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