }; 269 } 270 } 271 272 //------------------------------policy_peeling--------------------------------- 273 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 274 // make some loop-invariant test (usually a null-check) happen before the loop. 275 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 276 Node *test = ((IdealLoopTree*)this)->tail(); 277 int body_size = ((IdealLoopTree*)this)->_body.size(); 278 // Peeling does loop cloning which can result in O(N^2) node construction 279 if( body_size > 255 /* Prevent overflow for large body_size */ 280 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) { 281 return false; // too large to safely clone 282 } 283 284 // check for vectorized loops, any peeling done was already applied 285 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false; 286 287 while( test != _head ) { // Scan till run off top of loop 288 if( test->is_If() ) { // Test? 289 Node *ctrl = phase->get_ctrl(test->in(1)); 290 if (ctrl->is_top()) 291 return false; // Found dead test on live IF? No peeling! 292 // Standard IF only has one input value to check for loop invariance 293 assert(test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd || test->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 294 // Condition is not a member of this loop? 295 if( !is_member(phase->get_loop(ctrl)) && 296 is_loop_exit(test) ) 297 return true; // Found reason to peel! 298 } 299 // Walk up dominators to loop _head looking for test which is 300 // executed on every path thru loop. 301 test = phase->idom(test); 302 } 303 return false; 304 } 305 306 //------------------------------peeled_dom_test_elim--------------------------- 307 // If we got the effect of peeling, either by actually peeling or by making 308 // a pre-loop which must execute at least once, we can remove all 309 // loop-invariant dominated tests in the main body. 310 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 311 bool progress = true; 312 while( progress ) { 313 progress = false; // Reset for next iteration 314 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 315 Node *test = prev->in(0); 316 while( test != loop->_head ) { // Scan till run off top of loop 317 318 int p_op = prev->Opcode(); 319 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 320 test->is_If() && // Test? 321 !test->in(1)->is_Con() && // And not already obvious? 322 // Condition is not a member of this loop? 323 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 324 // Walk loop body looking for instances of this test 325 for( uint i = 0; i < loop->_body.size(); i++ ) { 326 Node *n = loop->_body.at(i); 327 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 328 // IfNode was dominated by version in peeled loop body 329 progress = true; 330 dominated_by( old_new[prev->_idx], n ); 331 } 332 } 333 } 334 prev = test; 335 test = idom(test); 336 } // End of scan tests in loop 337 338 } // End of while( progress ) 339 } 340 341 //------------------------------do_peeling------------------------------------- 342 // Peel the first iteration of the given loop. 343 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 344 // The pre-loop illegally has 2 control users (old & new loops). 345 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 346 // Do this by making the old-loop fall-in edges act as if they came 347 // around the loopback from the prior iteration (follow the old-loop 348 // backedges) and then map to the new peeled iteration. This leaves 349 // the pre-loop with only 1 user (the new peeled iteration), but the 350 // peeled-loop backedge has 2 users. 351 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 352 // extra backedge user. 353 // 354 // orig 355 // 356 // stmt1 357 // | 358 // v 359 // loop predicate 360 // | 361 // v 362 // loop<----+ 363 // | | 364 // stmt2 | 365 // | | 366 // v | 367 // if ^ 368 // / \ | 369 // / \ | 370 // v v | 371 // false true | 372 // / \ | 373 // / ----+ 374 // | 375 // v 376 // exit 377 // 378 // 379 // after clone loop 380 // 381 // stmt1 382 // | 383 // v 384 // loop predicate 385 // / \ 386 // clone / \ orig 387 // / \ 388 // / \ 389 // v v 390 // +---->loop clone loop<----+ 391 // | | | | 392 // | stmt2 clone stmt2 | 393 // | | | | 394 // | v v | 395 // ^ if clone If ^ 396 // | / \ / \ | 397 // | / \ / \ | 398 // | v v v v | 399 // | true false false true | 400 // | / \ / \ | 401 // +---- \ / ----+ 402 // \ / 403 // 1v v2 404 // region 405 // | 406 // v 407 // exit 408 // 409 // 410 // after peel and predicate move 411 // 412 // stmt1 413 // / 414 // / 415 // clone / orig 416 // / 417 // / +----------+ 418 // / | | 419 // / loop predicate | 420 // / | | 421 // v v | 422 // TOP-->loop clone loop<----+ | 423 // | | | | 424 // stmt2 clone stmt2 | | 425 // | | | ^ 426 // v v | | 427 // if clone If ^ | 428 // / \ / \ | | 429 // / \ / \ | | 430 // v v v v | | 431 // true false false true | | 432 // | \ / \ | | 433 // | \ / ----+ ^ 434 // | \ / | 435 // | 1v v2 | 436 // v region | 437 // | | | 438 // | v | 439 // | exit | 440 // | | 441 // +--------------->-----------------+ 442 // 443 // 444 // final graph 445 // 446 // stmt1 447 // | 448 // v 449 // stmt2 clone 450 // | 451 // v 452 // if clone 453 // / | 454 // / | 455 // v v 456 // false true 457 // | | 458 // | v 459 // | loop predicate 460 // | | 461 // | v 462 // | loop<----+ 463 // | | | 464 // | stmt2 | 465 // | | | 466 // | v | 467 // v if ^ 468 // | / \ | 469 // | / \ | 470 // | v v | 471 // | false true | 472 // | | \ | 473 // v v --+ 474 // region 475 // | 476 // v 477 // exit 478 // 479 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 480 481 C->set_major_progress(); 482 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 483 // 'pre' loop from the main and the 'pre' can no longer have its 484 // iterations adjusted. Therefore, we need to declare this loop as 485 // no longer a 'main' loop; it will need new pre and post loops before 486 // we can do further RCE. 487 #ifndef PRODUCT 488 if (TraceLoopOpts) { 489 tty->print("Peel "); 490 loop->dump_head(); 491 } 492 #endif 493 Node* head = loop->_head; 494 bool counted_loop = head->is_CountedLoop(); 495 if (counted_loop) { 496 CountedLoopNode *cl = head->as_CountedLoop(); 497 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 498 cl->set_trip_count(cl->trip_count() - 1); 499 if (cl->is_main_loop()) { 500 cl->set_normal_loop(); 501 #ifndef PRODUCT 502 if (PrintOpto && VerifyLoopOptimizations) { 503 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 504 loop->dump_head(); 505 } 506 #endif 507 } 508 } 509 Node* entry = head->in(LoopNode::EntryControl); 510 511 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 512 // The pre-loop illegally has 2 control users (old & new loops). 513 clone_loop( loop, old_new, dom_depth(head) ); 514 515 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 516 // Do this by making the old-loop fall-in edges act as if they came 517 // around the loopback from the prior iteration (follow the old-loop 518 // backedges) and then map to the new peeled iteration. This leaves 519 // the pre-loop with only 1 user (the new peeled iteration), but the 520 // peeled-loop backedge has 2 users. 521 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 522 _igvn.hash_delete(head); 523 head->set_req(LoopNode::EntryControl, new_entry); 524 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 525 Node* old = head->fast_out(j); 526 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 527 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 528 if (!new_exit_value ) // Backedge value is ALSO loop invariant? 529 // Then loop body backedge value remains the same. 530 new_exit_value = old->in(LoopNode::LoopBackControl); 531 _igvn.hash_delete(old); 532 old->set_req(LoopNode::EntryControl, new_exit_value); 533 } 534 } 535 536 537 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 538 // extra backedge user. 539 Node* new_head = old_new[head->_idx]; 540 _igvn.hash_delete(new_head); 541 new_head->set_req(LoopNode::LoopBackControl, C->top()); 542 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 543 Node* use = new_head->fast_out(j2); 544 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 545 _igvn.hash_delete(use); 546 use->set_req(LoopNode::LoopBackControl, C->top()); 547 } 548 } 549 550 551 // Step 4: Correct dom-depth info. Set to loop-head depth. 552 int dd = dom_depth(head); 553 set_idom(head, head->in(1), dd); 554 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 555 Node *old = loop->_body.at(j3); 556 Node *nnn = old_new[old->_idx]; 557 if (!has_ctrl(nnn)) 558 set_idom(nnn, idom(nnn), dd-1); 559 } 560 561 // Now force out all loop-invariant dominating tests. The optimizer 562 // finds some, but we _know_ they are all useless. 563 peeled_dom_test_elim(loop,old_new); 564 565 loop->record_for_igvn(); 566 } 567 568 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 569 570 //------------------------------policy_maximally_unroll------------------------ 571 // Calculate exact loop trip count and return true if loop can be maximally 572 // unrolled. 573 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 574 CountedLoopNode *cl = _head->as_CountedLoop(); 575 assert(cl->is_normal_loop(), ""); 576 if (!cl->is_valid_counted_loop()) 577 return false; // Malformed counted loop 578 579 if (!cl->has_exact_trip_count()) { 580 // Trip count is not exact. 581 return false; 582 } 583 584 uint trip_count = cl->trip_count(); 585 // Note, max_juint is used to indicate unknown trip count. 586 assert(trip_count > 1, "one iteration loop should be optimized out already"); 587 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 588 589 // Real policy: if we maximally unroll, does it get too big? 590 // Allow the unrolled mess to get larger than standard loop 591 // size. After all, it will no longer be a loop. 592 uint body_size = _body.size(); 593 uint unroll_limit = (uint)LoopUnrollLimit * 4; 594 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 595 if (trip_count > unroll_limit || body_size > unroll_limit) { 596 return false; 597 } 598 599 // Fully unroll a loop with few iterations regardless next 600 // conditions since following loop optimizations will split 601 // such loop anyway (pre-main-post). 602 if (trip_count <= 3) 603 return true; 604 605 // Take into account that after unroll conjoined heads and tails will fold, 606 // otherwise policy_unroll() may allow more unrolling than max unrolling. 607 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; 608 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; 609 if (body_size != tst_body_size) // Check for int overflow 610 return false; 611 if (new_body_size > unroll_limit || 612 // Unrolling can result in a large amount of node construction 613 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) { 614 return false; 615 } 616 617 // Do not unroll a loop with String intrinsics code. 618 // String intrinsics are large and have loops. 619 for (uint k = 0; k < _body.size(); k++) { 620 Node* n = _body.at(k); 621 switch (n->Opcode()) { 622 case Op_StrComp: 623 case Op_StrEquals: 624 case Op_StrIndexOf: 625 case Op_StrIndexOfChar: 626 case Op_EncodeISOArray: 627 case Op_AryEq: 628 case Op_HasNegatives: { 629 return false; 630 } 631 #if INCLUDE_RTM_OPT 632 case Op_FastLock: 633 case Op_FastUnlock: { 634 // Don't unroll RTM locking code because it is large. 635 if (UseRTMLocking) { 636 return false; 637 } 638 } 639 #endif 640 } // switch 641 } 642 643 return true; // Do maximally unroll 644 } 645 646 647 //------------------------------policy_unroll---------------------------------- 648 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 649 // the loop is a CountedLoop and the body is small enough. 650 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) { 651 652 CountedLoopNode *cl = _head->as_CountedLoop(); 653 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 654 655 if (!cl->is_valid_counted_loop()) 656 return false; // Malformed counted loop 657 658 // Protect against over-unrolling. 659 // After split at least one iteration will be executed in pre-loop. 660 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; 661 662 _local_loop_unroll_limit = LoopUnrollLimit; 663 _local_loop_unroll_factor = 4; 664 int future_unroll_ct = cl->unrolled_count() * 2; 665 if (!cl->do_unroll_only()) { 666 if (future_unroll_ct > LoopMaxUnroll) return false; 667 } else { 668 // obey user constraints on vector mapped loops with additional unrolling applied 669 if ((future_unroll_ct / cl->slp_max_unroll()) > LoopMaxUnroll) return false; 670 } 671 672 // Check for initial stride being a small enough constant 673 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; 674 675 // Don't unroll if the next round of unrolling would push us 676 // over the expected trip count of the loop. One is subtracted 677 // from the expected trip count because the pre-loop normally 678 // executes 1 iteration. 679 if (UnrollLimitForProfileCheck > 0 && 680 cl->profile_trip_cnt() != COUNT_UNKNOWN && 681 future_unroll_ct > UnrollLimitForProfileCheck && 682 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 683 return false; 684 } 685 686 // When unroll count is greater than LoopUnrollMin, don't unroll if: 687 // the residual iterations are more than 10% of the trip count 688 // and rounds of "unroll,optimize" are not making significant progress 689 // Progress defined as current size less than 20% larger than previous size. 690 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 691 future_unroll_ct > LoopUnrollMin && 692 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 693 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 694 return false; 695 } 696 697 Node *init_n = cl->init_trip(); 698 Node *limit_n = cl->limit(); 699 int stride_con = cl->stride_con(); 700 // Non-constant bounds. 701 // Protect against over-unrolling when init or/and limit are not constant 702 // (so that trip_count's init value is maxint) but iv range is known. 703 if (init_n == NULL || !init_n->is_Con() || 704 limit_n == NULL || !limit_n->is_Con()) { 705 Node* phi = cl->phi(); 706 if (phi != NULL) { 707 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 708 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 709 int next_stride = stride_con * 2; // stride after this unroll 710 if (next_stride > 0) { 711 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 712 iv_type->_lo + next_stride > iv_type->_hi) { 713 return false; // over-unrolling 714 } 715 } else if (next_stride < 0) { 716 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 717 iv_type->_hi + next_stride < iv_type->_lo) { 718 return false; // over-unrolling 719 } 720 } 721 } 722 } 723 724 // After unroll limit will be adjusted: new_limit = limit-stride. 725 // Bailout if adjustment overflow. 726 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 727 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || 728 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) 729 return false; // overflow 730 731 // Adjust body_size to determine if we unroll or not 732 uint body_size = _body.size(); 733 // Key test to unroll loop in CRC32 java code 734 int xors_in_loop = 0; 735 // Also count ModL, DivL and MulL which expand mightly 736 for (uint k = 0; k < _body.size(); k++) { 737 Node* n = _body.at(k); 738 switch (n->Opcode()) { 739 case Op_XorI: xors_in_loop++; break; // CRC32 java code 740 case Op_ModL: body_size += 30; break; 741 case Op_DivL: body_size += 30; break; 742 case Op_MulL: body_size += 10; break; 743 case Op_StrComp: 744 case Op_StrEquals: 745 case Op_StrIndexOf: 746 case Op_StrIndexOfChar: 747 case Op_EncodeISOArray: 748 case Op_AryEq: 749 case Op_HasNegatives: { 750 // Do not unroll a loop with String intrinsics code. 751 // String intrinsics are large and have loops. 752 return false; 753 } 754 #if INCLUDE_RTM_OPT 755 case Op_FastLock: 756 case Op_FastUnlock: { 757 // Don't unroll RTM locking code because it is large. 758 if (UseRTMLocking) { 759 return false; 760 } 761 } 762 #endif 763 } // switch 764 } 765 766 if (UseSuperWord) { 767 if (!cl->is_reduction_loop()) { 768 phase->mark_reductions(this); 769 } 770 771 // Only attempt slp analysis when user controls do not prohibit it 772 if (LoopMaxUnroll > _local_loop_unroll_factor) { 773 // Once policy_slp_analysis succeeds, mark the loop with the 774 // maximal unroll factor so that we minimize analysis passes 775 if (future_unroll_ct >= _local_loop_unroll_factor) { 776 policy_unroll_slp_analysis(cl, phase, future_unroll_ct); 777 } 778 } 779 } 780 781 int slp_max_unroll_factor = cl->slp_max_unroll(); 782 if (cl->has_passed_slp()) { 783 if (slp_max_unroll_factor >= future_unroll_ct) return true; 784 // Normal case: loop too big 785 return false; 786 } 787 788 // Check for being too big 789 if (body_size > (uint)_local_loop_unroll_limit) { 790 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 791 // Normal case: loop too big 792 return false; 793 } 794 795 if (cl->do_unroll_only()) { 796 if (TraceSuperWordLoopUnrollAnalysis) { 797 tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct); 798 } 799 } 800 801 // Unroll once! (Each trip will soon do double iterations) 802 return true; 803 } 804 805 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) { 806 // Enable this functionality target by target as needed 807 if (SuperWordLoopUnrollAnalysis) { 808 if (!cl->was_slp_analyzed()) { 809 SuperWord sw(phase); 810 sw.transform_loop(this, false); 811 812 // If the loop is slp canonical analyze it 813 if (sw.early_return() == false) { 814 sw.unrolling_analysis(_local_loop_unroll_factor); 815 } 816 } 817 818 if (cl->has_passed_slp()) { 819 int slp_max_unroll_factor = cl->slp_max_unroll(); 820 if (slp_max_unroll_factor >= future_unroll_ct) { 821 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor; 822 if (new_limit > LoopUnrollLimit) { 823 if (TraceSuperWordLoopUnrollAnalysis) { 824 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit); 825 } 826 _local_loop_unroll_limit = new_limit; 827 } 828 } 829 } 830 } 831 } 832 833 //------------------------------policy_align----------------------------------- 834 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 835 // expression that does the alignment. Note that only one array base can be 836 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 837 // if we vectorize short memory ops into longer memory ops, we may want to 838 // increase alignment. 839 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 840 return false; 841 } 842 843 //------------------------------policy_range_check----------------------------- 844 // Return TRUE or FALSE if the loop should be range-check-eliminated. 845 // Actually we do iteration-splitting, a more powerful form of RCE. 846 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 847 if (!RangeCheckElimination) return false; 848 849 CountedLoopNode *cl = _head->as_CountedLoop(); 850 // If we unrolled with no intention of doing RCE and we later 851 // changed our minds, we got no pre-loop. Either we need to 852 // make a new pre-loop, or we gotta disallow RCE. 853 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 854 Node *trip_counter = cl->phi(); 855 856 // check for vectorized loops, some opts are no longer needed 857 if (cl->do_unroll_only()) return false; 858 859 // Check loop body for tests of trip-counter plus loop-invariant vs 860 // loop-invariant. 861 for (uint i = 0; i < _body.size(); i++) { 862 Node *iff = _body[i]; 863 if (iff->Opcode() == Op_If || 864 iff->Opcode() == Op_RangeCheck) { // Test? 865 866 // Comparing trip+off vs limit 867 Node *bol = iff->in(1); 868 if (bol->req() != 2) continue; // dead constant test 869 if (!bol->is_Bool()) { 870 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 871 continue; 872 } 873 if (bol->as_Bool()->_test._test == BoolTest::ne) 874 continue; // not RC 875 876 Node *cmp = bol->in(1); 877 Node *rc_exp = cmp->in(1); 878 Node *limit = cmp->in(2); 879 880 Node *limit_c = phase->get_ctrl(limit); 881 if( limit_c == phase->C->top() ) 882 return false; // Found dead test on live IF? No RCE! 883 if( is_member(phase->get_loop(limit_c) ) ) { 884 // Compare might have operands swapped; commute them 885 rc_exp = cmp->in(2); 886 limit = cmp->in(1); 887 limit_c = phase->get_ctrl(limit); 888 if( is_member(phase->get_loop(limit_c) ) ) 889 continue; // Both inputs are loop varying; cannot RCE 890 } 891 892 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 893 continue; 894 } 895 // Yeah! Found a test like 'trip+off vs limit' 896 // Test is an IfNode, has 2 projections. If BOTH are in the loop 897 // we need loop unswitching instead of iteration splitting. 898 if( is_loop_exit(iff) ) 899 return true; // Found reason to split iterations 900 } // End of is IF 901 } 902 903 return false; 904 } 905 906 //------------------------------policy_peel_only------------------------------- 907 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 908 // for unrolling loops with NO array accesses. 909 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 910 // check for vectorized loops, any peeling done was already applied 911 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false; 912 913 for( uint i = 0; i < _body.size(); i++ ) 914 if( _body[i]->is_Mem() ) 915 return false; 916 917 // No memory accesses at all! 918 return true; 919 } 920 921 //------------------------------clone_up_backedge_goo-------------------------- 922 // If Node n lives in the back_ctrl block and cannot float, we clone a private 923 // version of n in preheader_ctrl block and return that, otherwise return n. 924 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { 925 if( get_ctrl(n) != back_ctrl ) return n; 926 927 // Only visit once 928 if (visited.test_set(n->_idx)) { 929 Node *x = clones.find(n->_idx); 930 if (x != NULL) 931 return x; 932 return n; 933 } 934 935 Node *x = NULL; // If required, a clone of 'n' 936 // Check for 'n' being pinned in the backedge. 937 if( n->in(0) && n->in(0) == back_ctrl ) { 938 assert(clones.find(n->_idx) == NULL, "dead loop"); 939 x = n->clone(); // Clone a copy of 'n' to preheader 940 clones.push(x, n->_idx); 941 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 942 } 943 944 // Recursive fixup any other input edges into x. 945 // If there are no changes we can just return 'n', otherwise 946 // we need to clone a private copy and change it. 947 for( uint i = 1; i < n->req(); i++ ) { 948 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); 949 if( g != n->in(i) ) { 950 if( !x ) { 951 assert(clones.find(n->_idx) == NULL, "dead loop"); 952 x = n->clone(); 953 clones.push(x, n->_idx); 954 } 955 x->set_req(i, g); 956 } 957 } 958 if( x ) { // x can legally float to pre-header location 959 register_new_node( x, preheader_ctrl ); 960 return x; 961 } else { // raise n to cover LCA of uses 962 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 963 } 964 return n; 965 } 966 967 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { 968 Node* castii = new CastIINode(incr, TypeInt::INT, true); 969 castii->set_req(0, ctrl); 970 register_new_node(castii, ctrl); 971 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { 972 Node* n = incr->fast_out(i); 973 if (n->is_Phi() && n->in(0) == loop) { 974 int nrep = n->replace_edge(incr, castii); 975 return true; 976 } 977 } 978 return false; 979 } 980 981 //------------------------------insert_pre_post_loops-------------------------- 982 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 983 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 984 // alignment. Useful to unroll loops that do no array accesses. 985 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 986 987 #ifndef PRODUCT 988 if (TraceLoopOpts) { 989 if (peel_only) 990 tty->print("PeelMainPost "); 991 else 992 tty->print("PreMainPost "); 993 loop->dump_head(); 994 } 995 #endif 996 C->set_major_progress(); 997 998 // Find common pieces of the loop being guarded with pre & post loops 999 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1000 assert( main_head->is_normal_loop(), "" ); 1001 CountedLoopEndNode *main_end = main_head->loopexit(); 1002 guarantee(main_end != NULL, "no loop exit node"); 1003 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 1004 uint dd_main_head = dom_depth(main_head); 1005 uint max = main_head->outcnt(); 1006 1007 Node *pre_header= main_head->in(LoopNode::EntryControl); 1008 Node *init = main_head->init_trip(); 1009 Node *incr = main_end ->incr(); 1010 Node *limit = main_end ->limit(); 1011 Node *stride = main_end ->stride(); 1012 Node *cmp = main_end ->cmp_node(); 1013 BoolTest::mask b_test = main_end->test_trip(); 1014 1015 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 1016 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 1017 if( bol->outcnt() != 1 ) { 1018 bol = bol->clone(); 1019 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 1020 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol); 1021 } 1022 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 1023 if( cmp->outcnt() != 1 ) { 1024 cmp = cmp->clone(); 1025 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 1026 _igvn.replace_input_of(bol, 1, cmp); 1027 } 1028 1029 //------------------------------ 1030 // Step A: Create Post-Loop. 1031 Node* main_exit = main_end->proj_out(false); 1032 assert( main_exit->Opcode() == Op_IfFalse, "" ); 1033 int dd_main_exit = dom_depth(main_exit); 1034 1035 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 1036 // loop pre-header illegally has 2 control users (old & new loops). 1037 clone_loop( loop, old_new, dd_main_exit ); 1038 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 1039 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 1040 post_head->set_post_loop(main_head); 1041 1042 // Reduce the post-loop trip count. 1043 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1044 post_end->_prob = PROB_FAIR; 1045 1046 // Build the main-loop normal exit. 1047 IfFalseNode *new_main_exit = new IfFalseNode(main_end); 1048 _igvn.register_new_node_with_optimizer( new_main_exit ); 1049 set_idom(new_main_exit, main_end, dd_main_exit ); 1050 set_loop(new_main_exit, loop->_parent); 1051 1052 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 1053 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 1054 // (the main-loop trip-counter exit value) because we will be changing 1055 // the exit value (via unrolling) so we cannot constant-fold away the zero 1056 // trip guard until all unrolling is done. 1057 Node *zer_opaq = new Opaque1Node(C, incr); 1058 Node *zer_cmp = new CmpINode( zer_opaq, limit ); 1059 Node *zer_bol = new BoolNode( zer_cmp, b_test ); 1060 register_new_node( zer_opaq, new_main_exit ); 1061 register_new_node( zer_cmp , new_main_exit ); 1062 register_new_node( zer_bol , new_main_exit ); 1063 1064 // Build the IfNode 1065 IfNode *zer_iff = new IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 1066 _igvn.register_new_node_with_optimizer( zer_iff ); 1067 set_idom(zer_iff, new_main_exit, dd_main_exit); 1068 set_loop(zer_iff, loop->_parent); 1069 1070 // Plug in the false-path, taken if we need to skip post-loop 1071 _igvn.replace_input_of(main_exit, 0, zer_iff); 1072 set_idom(main_exit, zer_iff, dd_main_exit); 1073 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 1074 // Make the true-path, must enter the post loop 1075 Node *zer_taken = new IfTrueNode( zer_iff ); 1076 _igvn.register_new_node_with_optimizer( zer_taken ); 1077 set_idom(zer_taken, zer_iff, dd_main_exit); 1078 set_loop(zer_taken, loop->_parent); 1079 // Plug in the true path 1080 _igvn.hash_delete( post_head ); 1081 post_head->set_req(LoopNode::EntryControl, zer_taken); 1082 set_idom(post_head, zer_taken, dd_main_exit); 1083 1084 Arena *a = Thread::current()->resource_area(); 1085 VectorSet visited(a); 1086 Node_Stack clones(a, main_head->back_control()->outcnt()); 1087 // Step A3: Make the fall-in values to the post-loop come from the 1088 // fall-out values of the main-loop. 1089 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 1090 Node* main_phi = main_head->fast_out(i); 1091 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 1092 Node *post_phi = old_new[main_phi->_idx]; 1093 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 1094 post_head->init_control(), 1095 main_phi->in(LoopNode::LoopBackControl), 1096 visited, clones); 1097 _igvn.hash_delete(post_phi); 1098 post_phi->set_req( LoopNode::EntryControl, fallmain ); 1099 } 1100 } 1101 1102 // Update local caches for next stanza 1103 main_exit = new_main_exit; 1104 1105 1106 //------------------------------ 1107 // Step B: Create Pre-Loop. 1108 1109 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 1110 // loop pre-header illegally has 2 control users (old & new loops). 1111 clone_loop( loop, old_new, dd_main_head ); 1112 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1113 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1114 pre_head->set_pre_loop(main_head); 1115 Node *pre_incr = old_new[incr->_idx]; 1116 1117 // Reduce the pre-loop trip count. 1118 pre_end->_prob = PROB_FAIR; 1119 1120 // Find the pre-loop normal exit. 1121 Node* pre_exit = pre_end->proj_out(false); 1122 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 1123 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end); 1124 _igvn.register_new_node_with_optimizer( new_pre_exit ); 1125 set_idom(new_pre_exit, pre_end, dd_main_head); 1126 set_loop(new_pre_exit, loop->_parent); 1127 1128 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1129 // pre-loop, the main-loop may not execute at all. Later in life this 1130 // zero-trip guard will become the minimum-trip guard when we unroll 1131 // the main-loop. 1132 Node *min_opaq = new Opaque1Node(C, limit); 1133 Node *min_cmp = new CmpINode( pre_incr, min_opaq ); 1134 Node *min_bol = new BoolNode( min_cmp, b_test ); 1135 register_new_node( min_opaq, new_pre_exit ); 1136 register_new_node( min_cmp , new_pre_exit ); 1137 register_new_node( min_bol , new_pre_exit ); 1138 1139 // Build the IfNode (assume the main-loop is executed always). 1140 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1141 _igvn.register_new_node_with_optimizer( min_iff ); 1142 set_idom(min_iff, new_pre_exit, dd_main_head); 1143 set_loop(min_iff, loop->_parent); 1144 1145 // Plug in the false-path, taken if we need to skip main-loop 1146 _igvn.hash_delete( pre_exit ); 1147 pre_exit->set_req(0, min_iff); 1148 set_idom(pre_exit, min_iff, dd_main_head); 1149 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1150 // Make the true-path, must enter the main loop 1151 Node *min_taken = new IfTrueNode( min_iff ); 1152 _igvn.register_new_node_with_optimizer( min_taken ); 1153 set_idom(min_taken, min_iff, dd_main_head); 1154 set_loop(min_taken, loop->_parent); 1155 // Plug in the true path 1156 _igvn.hash_delete( main_head ); 1157 main_head->set_req(LoopNode::EntryControl, min_taken); 1158 set_idom(main_head, min_taken, dd_main_head); 1159 1160 visited.Clear(); 1161 clones.clear(); 1162 // Step B3: Make the fall-in values to the main-loop come from the 1163 // fall-out values of the pre-loop. 1164 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1165 Node* main_phi = main_head->fast_out(i2); 1166 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1167 Node *pre_phi = old_new[main_phi->_idx]; 1168 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1169 main_head->init_control(), 1170 pre_phi->in(LoopNode::LoopBackControl), 1171 visited, clones); 1172 _igvn.hash_delete(main_phi); 1173 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1174 } 1175 } 1176 1177 // Nodes inside the loop may be control dependent on a predicate 1178 // that was moved before the preloop. If the back branch of the main 1179 // or post loops becomes dead, those nodes won't be dependent on the 1180 // test that guards that loop nest anymore which could lead to an 1181 // incorrect array access because it executes independently of the 1182 // test that was guarding the loop nest. We add a special CastII on 1183 // the if branch that enters the loop, between the input induction 1184 // variable value and the induction variable Phi to preserve correct 1185 // dependencies. 1186 1187 // CastII for the post loop: 1188 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); 1189 assert(inserted, "no castII inserted"); 1190 1191 // CastII for the main loop: 1192 inserted = cast_incr_before_loop(pre_incr, min_taken, main_head); 1193 assert(inserted, "no castII inserted"); 1194 1195 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1196 // RCE and alignment may change this later. 1197 Node *cmp_end = pre_end->cmp_node(); 1198 assert( cmp_end->in(2) == limit, "" ); 1199 Node *pre_limit = new AddINode( init, stride ); 1200 1201 // Save the original loop limit in this Opaque1 node for 1202 // use by range check elimination. 1203 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit); 1204 1205 register_new_node( pre_limit, pre_head->in(0) ); 1206 register_new_node( pre_opaq , pre_head->in(0) ); 1207 1208 // Since no other users of pre-loop compare, I can hack limit directly 1209 assert( cmp_end->outcnt() == 1, "no other users" ); 1210 _igvn.hash_delete(cmp_end); 1211 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1212 1213 // Special case for not-equal loop bounds: 1214 // Change pre loop test, main loop test, and the 1215 // main loop guard test to use lt or gt depending on stride 1216 // direction: 1217 // positive stride use < 1218 // negative stride use > 1219 // 1220 // not-equal test is kept for post loop to handle case 1221 // when init > limit when stride > 0 (and reverse). 1222 1223 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1224 1225 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1226 // Modify pre loop end condition 1227 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1228 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test); 1229 register_new_node( new_bol0, pre_head->in(0) ); 1230 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0); 1231 // Modify main loop guard condition 1232 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1233 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test); 1234 register_new_node( new_bol1, new_pre_exit ); 1235 _igvn.hash_delete(min_iff); 1236 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1237 // Modify main loop end condition 1238 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1239 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test); 1240 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1241 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2); 1242 } 1243 1244 // Flag main loop 1245 main_head->set_main_loop(); 1246 if( peel_only ) main_head->set_main_no_pre_loop(); 1247 1248 // Subtract a trip count for the pre-loop. 1249 main_head->set_trip_count(main_head->trip_count() - 1); 1250 1251 // It's difficult to be precise about the trip-counts 1252 // for the pre/post loops. They are usually very short, 1253 // so guess that 4 trips is a reasonable value. 1254 post_head->set_profile_trip_cnt(4.0); 1255 pre_head->set_profile_trip_cnt(4.0); 1256 1257 // Now force out all loop-invariant dominating tests. The optimizer 1258 // finds some, but we _know_ they are all useless. 1259 peeled_dom_test_elim(loop,old_new); 1260 loop->record_for_igvn(); 1261 } 1262 1263 //------------------------------is_invariant----------------------------- 1264 // Return true if n is invariant 1265 bool IdealLoopTree::is_invariant(Node* n) const { 1266 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1267 if (n_c->is_top()) return false; 1268 return !is_member(_phase->get_loop(n_c)); 1269 } 1270 1271 1272 //------------------------------do_unroll-------------------------------------- 1273 // Unroll the loop body one step - make each trip do 2 iterations. 1274 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1275 assert(LoopUnrollLimit, ""); 1276 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1277 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1278 assert(loop_end, ""); 1279 #ifndef PRODUCT 1280 if (PrintOpto && VerifyLoopOptimizations) { 1281 tty->print("Unrolling "); 1282 loop->dump_head(); 1283 } else if (TraceLoopOpts) { 1284 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1285 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1286 } else { 1287 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1288 } 1289 loop->dump_head(); 1290 } 1291 1292 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1293 Arena* arena = Thread::current()->resource_area(); 1294 Node_Stack stack(arena, C->live_nodes() >> 2); 1295 Node_List rpo_list; 1296 VectorSet visited(arena); 1297 visited.set(loop_head->_idx); 1298 rpo( loop_head, stack, visited, rpo_list ); 1299 dump(loop, rpo_list.size(), rpo_list ); 1300 } 1301 #endif 1302 1303 // Remember loop node count before unrolling to detect 1304 // if rounds of unroll,optimize are making progress 1305 loop_head->set_node_count_before_unroll(loop->_body.size()); 1306 1307 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1308 Node *limit = loop_head->limit(); 1309 Node *init = loop_head->init_trip(); 1310 Node *stride = loop_head->stride(); 1311 1312 Node *opaq = NULL; 1313 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1314 // Search for zero-trip guard. 1315 assert( loop_head->is_main_loop(), "" ); 1316 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1317 Node *iff = ctrl->in(0); 1318 assert( iff->Opcode() == Op_If, "" ); 1319 Node *bol = iff->in(1); 1320 assert( bol->Opcode() == Op_Bool, "" ); 1321 Node *cmp = bol->in(1); 1322 assert( cmp->Opcode() == Op_CmpI, "" ); 1323 opaq = cmp->in(2); 1324 // Occasionally it's possible for a zero-trip guard Opaque1 node to be 1325 // optimized away and then another round of loop opts attempted. 1326 // We can not optimize this particular loop in that case. 1327 if (opaq->Opcode() != Op_Opaque1) 1328 return; // Cannot find zero-trip guard! Bail out! 1329 // Zero-trip test uses an 'opaque' node which is not shared. 1330 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1331 } 1332 1333 C->set_major_progress(); 1334 1335 Node* new_limit = NULL; 1336 if (UnrollLimitCheck) { 1337 int stride_con = stride->get_int(); 1338 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1339 uint old_trip_count = loop_head->trip_count(); 1340 // Verify that unroll policy result is still valid. 1341 assert(old_trip_count > 1 && 1342 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1343 1344 // Adjust loop limit to keep valid iterations number after unroll. 1345 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1346 // which may overflow. 1347 if (!adjust_min_trip) { 1348 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1349 "odd trip count for maximally unroll"); 1350 // Don't need to adjust limit for maximally unroll since trip count is even. 1351 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1352 // Loop's limit is constant. Loop's init could be constant when pre-loop 1353 // become peeled iteration. 1354 jlong init_con = init->get_int(); 1355 // We can keep old loop limit if iterations count stays the same: 1356 // old_trip_count == new_trip_count * 2 1357 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1358 // so we also don't need to adjust zero trip test. 1359 jlong limit_con = limit->get_int(); 1360 // (stride_con*2) not overflow since stride_con <= 8. 1361 int new_stride_con = stride_con * 2; 1362 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1363 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1364 // New trip count should satisfy next conditions. 1365 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1366 uint new_trip_count = (uint)trip_count; 1367 adjust_min_trip = (old_trip_count != new_trip_count*2); 1368 } 1369 1370 if (adjust_min_trip) { 1371 // Step 2: Adjust the trip limit if it is called for. 1372 // The adjustment amount is -stride. Need to make sure if the 1373 // adjustment underflows or overflows, then the main loop is skipped. 1374 Node* cmp = loop_end->cmp_node(); 1375 assert(cmp->in(2) == limit, "sanity"); 1376 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1377 1378 // Verify that policy_unroll result is still valid. 1379 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1380 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1381 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); 1382 1383 if (limit->is_Con()) { 1384 // The check in policy_unroll and the assert above guarantee 1385 // no underflow if limit is constant. 1386 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1387 set_ctrl(new_limit, C->root()); 1388 } else { 1389 // Limit is not constant. 1390 if (loop_head->unrolled_count() == 1) { // only for first unroll 1391 // Separate limit by Opaque node in case it is an incremented 1392 // variable from previous loop to avoid using pre-incremented 1393 // value which could increase register pressure. 1394 // Otherwise reorg_offsets() optimization will create a separate 1395 // Opaque node for each use of trip-counter and as result 1396 // zero trip guard limit will be different from loop limit. 1397 assert(has_ctrl(opaq), "should have it"); 1398 Node* opaq_ctrl = get_ctrl(opaq); 1399 limit = new Opaque2Node( C, limit ); 1400 register_new_node( limit, opaq_ctrl ); 1401 } 1402 if (stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo) || 1403 stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi)) { 1404 // No underflow. 1405 new_limit = new SubINode(limit, stride); 1406 } else { 1407 // (limit - stride) may underflow. 1408 // Clamp the adjustment value with MININT or MAXINT: 1409 // 1410 // new_limit = limit-stride 1411 // if (stride > 0) 1412 // new_limit = (limit < new_limit) ? MININT : new_limit; 1413 // else 1414 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1415 // 1416 BoolTest::mask bt = loop_end->test_trip(); 1417 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1418 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1419 set_ctrl(adj_max, C->root()); 1420 Node* old_limit = NULL; 1421 Node* adj_limit = NULL; 1422 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1423 if (loop_head->unrolled_count() > 1 && 1424 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1425 limit->in(CMoveNode::IfTrue) == adj_max && 1426 bol->as_Bool()->_test._test == bt && 1427 bol->in(1)->Opcode() == Op_CmpI && 1428 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1429 // Loop was unrolled before. 1430 // Optimize the limit to avoid nested CMove: 1431 // use original limit as old limit. 1432 old_limit = bol->in(1)->in(1); 1433 // Adjust previous adjusted limit. 1434 adj_limit = limit->in(CMoveNode::IfFalse); 1435 adj_limit = new SubINode(adj_limit, stride); 1436 } else { 1437 old_limit = limit; 1438 adj_limit = new SubINode(limit, stride); 1439 } 1440 assert(old_limit != NULL && adj_limit != NULL, ""); 1441 register_new_node( adj_limit, ctrl ); // adjust amount 1442 Node* adj_cmp = new CmpINode(old_limit, adj_limit); 1443 register_new_node( adj_cmp, ctrl ); 1444 Node* adj_bool = new BoolNode(adj_cmp, bt); 1445 register_new_node( adj_bool, ctrl ); 1446 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1447 } 1448 register_new_node(new_limit, ctrl); 1449 } 1450 assert(new_limit != NULL, ""); 1451 // Replace in loop test. 1452 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1453 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1454 // Don't need to create new test since only one user. 1455 _igvn.hash_delete(cmp); 1456 cmp->set_req(2, new_limit); 1457 } else { 1458 // Create new test since it is shared. 1459 Node* ctrl2 = loop_end->in(0); 1460 Node* cmp2 = cmp->clone(); 1461 cmp2->set_req(2, new_limit); 1462 register_new_node(cmp2, ctrl2); 1463 Node* bol2 = loop_end->in(1)->clone(); 1464 bol2->set_req(1, cmp2); 1465 register_new_node(bol2, ctrl2); 1466 _igvn.replace_input_of(loop_end, 1, bol2); 1467 } 1468 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1469 // Make it a 1-trip test (means at least 2 trips). 1470 1471 // Guard test uses an 'opaque' node which is not shared. Hence I 1472 // can edit it's inputs directly. Hammer in the new limit for the 1473 // minimum-trip guard. 1474 assert(opaq->outcnt() == 1, ""); 1475 _igvn.replace_input_of(opaq, 1, new_limit); 1476 } 1477 1478 // Adjust max trip count. The trip count is intentionally rounded 1479 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1480 // the main, unrolled, part of the loop will never execute as it is protected 1481 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1482 // and later determined that part of the unrolled loop was dead. 1483 loop_head->set_trip_count(old_trip_count / 2); 1484 1485 // Double the count of original iterations in the unrolled loop body. 1486 loop_head->double_unrolled_count(); 1487 1488 } else { // LoopLimitCheck 1489 1490 // Adjust max trip count. The trip count is intentionally rounded 1491 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1492 // the main, unrolled, part of the loop will never execute as it is protected 1493 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1494 // and later determined that part of the unrolled loop was dead. 1495 loop_head->set_trip_count(loop_head->trip_count() / 2); 1496 1497 // Double the count of original iterations in the unrolled loop body. 1498 loop_head->double_unrolled_count(); 1499 1500 // ----------- 1501 // Step 2: Cut back the trip counter for an unroll amount of 2. 1502 // Loop will normally trip (limit - init)/stride_con. Since it's a 1503 // CountedLoop this is exact (stride divides limit-init exactly). 1504 // We are going to double the loop body, so we want to knock off any 1505 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 1506 Node *span = new SubINode( limit, init ); 1507 register_new_node( span, ctrl ); 1508 Node *trip = new DivINode( 0, span, stride ); 1509 register_new_node( trip, ctrl ); 1510 Node *mtwo = _igvn.intcon(-2); 1511 set_ctrl(mtwo, C->root()); 1512 Node *rond = new AndINode( trip, mtwo ); 1513 register_new_node( rond, ctrl ); 1514 Node *spn2 = new MulINode( rond, stride ); 1515 register_new_node( spn2, ctrl ); 1516 new_limit = new AddINode( spn2, init ); 1517 register_new_node( new_limit, ctrl ); 1518 1519 // Hammer in the new limit 1520 Node *ctrl2 = loop_end->in(0); 1521 Node *cmp2 = new CmpINode( loop_head->incr(), new_limit ); 1522 register_new_node( cmp2, ctrl2 ); 1523 Node *bol2 = new BoolNode( cmp2, loop_end->test_trip() ); 1524 register_new_node( bol2, ctrl2 ); 1525 _igvn.replace_input_of(loop_end, CountedLoopEndNode::TestValue, bol2); 1526 1527 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1528 // Make it a 1-trip test (means at least 2 trips). 1529 if( adjust_min_trip ) { 1530 assert( new_limit != NULL, "" ); 1531 // Guard test uses an 'opaque' node which is not shared. Hence I 1532 // can edit it's inputs directly. Hammer in the new limit for the 1533 // minimum-trip guard. 1534 assert( opaq->outcnt() == 1, "" ); 1535 _igvn.hash_delete(opaq); 1536 opaq->set_req(1, new_limit); 1537 } 1538 } // LoopLimitCheck 1539 1540 // --------- 1541 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1542 // represents the odd iterations; since the loop trips an even number of 1543 // times its backedge is never taken. Kill the backedge. 1544 uint dd = dom_depth(loop_head); 1545 clone_loop( loop, old_new, dd ); 1546 1547 // Make backedges of the clone equal to backedges of the original. 1548 // Make the fall-in from the original come from the fall-out of the clone. 1549 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1550 Node* phi = loop_head->fast_out(j); 1551 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1552 Node *newphi = old_new[phi->_idx]; 1553 _igvn.hash_delete( phi ); 1554 _igvn.hash_delete( newphi ); 1555 1556 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1557 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1558 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1559 } 1560 } 1561 Node *clone_head = old_new[loop_head->_idx]; 1562 _igvn.hash_delete( clone_head ); 1563 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1564 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1565 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1566 loop->_head = clone_head; // New loop header 1567 1568 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1569 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1570 1571 // Kill the clone's backedge 1572 Node *newcle = old_new[loop_end->_idx]; 1573 _igvn.hash_delete( newcle ); 1574 Node *one = _igvn.intcon(1); 1575 set_ctrl(one, C->root()); 1576 newcle->set_req(1, one); 1577 // Force clone into same loop body 1578 uint max = loop->_body.size(); 1579 for( uint k = 0; k < max; k++ ) { 1580 Node *old = loop->_body.at(k); 1581 Node *nnn = old_new[old->_idx]; 1582 loop->_body.push(nnn); 1583 if (!has_ctrl(old)) 1584 set_loop(nnn, loop); 1585 } 1586 1587 loop->record_for_igvn(); 1588 1589 #ifndef PRODUCT 1590 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1591 tty->print("\nnew loop after unroll\n"); loop->dump_head(); 1592 for (uint i = 0; i < loop->_body.size(); i++) { 1593 loop->_body.at(i)->dump(); 1594 } 1595 if(C->clone_map().is_debug()) { 1596 tty->print("\nCloneMap\n"); 1597 Dict* dict = C->clone_map().dict(); 1598 DictI i(dict); 1599 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size()); 1600 for (int ii = 0; i.test(); ++i, ++ii) { 1601 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key)); 1602 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen()); 1603 if (ii % 10 == 9) { 1604 tty->print_cr(" "); 1605 } 1606 } 1607 tty->print_cr(" "); 1608 } 1609 } 1610 #endif 1611 1612 } 1613 1614 //------------------------------do_maximally_unroll---------------------------- 1615 1616 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1617 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1618 assert(cl->has_exact_trip_count(), "trip count is not exact"); 1619 assert(cl->trip_count() > 0, ""); 1620 #ifndef PRODUCT 1621 if (TraceLoopOpts) { 1622 tty->print("MaxUnroll %d ", cl->trip_count()); 1623 loop->dump_head(); 1624 } 1625 #endif 1626 1627 // If loop is tripping an odd number of times, peel odd iteration 1628 if ((cl->trip_count() & 1) == 1) { 1629 do_peeling(loop, old_new); 1630 } 1631 1632 // Now its tripping an even number of times remaining. Double loop body. 1633 // Do not adjust pre-guards; they are not needed and do not exist. 1634 if (cl->trip_count() > 0) { 1635 assert((cl->trip_count() & 1) == 0, "missed peeling"); 1636 do_unroll(loop, old_new, false); 1637 } 1638 } 1639 1640 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) { 1641 if (SuperWordReductions == false) return; 1642 1643 CountedLoopNode* loop_head = loop->_head->as_CountedLoop(); 1644 if (loop_head->unrolled_count() > 1) { 1645 return; 1646 } 1647 1648 Node* trip_phi = loop_head->phi(); 1649 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) { 1650 Node* phi = loop_head->fast_out(i); 1651 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) { 1652 // For definitions which are loop inclusive and not tripcounts. 1653 Node* def_node = phi->in(LoopNode::LoopBackControl); 1654 1655 if (def_node != NULL) { 1656 Node* n_ctrl = get_ctrl(def_node); 1657 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) { 1658 // Now test it to see if it fits the standard pattern for a reduction operator. 1659 int opc = def_node->Opcode(); 1660 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) { 1661 if (!def_node->is_reduction()) { // Not marked yet 1662 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it 1663 bool ok = false; 1664 for (unsigned j = 1; j < def_node->req(); j++) { 1665 Node* in = def_node->in(j); 1666 if (in == phi) { 1667 ok = true; 1668 break; 1669 } 1670 } 1671 1672 // do nothing if we did not match the initial criteria 1673 if (ok == false) { 1674 continue; 1675 } 1676 1677 // The result of the reduction must not be used in the loop 1678 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) { 1679 Node* u = def_node->fast_out(i); 1680 if (has_ctrl(u) && !loop->is_member(get_loop(get_ctrl(u)))) { 1681 continue; 1682 } 1683 if (u == phi) { 1684 continue; 1685 } 1686 ok = false; 1687 } 1688 1689 // iff the uses conform 1690 if (ok) { 1691 def_node->add_flag(Node::Flag_is_reduction); 1692 loop_head->mark_has_reductions(); 1693 } 1694 } 1695 } 1696 } 1697 } 1698 } 1699 } 1700 } 1701 1702 //------------------------------dominates_backedge--------------------------------- 1703 // Returns true if ctrl is executed on every complete iteration 1704 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1705 assert(ctrl->is_CFG(), "must be control"); 1706 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1707 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1708 } 1709 1710 //------------------------------adjust_limit----------------------------------- 1711 // Helper function for add_constraint(). 1712 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { 1713 // Compute "I :: (limit-offset)/scale" 1714 Node *con = new SubINode(rc_limit, offset); 1715 register_new_node(con, pre_ctrl); 1716 Node *X = new DivINode(0, con, scale); 1717 register_new_node(X, pre_ctrl); 1718 1719 // Adjust loop limit 1720 loop_limit = (stride_con > 0) 1721 ? (Node*)(new MinINode(loop_limit, X)) 1722 : (Node*)(new MaxINode(loop_limit, X)); 1723 register_new_node(loop_limit, pre_ctrl); 1724 return loop_limit; 1725 } 1726 1727 //------------------------------add_constraint--------------------------------- 1728 // Constrain the main loop iterations so the conditions: 1729 // low_limit <= scale_con * I + offset < upper_limit 1730 // always holds true. That is, either increase the number of iterations in 1731 // the pre-loop or the post-loop until the condition holds true in the main 1732 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1733 // stride and scale are constants (offset and limit often are). 1734 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 ) { 1735 // For positive stride, the pre-loop limit always uses a MAX function 1736 // and the main loop a MIN function. For negative stride these are 1737 // reversed. 1738 1739 // Also for positive stride*scale the affine function is increasing, so the 1740 // pre-loop must check for underflow and the post-loop for overflow. 1741 // Negative stride*scale reverses this; pre-loop checks for overflow and 1742 // post-loop for underflow. 1743 1744 Node *scale = _igvn.intcon(scale_con); 1745 set_ctrl(scale, C->root()); 1746 1747 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 1748 // The overflow limit: scale*I+offset < upper_limit 1749 // For main-loop compute 1750 // ( if (scale > 0) /* and stride > 0 */ 1751 // I < (upper_limit-offset)/scale 1752 // else /* scale < 0 and stride < 0 */ 1753 // I > (upper_limit-offset)/scale 1754 // ) 1755 // 1756 // (upper_limit-offset) may overflow or underflow. 1757 // But it is fine since main loop will either have 1758 // less iterations or will be skipped in such case. 1759 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); 1760 1761 // The underflow limit: low_limit <= scale*I+offset. 1762 // For pre-loop compute 1763 // NOT(scale*I+offset >= low_limit) 1764 // scale*I+offset < low_limit 1765 // ( if (scale > 0) /* and stride > 0 */ 1766 // I < (low_limit-offset)/scale 1767 // else /* scale < 0 and stride < 0 */ 1768 // I > (low_limit-offset)/scale 1769 // ) 1770 1771 if (low_limit->get_int() == -max_jint) { 1772 if (!RangeLimitCheck) return; 1773 // We need this guard when scale*pre_limit+offset >= limit 1774 // due to underflow. So we need execute pre-loop until 1775 // scale*I+offset >= min_int. But (min_int-offset) will 1776 // underflow when offset > 0 and X will be > original_limit 1777 // when stride > 0. To avoid it we replace positive offset with 0. 1778 // 1779 // Also (min_int+1 == -max_int) is used instead of min_int here 1780 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1781 Node* shift = _igvn.intcon(31); 1782 set_ctrl(shift, C->root()); 1783 Node* sign = new RShiftINode(offset, shift); 1784 register_new_node(sign, pre_ctrl); 1785 offset = new AndINode(offset, sign); 1786 register_new_node(offset, pre_ctrl); 1787 } else { 1788 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1789 // The only problem we have here when offset == min_int 1790 // since (0-min_int) == min_int. It may be fine for stride > 0 1791 // but for stride < 0 X will be < original_limit. To avoid it 1792 // max(pre_limit, original_limit) is used in do_range_check(). 1793 } 1794 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1795 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); 1796 1797 } else { // stride_con*scale_con < 0 1798 // For negative stride*scale pre-loop checks for overflow and 1799 // post-loop for underflow. 1800 // 1801 // The overflow limit: scale*I+offset < upper_limit 1802 // For pre-loop compute 1803 // NOT(scale*I+offset < upper_limit) 1804 // scale*I+offset >= upper_limit 1805 // scale*I+offset+1 > upper_limit 1806 // ( if (scale < 0) /* and stride > 0 */ 1807 // I < (upper_limit-(offset+1))/scale 1808 // else /* scale > 0 and stride < 0 */ 1809 // I > (upper_limit-(offset+1))/scale 1810 // ) 1811 // 1812 // (upper_limit-offset-1) may underflow or overflow. 1813 // To avoid it min(pre_limit, original_limit) is used 1814 // in do_range_check() for stride > 0 and max() for < 0. 1815 Node *one = _igvn.intcon(1); 1816 set_ctrl(one, C->root()); 1817 1818 Node *plus_one = new AddINode(offset, one); 1819 register_new_node( plus_one, pre_ctrl ); 1820 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1821 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); 1822 1823 if (low_limit->get_int() == -max_jint) { 1824 if (!RangeLimitCheck) return; 1825 // We need this guard when scale*main_limit+offset >= limit 1826 // due to underflow. So we need execute main-loop while 1827 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 1828 // underflow when (offset+1) > 0 and X will be < main_limit 1829 // when scale < 0 (and stride > 0). To avoid it we replace 1830 // positive (offset+1) with 0. 1831 // 1832 // Also (min_int+1 == -max_int) is used instead of min_int here 1833 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1834 Node* shift = _igvn.intcon(31); 1835 set_ctrl(shift, C->root()); 1836 Node* sign = new RShiftINode(plus_one, shift); 1837 register_new_node(sign, pre_ctrl); 1838 plus_one = new AndINode(plus_one, sign); 1839 register_new_node(plus_one, pre_ctrl); 1840 } else { 1841 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1842 // The only problem we have here when offset == max_int 1843 // since (max_int+1) == min_int and (0-min_int) == min_int. 1844 // But it is fine since main loop will either have 1845 // less iterations or will be skipped in such case. 1846 } 1847 // The underflow limit: low_limit <= scale*I+offset. 1848 // For main-loop compute 1849 // scale*I+offset+1 > low_limit 1850 // ( if (scale < 0) /* and stride > 0 */ 1851 // I < (low_limit-(offset+1))/scale 1852 // else /* scale > 0 and stride < 0 */ 1853 // I > (low_limit-(offset+1))/scale 1854 // ) 1855 1856 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); 1857 } 1858 } 1859 1860 1861 //------------------------------is_scaled_iv--------------------------------- 1862 // Return true if exp is a constant times an induction var 1863 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1864 if (exp == iv) { 1865 if (p_scale != NULL) { 1866 *p_scale = 1; 1867 } 1868 return true; 1869 } 1870 int opc = exp->Opcode(); 1871 if (opc == Op_MulI) { 1872 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1873 if (p_scale != NULL) { 1874 *p_scale = exp->in(2)->get_int(); 1875 } 1876 return true; 1877 } 1878 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1879 if (p_scale != NULL) { 1880 *p_scale = exp->in(1)->get_int(); 1881 } 1882 return true; 1883 } 1884 } else if (opc == Op_LShiftI) { 1885 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1886 if (p_scale != NULL) { 1887 *p_scale = 1 << exp->in(2)->get_int(); 1888 } 1889 return true; 1890 } 1891 } 1892 return false; 1893 } 1894 1895 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1896 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1897 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1898 if (is_scaled_iv(exp, iv, p_scale)) { 1899 if (p_offset != NULL) { 1900 Node *zero = _igvn.intcon(0); 1901 set_ctrl(zero, C->root()); 1902 *p_offset = zero; 1903 } 1904 return true; 1905 } 1906 int opc = exp->Opcode(); 1907 if (opc == Op_AddI) { 1908 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1909 if (p_offset != NULL) { 1910 *p_offset = exp->in(2); 1911 } 1912 return true; 1913 } 1914 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1915 if (p_offset != NULL) { 1916 *p_offset = exp->in(1); 1917 } 1918 return true; 1919 } 1920 if (exp->in(2)->is_Con()) { 1921 Node* offset2 = NULL; 1922 if (depth < 2 && 1923 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1924 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1925 if (p_offset != NULL) { 1926 Node *ctrl_off2 = get_ctrl(offset2); 1927 Node* offset = new AddINode(offset2, exp->in(2)); 1928 register_new_node(offset, ctrl_off2); 1929 *p_offset = offset; 1930 } 1931 return true; 1932 } 1933 } 1934 } else if (opc == Op_SubI) { 1935 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1936 if (p_offset != NULL) { 1937 Node *zero = _igvn.intcon(0); 1938 set_ctrl(zero, C->root()); 1939 Node *ctrl_off = get_ctrl(exp->in(2)); 1940 Node* offset = new SubINode(zero, exp->in(2)); 1941 register_new_node(offset, ctrl_off); 1942 *p_offset = offset; 1943 } 1944 return true; 1945 } 1946 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1947 if (p_offset != NULL) { 1948 *p_scale *= -1; 1949 *p_offset = exp->in(1); 1950 } 1951 return true; 1952 } 1953 } 1954 return false; 1955 } 1956 1957 //------------------------------do_range_check--------------------------------- 1958 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1959 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1960 #ifndef PRODUCT 1961 if (PrintOpto && VerifyLoopOptimizations) { 1962 tty->print("Range Check Elimination "); 1963 loop->dump_head(); 1964 } else if (TraceLoopOpts) { 1965 tty->print("RangeCheck "); 1966 loop->dump_head(); 1967 } 1968 #endif 1969 assert(RangeCheckElimination, ""); 1970 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1971 assert(cl->is_main_loop(), ""); 1972 1973 // protect against stride not being a constant 1974 if (!cl->stride_is_con()) 1975 return; 1976 1977 // Find the trip counter; we are iteration splitting based on it 1978 Node *trip_counter = cl->phi(); 1979 // Find the main loop limit; we will trim it's iterations 1980 // to not ever trip end tests 1981 Node *main_limit = cl->limit(); 1982 1983 // Need to find the main-loop zero-trip guard 1984 Node *ctrl = cl->in(LoopNode::EntryControl); 1985 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 1986 Node *iffm = ctrl->in(0); 1987 assert(iffm->Opcode() == Op_If, ""); 1988 Node *bolzm = iffm->in(1); 1989 assert(bolzm->Opcode() == Op_Bool, ""); 1990 Node *cmpzm = bolzm->in(1); 1991 assert(cmpzm->is_Cmp(), ""); 1992 Node *opqzm = cmpzm->in(2); 1993 // Can not optimize a loop if zero-trip Opaque1 node is optimized 1994 // away and then another round of loop opts attempted. 1995 if (opqzm->Opcode() != Op_Opaque1) 1996 return; 1997 assert(opqzm->in(1) == main_limit, "do not understand situation"); 1998 1999 // Find the pre-loop limit; we will expand its iterations to 2000 // not ever trip low tests. 2001 Node *p_f = iffm->in(0); 2002 // pre loop may have been optimized out 2003 if (p_f->Opcode() != Op_IfFalse) { 2004 return; 2005 } 2006 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2007 assert(pre_end->loopnode()->is_pre_loop(), ""); 2008 Node *pre_opaq1 = pre_end->limit(); 2009 // Occasionally it's possible for a pre-loop Opaque1 node to be 2010 // optimized away and then another round of loop opts attempted. 2011 // We can not optimize this particular loop in that case. 2012 if (pre_opaq1->Opcode() != Op_Opaque1) 2013 return; 2014 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 2015 Node *pre_limit = pre_opaq->in(1); 2016 2017 // Where do we put new limit calculations 2018 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 2019 2020 // Ensure the original loop limit is available from the 2021 // pre-loop Opaque1 node. 2022 Node *orig_limit = pre_opaq->original_loop_limit(); 2023 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 2024 return; 2025 2026 // Must know if its a count-up or count-down loop 2027 2028 int stride_con = cl->stride_con(); 2029 Node *zero = _igvn.intcon(0); 2030 Node *one = _igvn.intcon(1); 2031 // Use symmetrical int range [-max_jint,max_jint] 2032 Node *mini = _igvn.intcon(-max_jint); 2033 set_ctrl(zero, C->root()); 2034 set_ctrl(one, C->root()); 2035 set_ctrl(mini, C->root()); 2036 2037 // Range checks that do not dominate the loop backedge (ie. 2038 // conditionally executed) can lengthen the pre loop limit beyond 2039 // the original loop limit. To prevent this, the pre limit is 2040 // (for stride > 0) MINed with the original loop limit (MAXed 2041 // stride < 0) when some range_check (rc) is conditionally 2042 // executed. 2043 bool conditional_rc = false; 2044 2045 // Check loop body for tests of trip-counter plus loop-invariant vs 2046 // loop-invariant. 2047 for( uint i = 0; i < loop->_body.size(); i++ ) { 2048 Node *iff = loop->_body[i]; 2049 if (iff->Opcode() == Op_If || 2050 iff->Opcode() == Op_RangeCheck) { // Test? 2051 // Test is an IfNode, has 2 projections. If BOTH are in the loop 2052 // we need loop unswitching instead of iteration splitting. 2053 Node *exit = loop->is_loop_exit(iff); 2054 if( !exit ) continue; 2055 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 2056 2057 // Get boolean condition to test 2058 Node *i1 = iff->in(1); 2059 if( !i1->is_Bool() ) continue; 2060 BoolNode *bol = i1->as_Bool(); 2061 BoolTest b_test = bol->_test; 2062 // Flip sense of test if exit condition is flipped 2063 if( flip ) 2064 b_test = b_test.negate(); 2065 2066 // Get compare 2067 Node *cmp = bol->in(1); 2068 2069 // Look for trip_counter + offset vs limit 2070 Node *rc_exp = cmp->in(1); 2071 Node *limit = cmp->in(2); 2072 jint scale_con= 1; // Assume trip counter not scaled 2073 2074 Node *limit_c = get_ctrl(limit); 2075 if( loop->is_member(get_loop(limit_c) ) ) { 2076 // Compare might have operands swapped; commute them 2077 b_test = b_test.commute(); 2078 rc_exp = cmp->in(2); 2079 limit = cmp->in(1); 2080 limit_c = get_ctrl(limit); 2081 if( loop->is_member(get_loop(limit_c) ) ) 2082 continue; // Both inputs are loop varying; cannot RCE 2083 } 2084 // Here we know 'limit' is loop invariant 2085 2086 // 'limit' maybe pinned below the zero trip test (probably from a 2087 // previous round of rce), in which case, it can't be used in the 2088 // zero trip test expression which must occur before the zero test's if. 2089 if( limit_c == ctrl ) { 2090 continue; // Don't rce this check but continue looking for other candidates. 2091 } 2092 2093 // Check for scaled induction variable plus an offset 2094 Node *offset = NULL; 2095 2096 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 2097 continue; 2098 } 2099 2100 Node *offset_c = get_ctrl(offset); 2101 if( loop->is_member( get_loop(offset_c) ) ) 2102 continue; // Offset is not really loop invariant 2103 // Here we know 'offset' is loop invariant. 2104 2105 // As above for the 'limit', the 'offset' maybe pinned below the 2106 // zero trip test. 2107 if( offset_c == ctrl ) { 2108 continue; // Don't rce this check but continue looking for other candidates. 2109 } 2110 #ifdef ASSERT 2111 if (TraceRangeLimitCheck) { 2112 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 2113 bol->dump(2); 2114 } 2115 #endif 2116 // At this point we have the expression as: 2117 // scale_con * trip_counter + offset :: limit 2118 // where scale_con, offset and limit are loop invariant. Trip_counter 2119 // monotonically increases by stride_con, a constant. Both (or either) 2120 // stride_con and scale_con can be negative which will flip about the 2121 // sense of the test. 2122 2123 // Adjust pre and main loop limits to guard the correct iteration set 2124 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 2125 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 2126 // The underflow and overflow limits: 0 <= scale*I+offset < limit 2127 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); 2128 if (!conditional_rc) { 2129 // (0-offset)/scale could be outside of loop iterations range. 2130 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2131 } 2132 } else { 2133 if (PrintOpto) { 2134 tty->print_cr("missed RCE opportunity"); 2135 } 2136 continue; // In release mode, ignore it 2137 } 2138 } else { // Otherwise work on normal compares 2139 switch( b_test._test ) { 2140 case BoolTest::gt: 2141 // Fall into GE case 2142 case BoolTest::ge: 2143 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 2144 scale_con = -scale_con; 2145 offset = new SubINode( zero, offset ); 2146 register_new_node( offset, pre_ctrl ); 2147 limit = new SubINode( zero, limit ); 2148 register_new_node( limit, pre_ctrl ); 2149 // Fall into LE case 2150 case BoolTest::le: 2151 if (b_test._test != BoolTest::gt) { 2152 // Convert X <= Y to X < Y+1 2153 limit = new AddINode( limit, one ); 2154 register_new_node( limit, pre_ctrl ); 2155 } 2156 // Fall into LT case 2157 case BoolTest::lt: 2158 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 2159 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 2160 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 2161 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); 2162 if (!conditional_rc) { 2163 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 2164 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 2165 // still be outside of loop range. 2166 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2167 } 2168 break; 2169 default: 2170 if (PrintOpto) { 2171 tty->print_cr("missed RCE opportunity"); 2172 } 2173 continue; // Unhandled case 2174 } 2175 } 2176 2177 // Kill the eliminated test 2178 C->set_major_progress(); 2179 Node *kill_con = _igvn.intcon( 1-flip ); 2180 set_ctrl(kill_con, C->root()); 2181 _igvn.replace_input_of(iff, 1, kill_con); 2182 // Find surviving projection 2183 assert(iff->is_If(), ""); 2184 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 2185 // Find loads off the surviving projection; remove their control edge 2186 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2187 Node* cd = dp->fast_out(i); // Control-dependent node 2188 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop 2189 // Allow the load to float around in the loop, or before it 2190 // but NOT before the pre-loop. 2191 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 2192 --i; 2193 --imax; 2194 } 2195 } 2196 2197 } // End of is IF 2198 2199 } 2200 2201 // Update loop limits 2202 if (conditional_rc) { 2203 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit) 2204 : (Node*)new MaxINode(pre_limit, orig_limit); 2205 register_new_node(pre_limit, pre_ctrl); 2206 } 2207 _igvn.replace_input_of(pre_opaq, 1, pre_limit); 2208 2209 // Note:: we are making the main loop limit no longer precise; 2210 // need to round up based on stride. 2211 cl->set_nonexact_trip_count(); 2212 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case 2213 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 2214 // Hopefully, compiler will optimize for powers of 2. 2215 Node *ctrl = get_ctrl(main_limit); 2216 Node *stride = cl->stride(); 2217 Node *init = cl->init_trip()->uncast(); 2218 Node *span = new SubINode(main_limit,init); 2219 register_new_node(span,ctrl); 2220 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 2221 Node *add = new AddINode(span,rndup); 2222 register_new_node(add,ctrl); 2223 Node *div = new DivINode(0,add,stride); 2224 register_new_node(div,ctrl); 2225 Node *mul = new MulINode(div,stride); 2226 register_new_node(mul,ctrl); 2227 Node *newlim = new AddINode(mul,init); 2228 register_new_node(newlim,ctrl); 2229 main_limit = newlim; 2230 } 2231 2232 Node *main_cle = cl->loopexit(); 2233 Node *main_bol = main_cle->in(1); 2234 // Hacking loop bounds; need private copies of exit test 2235 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 2236 main_bol = main_bol->clone();// Clone a private BoolNode 2237 register_new_node( main_bol, main_cle->in(0) ); 2238 _igvn.replace_input_of(main_cle, 1, main_bol); 2239 } 2240 Node *main_cmp = main_bol->in(1); 2241 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 2242 main_cmp = main_cmp->clone();// Clone a private CmpNode 2243 register_new_node( main_cmp, main_cle->in(0) ); 2244 _igvn.replace_input_of(main_bol, 1, main_cmp); 2245 } 2246 // Hack the now-private loop bounds 2247 _igvn.replace_input_of(main_cmp, 2, main_limit); 2248 // The OpaqueNode is unshared by design 2249 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 2250 _igvn.replace_input_of(opqzm, 1, main_limit); 2251 } 2252 2253 //------------------------------DCE_loop_body---------------------------------- 2254 // Remove simplistic dead code from loop body 2255 void IdealLoopTree::DCE_loop_body() { 2256 for( uint i = 0; i < _body.size(); i++ ) 2257 if( _body.at(i)->outcnt() == 0 ) 2258 _body.map( i--, _body.pop() ); 2259 } 2260 2261 2262 //------------------------------adjust_loop_exit_prob-------------------------- 2263 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2264 // Replace with a 1-in-10 exit guess. 2265 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 2266 Node *test = tail(); 2267 while( test != _head ) { 2268 uint top = test->Opcode(); 2269 if( top == Op_IfTrue || top == Op_IfFalse ) { 2270 int test_con = ((ProjNode*)test)->_con; 2271 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2272 IfNode *iff = test->in(0)->as_If(); 2273 if( iff->outcnt() == 2 ) { // Ignore dead tests 2274 Node *bol = iff->in(1); 2275 if( bol && bol->req() > 1 && bol->in(1) && 2276 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 2277 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 2278 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 2279 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 2280 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 2281 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 2282 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 2283 return; // Allocation loops RARELY take backedge 2284 // Find the OTHER exit path from the IF 2285 Node* ex = iff->proj_out(1-test_con); 2286 float p = iff->_prob; 2287 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 2288 if( top == Op_IfTrue ) { 2289 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2290 iff->_prob = PROB_STATIC_FREQUENT; 2291 } 2292 } else { 2293 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2294 iff->_prob = PROB_STATIC_INFREQUENT; 2295 } 2296 } 2297 } 2298 } 2299 } 2300 test = phase->idom(test); 2301 } 2302 } 2303 2304 #ifdef ASSERT 2305 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) { 2306 Node *ctrl = cl->in(LoopNode::EntryControl); 2307 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 2308 Node *iffm = ctrl->in(0); 2309 assert(iffm->Opcode() == Op_If, ""); 2310 Node *p_f = iffm->in(0); 2311 assert(p_f->Opcode() == Op_IfFalse, ""); 2312 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2313 assert(pre_end->loopnode()->is_pre_loop(), ""); 2314 return pre_end->loopnode(); 2315 } 2316 #endif 2317 2318 // Remove the main and post loops and make the pre loop execute all 2319 // iterations. Useful when the pre loop is found empty. 2320 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) { 2321 CountedLoopEndNode* pre_end = cl->loopexit(); 2322 Node* pre_cmp = pre_end->cmp_node(); 2323 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) { 2324 // Only safe to remove the main loop if the compiler optimized it 2325 // out based on an unknown number of iterations 2326 return; 2327 } 2328 2329 // Can we find the main loop? 2330 if (_next == NULL) { 2331 return; 2332 } 2333 2334 Node* next_head = _next->_head; 2335 if (!next_head->is_CountedLoop()) { 2336 return; 2337 } 2338 2339 CountedLoopNode* main_head = next_head->as_CountedLoop(); 2340 if (!main_head->is_main_loop()) { 2341 return; 2342 } 2343 2344 assert(locate_pre_from_main(main_head) == cl, "bad main loop"); 2345 Node* main_iff = main_head->in(LoopNode::EntryControl)->in(0); 2346 2347 // Remove the Opaque1Node of the pre loop and make it execute all iterations 2348 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2)); 2349 // Remove the Opaque1Node of the main loop so it can be optimized out 2350 Node* main_cmp = main_iff->in(1)->in(1); 2351 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?"); 2352 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1)); 2353 } 2354 2355 //------------------------------policy_do_remove_empty_loop-------------------- 2356 // Micro-benchmark spamming. Policy is to always remove empty loops. 2357 // The 'DO' part is to replace the trip counter with the value it will 2358 // have on the last iteration. This will break the loop. 2359 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 2360 // Minimum size must be empty loop 2361 if (_body.size() > EMPTY_LOOP_SIZE) 2362 return false; 2363 2364 if (!_head->is_CountedLoop()) 2365 return false; // Dead loop 2366 CountedLoopNode *cl = _head->as_CountedLoop(); 2367 if (!cl->is_valid_counted_loop()) 2368 return false; // Malformed loop 2369 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 2370 return false; // Infinite loop 2371 2372 if (cl->is_pre_loop()) { 2373 // If the loop we are removing is a pre-loop then the main and 2374 // post loop can be removed as well 2375 remove_main_post_loops(cl, phase); 2376 } 2377 2378 #ifdef ASSERT 2379 // Ensure only one phi which is the iv. 2380 Node* iv = NULL; 2381 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 2382 Node* n = cl->fast_out(i); 2383 if (n->Opcode() == Op_Phi) { 2384 assert(iv == NULL, "Too many phis" ); 2385 iv = n; 2386 } 2387 } 2388 assert(iv == cl->phi(), "Wrong phi" ); 2389 #endif 2390 2391 // main and post loops have explicitly created zero trip guard 2392 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 2393 if (needs_guard) { 2394 // Skip guard if values not overlap. 2395 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 2396 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 2397 int stride_con = cl->stride_con(); 2398 if (stride_con > 0) { 2399 needs_guard = (init_t->_hi >= limit_t->_lo); 2400 } else { 2401 needs_guard = (init_t->_lo <= limit_t->_hi); 2402 } 2403 } 2404 if (needs_guard) { 2405 // Check for an obvious zero trip guard. 2406 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 2407 if (inctrl->Opcode() == Op_IfTrue) { 2408 // The test should look like just the backedge of a CountedLoop 2409 Node* iff = inctrl->in(0); 2410 if (iff->is_If()) { 2411 Node* bol = iff->in(1); 2412 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { 2413 Node* cmp = bol->in(1); 2414 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { 2415 needs_guard = false; 2416 } 2417 } 2418 } 2419 } 2420 } 2421 2422 #ifndef PRODUCT 2423 if (PrintOpto) { 2424 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 2425 this->dump_head(); 2426 } else if (TraceLoopOpts) { 2427 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 2428 this->dump_head(); 2429 } 2430 #endif 2431 2432 if (needs_guard) { 2433 // Peel the loop to ensure there's a zero trip guard 2434 Node_List old_new; 2435 phase->do_peeling(this, old_new); 2436 } 2437 2438 // Replace the phi at loop head with the final value of the last 2439 // iteration. Then the CountedLoopEnd will collapse (backedge never 2440 // taken) and all loop-invariant uses of the exit values will be correct. 2441 Node *phi = cl->phi(); 2442 Node *exact_limit = phase->exact_limit(this); 2443 if (exact_limit != cl->limit()) { 2444 // We also need to replace the original limit to collapse loop exit. 2445 Node* cmp = cl->loopexit()->cmp_node(); 2446 assert(cl->limit() == cmp->in(2), "sanity"); 2447 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 2448 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 2449 } 2450 // Note: the final value after increment should not overflow since 2451 // counted loop has limit check predicate. 2452 Node *final = new SubINode( exact_limit, cl->stride() ); 2453 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 2454 phase->_igvn.replace_node(phi,final); 2455 phase->C->set_major_progress(); 2456 return true; 2457 } 2458 2459 //------------------------------policy_do_one_iteration_loop------------------- 2460 // Convert one iteration loop into normal code. 2461 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 2462 if (!_head->as_Loop()->is_valid_counted_loop()) 2463 return false; // Only for counted loop 2464 2465 CountedLoopNode *cl = _head->as_CountedLoop(); 2466 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 2467 return false; 2468 } 2469 2470 #ifndef PRODUCT 2471 if(TraceLoopOpts) { 2472 tty->print("OneIteration "); 2473 this->dump_head(); 2474 } 2475 #endif 2476 2477 Node *init_n = cl->init_trip(); 2478 #ifdef ASSERT 2479 // Loop boundaries should be constant since trip count is exact. 2480 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 2481 #endif 2482 // Replace the phi at loop head with the value of the init_trip. 2483 // Then the CountedLoopEnd will collapse (backedge will not be taken) 2484 // and all loop-invariant uses of the exit values will be correct. 2485 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 2486 phase->C->set_major_progress(); 2487 return true; 2488 } 2489 2490 //============================================================================= 2491 //------------------------------iteration_split_impl--------------------------- 2492 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 2493 // Compute exact loop trip count if possible. 2494 compute_exact_trip_count(phase); 2495 2496 // Convert one iteration loop into normal code. 2497 if (policy_do_one_iteration_loop(phase)) 2498 return true; 2499 2500 // Check and remove empty loops (spam micro-benchmarks) 2501 if (policy_do_remove_empty_loop(phase)) 2502 return true; // Here we removed an empty loop 2503 2504 bool should_peel = policy_peeling(phase); // Should we peel? 2505 2506 bool should_unswitch = policy_unswitching(phase); 2507 2508 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 2509 // This removes loop-invariant tests (usually null checks). 2510 if (!_head->is_CountedLoop()) { // Non-counted loop 2511 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 2512 // Partial peel succeeded so terminate this round of loop opts 2513 return false; 2514 } 2515 if (should_peel) { // Should we peel? 2516 if (PrintOpto) { tty->print_cr("should_peel"); } 2517 phase->do_peeling(this,old_new); 2518 } else if (should_unswitch) { 2519 phase->do_unswitching(this, old_new); 2520 } 2521 return true; 2522 } 2523 CountedLoopNode *cl = _head->as_CountedLoop(); 2524 2525 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 2526 2527 // Do nothing special to pre- and post- loops 2528 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 2529 2530 // Compute loop trip count from profile data 2531 compute_profile_trip_cnt(phase); 2532 2533 // Before attempting fancy unrolling, RCE or alignment, see if we want 2534 // to completely unroll this loop or do loop unswitching. 2535 if (cl->is_normal_loop()) { 2536 if (should_unswitch) { 2537 phase->do_unswitching(this, old_new); 2538 return true; 2539 } 2540 bool should_maximally_unroll = policy_maximally_unroll(phase); 2541 if (should_maximally_unroll) { 2542 // Here we did some unrolling and peeling. Eventually we will 2543 // completely unroll this loop and it will no longer be a loop. 2544 phase->do_maximally_unroll(this,old_new); 2545 return true; 2546 } 2547 } 2548 2549 // Skip next optimizations if running low on nodes. Note that 2550 // policy_unswitching and policy_maximally_unroll have this check. 2551 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes(); 2552 if ((int)(2 * _body.size()) > nodes_left) { 2553 return true; 2554 } 2555 2556 // Counted loops may be peeled, may need some iterations run up 2557 // front for RCE, and may want to align loop refs to a cache 2558 // line. Thus we clone a full loop up front whose trip count is 2559 // at least 1 (if peeling), but may be several more. 2560 2561 // The main loop will start cache-line aligned with at least 1 2562 // iteration of the unrolled body (zero-trip test required) and 2563 // will have some range checks removed. 2564 2565 // A post-loop will finish any odd iterations (leftover after 2566 // unrolling), plus any needed for RCE purposes. 2567 2568 bool should_unroll = policy_unroll(phase); 2569 2570 bool should_rce = policy_range_check(phase); 2571 2572 bool should_align = policy_align(phase); 2573 2574 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2575 // need a pre-loop. We may still need to peel an initial iteration but 2576 // we will not be needing an unknown number of pre-iterations. 2577 // 2578 // Basically, if may_rce_align reports FALSE first time through, 2579 // we will not be able to later do RCE or Aligning on this loop. 2580 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2581 2582 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2583 // we switch to the pre-/main-/post-loop model. This model also covers 2584 // peeling. 2585 if (should_rce || should_align || should_unroll) { 2586 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2587 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2588 2589 // Adjust the pre- and main-loop limits to let the pre and post loops run 2590 // with full checks, but the main-loop with no checks. Remove said 2591 // checks from the main body. 2592 if (should_rce) 2593 phase->do_range_check(this,old_new); 2594 2595 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2596 // twice as many iterations as before) and the main body limit (only do 2597 // an even number of trips). If we are peeling, we might enable some RCE 2598 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2599 // peeling. 2600 if (should_unroll && !should_peel) { 2601 phase->do_unroll(this, old_new, true); 2602 } 2603 2604 // Adjust the pre-loop limits to align the main body 2605 // iterations. 2606 if (should_align) 2607 Unimplemented(); 2608 2609 } else { // Else we have an unchanged counted loop 2610 if (should_peel) // Might want to peel but do nothing else 2611 phase->do_peeling(this,old_new); 2612 } 2613 return true; 2614 } 2615 2616 2617 //============================================================================= 2618 //------------------------------iteration_split-------------------------------- 2619 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2620 // Recursively iteration split nested loops 2621 if (_child && !_child->iteration_split(phase, old_new)) 2622 return false; 2623 2624 // Clean out prior deadwood 2625 DCE_loop_body(); 2626 2627 2628 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2629 // Replace with a 1-in-10 exit guess. 2630 if (_parent /*not the root loop*/ && 2631 !_irreducible && 2632 // Also ignore the occasional dead backedge 2633 !tail()->is_top()) { 2634 adjust_loop_exit_prob(phase); 2635 } 2636 2637 // Gate unrolling, RCE and peeling efforts. 2638 if (!_child && // If not an inner loop, do not split 2639 !_irreducible && 2640 _allow_optimizations && 2641 !tail()->is_top()) { // Also ignore the occasional dead backedge 2642 if (!_has_call) { 2643 if (!iteration_split_impl(phase, old_new)) { 2644 return false; 2645 } 2646 } else if (policy_unswitching(phase)) { 2647 phase->do_unswitching(this, old_new); 2648 } 2649 } 2650 2651 // Minor offset re-organization to remove loop-fallout uses of 2652 // trip counter when there was no major reshaping. 2653 phase->reorg_offsets(this); 2654 2655 if (_next && !_next->iteration_split(phase, old_new)) 2656 return false; 2657 return true; 2658 } 2659 2660 2661 //============================================================================= 2662 // Process all the loops in the loop tree and replace any fill 2663 // patterns with an intrinsic version. 2664 bool PhaseIdealLoop::do_intrinsify_fill() { 2665 bool changed = false; 2666 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2667 IdealLoopTree* lpt = iter.current(); 2668 changed |= intrinsify_fill(lpt); 2669 } 2670 return changed; 2671 } 2672 2673 2674 // Examine an inner loop looking for a a single store of an invariant 2675 // value in a unit stride loop, 2676 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2677 Node*& shift, Node*& con) { 2678 const char* msg = NULL; 2679 Node* msg_node = NULL; 2680 2681 store_value = NULL; 2682 con = NULL; 2683 shift = NULL; 2684 2685 // Process the loop looking for stores. If there are multiple 2686 // stores or extra control flow give at this point. 2687 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2688 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2689 Node* n = lpt->_body.at(i); 2690 if (n->outcnt() == 0) continue; // Ignore dead 2691 if (n->is_Store()) { 2692 if (store != NULL) { 2693 msg = "multiple stores"; 2694 break; 2695 } 2696 int opc = n->Opcode(); 2697 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 2698 msg = "oop fills not handled"; 2699 break; 2700 } 2701 Node* value = n->in(MemNode::ValueIn); 2702 if (!lpt->is_invariant(value)) { 2703 msg = "variant store value"; 2704 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2705 msg = "not array address"; 2706 } 2707 store = n; 2708 store_value = value; 2709 } else if (n->is_If() && n != head->loopexit()) { 2710 msg = "extra control flow"; 2711 msg_node = n; 2712 } 2713 } 2714 2715 if (store == NULL) { 2716 // No store in loop 2717 return false; 2718 } 2719 2720 if (msg == NULL && head->stride_con() != 1) { 2721 // could handle negative strides too 2722 if (head->stride_con() < 0) { 2723 msg = "negative stride"; 2724 } else { 2725 msg = "non-unit stride"; 2726 } 2727 } 2728 2729 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2730 msg = "can't handle store address"; 2731 msg_node = store->in(MemNode::Address); 2732 } 2733 2734 if (msg == NULL && 2735 (!store->in(MemNode::Memory)->is_Phi() || 2736 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2737 msg = "store memory isn't proper phi"; 2738 msg_node = store->in(MemNode::Memory); 2739 } 2740 2741 // Make sure there is an appropriate fill routine 2742 BasicType t = store->as_Mem()->memory_type(); 2743 const char* fill_name; 2744 if (msg == NULL && 2745 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2746 msg = "unsupported store"; 2747 msg_node = store; 2748 } 2749 2750 if (msg != NULL) { 2751 #ifndef PRODUCT 2752 if (TraceOptimizeFill) { 2753 tty->print_cr("not fill intrinsic candidate: %s", msg); 2754 if (msg_node != NULL) msg_node->dump(); 2755 } 2756 #endif 2757 return false; 2758 } 2759 2760 // Make sure the address expression can be handled. It should be 2761 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()). 2762 Node* elements[4]; 2763 Node* cast = NULL; 2764 Node* conv = NULL; 2765 bool found_index = false; 2766 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2767 for (int e = 0; e < count; e++) { 2768 Node* n = elements[e]; 2769 if (n->is_Con() && con == NULL) { 2770 con = n; 2771 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2772 Node* value = n->in(1); 2773 #ifdef _LP64 2774 if (value->Opcode() == Op_ConvI2L) { 2775 conv = value; 2776 value = value->in(1); 2777 } 2778 if (value->Opcode() == Op_CastII && 2779 value->as_CastII()->has_range_check()) { 2780 // Skip range check dependent CastII nodes 2781 cast = value; 2782 value = value->in(1); 2783 } 2784 #endif 2785 if (value != head->phi()) { 2786 msg = "unhandled shift in address"; 2787 } else { 2788 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 2789 msg = "scale doesn't match"; 2790 } else { 2791 found_index = true; 2792 shift = n; 2793 } 2794 } 2795 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2796 conv = n; 2797 n = n->in(1); 2798 if (n->Opcode() == Op_CastII && 2799 n->as_CastII()->has_range_check()) { 2800 // Skip range check dependent CastII nodes 2801 cast = n; 2802 n = n->in(1); 2803 } 2804 if (n == head->phi()) { 2805 found_index = true; 2806 } else { 2807 msg = "unhandled input to ConvI2L"; 2808 } 2809 } else if (n == head->phi()) { 2810 // no shift, check below for allowed cases 2811 found_index = true; 2812 } else { 2813 msg = "unhandled node in address"; 2814 msg_node = n; 2815 } 2816 } 2817 2818 if (count == -1) { 2819 msg = "malformed address expression"; 2820 msg_node = store; 2821 } 2822 2823 if (!found_index) { 2824 msg = "missing use of index"; 2825 } 2826 2827 // byte sized items won't have a shift 2828 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2829 msg = "can't find shift"; 2830 msg_node = store; 2831 } 2832 2833 if (msg != NULL) { 2834 #ifndef PRODUCT 2835 if (TraceOptimizeFill) { 2836 tty->print_cr("not fill intrinsic: %s", msg); 2837 if (msg_node != NULL) msg_node->dump(); 2838 } 2839 #endif 2840 return false; 2841 } 2842 2843 // No make sure all the other nodes in the loop can be handled 2844 VectorSet ok(Thread::current()->resource_area()); 2845 2846 // store related values are ok 2847 ok.set(store->_idx); 2848 ok.set(store->in(MemNode::Memory)->_idx); 2849 2850 CountedLoopEndNode* loop_exit = head->loopexit(); 2851 guarantee(loop_exit != NULL, "no loop exit node"); 2852 2853 // Loop structure is ok 2854 ok.set(head->_idx); 2855 ok.set(loop_exit->_idx); 2856 ok.set(head->phi()->_idx); 2857 ok.set(head->incr()->_idx); 2858 ok.set(loop_exit->cmp_node()->_idx); 2859 ok.set(loop_exit->in(1)->_idx); 2860 2861 // Address elements are ok 2862 if (con) ok.set(con->_idx); 2863 if (shift) ok.set(shift->_idx); 2864 if (cast) ok.set(cast->_idx); 2865 if (conv) ok.set(conv->_idx); 2866 2867 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2868 Node* n = lpt->_body.at(i); 2869 if (n->outcnt() == 0) continue; // Ignore dead 2870 if (ok.test(n->_idx)) continue; 2871 // Backedge projection is ok 2872 if (n->is_IfTrue() && n->in(0) == loop_exit) continue; 2873 if (!n->is_AddP()) { 2874 msg = "unhandled node"; 2875 msg_node = n; 2876 break; 2877 } 2878 } 2879 2880 // Make sure no unexpected values are used outside the loop 2881 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2882 Node* n = lpt->_body.at(i); 2883 // These values can be replaced with other nodes if they are used 2884 // outside the loop. 2885 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2886 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2887 Node* use = iter.get(); 2888 if (!lpt->_body.contains(use)) { 2889 msg = "node is used outside loop"; 2890 // lpt->_body.dump(); 2891 msg_node = n; 2892 break; 2893 } 2894 } 2895 } 2896 2897 #ifdef ASSERT 2898 if (TraceOptimizeFill) { 2899 if (msg != NULL) { 2900 tty->print_cr("no fill intrinsic: %s", msg); 2901 if (msg_node != NULL) msg_node->dump(); 2902 } else { 2903 tty->print_cr("fill intrinsic for:"); 2904 } 2905 store->dump(); 2906 if (Verbose) { 2907 lpt->_body.dump(); 2908 } 2909 } 2910 #endif 2911 2912 return msg == NULL; 2913 } 2914 2915 2916 2917 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2918 // Only for counted inner loops 2919 if (!lpt->is_counted() || !lpt->is_inner()) { 2920 return false; 2921 } 2922 2923 // Must have constant stride 2924 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2925 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 2926 return false; 2927 } 2928 2929 // Check that the body only contains a store of a loop invariant 2930 // value that is indexed by the loop phi. 2931 Node* store = NULL; 2932 Node* store_value = NULL; 2933 Node* shift = NULL; 2934 Node* offset = NULL; 2935 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2936 return false; 2937 } 2938 2939 #ifndef PRODUCT 2940 if (TraceLoopOpts) { 2941 tty->print("ArrayFill "); 2942 lpt->dump_head(); 2943 } 2944 #endif 2945 2946 // Now replace the whole loop body by a call to a fill routine that 2947 // covers the same region as the loop. 2948 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2949 2950 // Build an expression for the beginning of the copy region 2951 Node* index = head->init_trip(); 2952 #ifdef _LP64 2953 index = new ConvI2LNode(index); 2954 _igvn.register_new_node_with_optimizer(index); 2955 #endif 2956 if (shift != NULL) { 2957 // byte arrays don't require a shift but others do. 2958 index = new LShiftXNode(index, shift->in(2)); 2959 _igvn.register_new_node_with_optimizer(index); 2960 } 2961 index = new AddPNode(base, base, index); 2962 _igvn.register_new_node_with_optimizer(index); 2963 Node* from = new AddPNode(base, index, offset); 2964 _igvn.register_new_node_with_optimizer(from); 2965 // Compute the number of elements to copy 2966 Node* len = new SubINode(head->limit(), head->init_trip()); 2967 _igvn.register_new_node_with_optimizer(len); 2968 2969 BasicType t = store->as_Mem()->memory_type(); 2970 bool aligned = false; 2971 if (offset != NULL && head->init_trip()->is_Con()) { 2972 int element_size = type2aelembytes(t); 2973 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2974 } 2975 2976 // Build a call to the fill routine 2977 const char* fill_name; 2978 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2979 assert(fill != NULL, "what?"); 2980 2981 // Convert float/double to int/long for fill routines 2982 if (t == T_FLOAT) { 2983 store_value = new MoveF2INode(store_value); 2984 _igvn.register_new_node_with_optimizer(store_value); 2985 } else if (t == T_DOUBLE) { 2986 store_value = new MoveD2LNode(store_value); 2987 _igvn.register_new_node_with_optimizer(store_value); 2988 } 2989 2990 Node* mem_phi = store->in(MemNode::Memory); 2991 Node* result_ctrl; 2992 Node* result_mem; 2993 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2994 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill, 2995 fill_name, TypeAryPtr::get_array_body_type(t)); 2996 uint cnt = 0; 2997 call->init_req(TypeFunc::Parms + cnt++, from); 2998 call->init_req(TypeFunc::Parms + cnt++, store_value); 2999 #ifdef _LP64 3000 len = new ConvI2LNode(len); 3001 _igvn.register_new_node_with_optimizer(len); 3002 #endif 3003 call->init_req(TypeFunc::Parms + cnt++, len); 3004 #ifdef _LP64 3005 call->init_req(TypeFunc::Parms + cnt++, C->top()); 3006 #endif 3007 call->init_req(TypeFunc::Control, head->init_control()); 3008 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. 3009 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); 3010 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr)); 3011 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr)); 3012 _igvn.register_new_node_with_optimizer(call); 3013 result_ctrl = new ProjNode(call,TypeFunc::Control); 3014 _igvn.register_new_node_with_optimizer(result_ctrl); 3015 result_mem = new ProjNode(call,TypeFunc::Memory); 3016 _igvn.register_new_node_with_optimizer(result_mem); 3017 3018 /* Disable following optimization until proper fix (add missing checks). 3019 3020 // If this fill is tightly coupled to an allocation and overwrites 3021 // the whole body, allow it to take over the zeroing. 3022 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 3023 if (alloc != NULL && alloc->is_AllocateArray()) { 3024 Node* length = alloc->as_AllocateArray()->Ideal_length(); 3025 if (head->limit() == length && 3026 head->init_trip() == _igvn.intcon(0)) { 3027 if (TraceOptimizeFill) { 3028 tty->print_cr("Eliminated zeroing in allocation"); 3029 } 3030 alloc->maybe_set_complete(&_igvn); 3031 } else { 3032 #ifdef ASSERT 3033 if (TraceOptimizeFill) { 3034 tty->print_cr("filling array but bounds don't match"); 3035 alloc->dump(); 3036 head->init_trip()->dump(); 3037 head->limit()->dump(); 3038 length->dump(); 3039 } 3040 #endif 3041 } 3042 } 3043 */ 3044 3045 // Redirect the old control and memory edges that are outside the loop. 3046 Node* exit = head->loopexit()->proj_out(0); 3047 // Sometimes the memory phi of the head is used as the outgoing 3048 // state of the loop. It's safe in this case to replace it with the 3049 // result_mem. 3050 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 3051 _igvn.replace_node(exit, result_ctrl); 3052 _igvn.replace_node(store, result_mem); 3053 // Any uses the increment outside of the loop become the loop limit. 3054 _igvn.replace_node(head->incr(), head->limit()); 3055 3056 // Disconnect the head from the loop. 3057 for (uint i = 0; i < lpt->_body.size(); i++) { 3058 Node* n = lpt->_body.at(i); 3059 _igvn.replace_node(n, C->top()); 3060 } 3061 3062 return true; 3063 }