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, "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 NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct)); 797 } 798 799 // Unroll once! (Each trip will soon do double iterations) 800 return true; 801 } 802 803 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) { 804 // Enable this functionality target by target as needed 805 if (SuperWordLoopUnrollAnalysis) { 806 if (!cl->was_slp_analyzed()) { 807 SuperWord sw(phase); 808 sw.transform_loop(this, false); 809 810 // If the loop is slp canonical analyze it 811 if (sw.early_return() == false) { 812 sw.unrolling_analysis(_local_loop_unroll_factor); 813 } 814 } 815 816 if (cl->has_passed_slp()) { 817 int slp_max_unroll_factor = cl->slp_max_unroll(); 818 if (slp_max_unroll_factor >= future_unroll_ct) { 819 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor; 820 if (new_limit > LoopUnrollLimit) { 821 NOT_PRODUCT(if (TraceSuperWordLoopUnrollAnalysis) tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit)); 822 _local_loop_unroll_limit = new_limit; 823 } 824 } 825 } 826 } 827 } 828 829 //------------------------------policy_align----------------------------------- 830 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 831 // expression that does the alignment. Note that only one array base can be 832 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 833 // if we vectorize short memory ops into longer memory ops, we may want to 834 // increase alignment. 835 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 836 return false; 837 } 838 839 //------------------------------policy_range_check----------------------------- 840 // Return TRUE or FALSE if the loop should be range-check-eliminated. 841 // Actually we do iteration-splitting, a more powerful form of RCE. 842 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 843 if (!RangeCheckElimination) return false; 844 845 CountedLoopNode *cl = _head->as_CountedLoop(); 846 // If we unrolled with no intention of doing RCE and we later 847 // changed our minds, we got no pre-loop. Either we need to 848 // make a new pre-loop, or we gotta disallow RCE. 849 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 850 Node *trip_counter = cl->phi(); 851 852 // check for vectorized loops, some opts are no longer needed 853 if (cl->do_unroll_only()) return false; 854 855 // Check loop body for tests of trip-counter plus loop-invariant vs 856 // loop-invariant. 857 for (uint i = 0; i < _body.size(); i++) { 858 Node *iff = _body[i]; 859 if (iff->Opcode() == Op_If) { // Test? 860 861 // Comparing trip+off vs limit 862 Node *bol = iff->in(1); 863 if (bol->req() != 2) continue; // dead constant test 864 if (!bol->is_Bool()) { 865 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 866 continue; 867 } 868 if (bol->as_Bool()->_test._test == BoolTest::ne) 869 continue; // not RC 870 871 Node *cmp = bol->in(1); 872 Node *rc_exp = cmp->in(1); 873 Node *limit = cmp->in(2); 874 875 Node *limit_c = phase->get_ctrl(limit); 876 if( limit_c == phase->C->top() ) 877 return false; // Found dead test on live IF? No RCE! 878 if( is_member(phase->get_loop(limit_c) ) ) { 879 // Compare might have operands swapped; commute them 880 rc_exp = cmp->in(2); 881 limit = cmp->in(1); 882 limit_c = phase->get_ctrl(limit); 883 if( is_member(phase->get_loop(limit_c) ) ) 884 continue; // Both inputs are loop varying; cannot RCE 885 } 886 887 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 888 continue; 889 } 890 // Yeah! Found a test like 'trip+off vs limit' 891 // Test is an IfNode, has 2 projections. If BOTH are in the loop 892 // we need loop unswitching instead of iteration splitting. 893 if( is_loop_exit(iff) ) 894 return true; // Found reason to split iterations 895 } // End of is IF 896 } 897 898 return false; 899 } 900 901 //------------------------------policy_peel_only------------------------------- 902 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 903 // for unrolling loops with NO array accesses. 904 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 905 // check for vectorized loops, any peeling done was already applied 906 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false; 907 908 for( uint i = 0; i < _body.size(); i++ ) 909 if( _body[i]->is_Mem() ) 910 return false; 911 912 // No memory accesses at all! 913 return true; 914 } 915 916 //------------------------------clone_up_backedge_goo-------------------------- 917 // If Node n lives in the back_ctrl block and cannot float, we clone a private 918 // version of n in preheader_ctrl block and return that, otherwise return n. 919 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { 920 if( get_ctrl(n) != back_ctrl ) return n; 921 922 // Only visit once 923 if (visited.test_set(n->_idx)) { 924 Node *x = clones.find(n->_idx); 925 if (x != NULL) 926 return x; 927 return n; 928 } 929 930 Node *x = NULL; // If required, a clone of 'n' 931 // Check for 'n' being pinned in the backedge. 932 if( n->in(0) && n->in(0) == back_ctrl ) { 933 assert(clones.find(n->_idx) == NULL, "dead loop"); 934 x = n->clone(); // Clone a copy of 'n' to preheader 935 clones.push(x, n->_idx); 936 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 937 } 938 939 // Recursive fixup any other input edges into x. 940 // If there are no changes we can just return 'n', otherwise 941 // we need to clone a private copy and change it. 942 for( uint i = 1; i < n->req(); i++ ) { 943 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); 944 if( g != n->in(i) ) { 945 if( !x ) { 946 assert(clones.find(n->_idx) == NULL, "dead loop"); 947 x = n->clone(); 948 clones.push(x, n->_idx); 949 } 950 x->set_req(i, g); 951 } 952 } 953 if( x ) { // x can legally float to pre-header location 954 register_new_node( x, preheader_ctrl ); 955 return x; 956 } else { // raise n to cover LCA of uses 957 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 958 } 959 return n; 960 } 961 962 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { 963 Node* castii = new CastIINode(incr, TypeInt::INT, true); 964 castii->set_req(0, ctrl); 965 register_new_node(castii, ctrl); 966 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { 967 Node* n = incr->fast_out(i); 968 if (n->is_Phi() && n->in(0) == loop) { 969 int nrep = n->replace_edge(incr, castii); 970 return true; 971 } 972 } 973 return false; 974 } 975 976 //------------------------------insert_pre_post_loops-------------------------- 977 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 978 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 979 // alignment. Useful to unroll loops that do no array accesses. 980 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 981 982 #ifndef PRODUCT 983 if (TraceLoopOpts) { 984 if (peel_only) 985 tty->print("PeelMainPost "); 986 else 987 tty->print("PreMainPost "); 988 loop->dump_head(); 989 } 990 #endif 991 C->set_major_progress(); 992 993 // Find common pieces of the loop being guarded with pre & post loops 994 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 995 assert( main_head->is_normal_loop(), "" ); 996 CountedLoopEndNode *main_end = main_head->loopexit(); 997 guarantee(main_end != NULL, "no loop exit node"); 998 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 999 uint dd_main_head = dom_depth(main_head); 1000 uint max = main_head->outcnt(); 1001 1002 Node *pre_header= main_head->in(LoopNode::EntryControl); 1003 Node *init = main_head->init_trip(); 1004 Node *incr = main_end ->incr(); 1005 Node *limit = main_end ->limit(); 1006 Node *stride = main_end ->stride(); 1007 Node *cmp = main_end ->cmp_node(); 1008 BoolTest::mask b_test = main_end->test_trip(); 1009 1010 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 1011 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 1012 if( bol->outcnt() != 1 ) { 1013 bol = bol->clone(); 1014 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 1015 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol); 1016 } 1017 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 1018 if( cmp->outcnt() != 1 ) { 1019 cmp = cmp->clone(); 1020 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 1021 _igvn.replace_input_of(bol, 1, cmp); 1022 } 1023 1024 //------------------------------ 1025 // Step A: Create Post-Loop. 1026 Node* main_exit = main_end->proj_out(false); 1027 assert( main_exit->Opcode() == Op_IfFalse, "" ); 1028 int dd_main_exit = dom_depth(main_exit); 1029 1030 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 1031 // loop pre-header illegally has 2 control users (old & new loops). 1032 clone_loop( loop, old_new, dd_main_exit ); 1033 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 1034 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 1035 post_head->set_post_loop(main_head); 1036 1037 // Reduce the post-loop trip count. 1038 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1039 post_end->_prob = PROB_FAIR; 1040 1041 // Build the main-loop normal exit. 1042 IfFalseNode *new_main_exit = new IfFalseNode(main_end); 1043 _igvn.register_new_node_with_optimizer( new_main_exit ); 1044 set_idom(new_main_exit, main_end, dd_main_exit ); 1045 set_loop(new_main_exit, loop->_parent); 1046 1047 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 1048 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 1049 // (the main-loop trip-counter exit value) because we will be changing 1050 // the exit value (via unrolling) so we cannot constant-fold away the zero 1051 // trip guard until all unrolling is done. 1052 Node *zer_opaq = new Opaque1Node(C, incr); 1053 Node *zer_cmp = new CmpINode( zer_opaq, limit ); 1054 Node *zer_bol = new BoolNode( zer_cmp, b_test ); 1055 register_new_node( zer_opaq, new_main_exit ); 1056 register_new_node( zer_cmp , new_main_exit ); 1057 register_new_node( zer_bol , new_main_exit ); 1058 1059 // Build the IfNode 1060 IfNode *zer_iff = new IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 1061 _igvn.register_new_node_with_optimizer( zer_iff ); 1062 set_idom(zer_iff, new_main_exit, dd_main_exit); 1063 set_loop(zer_iff, loop->_parent); 1064 1065 // Plug in the false-path, taken if we need to skip post-loop 1066 _igvn.replace_input_of(main_exit, 0, zer_iff); 1067 set_idom(main_exit, zer_iff, dd_main_exit); 1068 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 1069 // Make the true-path, must enter the post loop 1070 Node *zer_taken = new IfTrueNode( zer_iff ); 1071 _igvn.register_new_node_with_optimizer( zer_taken ); 1072 set_idom(zer_taken, zer_iff, dd_main_exit); 1073 set_loop(zer_taken, loop->_parent); 1074 // Plug in the true path 1075 _igvn.hash_delete( post_head ); 1076 post_head->set_req(LoopNode::EntryControl, zer_taken); 1077 set_idom(post_head, zer_taken, dd_main_exit); 1078 1079 Arena *a = Thread::current()->resource_area(); 1080 VectorSet visited(a); 1081 Node_Stack clones(a, main_head->back_control()->outcnt()); 1082 // Step A3: Make the fall-in values to the post-loop come from the 1083 // fall-out values of the main-loop. 1084 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 1085 Node* main_phi = main_head->fast_out(i); 1086 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 1087 Node *post_phi = old_new[main_phi->_idx]; 1088 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 1089 post_head->init_control(), 1090 main_phi->in(LoopNode::LoopBackControl), 1091 visited, clones); 1092 _igvn.hash_delete(post_phi); 1093 post_phi->set_req( LoopNode::EntryControl, fallmain ); 1094 } 1095 } 1096 1097 // Update local caches for next stanza 1098 main_exit = new_main_exit; 1099 1100 1101 //------------------------------ 1102 // Step B: Create Pre-Loop. 1103 1104 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 1105 // loop pre-header illegally has 2 control users (old & new loops). 1106 clone_loop( loop, old_new, dd_main_head ); 1107 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1108 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1109 pre_head->set_pre_loop(main_head); 1110 Node *pre_incr = old_new[incr->_idx]; 1111 1112 // Reduce the pre-loop trip count. 1113 pre_end->_prob = PROB_FAIR; 1114 1115 // Find the pre-loop normal exit. 1116 Node* pre_exit = pre_end->proj_out(false); 1117 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 1118 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end); 1119 _igvn.register_new_node_with_optimizer( new_pre_exit ); 1120 set_idom(new_pre_exit, pre_end, dd_main_head); 1121 set_loop(new_pre_exit, loop->_parent); 1122 1123 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1124 // pre-loop, the main-loop may not execute at all. Later in life this 1125 // zero-trip guard will become the minimum-trip guard when we unroll 1126 // the main-loop. 1127 Node *min_opaq = new Opaque1Node(C, limit); 1128 Node *min_cmp = new CmpINode( pre_incr, min_opaq ); 1129 Node *min_bol = new BoolNode( min_cmp, b_test ); 1130 register_new_node( min_opaq, new_pre_exit ); 1131 register_new_node( min_cmp , new_pre_exit ); 1132 register_new_node( min_bol , new_pre_exit ); 1133 1134 // Build the IfNode (assume the main-loop is executed always). 1135 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1136 _igvn.register_new_node_with_optimizer( min_iff ); 1137 set_idom(min_iff, new_pre_exit, dd_main_head); 1138 set_loop(min_iff, loop->_parent); 1139 1140 // Plug in the false-path, taken if we need to skip main-loop 1141 _igvn.hash_delete( pre_exit ); 1142 pre_exit->set_req(0, min_iff); 1143 set_idom(pre_exit, min_iff, dd_main_head); 1144 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1145 // Make the true-path, must enter the main loop 1146 Node *min_taken = new IfTrueNode( min_iff ); 1147 _igvn.register_new_node_with_optimizer( min_taken ); 1148 set_idom(min_taken, min_iff, dd_main_head); 1149 set_loop(min_taken, loop->_parent); 1150 // Plug in the true path 1151 _igvn.hash_delete( main_head ); 1152 main_head->set_req(LoopNode::EntryControl, min_taken); 1153 set_idom(main_head, min_taken, dd_main_head); 1154 1155 visited.Clear(); 1156 clones.clear(); 1157 // Step B3: Make the fall-in values to the main-loop come from the 1158 // fall-out values of the pre-loop. 1159 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1160 Node* main_phi = main_head->fast_out(i2); 1161 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1162 Node *pre_phi = old_new[main_phi->_idx]; 1163 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1164 main_head->init_control(), 1165 pre_phi->in(LoopNode::LoopBackControl), 1166 visited, clones); 1167 _igvn.hash_delete(main_phi); 1168 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1169 } 1170 } 1171 1172 // Nodes inside the loop may be control dependent on a predicate 1173 // that was moved before the preloop. If the back branch of the main 1174 // or post loops becomes dead, those nodes won't be dependent on the 1175 // test that guards that loop nest anymore which could lead to an 1176 // incorrect array access because it executes independently of the 1177 // test that was guarding the loop nest. We add a special CastII on 1178 // the if branch that enters the loop, between the input induction 1179 // variable value and the induction variable Phi to preserve correct 1180 // dependencies. 1181 1182 // CastII for the post loop: 1183 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); 1184 assert(inserted, "no castII inserted"); 1185 1186 // CastII for the main loop: 1187 inserted = cast_incr_before_loop(pre_incr, min_taken, main_head); 1188 assert(inserted, "no castII inserted"); 1189 1190 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1191 // RCE and alignment may change this later. 1192 Node *cmp_end = pre_end->cmp_node(); 1193 assert( cmp_end->in(2) == limit, "" ); 1194 Node *pre_limit = new AddINode( init, stride ); 1195 1196 // Save the original loop limit in this Opaque1 node for 1197 // use by range check elimination. 1198 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit); 1199 1200 register_new_node( pre_limit, pre_head->in(0) ); 1201 register_new_node( pre_opaq , pre_head->in(0) ); 1202 1203 // Since no other users of pre-loop compare, I can hack limit directly 1204 assert( cmp_end->outcnt() == 1, "no other users" ); 1205 _igvn.hash_delete(cmp_end); 1206 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1207 1208 // Special case for not-equal loop bounds: 1209 // Change pre loop test, main loop test, and the 1210 // main loop guard test to use lt or gt depending on stride 1211 // direction: 1212 // positive stride use < 1213 // negative stride use > 1214 // 1215 // not-equal test is kept for post loop to handle case 1216 // when init > limit when stride > 0 (and reverse). 1217 1218 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1219 1220 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1221 // Modify pre loop end condition 1222 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1223 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test); 1224 register_new_node( new_bol0, pre_head->in(0) ); 1225 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0); 1226 // Modify main loop guard condition 1227 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1228 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test); 1229 register_new_node( new_bol1, new_pre_exit ); 1230 _igvn.hash_delete(min_iff); 1231 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1232 // Modify main loop end condition 1233 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1234 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test); 1235 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1236 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2); 1237 } 1238 1239 // Flag main loop 1240 main_head->set_main_loop(); 1241 if( peel_only ) main_head->set_main_no_pre_loop(); 1242 1243 // Subtract a trip count for the pre-loop. 1244 main_head->set_trip_count(main_head->trip_count() - 1); 1245 1246 // It's difficult to be precise about the trip-counts 1247 // for the pre/post loops. They are usually very short, 1248 // so guess that 4 trips is a reasonable value. 1249 post_head->set_profile_trip_cnt(4.0); 1250 pre_head->set_profile_trip_cnt(4.0); 1251 1252 // Now force out all loop-invariant dominating tests. The optimizer 1253 // finds some, but we _know_ they are all useless. 1254 peeled_dom_test_elim(loop,old_new); 1255 loop->record_for_igvn(); 1256 } 1257 1258 //------------------------------is_invariant----------------------------- 1259 // Return true if n is invariant 1260 bool IdealLoopTree::is_invariant(Node* n) const { 1261 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1262 if (n_c->is_top()) return false; 1263 return !is_member(_phase->get_loop(n_c)); 1264 } 1265 1266 1267 //------------------------------do_unroll-------------------------------------- 1268 // Unroll the loop body one step - make each trip do 2 iterations. 1269 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1270 assert(LoopUnrollLimit, ""); 1271 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1272 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1273 assert(loop_end, ""); 1274 #ifndef PRODUCT 1275 if (PrintOpto && VerifyLoopOptimizations) { 1276 tty->print("Unrolling "); 1277 loop->dump_head(); 1278 } else if (TraceLoopOpts) { 1279 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1280 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1281 } else { 1282 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1283 } 1284 loop->dump_head(); 1285 } 1286 1287 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1288 Arena* arena = Thread::current()->resource_area(); 1289 Node_Stack stack(arena, C->live_nodes() >> 2); 1290 Node_List rpo_list; 1291 VectorSet visited(arena); 1292 visited.set(loop_head->_idx); 1293 rpo( loop_head, stack, visited, rpo_list ); 1294 dump(loop, rpo_list.size(), rpo_list ); 1295 } 1296 #endif 1297 1298 // Remember loop node count before unrolling to detect 1299 // if rounds of unroll,optimize are making progress 1300 loop_head->set_node_count_before_unroll(loop->_body.size()); 1301 1302 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1303 Node *limit = loop_head->limit(); 1304 Node *init = loop_head->init_trip(); 1305 Node *stride = loop_head->stride(); 1306 1307 Node *opaq = NULL; 1308 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1309 // Search for zero-trip guard. 1310 assert( loop_head->is_main_loop(), "" ); 1311 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1312 Node *iff = ctrl->in(0); 1313 assert( iff->Opcode() == Op_If, "" ); 1314 Node *bol = iff->in(1); 1315 assert( bol->Opcode() == Op_Bool, "" ); 1316 Node *cmp = bol->in(1); 1317 assert( cmp->Opcode() == Op_CmpI, "" ); 1318 opaq = cmp->in(2); 1319 // Occasionally it's possible for a zero-trip guard Opaque1 node to be 1320 // optimized away and then another round of loop opts attempted. 1321 // We can not optimize this particular loop in that case. 1322 if (opaq->Opcode() != Op_Opaque1) 1323 return; // Cannot find zero-trip guard! Bail out! 1324 // Zero-trip test uses an 'opaque' node which is not shared. 1325 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1326 } 1327 1328 C->set_major_progress(); 1329 1330 Node* new_limit = NULL; 1331 if (UnrollLimitCheck) { 1332 int stride_con = stride->get_int(); 1333 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1334 uint old_trip_count = loop_head->trip_count(); 1335 // Verify that unroll policy result is still valid. 1336 assert(old_trip_count > 1 && 1337 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1338 1339 // Adjust loop limit to keep valid iterations number after unroll. 1340 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1341 // which may overflow. 1342 if (!adjust_min_trip) { 1343 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1344 "odd trip count for maximally unroll"); 1345 // Don't need to adjust limit for maximally unroll since trip count is even. 1346 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1347 // Loop's limit is constant. Loop's init could be constant when pre-loop 1348 // become peeled iteration. 1349 jlong init_con = init->get_int(); 1350 // We can keep old loop limit if iterations count stays the same: 1351 // old_trip_count == new_trip_count * 2 1352 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1353 // so we also don't need to adjust zero trip test. 1354 jlong limit_con = limit->get_int(); 1355 // (stride_con*2) not overflow since stride_con <= 8. 1356 int new_stride_con = stride_con * 2; 1357 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1358 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1359 // New trip count should satisfy next conditions. 1360 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1361 uint new_trip_count = (uint)trip_count; 1362 adjust_min_trip = (old_trip_count != new_trip_count*2); 1363 } 1364 1365 if (adjust_min_trip) { 1366 // Step 2: Adjust the trip limit if it is called for. 1367 // The adjustment amount is -stride. Need to make sure if the 1368 // adjustment underflows or overflows, then the main loop is skipped. 1369 Node* cmp = loop_end->cmp_node(); 1370 assert(cmp->in(2) == limit, "sanity"); 1371 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1372 1373 // Verify that policy_unroll result is still valid. 1374 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1375 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1376 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); 1377 1378 if (limit->is_Con()) { 1379 // The check in policy_unroll and the assert above guarantee 1380 // no underflow if limit is constant. 1381 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1382 set_ctrl(new_limit, C->root()); 1383 } else { 1384 // Limit is not constant. 1385 if (loop_head->unrolled_count() == 1) { // only for first unroll 1386 // Separate limit by Opaque node in case it is an incremented 1387 // variable from previous loop to avoid using pre-incremented 1388 // value which could increase register pressure. 1389 // Otherwise reorg_offsets() optimization will create a separate 1390 // Opaque node for each use of trip-counter and as result 1391 // zero trip guard limit will be different from loop limit. 1392 assert(has_ctrl(opaq), "should have it"); 1393 Node* opaq_ctrl = get_ctrl(opaq); 1394 limit = new Opaque2Node( C, limit ); 1395 register_new_node( limit, opaq_ctrl ); 1396 } 1397 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) || 1398 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) { 1399 // No underflow. 1400 new_limit = new SubINode(limit, stride); 1401 } else { 1402 // (limit - stride) may underflow. 1403 // Clamp the adjustment value with MININT or MAXINT: 1404 // 1405 // new_limit = limit-stride 1406 // if (stride > 0) 1407 // new_limit = (limit < new_limit) ? MININT : new_limit; 1408 // else 1409 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1410 // 1411 BoolTest::mask bt = loop_end->test_trip(); 1412 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1413 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1414 set_ctrl(adj_max, C->root()); 1415 Node* old_limit = NULL; 1416 Node* adj_limit = NULL; 1417 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1418 if (loop_head->unrolled_count() > 1 && 1419 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1420 limit->in(CMoveNode::IfTrue) == adj_max && 1421 bol->as_Bool()->_test._test == bt && 1422 bol->in(1)->Opcode() == Op_CmpI && 1423 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1424 // Loop was unrolled before. 1425 // Optimize the limit to avoid nested CMove: 1426 // use original limit as old limit. 1427 old_limit = bol->in(1)->in(1); 1428 // Adjust previous adjusted limit. 1429 adj_limit = limit->in(CMoveNode::IfFalse); 1430 adj_limit = new SubINode(adj_limit, stride); 1431 } else { 1432 old_limit = limit; 1433 adj_limit = new SubINode(limit, stride); 1434 } 1435 assert(old_limit != NULL && adj_limit != NULL, ""); 1436 register_new_node( adj_limit, ctrl ); // adjust amount 1437 Node* adj_cmp = new CmpINode(old_limit, adj_limit); 1438 register_new_node( adj_cmp, ctrl ); 1439 Node* adj_bool = new BoolNode(adj_cmp, bt); 1440 register_new_node( adj_bool, ctrl ); 1441 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1442 } 1443 register_new_node(new_limit, ctrl); 1444 } 1445 assert(new_limit != NULL, ""); 1446 // Replace in loop test. 1447 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1448 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1449 // Don't need to create new test since only one user. 1450 _igvn.hash_delete(cmp); 1451 cmp->set_req(2, new_limit); 1452 } else { 1453 // Create new test since it is shared. 1454 Node* ctrl2 = loop_end->in(0); 1455 Node* cmp2 = cmp->clone(); 1456 cmp2->set_req(2, new_limit); 1457 register_new_node(cmp2, ctrl2); 1458 Node* bol2 = loop_end->in(1)->clone(); 1459 bol2->set_req(1, cmp2); 1460 register_new_node(bol2, ctrl2); 1461 _igvn.replace_input_of(loop_end, 1, bol2); 1462 } 1463 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1464 // Make it a 1-trip test (means at least 2 trips). 1465 1466 // Guard test uses an 'opaque' node which is not shared. Hence I 1467 // can edit it's inputs directly. Hammer in the new limit for the 1468 // minimum-trip guard. 1469 assert(opaq->outcnt() == 1, ""); 1470 _igvn.replace_input_of(opaq, 1, new_limit); 1471 } 1472 1473 // Adjust max trip count. The trip count is intentionally rounded 1474 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1475 // the main, unrolled, part of the loop will never execute as it is protected 1476 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1477 // and later determined that part of the unrolled loop was dead. 1478 loop_head->set_trip_count(old_trip_count / 2); 1479 1480 // Double the count of original iterations in the unrolled loop body. 1481 loop_head->double_unrolled_count(); 1482 1483 } else { // LoopLimitCheck 1484 1485 // Adjust max trip count. The trip count is intentionally rounded 1486 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1487 // the main, unrolled, part of the loop will never execute as it is protected 1488 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1489 // and later determined that part of the unrolled loop was dead. 1490 loop_head->set_trip_count(loop_head->trip_count() / 2); 1491 1492 // Double the count of original iterations in the unrolled loop body. 1493 loop_head->double_unrolled_count(); 1494 1495 // ----------- 1496 // Step 2: Cut back the trip counter for an unroll amount of 2. 1497 // Loop will normally trip (limit - init)/stride_con. Since it's a 1498 // CountedLoop this is exact (stride divides limit-init exactly). 1499 // We are going to double the loop body, so we want to knock off any 1500 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 1501 Node *span = new SubINode( limit, init ); 1502 register_new_node( span, ctrl ); 1503 Node *trip = new DivINode( 0, span, stride ); 1504 register_new_node( trip, ctrl ); 1505 Node *mtwo = _igvn.intcon(-2); 1506 set_ctrl(mtwo, C->root()); 1507 Node *rond = new AndINode( trip, mtwo ); 1508 register_new_node( rond, ctrl ); 1509 Node *spn2 = new MulINode( rond, stride ); 1510 register_new_node( spn2, ctrl ); 1511 new_limit = new AddINode( spn2, init ); 1512 register_new_node( new_limit, ctrl ); 1513 1514 // Hammer in the new limit 1515 Node *ctrl2 = loop_end->in(0); 1516 Node *cmp2 = new CmpINode( loop_head->incr(), new_limit ); 1517 register_new_node( cmp2, ctrl2 ); 1518 Node *bol2 = new BoolNode( cmp2, loop_end->test_trip() ); 1519 register_new_node( bol2, ctrl2 ); 1520 _igvn.replace_input_of(loop_end, CountedLoopEndNode::TestValue, bol2); 1521 1522 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1523 // Make it a 1-trip test (means at least 2 trips). 1524 if( adjust_min_trip ) { 1525 assert( new_limit != NULL, "" ); 1526 // Guard test uses an 'opaque' node which is not shared. Hence I 1527 // can edit it's inputs directly. Hammer in the new limit for the 1528 // minimum-trip guard. 1529 assert( opaq->outcnt() == 1, "" ); 1530 _igvn.hash_delete(opaq); 1531 opaq->set_req(1, new_limit); 1532 } 1533 } // LoopLimitCheck 1534 1535 // --------- 1536 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1537 // represents the odd iterations; since the loop trips an even number of 1538 // times its backedge is never taken. Kill the backedge. 1539 uint dd = dom_depth(loop_head); 1540 clone_loop( loop, old_new, dd ); 1541 1542 // Make backedges of the clone equal to backedges of the original. 1543 // Make the fall-in from the original come from the fall-out of the clone. 1544 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1545 Node* phi = loop_head->fast_out(j); 1546 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1547 Node *newphi = old_new[phi->_idx]; 1548 _igvn.hash_delete( phi ); 1549 _igvn.hash_delete( newphi ); 1550 1551 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1552 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1553 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1554 } 1555 } 1556 Node *clone_head = old_new[loop_head->_idx]; 1557 _igvn.hash_delete( clone_head ); 1558 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1559 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1560 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1561 loop->_head = clone_head; // New loop header 1562 1563 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1564 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1565 1566 // Kill the clone's backedge 1567 Node *newcle = old_new[loop_end->_idx]; 1568 _igvn.hash_delete( newcle ); 1569 Node *one = _igvn.intcon(1); 1570 set_ctrl(one, C->root()); 1571 newcle->set_req(1, one); 1572 // Force clone into same loop body 1573 uint max = loop->_body.size(); 1574 for( uint k = 0; k < max; k++ ) { 1575 Node *old = loop->_body.at(k); 1576 Node *nnn = old_new[old->_idx]; 1577 loop->_body.push(nnn); 1578 if (!has_ctrl(old)) 1579 set_loop(nnn, loop); 1580 } 1581 1582 loop->record_for_igvn(); 1583 1584 #ifndef PRODUCT 1585 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1586 tty->print("\nnew loop after unroll\n"); loop->dump_head(); 1587 for (uint i = 0; i < loop->_body.size(); i++) { 1588 loop->_body.at(i)->dump(); 1589 } 1590 if(C->clone_map().is_debug()) { 1591 tty->print("\nCloneMap\n"); 1592 Dict* dict = C->clone_map().dict(); 1593 DictI i(dict); 1594 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size()); 1595 for (int ii = 0; i.test(); ++i, ++ii) { 1596 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key)); 1597 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen()); 1598 if (ii % 10 == 9) { 1599 tty->print_cr(" "); 1600 } 1601 } 1602 tty->print_cr(" "); 1603 } 1604 } 1605 #endif 1606 1607 } 1608 1609 //------------------------------do_maximally_unroll---------------------------- 1610 1611 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1612 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1613 assert(cl->has_exact_trip_count(), "trip count is not exact"); 1614 assert(cl->trip_count() > 0, ""); 1615 #ifndef PRODUCT 1616 if (TraceLoopOpts) { 1617 tty->print("MaxUnroll %d ", cl->trip_count()); 1618 loop->dump_head(); 1619 } 1620 #endif 1621 1622 // If loop is tripping an odd number of times, peel odd iteration 1623 if ((cl->trip_count() & 1) == 1) { 1624 do_peeling(loop, old_new); 1625 } 1626 1627 // Now its tripping an even number of times remaining. Double loop body. 1628 // Do not adjust pre-guards; they are not needed and do not exist. 1629 if (cl->trip_count() > 0) { 1630 assert((cl->trip_count() & 1) == 0, "missed peeling"); 1631 do_unroll(loop, old_new, false); 1632 } 1633 } 1634 1635 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) { 1636 if (SuperWordReductions == false) return; 1637 1638 CountedLoopNode* loop_head = loop->_head->as_CountedLoop(); 1639 if (loop_head->unrolled_count() > 1) { 1640 return; 1641 } 1642 1643 Node* trip_phi = loop_head->phi(); 1644 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) { 1645 Node* phi = loop_head->fast_out(i); 1646 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) { 1647 // For definitions which are loop inclusive and not tripcounts. 1648 Node* def_node = phi->in(LoopNode::LoopBackControl); 1649 1650 if (def_node != NULL) { 1651 Node* n_ctrl = get_ctrl(def_node); 1652 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) { 1653 // Now test it to see if it fits the standard pattern for a reduction operator. 1654 int opc = def_node->Opcode(); 1655 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) { 1656 if (!def_node->is_reduction()) { // Not marked yet 1657 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it 1658 bool ok = false; 1659 for (unsigned j = 1; j < def_node->req(); j++) { 1660 Node* in = def_node->in(j); 1661 if (in == phi) { 1662 ok = true; 1663 break; 1664 } 1665 } 1666 1667 // do nothing if we did not match the initial criteria 1668 if (ok == false) { 1669 continue; 1670 } 1671 1672 // The result of the reduction must not be used in the loop 1673 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) { 1674 Node* u = def_node->fast_out(i); 1675 if (has_ctrl(u) && !loop->is_member(get_loop(get_ctrl(u)))) { 1676 continue; 1677 } 1678 if (u == phi) { 1679 continue; 1680 } 1681 ok = false; 1682 } 1683 1684 // iff the uses conform 1685 if (ok) { 1686 def_node->add_flag(Node::Flag_is_reduction); 1687 loop_head->mark_has_reductions(); 1688 } 1689 } 1690 } 1691 } 1692 } 1693 } 1694 } 1695 } 1696 1697 //------------------------------dominates_backedge--------------------------------- 1698 // Returns true if ctrl is executed on every complete iteration 1699 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1700 assert(ctrl->is_CFG(), "must be control"); 1701 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1702 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1703 } 1704 1705 //------------------------------adjust_limit----------------------------------- 1706 // Helper function for add_constraint(). 1707 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { 1708 // Compute "I :: (limit-offset)/scale" 1709 Node *con = new SubINode(rc_limit, offset); 1710 register_new_node(con, pre_ctrl); 1711 Node *X = new DivINode(0, con, scale); 1712 register_new_node(X, pre_ctrl); 1713 1714 // Adjust loop limit 1715 loop_limit = (stride_con > 0) 1716 ? (Node*)(new MinINode(loop_limit, X)) 1717 : (Node*)(new MaxINode(loop_limit, X)); 1718 register_new_node(loop_limit, pre_ctrl); 1719 return loop_limit; 1720 } 1721 1722 //------------------------------add_constraint--------------------------------- 1723 // Constrain the main loop iterations so the conditions: 1724 // low_limit <= scale_con * I + offset < upper_limit 1725 // always holds true. That is, either increase the number of iterations in 1726 // the pre-loop or the post-loop until the condition holds true in the main 1727 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1728 // stride and scale are constants (offset and limit often are). 1729 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 ) { 1730 // For positive stride, the pre-loop limit always uses a MAX function 1731 // and the main loop a MIN function. For negative stride these are 1732 // reversed. 1733 1734 // Also for positive stride*scale the affine function is increasing, so the 1735 // pre-loop must check for underflow and the post-loop for overflow. 1736 // Negative stride*scale reverses this; pre-loop checks for overflow and 1737 // post-loop for underflow. 1738 1739 Node *scale = _igvn.intcon(scale_con); 1740 set_ctrl(scale, C->root()); 1741 1742 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 1743 // The overflow limit: scale*I+offset < upper_limit 1744 // For main-loop compute 1745 // ( if (scale > 0) /* and stride > 0 */ 1746 // I < (upper_limit-offset)/scale 1747 // else /* scale < 0 and stride < 0 */ 1748 // I > (upper_limit-offset)/scale 1749 // ) 1750 // 1751 // (upper_limit-offset) may overflow or underflow. 1752 // But it is fine since main loop will either have 1753 // less iterations or will be skipped in such case. 1754 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); 1755 1756 // The underflow limit: low_limit <= scale*I+offset. 1757 // For pre-loop compute 1758 // NOT(scale*I+offset >= low_limit) 1759 // scale*I+offset < low_limit 1760 // ( if (scale > 0) /* and stride > 0 */ 1761 // I < (low_limit-offset)/scale 1762 // else /* scale < 0 and stride < 0 */ 1763 // I > (low_limit-offset)/scale 1764 // ) 1765 1766 if (low_limit->get_int() == -max_jint) { 1767 if (!RangeLimitCheck) return; 1768 // We need this guard when scale*pre_limit+offset >= limit 1769 // due to underflow. So we need execute pre-loop until 1770 // scale*I+offset >= min_int. But (min_int-offset) will 1771 // underflow when offset > 0 and X will be > original_limit 1772 // when stride > 0. To avoid it we replace positive offset with 0. 1773 // 1774 // Also (min_int+1 == -max_int) is used instead of min_int here 1775 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1776 Node* shift = _igvn.intcon(31); 1777 set_ctrl(shift, C->root()); 1778 Node* sign = new RShiftINode(offset, shift); 1779 register_new_node(sign, pre_ctrl); 1780 offset = new AndINode(offset, sign); 1781 register_new_node(offset, pre_ctrl); 1782 } else { 1783 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1784 // The only problem we have here when offset == min_int 1785 // since (0-min_int) == min_int. It may be fine for stride > 0 1786 // but for stride < 0 X will be < original_limit. To avoid it 1787 // max(pre_limit, original_limit) is used in do_range_check(). 1788 } 1789 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1790 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); 1791 1792 } else { // stride_con*scale_con < 0 1793 // For negative stride*scale pre-loop checks for overflow and 1794 // post-loop for underflow. 1795 // 1796 // The overflow limit: scale*I+offset < upper_limit 1797 // For pre-loop compute 1798 // NOT(scale*I+offset < upper_limit) 1799 // scale*I+offset >= upper_limit 1800 // scale*I+offset+1 > upper_limit 1801 // ( if (scale < 0) /* and stride > 0 */ 1802 // I < (upper_limit-(offset+1))/scale 1803 // else /* scale > 0 and stride < 0 */ 1804 // I > (upper_limit-(offset+1))/scale 1805 // ) 1806 // 1807 // (upper_limit-offset-1) may underflow or overflow. 1808 // To avoid it min(pre_limit, original_limit) is used 1809 // in do_range_check() for stride > 0 and max() for < 0. 1810 Node *one = _igvn.intcon(1); 1811 set_ctrl(one, C->root()); 1812 1813 Node *plus_one = new AddINode(offset, one); 1814 register_new_node( plus_one, pre_ctrl ); 1815 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1816 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); 1817 1818 if (low_limit->get_int() == -max_jint) { 1819 if (!RangeLimitCheck) return; 1820 // We need this guard when scale*main_limit+offset >= limit 1821 // due to underflow. So we need execute main-loop while 1822 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 1823 // underflow when (offset+1) > 0 and X will be < main_limit 1824 // when scale < 0 (and stride > 0). To avoid it we replace 1825 // positive (offset+1) with 0. 1826 // 1827 // Also (min_int+1 == -max_int) is used instead of min_int here 1828 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1829 Node* shift = _igvn.intcon(31); 1830 set_ctrl(shift, C->root()); 1831 Node* sign = new RShiftINode(plus_one, shift); 1832 register_new_node(sign, pre_ctrl); 1833 plus_one = new AndINode(plus_one, sign); 1834 register_new_node(plus_one, pre_ctrl); 1835 } else { 1836 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1837 // The only problem we have here when offset == max_int 1838 // since (max_int+1) == min_int and (0-min_int) == min_int. 1839 // But it is fine since main loop will either have 1840 // less iterations or will be skipped in such case. 1841 } 1842 // The underflow limit: low_limit <= scale*I+offset. 1843 // For main-loop compute 1844 // scale*I+offset+1 > low_limit 1845 // ( if (scale < 0) /* and stride > 0 */ 1846 // I < (low_limit-(offset+1))/scale 1847 // else /* scale > 0 and stride < 0 */ 1848 // I > (low_limit-(offset+1))/scale 1849 // ) 1850 1851 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); 1852 } 1853 } 1854 1855 1856 //------------------------------is_scaled_iv--------------------------------- 1857 // Return true if exp is a constant times an induction var 1858 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1859 if (exp == iv) { 1860 if (p_scale != NULL) { 1861 *p_scale = 1; 1862 } 1863 return true; 1864 } 1865 int opc = exp->Opcode(); 1866 if (opc == Op_MulI) { 1867 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1868 if (p_scale != NULL) { 1869 *p_scale = exp->in(2)->get_int(); 1870 } 1871 return true; 1872 } 1873 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1874 if (p_scale != NULL) { 1875 *p_scale = exp->in(1)->get_int(); 1876 } 1877 return true; 1878 } 1879 } else if (opc == Op_LShiftI) { 1880 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1881 if (p_scale != NULL) { 1882 *p_scale = 1 << exp->in(2)->get_int(); 1883 } 1884 return true; 1885 } 1886 } 1887 return false; 1888 } 1889 1890 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1891 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1892 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1893 if (is_scaled_iv(exp, iv, p_scale)) { 1894 if (p_offset != NULL) { 1895 Node *zero = _igvn.intcon(0); 1896 set_ctrl(zero, C->root()); 1897 *p_offset = zero; 1898 } 1899 return true; 1900 } 1901 int opc = exp->Opcode(); 1902 if (opc == Op_AddI) { 1903 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1904 if (p_offset != NULL) { 1905 *p_offset = exp->in(2); 1906 } 1907 return true; 1908 } 1909 if (exp->in(2)->is_Con()) { 1910 Node* offset2 = NULL; 1911 if (depth < 2 && 1912 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1913 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1914 if (p_offset != NULL) { 1915 Node *ctrl_off2 = get_ctrl(offset2); 1916 Node* offset = new AddINode(offset2, exp->in(2)); 1917 register_new_node(offset, ctrl_off2); 1918 *p_offset = offset; 1919 } 1920 return true; 1921 } 1922 } 1923 } else if (opc == Op_SubI) { 1924 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1925 if (p_offset != NULL) { 1926 Node *zero = _igvn.intcon(0); 1927 set_ctrl(zero, C->root()); 1928 Node *ctrl_off = get_ctrl(exp->in(2)); 1929 Node* offset = new SubINode(zero, exp->in(2)); 1930 register_new_node(offset, ctrl_off); 1931 *p_offset = offset; 1932 } 1933 return true; 1934 } 1935 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1936 if (p_offset != NULL) { 1937 *p_scale *= -1; 1938 *p_offset = exp->in(1); 1939 } 1940 return true; 1941 } 1942 } 1943 return false; 1944 } 1945 1946 //------------------------------do_range_check--------------------------------- 1947 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1948 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1949 #ifndef PRODUCT 1950 if (PrintOpto && VerifyLoopOptimizations) { 1951 tty->print("Range Check Elimination "); 1952 loop->dump_head(); 1953 } else if (TraceLoopOpts) { 1954 tty->print("RangeCheck "); 1955 loop->dump_head(); 1956 } 1957 #endif 1958 assert(RangeCheckElimination, ""); 1959 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1960 assert(cl->is_main_loop(), ""); 1961 1962 // protect against stride not being a constant 1963 if (!cl->stride_is_con()) 1964 return; 1965 1966 // Find the trip counter; we are iteration splitting based on it 1967 Node *trip_counter = cl->phi(); 1968 // Find the main loop limit; we will trim it's iterations 1969 // to not ever trip end tests 1970 Node *main_limit = cl->limit(); 1971 1972 // Need to find the main-loop zero-trip guard 1973 Node *ctrl = cl->in(LoopNode::EntryControl); 1974 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 1975 Node *iffm = ctrl->in(0); 1976 assert(iffm->Opcode() == Op_If, ""); 1977 Node *bolzm = iffm->in(1); 1978 assert(bolzm->Opcode() == Op_Bool, ""); 1979 Node *cmpzm = bolzm->in(1); 1980 assert(cmpzm->is_Cmp(), ""); 1981 Node *opqzm = cmpzm->in(2); 1982 // Can not optimize a loop if zero-trip Opaque1 node is optimized 1983 // away and then another round of loop opts attempted. 1984 if (opqzm->Opcode() != Op_Opaque1) 1985 return; 1986 assert(opqzm->in(1) == main_limit, "do not understand situation"); 1987 1988 // Find the pre-loop limit; we will expand its iterations to 1989 // not ever trip low tests. 1990 Node *p_f = iffm->in(0); 1991 // pre loop may have been optimized out 1992 if (p_f->Opcode() != Op_IfFalse) { 1993 return; 1994 } 1995 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 1996 assert(pre_end->loopnode()->is_pre_loop(), ""); 1997 Node *pre_opaq1 = pre_end->limit(); 1998 // Occasionally it's possible for a pre-loop Opaque1 node to be 1999 // optimized away and then another round of loop opts attempted. 2000 // We can not optimize this particular loop in that case. 2001 if (pre_opaq1->Opcode() != Op_Opaque1) 2002 return; 2003 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 2004 Node *pre_limit = pre_opaq->in(1); 2005 2006 // Where do we put new limit calculations 2007 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 2008 2009 // Ensure the original loop limit is available from the 2010 // pre-loop Opaque1 node. 2011 Node *orig_limit = pre_opaq->original_loop_limit(); 2012 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 2013 return; 2014 2015 // Must know if its a count-up or count-down loop 2016 2017 int stride_con = cl->stride_con(); 2018 Node *zero = _igvn.intcon(0); 2019 Node *one = _igvn.intcon(1); 2020 // Use symmetrical int range [-max_jint,max_jint] 2021 Node *mini = _igvn.intcon(-max_jint); 2022 set_ctrl(zero, C->root()); 2023 set_ctrl(one, C->root()); 2024 set_ctrl(mini, C->root()); 2025 2026 // Range checks that do not dominate the loop backedge (ie. 2027 // conditionally executed) can lengthen the pre loop limit beyond 2028 // the original loop limit. To prevent this, the pre limit is 2029 // (for stride > 0) MINed with the original loop limit (MAXed 2030 // stride < 0) when some range_check (rc) is conditionally 2031 // executed. 2032 bool conditional_rc = false; 2033 2034 // Check loop body for tests of trip-counter plus loop-invariant vs 2035 // loop-invariant. 2036 for( uint i = 0; i < loop->_body.size(); i++ ) { 2037 Node *iff = loop->_body[i]; 2038 if( iff->Opcode() == Op_If ) { // Test? 2039 2040 // Test is an IfNode, has 2 projections. If BOTH are in the loop 2041 // we need loop unswitching instead of iteration splitting. 2042 Node *exit = loop->is_loop_exit(iff); 2043 if( !exit ) continue; 2044 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 2045 2046 // Get boolean condition to test 2047 Node *i1 = iff->in(1); 2048 if( !i1->is_Bool() ) continue; 2049 BoolNode *bol = i1->as_Bool(); 2050 BoolTest b_test = bol->_test; 2051 // Flip sense of test if exit condition is flipped 2052 if( flip ) 2053 b_test = b_test.negate(); 2054 2055 // Get compare 2056 Node *cmp = bol->in(1); 2057 2058 // Look for trip_counter + offset vs limit 2059 Node *rc_exp = cmp->in(1); 2060 Node *limit = cmp->in(2); 2061 jint scale_con= 1; // Assume trip counter not scaled 2062 2063 Node *limit_c = get_ctrl(limit); 2064 if( loop->is_member(get_loop(limit_c) ) ) { 2065 // Compare might have operands swapped; commute them 2066 b_test = b_test.commute(); 2067 rc_exp = cmp->in(2); 2068 limit = cmp->in(1); 2069 limit_c = get_ctrl(limit); 2070 if( loop->is_member(get_loop(limit_c) ) ) 2071 continue; // Both inputs are loop varying; cannot RCE 2072 } 2073 // Here we know 'limit' is loop invariant 2074 2075 // 'limit' maybe pinned below the zero trip test (probably from a 2076 // previous round of rce), in which case, it can't be used in the 2077 // zero trip test expression which must occur before the zero test's if. 2078 if( limit_c == ctrl ) { 2079 continue; // Don't rce this check but continue looking for other candidates. 2080 } 2081 2082 // Check for scaled induction variable plus an offset 2083 Node *offset = NULL; 2084 2085 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 2086 continue; 2087 } 2088 2089 Node *offset_c = get_ctrl(offset); 2090 if( loop->is_member( get_loop(offset_c) ) ) 2091 continue; // Offset is not really loop invariant 2092 // Here we know 'offset' is loop invariant. 2093 2094 // As above for the 'limit', the 'offset' maybe pinned below the 2095 // zero trip test. 2096 if( offset_c == ctrl ) { 2097 continue; // Don't rce this check but continue looking for other candidates. 2098 } 2099 #ifdef ASSERT 2100 if (TraceRangeLimitCheck) { 2101 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 2102 bol->dump(2); 2103 } 2104 #endif 2105 // At this point we have the expression as: 2106 // scale_con * trip_counter + offset :: limit 2107 // where scale_con, offset and limit are loop invariant. Trip_counter 2108 // monotonically increases by stride_con, a constant. Both (or either) 2109 // stride_con and scale_con can be negative which will flip about the 2110 // sense of the test. 2111 2112 // Adjust pre and main loop limits to guard the correct iteration set 2113 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 2114 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 2115 // The underflow and overflow limits: 0 <= scale*I+offset < limit 2116 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); 2117 if (!conditional_rc) { 2118 // (0-offset)/scale could be outside of loop iterations range. 2119 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2120 } 2121 } else { 2122 #ifndef PRODUCT 2123 if( PrintOpto ) 2124 tty->print_cr("missed RCE opportunity"); 2125 #endif 2126 continue; // In release mode, ignore it 2127 } 2128 } else { // Otherwise work on normal compares 2129 switch( b_test._test ) { 2130 case BoolTest::gt: 2131 // Fall into GE case 2132 case BoolTest::ge: 2133 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 2134 scale_con = -scale_con; 2135 offset = new SubINode( zero, offset ); 2136 register_new_node( offset, pre_ctrl ); 2137 limit = new SubINode( zero, limit ); 2138 register_new_node( limit, pre_ctrl ); 2139 // Fall into LE case 2140 case BoolTest::le: 2141 if (b_test._test != BoolTest::gt) { 2142 // Convert X <= Y to X < Y+1 2143 limit = new AddINode( limit, one ); 2144 register_new_node( limit, pre_ctrl ); 2145 } 2146 // Fall into LT case 2147 case BoolTest::lt: 2148 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 2149 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 2150 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 2151 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); 2152 if (!conditional_rc) { 2153 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 2154 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 2155 // still be outside of loop range. 2156 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2157 } 2158 break; 2159 default: 2160 #ifndef PRODUCT 2161 if( PrintOpto ) 2162 tty->print_cr("missed RCE opportunity"); 2163 #endif 2164 continue; // Unhandled case 2165 } 2166 } 2167 2168 // Kill the eliminated test 2169 C->set_major_progress(); 2170 Node *kill_con = _igvn.intcon( 1-flip ); 2171 set_ctrl(kill_con, C->root()); 2172 _igvn.replace_input_of(iff, 1, kill_con); 2173 // Find surviving projection 2174 assert(iff->is_If(), ""); 2175 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 2176 // Find loads off the surviving projection; remove their control edge 2177 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2178 Node* cd = dp->fast_out(i); // Control-dependent node 2179 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop 2180 // Allow the load to float around in the loop, or before it 2181 // but NOT before the pre-loop. 2182 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 2183 --i; 2184 --imax; 2185 } 2186 } 2187 2188 } // End of is IF 2189 2190 } 2191 2192 // Update loop limits 2193 if (conditional_rc) { 2194 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit) 2195 : (Node*)new MaxINode(pre_limit, orig_limit); 2196 register_new_node(pre_limit, pre_ctrl); 2197 } 2198 _igvn.replace_input_of(pre_opaq, 1, pre_limit); 2199 2200 // Note:: we are making the main loop limit no longer precise; 2201 // need to round up based on stride. 2202 cl->set_nonexact_trip_count(); 2203 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case 2204 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 2205 // Hopefully, compiler will optimize for powers of 2. 2206 Node *ctrl = get_ctrl(main_limit); 2207 Node *stride = cl->stride(); 2208 Node *init = cl->init_trip()->uncast(); 2209 Node *span = new SubINode(main_limit,init); 2210 register_new_node(span,ctrl); 2211 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 2212 Node *add = new AddINode(span,rndup); 2213 register_new_node(add,ctrl); 2214 Node *div = new DivINode(0,add,stride); 2215 register_new_node(div,ctrl); 2216 Node *mul = new MulINode(div,stride); 2217 register_new_node(mul,ctrl); 2218 Node *newlim = new AddINode(mul,init); 2219 register_new_node(newlim,ctrl); 2220 main_limit = newlim; 2221 } 2222 2223 Node *main_cle = cl->loopexit(); 2224 Node *main_bol = main_cle->in(1); 2225 // Hacking loop bounds; need private copies of exit test 2226 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 2227 main_bol = main_bol->clone();// Clone a private BoolNode 2228 register_new_node( main_bol, main_cle->in(0) ); 2229 _igvn.replace_input_of(main_cle, 1, main_bol); 2230 } 2231 Node *main_cmp = main_bol->in(1); 2232 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 2233 main_cmp = main_cmp->clone();// Clone a private CmpNode 2234 register_new_node( main_cmp, main_cle->in(0) ); 2235 _igvn.replace_input_of(main_bol, 1, main_cmp); 2236 } 2237 // Hack the now-private loop bounds 2238 _igvn.replace_input_of(main_cmp, 2, main_limit); 2239 // The OpaqueNode is unshared by design 2240 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 2241 _igvn.replace_input_of(opqzm, 1, main_limit); 2242 } 2243 2244 //------------------------------DCE_loop_body---------------------------------- 2245 // Remove simplistic dead code from loop body 2246 void IdealLoopTree::DCE_loop_body() { 2247 for( uint i = 0; i < _body.size(); i++ ) 2248 if( _body.at(i)->outcnt() == 0 ) 2249 _body.map( i--, _body.pop() ); 2250 } 2251 2252 2253 //------------------------------adjust_loop_exit_prob-------------------------- 2254 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2255 // Replace with a 1-in-10 exit guess. 2256 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 2257 Node *test = tail(); 2258 while( test != _head ) { 2259 uint top = test->Opcode(); 2260 if( top == Op_IfTrue || top == Op_IfFalse ) { 2261 int test_con = ((ProjNode*)test)->_con; 2262 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2263 IfNode *iff = test->in(0)->as_If(); 2264 if( iff->outcnt() == 2 ) { // Ignore dead tests 2265 Node *bol = iff->in(1); 2266 if( bol && bol->req() > 1 && bol->in(1) && 2267 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 2268 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 2269 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 2270 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 2271 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 2272 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 2273 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 2274 return; // Allocation loops RARELY take backedge 2275 // Find the OTHER exit path from the IF 2276 Node* ex = iff->proj_out(1-test_con); 2277 float p = iff->_prob; 2278 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 2279 if( top == Op_IfTrue ) { 2280 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2281 iff->_prob = PROB_STATIC_FREQUENT; 2282 } 2283 } else { 2284 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2285 iff->_prob = PROB_STATIC_INFREQUENT; 2286 } 2287 } 2288 } 2289 } 2290 } 2291 test = phase->idom(test); 2292 } 2293 } 2294 2295 #ifdef ASSERT 2296 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) { 2297 Node *ctrl = cl->in(LoopNode::EntryControl); 2298 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 2299 Node *iffm = ctrl->in(0); 2300 assert(iffm->Opcode() == Op_If, ""); 2301 Node *p_f = iffm->in(0); 2302 assert(p_f->Opcode() == Op_IfFalse, ""); 2303 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2304 assert(pre_end->loopnode()->is_pre_loop(), ""); 2305 return pre_end->loopnode(); 2306 } 2307 #endif 2308 2309 // Remove the main and post loops and make the pre loop execute all 2310 // iterations. Useful when the pre loop is found empty. 2311 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) { 2312 CountedLoopEndNode* pre_end = cl->loopexit(); 2313 Node* pre_cmp = pre_end->cmp_node(); 2314 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) { 2315 // Only safe to remove the main loop if the compiler optimized it 2316 // out based on an unknown number of iterations 2317 return; 2318 } 2319 2320 // Can we find the main loop? 2321 if (_next == NULL) { 2322 return; 2323 } 2324 2325 Node* next_head = _next->_head; 2326 if (!next_head->is_CountedLoop()) { 2327 return; 2328 } 2329 2330 CountedLoopNode* main_head = next_head->as_CountedLoop(); 2331 if (!main_head->is_main_loop()) { 2332 return; 2333 } 2334 2335 assert(locate_pre_from_main(main_head) == cl, "bad main loop"); 2336 Node* main_iff = main_head->in(LoopNode::EntryControl)->in(0); 2337 2338 // Remove the Opaque1Node of the pre loop and make it execute all iterations 2339 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2)); 2340 // Remove the Opaque1Node of the main loop so it can be optimized out 2341 Node* main_cmp = main_iff->in(1)->in(1); 2342 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?"); 2343 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1)); 2344 } 2345 2346 //------------------------------policy_do_remove_empty_loop-------------------- 2347 // Micro-benchmark spamming. Policy is to always remove empty loops. 2348 // The 'DO' part is to replace the trip counter with the value it will 2349 // have on the last iteration. This will break the loop. 2350 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 2351 // Minimum size must be empty loop 2352 if (_body.size() > EMPTY_LOOP_SIZE) 2353 return false; 2354 2355 if (!_head->is_CountedLoop()) 2356 return false; // Dead loop 2357 CountedLoopNode *cl = _head->as_CountedLoop(); 2358 if (!cl->is_valid_counted_loop()) 2359 return false; // Malformed loop 2360 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 2361 return false; // Infinite loop 2362 2363 if (cl->is_pre_loop()) { 2364 // If the loop we are removing is a pre-loop then the main and 2365 // post loop can be removed as well 2366 remove_main_post_loops(cl, phase); 2367 } 2368 2369 #ifdef ASSERT 2370 // Ensure only one phi which is the iv. 2371 Node* iv = NULL; 2372 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 2373 Node* n = cl->fast_out(i); 2374 if (n->Opcode() == Op_Phi) { 2375 assert(iv == NULL, "Too many phis" ); 2376 iv = n; 2377 } 2378 } 2379 assert(iv == cl->phi(), "Wrong phi" ); 2380 #endif 2381 2382 // main and post loops have explicitly created zero trip guard 2383 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 2384 if (needs_guard) { 2385 // Skip guard if values not overlap. 2386 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 2387 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 2388 int stride_con = cl->stride_con(); 2389 if (stride_con > 0) { 2390 needs_guard = (init_t->_hi >= limit_t->_lo); 2391 } else { 2392 needs_guard = (init_t->_lo <= limit_t->_hi); 2393 } 2394 } 2395 if (needs_guard) { 2396 // Check for an obvious zero trip guard. 2397 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 2398 if (inctrl->Opcode() == Op_IfTrue) { 2399 // The test should look like just the backedge of a CountedLoop 2400 Node* iff = inctrl->in(0); 2401 if (iff->is_If()) { 2402 Node* bol = iff->in(1); 2403 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { 2404 Node* cmp = bol->in(1); 2405 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { 2406 needs_guard = false; 2407 } 2408 } 2409 } 2410 } 2411 } 2412 2413 #ifndef PRODUCT 2414 if (PrintOpto) { 2415 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 2416 this->dump_head(); 2417 } else if (TraceLoopOpts) { 2418 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 2419 this->dump_head(); 2420 } 2421 #endif 2422 2423 if (needs_guard) { 2424 // Peel the loop to ensure there's a zero trip guard 2425 Node_List old_new; 2426 phase->do_peeling(this, old_new); 2427 } 2428 2429 // Replace the phi at loop head with the final value of the last 2430 // iteration. Then the CountedLoopEnd will collapse (backedge never 2431 // taken) and all loop-invariant uses of the exit values will be correct. 2432 Node *phi = cl->phi(); 2433 Node *exact_limit = phase->exact_limit(this); 2434 if (exact_limit != cl->limit()) { 2435 // We also need to replace the original limit to collapse loop exit. 2436 Node* cmp = cl->loopexit()->cmp_node(); 2437 assert(cl->limit() == cmp->in(2), "sanity"); 2438 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 2439 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 2440 } 2441 // Note: the final value after increment should not overflow since 2442 // counted loop has limit check predicate. 2443 Node *final = new SubINode( exact_limit, cl->stride() ); 2444 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 2445 phase->_igvn.replace_node(phi,final); 2446 phase->C->set_major_progress(); 2447 return true; 2448 } 2449 2450 //------------------------------policy_do_one_iteration_loop------------------- 2451 // Convert one iteration loop into normal code. 2452 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 2453 if (!_head->as_Loop()->is_valid_counted_loop()) 2454 return false; // Only for counted loop 2455 2456 CountedLoopNode *cl = _head->as_CountedLoop(); 2457 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 2458 return false; 2459 } 2460 2461 #ifndef PRODUCT 2462 if(TraceLoopOpts) { 2463 tty->print("OneIteration "); 2464 this->dump_head(); 2465 } 2466 #endif 2467 2468 Node *init_n = cl->init_trip(); 2469 #ifdef ASSERT 2470 // Loop boundaries should be constant since trip count is exact. 2471 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 2472 #endif 2473 // Replace the phi at loop head with the value of the init_trip. 2474 // Then the CountedLoopEnd will collapse (backedge will not be taken) 2475 // and all loop-invariant uses of the exit values will be correct. 2476 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 2477 phase->C->set_major_progress(); 2478 return true; 2479 } 2480 2481 //============================================================================= 2482 //------------------------------iteration_split_impl--------------------------- 2483 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 2484 // Compute exact loop trip count if possible. 2485 compute_exact_trip_count(phase); 2486 2487 // Convert one iteration loop into normal code. 2488 if (policy_do_one_iteration_loop(phase)) 2489 return true; 2490 2491 // Check and remove empty loops (spam micro-benchmarks) 2492 if (policy_do_remove_empty_loop(phase)) 2493 return true; // Here we removed an empty loop 2494 2495 bool should_peel = policy_peeling(phase); // Should we peel? 2496 2497 bool should_unswitch = policy_unswitching(phase); 2498 2499 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 2500 // This removes loop-invariant tests (usually null checks). 2501 if (!_head->is_CountedLoop()) { // Non-counted loop 2502 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 2503 // Partial peel succeeded so terminate this round of loop opts 2504 return false; 2505 } 2506 if (should_peel) { // Should we peel? 2507 #ifndef PRODUCT 2508 if (PrintOpto) tty->print_cr("should_peel"); 2509 #endif 2510 phase->do_peeling(this,old_new); 2511 } else if (should_unswitch) { 2512 phase->do_unswitching(this, old_new); 2513 } 2514 return true; 2515 } 2516 CountedLoopNode *cl = _head->as_CountedLoop(); 2517 2518 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 2519 2520 // Do nothing special to pre- and post- loops 2521 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 2522 2523 // Compute loop trip count from profile data 2524 compute_profile_trip_cnt(phase); 2525 2526 // Before attempting fancy unrolling, RCE or alignment, see if we want 2527 // to completely unroll this loop or do loop unswitching. 2528 if (cl->is_normal_loop()) { 2529 if (should_unswitch) { 2530 phase->do_unswitching(this, old_new); 2531 return true; 2532 } 2533 bool should_maximally_unroll = policy_maximally_unroll(phase); 2534 if (should_maximally_unroll) { 2535 // Here we did some unrolling and peeling. Eventually we will 2536 // completely unroll this loop and it will no longer be a loop. 2537 phase->do_maximally_unroll(this,old_new); 2538 return true; 2539 } 2540 } 2541 2542 // Skip next optimizations if running low on nodes. Note that 2543 // policy_unswitching and policy_maximally_unroll have this check. 2544 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes(); 2545 if ((int)(2 * _body.size()) > nodes_left) { 2546 return true; 2547 } 2548 2549 // Counted loops may be peeled, may need some iterations run up 2550 // front for RCE, and may want to align loop refs to a cache 2551 // line. Thus we clone a full loop up front whose trip count is 2552 // at least 1 (if peeling), but may be several more. 2553 2554 // The main loop will start cache-line aligned with at least 1 2555 // iteration of the unrolled body (zero-trip test required) and 2556 // will have some range checks removed. 2557 2558 // A post-loop will finish any odd iterations (leftover after 2559 // unrolling), plus any needed for RCE purposes. 2560 2561 bool should_unroll = policy_unroll(phase); 2562 2563 bool should_rce = policy_range_check(phase); 2564 2565 bool should_align = policy_align(phase); 2566 2567 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2568 // need a pre-loop. We may still need to peel an initial iteration but 2569 // we will not be needing an unknown number of pre-iterations. 2570 // 2571 // Basically, if may_rce_align reports FALSE first time through, 2572 // we will not be able to later do RCE or Aligning on this loop. 2573 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2574 2575 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2576 // we switch to the pre-/main-/post-loop model. This model also covers 2577 // peeling. 2578 if (should_rce || should_align || should_unroll) { 2579 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2580 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2581 2582 // Adjust the pre- and main-loop limits to let the pre and post loops run 2583 // with full checks, but the main-loop with no checks. Remove said 2584 // checks from the main body. 2585 if (should_rce) 2586 phase->do_range_check(this,old_new); 2587 2588 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2589 // twice as many iterations as before) and the main body limit (only do 2590 // an even number of trips). If we are peeling, we might enable some RCE 2591 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2592 // peeling. 2593 if (should_unroll && !should_peel) { 2594 phase->do_unroll(this, old_new, true); 2595 } 2596 2597 // Adjust the pre-loop limits to align the main body 2598 // iterations. 2599 if (should_align) 2600 Unimplemented(); 2601 2602 } else { // Else we have an unchanged counted loop 2603 if (should_peel) // Might want to peel but do nothing else 2604 phase->do_peeling(this,old_new); 2605 } 2606 return true; 2607 } 2608 2609 2610 //============================================================================= 2611 //------------------------------iteration_split-------------------------------- 2612 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2613 // Recursively iteration split nested loops 2614 if (_child && !_child->iteration_split(phase, old_new)) 2615 return false; 2616 2617 // Clean out prior deadwood 2618 DCE_loop_body(); 2619 2620 2621 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2622 // Replace with a 1-in-10 exit guess. 2623 if (_parent /*not the root loop*/ && 2624 !_irreducible && 2625 // Also ignore the occasional dead backedge 2626 !tail()->is_top()) { 2627 adjust_loop_exit_prob(phase); 2628 } 2629 2630 // Gate unrolling, RCE and peeling efforts. 2631 if (!_child && // If not an inner loop, do not split 2632 !_irreducible && 2633 _allow_optimizations && 2634 !tail()->is_top()) { // Also ignore the occasional dead backedge 2635 if (!_has_call) { 2636 if (!iteration_split_impl(phase, old_new)) { 2637 return false; 2638 } 2639 } else if (policy_unswitching(phase)) { 2640 phase->do_unswitching(this, old_new); 2641 } 2642 } 2643 2644 // Minor offset re-organization to remove loop-fallout uses of 2645 // trip counter when there was no major reshaping. 2646 phase->reorg_offsets(this); 2647 2648 if (_next && !_next->iteration_split(phase, old_new)) 2649 return false; 2650 return true; 2651 } 2652 2653 2654 //============================================================================= 2655 // Process all the loops in the loop tree and replace any fill 2656 // patterns with an intrisc version. 2657 bool PhaseIdealLoop::do_intrinsify_fill() { 2658 bool changed = false; 2659 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2660 IdealLoopTree* lpt = iter.current(); 2661 changed |= intrinsify_fill(lpt); 2662 } 2663 return changed; 2664 } 2665 2666 2667 // Examine an inner loop looking for a a single store of an invariant 2668 // value in a unit stride loop, 2669 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2670 Node*& shift, Node*& con) { 2671 const char* msg = NULL; 2672 Node* msg_node = NULL; 2673 2674 store_value = NULL; 2675 con = NULL; 2676 shift = NULL; 2677 2678 // Process the loop looking for stores. If there are multiple 2679 // stores or extra control flow give at this point. 2680 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2681 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2682 Node* n = lpt->_body.at(i); 2683 if (n->outcnt() == 0) continue; // Ignore dead 2684 if (n->is_Store()) { 2685 if (store != NULL) { 2686 msg = "multiple stores"; 2687 break; 2688 } 2689 int opc = n->Opcode(); 2690 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 2691 msg = "oop fills not handled"; 2692 break; 2693 } 2694 Node* value = n->in(MemNode::ValueIn); 2695 if (!lpt->is_invariant(value)) { 2696 msg = "variant store value"; 2697 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2698 msg = "not array address"; 2699 } 2700 store = n; 2701 store_value = value; 2702 } else if (n->is_If() && n != head->loopexit()) { 2703 msg = "extra control flow"; 2704 msg_node = n; 2705 } 2706 } 2707 2708 if (store == NULL) { 2709 // No store in loop 2710 return false; 2711 } 2712 2713 if (msg == NULL && head->stride_con() != 1) { 2714 // could handle negative strides too 2715 if (head->stride_con() < 0) { 2716 msg = "negative stride"; 2717 } else { 2718 msg = "non-unit stride"; 2719 } 2720 } 2721 2722 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2723 msg = "can't handle store address"; 2724 msg_node = store->in(MemNode::Address); 2725 } 2726 2727 if (msg == NULL && 2728 (!store->in(MemNode::Memory)->is_Phi() || 2729 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2730 msg = "store memory isn't proper phi"; 2731 msg_node = store->in(MemNode::Memory); 2732 } 2733 2734 // Make sure there is an appropriate fill routine 2735 BasicType t = store->as_Mem()->memory_type(); 2736 const char* fill_name; 2737 if (msg == NULL && 2738 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2739 msg = "unsupported store"; 2740 msg_node = store; 2741 } 2742 2743 if (msg != NULL) { 2744 #ifndef PRODUCT 2745 if (TraceOptimizeFill) { 2746 tty->print_cr("not fill intrinsic candidate: %s", msg); 2747 if (msg_node != NULL) msg_node->dump(); 2748 } 2749 #endif 2750 return false; 2751 } 2752 2753 // Make sure the address expression can be handled. It should be 2754 // head->phi * elsize + con. head->phi might have a ConvI2L. 2755 Node* elements[4]; 2756 Node* conv = NULL; 2757 bool found_index = false; 2758 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2759 for (int e = 0; e < count; e++) { 2760 Node* n = elements[e]; 2761 if (n->is_Con() && con == NULL) { 2762 con = n; 2763 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2764 Node* value = n->in(1); 2765 #ifdef _LP64 2766 if (value->Opcode() == Op_ConvI2L) { 2767 conv = value; 2768 value = value->in(1); 2769 } 2770 #endif 2771 if (value != head->phi()) { 2772 msg = "unhandled shift in address"; 2773 } else { 2774 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 2775 msg = "scale doesn't match"; 2776 } else { 2777 found_index = true; 2778 shift = n; 2779 } 2780 } 2781 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2782 if (n->in(1) == head->phi()) { 2783 found_index = true; 2784 conv = n; 2785 } else { 2786 msg = "unhandled input to ConvI2L"; 2787 } 2788 } else if (n == head->phi()) { 2789 // no shift, check below for allowed cases 2790 found_index = true; 2791 } else { 2792 msg = "unhandled node in address"; 2793 msg_node = n; 2794 } 2795 } 2796 2797 if (count == -1) { 2798 msg = "malformed address expression"; 2799 msg_node = store; 2800 } 2801 2802 if (!found_index) { 2803 msg = "missing use of index"; 2804 } 2805 2806 // byte sized items won't have a shift 2807 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2808 msg = "can't find shift"; 2809 msg_node = store; 2810 } 2811 2812 if (msg != NULL) { 2813 #ifndef PRODUCT 2814 if (TraceOptimizeFill) { 2815 tty->print_cr("not fill intrinsic: %s", msg); 2816 if (msg_node != NULL) msg_node->dump(); 2817 } 2818 #endif 2819 return false; 2820 } 2821 2822 // No make sure all the other nodes in the loop can be handled 2823 VectorSet ok(Thread::current()->resource_area()); 2824 2825 // store related values are ok 2826 ok.set(store->_idx); 2827 ok.set(store->in(MemNode::Memory)->_idx); 2828 2829 CountedLoopEndNode* loop_exit = head->loopexit(); 2830 guarantee(loop_exit != NULL, "no loop exit node"); 2831 2832 // Loop structure is ok 2833 ok.set(head->_idx); 2834 ok.set(loop_exit->_idx); 2835 ok.set(head->phi()->_idx); 2836 ok.set(head->incr()->_idx); 2837 ok.set(loop_exit->cmp_node()->_idx); 2838 ok.set(loop_exit->in(1)->_idx); 2839 2840 // Address elements are ok 2841 if (con) ok.set(con->_idx); 2842 if (shift) ok.set(shift->_idx); 2843 if (conv) ok.set(conv->_idx); 2844 2845 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2846 Node* n = lpt->_body.at(i); 2847 if (n->outcnt() == 0) continue; // Ignore dead 2848 if (ok.test(n->_idx)) continue; 2849 // Backedge projection is ok 2850 if (n->is_IfTrue() && n->in(0) == loop_exit) continue; 2851 if (!n->is_AddP()) { 2852 msg = "unhandled node"; 2853 msg_node = n; 2854 break; 2855 } 2856 } 2857 2858 // Make sure no unexpected values are used outside the loop 2859 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2860 Node* n = lpt->_body.at(i); 2861 // These values can be replaced with other nodes if they are used 2862 // outside the loop. 2863 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2864 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2865 Node* use = iter.get(); 2866 if (!lpt->_body.contains(use)) { 2867 msg = "node is used outside loop"; 2868 // lpt->_body.dump(); 2869 msg_node = n; 2870 break; 2871 } 2872 } 2873 } 2874 2875 #ifdef ASSERT 2876 if (TraceOptimizeFill) { 2877 if (msg != NULL) { 2878 tty->print_cr("no fill intrinsic: %s", msg); 2879 if (msg_node != NULL) msg_node->dump(); 2880 } else { 2881 tty->print_cr("fill intrinsic for:"); 2882 } 2883 store->dump(); 2884 if (Verbose) { 2885 lpt->_body.dump(); 2886 } 2887 } 2888 #endif 2889 2890 return msg == NULL; 2891 } 2892 2893 2894 2895 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2896 // Only for counted inner loops 2897 if (!lpt->is_counted() || !lpt->is_inner()) { 2898 return false; 2899 } 2900 2901 // Must have constant stride 2902 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2903 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 2904 return false; 2905 } 2906 2907 // Check that the body only contains a store of a loop invariant 2908 // value that is indexed by the loop phi. 2909 Node* store = NULL; 2910 Node* store_value = NULL; 2911 Node* shift = NULL; 2912 Node* offset = NULL; 2913 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2914 return false; 2915 } 2916 2917 #ifndef PRODUCT 2918 if (TraceLoopOpts) { 2919 tty->print("ArrayFill "); 2920 lpt->dump_head(); 2921 } 2922 #endif 2923 2924 // Now replace the whole loop body by a call to a fill routine that 2925 // covers the same region as the loop. 2926 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2927 2928 // Build an expression for the beginning of the copy region 2929 Node* index = head->init_trip(); 2930 #ifdef _LP64 2931 index = new ConvI2LNode(index); 2932 _igvn.register_new_node_with_optimizer(index); 2933 #endif 2934 if (shift != NULL) { 2935 // byte arrays don't require a shift but others do. 2936 index = new LShiftXNode(index, shift->in(2)); 2937 _igvn.register_new_node_with_optimizer(index); 2938 } 2939 index = new AddPNode(base, base, index); 2940 _igvn.register_new_node_with_optimizer(index); 2941 Node* from = new AddPNode(base, index, offset); 2942 _igvn.register_new_node_with_optimizer(from); 2943 // Compute the number of elements to copy 2944 Node* len = new SubINode(head->limit(), head->init_trip()); 2945 _igvn.register_new_node_with_optimizer(len); 2946 2947 BasicType t = store->as_Mem()->memory_type(); 2948 bool aligned = false; 2949 if (offset != NULL && head->init_trip()->is_Con()) { 2950 int element_size = type2aelembytes(t); 2951 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2952 } 2953 2954 // Build a call to the fill routine 2955 const char* fill_name; 2956 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2957 assert(fill != NULL, "what?"); 2958 2959 // Convert float/double to int/long for fill routines 2960 if (t == T_FLOAT) { 2961 store_value = new MoveF2INode(store_value); 2962 _igvn.register_new_node_with_optimizer(store_value); 2963 } else if (t == T_DOUBLE) { 2964 store_value = new MoveD2LNode(store_value); 2965 _igvn.register_new_node_with_optimizer(store_value); 2966 } 2967 2968 Node* mem_phi = store->in(MemNode::Memory); 2969 Node* result_ctrl; 2970 Node* result_mem; 2971 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2972 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill, 2973 fill_name, TypeAryPtr::get_array_body_type(t)); 2974 uint cnt = 0; 2975 call->init_req(TypeFunc::Parms + cnt++, from); 2976 call->init_req(TypeFunc::Parms + cnt++, store_value); 2977 #ifdef _LP64 2978 len = new ConvI2LNode(len); 2979 _igvn.register_new_node_with_optimizer(len); 2980 #endif 2981 call->init_req(TypeFunc::Parms + cnt++, len); 2982 #ifdef _LP64 2983 call->init_req(TypeFunc::Parms + cnt++, C->top()); 2984 #endif 2985 call->init_req(TypeFunc::Control, head->init_control()); 2986 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. 2987 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); 2988 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr)); 2989 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr)); 2990 _igvn.register_new_node_with_optimizer(call); 2991 result_ctrl = new ProjNode(call,TypeFunc::Control); 2992 _igvn.register_new_node_with_optimizer(result_ctrl); 2993 result_mem = new ProjNode(call,TypeFunc::Memory); 2994 _igvn.register_new_node_with_optimizer(result_mem); 2995 2996 /* Disable following optimization until proper fix (add missing checks). 2997 2998 // If this fill is tightly coupled to an allocation and overwrites 2999 // the whole body, allow it to take over the zeroing. 3000 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 3001 if (alloc != NULL && alloc->is_AllocateArray()) { 3002 Node* length = alloc->as_AllocateArray()->Ideal_length(); 3003 if (head->limit() == length && 3004 head->init_trip() == _igvn.intcon(0)) { 3005 if (TraceOptimizeFill) { 3006 tty->print_cr("Eliminated zeroing in allocation"); 3007 } 3008 alloc->maybe_set_complete(&_igvn); 3009 } else { 3010 #ifdef ASSERT 3011 if (TraceOptimizeFill) { 3012 tty->print_cr("filling array but bounds don't match"); 3013 alloc->dump(); 3014 head->init_trip()->dump(); 3015 head->limit()->dump(); 3016 length->dump(); 3017 } 3018 #endif 3019 } 3020 } 3021 */ 3022 3023 // Redirect the old control and memory edges that are outside the loop. 3024 Node* exit = head->loopexit()->proj_out(0); 3025 // Sometimes the memory phi of the head is used as the outgoing 3026 // state of the loop. It's safe in this case to replace it with the 3027 // result_mem. 3028 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 3029 _igvn.replace_node(exit, result_ctrl); 3030 _igvn.replace_node(store, result_mem); 3031 // Any uses the increment outside of the loop become the loop limit. 3032 _igvn.replace_node(head->incr(), head->limit()); 3033 3034 // Disconnect the head from the loop. 3035 for (uint i = 0; i < lpt->_body.size(); i++) { 3036 Node* n = lpt->_body.at(i); 3037 _igvn.replace_node(n, C->top()); 3038 } 3039 3040 return true; 3041 }