1 /* 2 * Copyright (c) 2000, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/convertnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opaquenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 #include "opto/superword.hpp" 42 #include "opto/vectornode.hpp" 43 44 //------------------------------is_loop_exit----------------------------------- 45 // Given an IfNode, return the loop-exiting projection or NULL if both 46 // arms remain in the loop. 47 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 48 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 49 PhaseIdealLoop *phase = _phase; 50 // Test is an IfNode, has 2 projections. If BOTH are in the loop 51 // we need loop unswitching instead of peeling. 52 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 53 return iff->raw_out(0); 54 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 55 return iff->raw_out(1); 56 return NULL; 57 } 58 59 60 //============================================================================= 61 62 63 //------------------------------record_for_igvn---------------------------- 64 // Put loop body on igvn work list 65 void IdealLoopTree::record_for_igvn() { 66 for( uint i = 0; i < _body.size(); i++ ) { 67 Node *n = _body.at(i); 68 _phase->_igvn._worklist.push(n); 69 } 70 } 71 72 //------------------------------compute_exact_trip_count----------------------- 73 // Compute loop exact trip count if possible. Do not recalculate trip count for 74 // split loops (pre-main-post) which have their limits and inits behind Opaque node. 75 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) { 76 if (!_head->as_Loop()->is_valid_counted_loop()) { 77 return; 78 } 79 CountedLoopNode* cl = _head->as_CountedLoop(); 80 // Trip count may become nonexact for iteration split loops since 81 // RCE modifies limits. Note, _trip_count value is not reset since 82 // it is used to limit unrolling of main loop. 83 cl->set_nonexact_trip_count(); 84 85 // Loop's test should be part of loop. 86 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 87 return; // Infinite loop 88 89 #ifdef ASSERT 90 BoolTest::mask bt = cl->loopexit()->test_trip(); 91 assert(bt == BoolTest::lt || bt == BoolTest::gt || 92 bt == BoolTest::ne, "canonical test is expected"); 93 #endif 94 95 Node* init_n = cl->init_trip(); 96 Node* limit_n = cl->limit(); 97 if (init_n != NULL && init_n->is_Con() && 98 limit_n != NULL && limit_n->is_Con()) { 99 // Use longs to avoid integer overflow. 100 int stride_con = cl->stride_con(); 101 jlong init_con = cl->init_trip()->get_int(); 102 jlong limit_con = cl->limit()->get_int(); 103 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 104 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 105 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { 106 // Set exact trip count. 107 cl->set_exact_trip_count((uint)trip_count); 108 } 109 } 110 } 111 112 //------------------------------compute_profile_trip_cnt---------------------------- 113 // Compute loop trip count from profile data as 114 // (backedge_count + loop_exit_count) / loop_exit_count 115 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 116 if (!_head->is_CountedLoop()) { 117 return; 118 } 119 CountedLoopNode* head = _head->as_CountedLoop(); 120 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 121 return; // Already computed 122 } 123 float trip_cnt = (float)max_jint; // default is big 124 125 Node* back = head->in(LoopNode::LoopBackControl); 126 while (back != head) { 127 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 128 back->in(0) && 129 back->in(0)->is_If() && 130 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 131 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 132 break; 133 } 134 back = phase->idom(back); 135 } 136 if (back != head) { 137 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 138 back->in(0), "if-projection exists"); 139 IfNode* back_if = back->in(0)->as_If(); 140 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 141 142 // Now compute a loop exit count 143 float loop_exit_cnt = 0.0f; 144 for( uint i = 0; i < _body.size(); i++ ) { 145 Node *n = _body[i]; 146 if( n->is_If() ) { 147 IfNode *iff = n->as_If(); 148 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 149 Node *exit = is_loop_exit(iff); 150 if( exit ) { 151 float exit_prob = iff->_prob; 152 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 153 if (exit_prob > PROB_MIN) { 154 float exit_cnt = iff->_fcnt * exit_prob; 155 loop_exit_cnt += exit_cnt; 156 } 157 } 158 } 159 } 160 } 161 if (loop_exit_cnt > 0.0f) { 162 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 163 } else { 164 // No exit count so use 165 trip_cnt = loop_back_cnt; 166 } 167 } 168 #ifndef PRODUCT 169 if (TraceProfileTripCount) { 170 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 171 } 172 #endif 173 head->set_profile_trip_cnt(trip_cnt); 174 } 175 176 //---------------------is_invariant_addition----------------------------- 177 // Return nonzero index of invariant operand for an Add or Sub 178 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 179 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 180 int op = n->Opcode(); 181 if (op == Op_AddI || op == Op_SubI) { 182 bool in1_invar = this->is_invariant(n->in(1)); 183 bool in2_invar = this->is_invariant(n->in(2)); 184 if (in1_invar && !in2_invar) return 1; 185 if (!in1_invar && in2_invar) return 2; 186 } 187 return 0; 188 } 189 190 //---------------------reassociate_add_sub----------------------------- 191 // Reassociate invariant add and subtract expressions: 192 // 193 // inv1 + (x + inv2) => ( inv1 + inv2) + x 194 // (x + inv2) + inv1 => ( inv1 + inv2) + x 195 // inv1 + (x - inv2) => ( inv1 - inv2) + x 196 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 197 // (x + inv2) - inv1 => (-inv1 + inv2) + x 198 // (x - inv2) + inv1 => ( inv1 - inv2) + x 199 // (x - inv2) - inv1 => (-inv1 - inv2) + x 200 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 201 // inv1 - (x - inv2) => ( inv1 + inv2) - x 202 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 203 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 204 // inv1 - (x + inv2) => ( inv1 - inv2) - x 205 // 206 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 207 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 208 if (is_invariant(n1)) return NULL; 209 int inv1_idx = is_invariant_addition(n1, phase); 210 if (!inv1_idx) return NULL; 211 // Don't mess with add of constant (igvn moves them to expression tree root.) 212 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 213 Node* inv1 = n1->in(inv1_idx); 214 Node* n2 = n1->in(3 - inv1_idx); 215 int inv2_idx = is_invariant_addition(n2, phase); 216 if (!inv2_idx) return NULL; 217 Node* x = n2->in(3 - inv2_idx); 218 Node* inv2 = n2->in(inv2_idx); 219 220 bool neg_x = n2->is_Sub() && inv2_idx == 1; 221 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 222 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 223 if (n1->is_Sub() && inv1_idx == 1) { 224 neg_x = !neg_x; 225 neg_inv2 = !neg_inv2; 226 } 227 Node* inv1_c = phase->get_ctrl(inv1); 228 Node* inv2_c = phase->get_ctrl(inv2); 229 Node* n_inv1; 230 if (neg_inv1) { 231 Node *zero = phase->_igvn.intcon(0); 232 phase->set_ctrl(zero, phase->C->root()); 233 n_inv1 = new SubINode(zero, inv1); 234 phase->register_new_node(n_inv1, inv1_c); 235 } else { 236 n_inv1 = inv1; 237 } 238 Node* inv; 239 if (neg_inv2) { 240 inv = new SubINode(n_inv1, inv2); 241 } else { 242 inv = new AddINode(n_inv1, inv2); 243 } 244 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 245 246 Node* addx; 247 if (neg_x) { 248 addx = new SubINode(inv, x); 249 } else { 250 addx = new AddINode(x, inv); 251 } 252 phase->register_new_node(addx, phase->get_ctrl(x)); 253 phase->_igvn.replace_node(n1, addx); 254 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 255 _body.yank(n1); 256 return addx; 257 } 258 259 //---------------------reassociate_invariants----------------------------- 260 // Reassociate invariant expressions: 261 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 262 for (int i = _body.size() - 1; i >= 0; i--) { 263 Node *n = _body.at(i); 264 for (int j = 0; j < 5; j++) { 265 Node* nn = reassociate_add_sub(n, phase); 266 if (nn == NULL) break; 267 n = nn; // again 268 }; 269 } 270 } 271 272 //------------------------------policy_peeling--------------------------------- 273 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 274 // make some loop-invariant test (usually a null-check) happen before the loop. 275 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 276 Node *test = ((IdealLoopTree*)this)->tail(); 277 int body_size = ((IdealLoopTree*)this)->_body.size(); 278 // Peeling does loop cloning which can result in O(N^2) node construction 279 if( body_size > 255 /* Prevent overflow for large body_size */ 280 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) { 281 return false; // too large to safely clone 282 } 283 284 // check for vectorized loops, any peeling done was already applied 285 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false; 286 287 while( test != _head ) { // Scan till run off top of loop 288 if( test->is_If() ) { // Test? 289 Node *ctrl = phase->get_ctrl(test->in(1)); 290 if (ctrl->is_top()) 291 return false; // Found dead test on live IF? No peeling! 292 // Standard IF only has one input value to check for loop invariance 293 assert(test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd || test->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 294 // Condition is not a member of this loop? 295 if( !is_member(phase->get_loop(ctrl)) && 296 is_loop_exit(test) ) 297 return true; // Found reason to peel! 298 } 299 // Walk up dominators to loop _head looking for test which is 300 // executed on every path thru loop. 301 test = phase->idom(test); 302 } 303 return false; 304 } 305 306 //------------------------------peeled_dom_test_elim--------------------------- 307 // If we got the effect of peeling, either by actually peeling or by making 308 // a pre-loop which must execute at least once, we can remove all 309 // loop-invariant dominated tests in the main body. 310 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 311 bool progress = true; 312 while( progress ) { 313 progress = false; // Reset for next iteration 314 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 315 Node *test = prev->in(0); 316 while( test != loop->_head ) { // Scan till run off top of loop 317 318 int p_op = prev->Opcode(); 319 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 320 test->is_If() && // Test? 321 !test->in(1)->is_Con() && // And not already obvious? 322 // Condition is not a member of this loop? 323 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 324 // Walk loop body looking for instances of this test 325 for( uint i = 0; i < loop->_body.size(); i++ ) { 326 Node *n = loop->_body.at(i); 327 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 328 // IfNode was dominated by version in peeled loop body 329 progress = true; 330 dominated_by( old_new[prev->_idx], n ); 331 } 332 } 333 } 334 prev = test; 335 test = idom(test); 336 } // End of scan tests in loop 337 338 } // End of while( progress ) 339 } 340 341 //------------------------------do_peeling------------------------------------- 342 // Peel the first iteration of the given loop. 343 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 344 // The pre-loop illegally has 2 control users (old & new loops). 345 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 346 // Do this by making the old-loop fall-in edges act as if they came 347 // around the loopback from the prior iteration (follow the old-loop 348 // backedges) and then map to the new peeled iteration. This leaves 349 // the pre-loop with only 1 user (the new peeled iteration), but the 350 // peeled-loop backedge has 2 users. 351 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 352 // extra backedge user. 353 // 354 // orig 355 // 356 // stmt1 357 // | 358 // v 359 // loop predicate 360 // | 361 // v 362 // loop<----+ 363 // | | 364 // stmt2 | 365 // | | 366 // v | 367 // if ^ 368 // / \ | 369 // / \ | 370 // v v | 371 // false true | 372 // / \ | 373 // / ----+ 374 // | 375 // v 376 // exit 377 // 378 // 379 // after clone loop 380 // 381 // stmt1 382 // | 383 // v 384 // loop predicate 385 // / \ 386 // clone / \ orig 387 // / \ 388 // / \ 389 // v v 390 // +---->loop clone loop<----+ 391 // | | | | 392 // | stmt2 clone stmt2 | 393 // | | | | 394 // | v v | 395 // ^ if clone If ^ 396 // | / \ / \ | 397 // | / \ / \ | 398 // | v v v v | 399 // | true false false true | 400 // | / \ / \ | 401 // +---- \ / ----+ 402 // \ / 403 // 1v v2 404 // region 405 // | 406 // v 407 // exit 408 // 409 // 410 // after peel and predicate move 411 // 412 // stmt1 413 // / 414 // / 415 // clone / orig 416 // / 417 // / +----------+ 418 // / | | 419 // / loop predicate | 420 // / | | 421 // v v | 422 // TOP-->loop clone loop<----+ | 423 // | | | | 424 // stmt2 clone stmt2 | | 425 // | | | ^ 426 // v v | | 427 // if clone If ^ | 428 // / \ / \ | | 429 // / \ / \ | | 430 // v v v v | | 431 // true false false true | | 432 // | \ / \ | | 433 // | \ / ----+ ^ 434 // | \ / | 435 // | 1v v2 | 436 // v region | 437 // | | | 438 // | v | 439 // | exit | 440 // | | 441 // +--------------->-----------------+ 442 // 443 // 444 // final graph 445 // 446 // stmt1 447 // | 448 // v 449 // stmt2 clone 450 // | 451 // v 452 // if clone 453 // / | 454 // / | 455 // v v 456 // false true 457 // | | 458 // | v 459 // | loop predicate 460 // | | 461 // | v 462 // | loop<----+ 463 // | | | 464 // | stmt2 | 465 // | | | 466 // | v | 467 // v if ^ 468 // | / \ | 469 // | / \ | 470 // | v v | 471 // | false true | 472 // | | \ | 473 // v v --+ 474 // region 475 // | 476 // v 477 // exit 478 // 479 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 480 481 C->set_major_progress(); 482 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 483 // 'pre' loop from the main and the 'pre' can no longer have its 484 // iterations adjusted. Therefore, we need to declare this loop as 485 // no longer a 'main' loop; it will need new pre and post loops before 486 // we can do further RCE. 487 #ifndef PRODUCT 488 if (TraceLoopOpts) { 489 tty->print("Peel "); 490 loop->dump_head(); 491 } 492 #endif 493 Node* head = loop->_head; 494 bool counted_loop = head->is_CountedLoop(); 495 if (counted_loop) { 496 CountedLoopNode *cl = head->as_CountedLoop(); 497 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 498 cl->set_trip_count(cl->trip_count() - 1); 499 if (cl->is_main_loop()) { 500 cl->set_normal_loop(); 501 #ifndef PRODUCT 502 if (PrintOpto && VerifyLoopOptimizations) { 503 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 504 loop->dump_head(); 505 } 506 #endif 507 } 508 } 509 Node* entry = head->in(LoopNode::EntryControl); 510 511 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 512 // The pre-loop illegally has 2 control users (old & new loops). 513 clone_loop( loop, old_new, dom_depth(head) ); 514 515 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 516 // Do this by making the old-loop fall-in edges act as if they came 517 // around the loopback from the prior iteration (follow the old-loop 518 // backedges) and then map to the new peeled iteration. This leaves 519 // the pre-loop with only 1 user (the new peeled iteration), but the 520 // peeled-loop backedge has 2 users. 521 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 522 _igvn.hash_delete(head); 523 head->set_req(LoopNode::EntryControl, new_entry); 524 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 525 Node* old = head->fast_out(j); 526 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 527 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 528 if (!new_exit_value ) // Backedge value is ALSO loop invariant? 529 // Then loop body backedge value remains the same. 530 new_exit_value = old->in(LoopNode::LoopBackControl); 531 _igvn.hash_delete(old); 532 old->set_req(LoopNode::EntryControl, new_exit_value); 533 } 534 } 535 536 537 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 538 // extra backedge user. 539 Node* new_head = old_new[head->_idx]; 540 _igvn.hash_delete(new_head); 541 new_head->set_req(LoopNode::LoopBackControl, C->top()); 542 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 543 Node* use = new_head->fast_out(j2); 544 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 545 _igvn.hash_delete(use); 546 use->set_req(LoopNode::LoopBackControl, C->top()); 547 } 548 } 549 550 551 // Step 4: Correct dom-depth info. Set to loop-head depth. 552 int dd = dom_depth(head); 553 set_idom(head, head->in(1), dd); 554 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 555 Node *old = loop->_body.at(j3); 556 Node *nnn = old_new[old->_idx]; 557 if (!has_ctrl(nnn)) 558 set_idom(nnn, idom(nnn), dd-1); 559 } 560 561 // Now force out all loop-invariant dominating tests. The optimizer 562 // finds some, but we _know_ they are all useless. 563 peeled_dom_test_elim(loop,old_new); 564 565 loop->record_for_igvn(); 566 } 567 568 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 569 570 //------------------------------policy_maximally_unroll------------------------ 571 // Calculate exact loop trip count and return true if loop can be maximally 572 // unrolled. 573 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 574 CountedLoopNode *cl = _head->as_CountedLoop(); 575 assert(cl->is_normal_loop(), ""); 576 if (!cl->is_valid_counted_loop()) 577 return false; // Malformed counted loop 578 579 if (!cl->has_exact_trip_count()) { 580 // Trip count is not exact. 581 return false; 582 } 583 584 uint trip_count = cl->trip_count(); 585 // Note, max_juint is used to indicate unknown trip count. 586 assert(trip_count > 1, "one iteration loop should be optimized out already"); 587 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 588 589 // Real policy: if we maximally unroll, does it get too big? 590 // Allow the unrolled mess to get larger than standard loop 591 // size. After all, it will no longer be a loop. 592 uint body_size = _body.size(); 593 uint unroll_limit = (uint)LoopUnrollLimit * 4; 594 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 595 if (trip_count > unroll_limit || body_size > unroll_limit) { 596 return false; 597 } 598 599 // Fully unroll a loop with few iterations regardless next 600 // conditions since following loop optimizations will split 601 // such loop anyway (pre-main-post). 602 if (trip_count <= 3) 603 return true; 604 605 // Take into account that after unroll conjoined heads and tails will fold, 606 // otherwise policy_unroll() may allow more unrolling than max unrolling. 607 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; 608 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; 609 if (body_size != tst_body_size) // Check for int overflow 610 return false; 611 if (new_body_size > unroll_limit || 612 // Unrolling can result in a large amount of node construction 613 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) { 614 return false; 615 } 616 617 // Do not unroll a loop with String intrinsics code. 618 // String intrinsics are large and have loops. 619 for (uint k = 0; k < _body.size(); k++) { 620 Node* n = _body.at(k); 621 switch (n->Opcode()) { 622 case Op_StrComp: 623 case Op_StrEquals: 624 case Op_StrIndexOf: 625 case Op_StrIndexOfChar: 626 case Op_EncodeISOArray: 627 case Op_AryEq: 628 case Op_HasNegatives: { 629 return false; 630 } 631 #if INCLUDE_RTM_OPT 632 case Op_FastLock: 633 case Op_FastUnlock: { 634 // Don't unroll RTM locking code because it is large. 635 if (UseRTMLocking) { 636 return false; 637 } 638 } 639 #endif 640 } // switch 641 } 642 643 return true; // Do maximally unroll 644 } 645 646 647 //------------------------------policy_unroll---------------------------------- 648 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 649 // the loop is a CountedLoop and the body is small enough. 650 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) { 651 652 CountedLoopNode *cl = _head->as_CountedLoop(); 653 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 654 655 if (!cl->is_valid_counted_loop()) 656 return false; // Malformed counted loop 657 658 // Protect against over-unrolling. 659 // After split at least one iteration will be executed in pre-loop. 660 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; 661 662 _local_loop_unroll_limit = LoopUnrollLimit; 663 _local_loop_unroll_factor = 4; 664 int future_unroll_ct = cl->unrolled_count() * 2; 665 if (!cl->do_unroll_only()) { 666 if (future_unroll_ct > LoopMaxUnroll) return false; 667 } else { 668 // obey user constraints on vector mapped loops with additional unrolling applied 669 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1; 670 if ((future_unroll_ct / unroll_constraint) > LoopMaxUnroll) return false; 671 } 672 673 // Check for initial stride being a small enough constant 674 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; 675 676 // Don't unroll if the next round of unrolling would push us 677 // over the expected trip count of the loop. One is subtracted 678 // from the expected trip count because the pre-loop normally 679 // executes 1 iteration. 680 if (UnrollLimitForProfileCheck > 0 && 681 cl->profile_trip_cnt() != COUNT_UNKNOWN && 682 future_unroll_ct > UnrollLimitForProfileCheck && 683 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 684 return false; 685 } 686 687 // When unroll count is greater than LoopUnrollMin, don't unroll if: 688 // the residual iterations are more than 10% of the trip count 689 // and rounds of "unroll,optimize" are not making significant progress 690 // Progress defined as current size less than 20% larger than previous size. 691 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 692 future_unroll_ct > LoopUnrollMin && 693 (future_unroll_ct - 1) * (100 / LoopPercentProfileLimit) > cl->profile_trip_cnt() && 694 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 695 return false; 696 } 697 698 Node *init_n = cl->init_trip(); 699 Node *limit_n = cl->limit(); 700 int stride_con = cl->stride_con(); 701 // Non-constant bounds. 702 // Protect against over-unrolling when init or/and limit are not constant 703 // (so that trip_count's init value is maxint) but iv range is known. 704 if (init_n == NULL || !init_n->is_Con() || 705 limit_n == NULL || !limit_n->is_Con()) { 706 Node* phi = cl->phi(); 707 if (phi != NULL) { 708 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 709 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 710 int next_stride = stride_con * 2; // stride after this unroll 711 if (next_stride > 0) { 712 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 713 iv_type->_lo + next_stride > iv_type->_hi) { 714 return false; // over-unrolling 715 } 716 } else if (next_stride < 0) { 717 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 718 iv_type->_hi + next_stride < iv_type->_lo) { 719 return false; // over-unrolling 720 } 721 } 722 } 723 } 724 725 // After unroll limit will be adjusted: new_limit = limit-stride. 726 // Bailout if adjustment overflow. 727 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 728 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || 729 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) 730 return false; // overflow 731 732 // Adjust body_size to determine if we unroll or not 733 uint body_size = _body.size(); 734 // Key test to unroll loop in CRC32 java code 735 int xors_in_loop = 0; 736 // Also count ModL, DivL and MulL which expand mightly 737 for (uint k = 0; k < _body.size(); k++) { 738 Node* n = _body.at(k); 739 switch (n->Opcode()) { 740 case Op_XorI: xors_in_loop++; break; // CRC32 java code 741 case Op_ModL: body_size += 30; break; 742 case Op_DivL: body_size += 30; break; 743 case Op_MulL: body_size += 10; break; 744 case Op_StrComp: 745 case Op_StrEquals: 746 case Op_StrIndexOf: 747 case Op_StrIndexOfChar: 748 case Op_EncodeISOArray: 749 case Op_AryEq: 750 case Op_HasNegatives: { 751 // Do not unroll a loop with String intrinsics code. 752 // String intrinsics are large and have loops. 753 return false; 754 } 755 #if INCLUDE_RTM_OPT 756 case Op_FastLock: 757 case Op_FastUnlock: { 758 // Don't unroll RTM locking code because it is large. 759 if (UseRTMLocking) { 760 return false; 761 } 762 } 763 #endif 764 } // switch 765 } 766 767 if (UseSuperWord) { 768 if (!cl->is_reduction_loop()) { 769 phase->mark_reductions(this); 770 } 771 772 // Only attempt slp analysis when user controls do not prohibit it 773 if (LoopMaxUnroll > _local_loop_unroll_factor) { 774 // Once policy_slp_analysis succeeds, mark the loop with the 775 // maximal unroll factor so that we minimize analysis passes 776 if (future_unroll_ct >= _local_loop_unroll_factor) { 777 policy_unroll_slp_analysis(cl, phase, future_unroll_ct); 778 } 779 } 780 } 781 782 int slp_max_unroll_factor = cl->slp_max_unroll(); 783 if (cl->has_passed_slp()) { 784 if (slp_max_unroll_factor >= future_unroll_ct) return true; 785 // Normal case: loop too big 786 return false; 787 } 788 789 // Check for being too big 790 if (body_size > (uint)_local_loop_unroll_limit) { 791 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 792 // Normal case: loop too big 793 return false; 794 } 795 796 if (cl->do_unroll_only()) { 797 if (TraceSuperWordLoopUnrollAnalysis) { 798 tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct); 799 } 800 } 801 802 // Unroll once! (Each trip will soon do double iterations) 803 return true; 804 } 805 806 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) { 807 // Enable this functionality target by target as needed 808 if (SuperWordLoopUnrollAnalysis) { 809 if (!cl->was_slp_analyzed()) { 810 SuperWord sw(phase); 811 sw.transform_loop(this, false); 812 813 // If the loop is slp canonical analyze it 814 if (sw.early_return() == false) { 815 sw.unrolling_analysis(_local_loop_unroll_factor); 816 } 817 } 818 819 if (cl->has_passed_slp()) { 820 int slp_max_unroll_factor = cl->slp_max_unroll(); 821 if (slp_max_unroll_factor >= future_unroll_ct) { 822 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor; 823 if (new_limit > LoopUnrollLimit) { 824 if (TraceSuperWordLoopUnrollAnalysis) { 825 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit); 826 } 827 _local_loop_unroll_limit = new_limit; 828 } 829 } 830 } 831 } 832 } 833 834 //------------------------------policy_align----------------------------------- 835 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 836 // expression that does the alignment. Note that only one array base can be 837 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 838 // if we vectorize short memory ops into longer memory ops, we may want to 839 // increase alignment. 840 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 841 return false; 842 } 843 844 //------------------------------policy_range_check----------------------------- 845 // Return TRUE or FALSE if the loop should be range-check-eliminated. 846 // Actually we do iteration-splitting, a more powerful form of RCE. 847 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 848 if (!RangeCheckElimination) return false; 849 850 CountedLoopNode *cl = _head->as_CountedLoop(); 851 // If we unrolled with no intention of doing RCE and we later 852 // changed our minds, we got no pre-loop. Either we need to 853 // make a new pre-loop, or we gotta disallow RCE. 854 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 855 Node *trip_counter = cl->phi(); 856 857 // check for vectorized loops, some opts are no longer needed 858 if (cl->do_unroll_only()) return false; 859 860 // Check loop body for tests of trip-counter plus loop-invariant vs 861 // loop-invariant. 862 for (uint i = 0; i < _body.size(); i++) { 863 Node *iff = _body[i]; 864 if (iff->Opcode() == Op_If || 865 iff->Opcode() == Op_RangeCheck) { // Test? 866 867 // Comparing trip+off vs limit 868 Node *bol = iff->in(1); 869 if (bol->req() != 2) continue; // dead constant test 870 if (!bol->is_Bool()) { 871 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 872 continue; 873 } 874 if (bol->as_Bool()->_test._test == BoolTest::ne) 875 continue; // not RC 876 877 Node *cmp = bol->in(1); 878 Node *rc_exp = cmp->in(1); 879 Node *limit = cmp->in(2); 880 881 Node *limit_c = phase->get_ctrl(limit); 882 if( limit_c == phase->C->top() ) 883 return false; // Found dead test on live IF? No RCE! 884 if( is_member(phase->get_loop(limit_c) ) ) { 885 // Compare might have operands swapped; commute them 886 rc_exp = cmp->in(2); 887 limit = cmp->in(1); 888 limit_c = phase->get_ctrl(limit); 889 if( is_member(phase->get_loop(limit_c) ) ) 890 continue; // Both inputs are loop varying; cannot RCE 891 } 892 893 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 894 continue; 895 } 896 // Yeah! Found a test like 'trip+off vs limit' 897 // Test is an IfNode, has 2 projections. If BOTH are in the loop 898 // we need loop unswitching instead of iteration splitting. 899 if( is_loop_exit(iff) ) 900 return true; // Found reason to split iterations 901 } // End of is IF 902 } 903 904 return false; 905 } 906 907 //------------------------------policy_peel_only------------------------------- 908 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 909 // for unrolling loops with NO array accesses. 910 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 911 // check for vectorized loops, any peeling done was already applied 912 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false; 913 914 for( uint i = 0; i < _body.size(); i++ ) 915 if( _body[i]->is_Mem() ) 916 return false; 917 918 // No memory accesses at all! 919 return true; 920 } 921 922 //------------------------------clone_up_backedge_goo-------------------------- 923 // If Node n lives in the back_ctrl block and cannot float, we clone a private 924 // version of n in preheader_ctrl block and return that, otherwise return n. 925 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { 926 if( get_ctrl(n) != back_ctrl ) return n; 927 928 // Only visit once 929 if (visited.test_set(n->_idx)) { 930 Node *x = clones.find(n->_idx); 931 if (x != NULL) 932 return x; 933 return n; 934 } 935 936 Node *x = NULL; // If required, a clone of 'n' 937 // Check for 'n' being pinned in the backedge. 938 if( n->in(0) && n->in(0) == back_ctrl ) { 939 assert(clones.find(n->_idx) == NULL, "dead loop"); 940 x = n->clone(); // Clone a copy of 'n' to preheader 941 clones.push(x, n->_idx); 942 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 943 } 944 945 // Recursive fixup any other input edges into x. 946 // If there are no changes we can just return 'n', otherwise 947 // we need to clone a private copy and change it. 948 for( uint i = 1; i < n->req(); i++ ) { 949 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); 950 if( g != n->in(i) ) { 951 if( !x ) { 952 assert(clones.find(n->_idx) == NULL, "dead loop"); 953 x = n->clone(); 954 clones.push(x, n->_idx); 955 } 956 x->set_req(i, g); 957 } 958 } 959 if( x ) { // x can legally float to pre-header location 960 register_new_node( x, preheader_ctrl ); 961 return x; 962 } else { // raise n to cover LCA of uses 963 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 964 } 965 return n; 966 } 967 968 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { 969 Node* castii = new CastIINode(incr, TypeInt::INT, true); 970 castii->set_req(0, ctrl); 971 register_new_node(castii, ctrl); 972 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { 973 Node* n = incr->fast_out(i); 974 if (n->is_Phi() && n->in(0) == loop) { 975 int nrep = n->replace_edge(incr, castii); 976 return true; 977 } 978 } 979 return false; 980 } 981 982 //------------------------------insert_pre_post_loops-------------------------- 983 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 984 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 985 // alignment. Useful to unroll loops that do no array accesses. 986 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 987 988 #ifndef PRODUCT 989 if (TraceLoopOpts) { 990 if (peel_only) 991 tty->print("PeelMainPost "); 992 else 993 tty->print("PreMainPost "); 994 loop->dump_head(); 995 } 996 #endif 997 C->set_major_progress(); 998 999 // Find common pieces of the loop being guarded with pre & post loops 1000 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1001 assert( main_head->is_normal_loop(), "" ); 1002 CountedLoopEndNode *main_end = main_head->loopexit(); 1003 guarantee(main_end != NULL, "no loop exit node"); 1004 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 1005 uint dd_main_head = dom_depth(main_head); 1006 uint max = main_head->outcnt(); 1007 1008 Node *pre_header= main_head->in(LoopNode::EntryControl); 1009 Node *init = main_head->init_trip(); 1010 Node *incr = main_end ->incr(); 1011 Node *limit = main_end ->limit(); 1012 Node *stride = main_end ->stride(); 1013 Node *cmp = main_end ->cmp_node(); 1014 BoolTest::mask b_test = main_end->test_trip(); 1015 1016 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 1017 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 1018 if( bol->outcnt() != 1 ) { 1019 bol = bol->clone(); 1020 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 1021 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol); 1022 } 1023 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 1024 if( cmp->outcnt() != 1 ) { 1025 cmp = cmp->clone(); 1026 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 1027 _igvn.replace_input_of(bol, 1, cmp); 1028 } 1029 1030 // Add the post loop 1031 CountedLoopNode *post_head = NULL; 1032 Node *main_exit = insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1033 1034 //------------------------------ 1035 // Step B: Create Pre-Loop. 1036 1037 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 1038 // loop pre-header illegally has 2 control users (old & new loops). 1039 clone_loop( loop, old_new, dd_main_head ); 1040 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1041 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1042 pre_head->set_pre_loop(main_head); 1043 Node *pre_incr = old_new[incr->_idx]; 1044 1045 // Reduce the pre-loop trip count. 1046 pre_end->_prob = PROB_FAIR; 1047 1048 // Find the pre-loop normal exit. 1049 Node* pre_exit = pre_end->proj_out(false); 1050 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 1051 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end); 1052 _igvn.register_new_node_with_optimizer( new_pre_exit ); 1053 set_idom(new_pre_exit, pre_end, dd_main_head); 1054 set_loop(new_pre_exit, loop->_parent); 1055 1056 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1057 // pre-loop, the main-loop may not execute at all. Later in life this 1058 // zero-trip guard will become the minimum-trip guard when we unroll 1059 // the main-loop. 1060 Node *min_opaq = new Opaque1Node(C, limit); 1061 Node *min_cmp = new CmpINode( pre_incr, min_opaq ); 1062 Node *min_bol = new BoolNode( min_cmp, b_test ); 1063 register_new_node( min_opaq, new_pre_exit ); 1064 register_new_node( min_cmp , new_pre_exit ); 1065 register_new_node( min_bol , new_pre_exit ); 1066 1067 // Build the IfNode (assume the main-loop is executed always). 1068 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1069 _igvn.register_new_node_with_optimizer( min_iff ); 1070 set_idom(min_iff, new_pre_exit, dd_main_head); 1071 set_loop(min_iff, loop->_parent); 1072 1073 // Plug in the false-path, taken if we need to skip main-loop 1074 _igvn.hash_delete( pre_exit ); 1075 pre_exit->set_req(0, min_iff); 1076 set_idom(pre_exit, min_iff, dd_main_head); 1077 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1078 // Make the true-path, must enter the main loop 1079 Node *min_taken = new IfTrueNode( min_iff ); 1080 _igvn.register_new_node_with_optimizer( min_taken ); 1081 set_idom(min_taken, min_iff, dd_main_head); 1082 set_loop(min_taken, loop->_parent); 1083 // Plug in the true path 1084 _igvn.hash_delete( main_head ); 1085 main_head->set_req(LoopNode::EntryControl, min_taken); 1086 set_idom(main_head, min_taken, dd_main_head); 1087 1088 Arena *a = Thread::current()->resource_area(); 1089 VectorSet visited(a); 1090 Node_Stack clones(a, main_head->back_control()->outcnt()); 1091 // Step B3: Make the fall-in values to the main-loop come from the 1092 // fall-out values of the pre-loop. 1093 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1094 Node* main_phi = main_head->fast_out(i2); 1095 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1096 Node *pre_phi = old_new[main_phi->_idx]; 1097 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1098 main_head->init_control(), 1099 pre_phi->in(LoopNode::LoopBackControl), 1100 visited, clones); 1101 _igvn.hash_delete(main_phi); 1102 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1103 } 1104 } 1105 1106 // Nodes inside the loop may be control dependent on a predicate 1107 // that was moved before the preloop. If the back branch of the main 1108 // or post loops becomes dead, those nodes won't be dependent on the 1109 // test that guards that loop nest anymore which could lead to an 1110 // incorrect array access because it executes independently of the 1111 // test that was guarding the loop nest. We add a special CastII on 1112 // the if branch that enters the loop, between the input induction 1113 // variable value and the induction variable Phi to preserve correct 1114 // dependencies. 1115 1116 // CastII for the main loop: 1117 bool inserted = cast_incr_before_loop( pre_incr, min_taken, main_head ); 1118 assert(inserted, "no castII inserted"); 1119 1120 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1121 // RCE and alignment may change this later. 1122 Node *cmp_end = pre_end->cmp_node(); 1123 assert( cmp_end->in(2) == limit, "" ); 1124 Node *pre_limit = new AddINode( init, stride ); 1125 1126 // Save the original loop limit in this Opaque1 node for 1127 // use by range check elimination. 1128 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit); 1129 1130 register_new_node( pre_limit, pre_head->in(0) ); 1131 register_new_node( pre_opaq , pre_head->in(0) ); 1132 1133 // Since no other users of pre-loop compare, I can hack limit directly 1134 assert( cmp_end->outcnt() == 1, "no other users" ); 1135 _igvn.hash_delete(cmp_end); 1136 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1137 1138 // Special case for not-equal loop bounds: 1139 // Change pre loop test, main loop test, and the 1140 // main loop guard test to use lt or gt depending on stride 1141 // direction: 1142 // positive stride use < 1143 // negative stride use > 1144 // 1145 // not-equal test is kept for post loop to handle case 1146 // when init > limit when stride > 0 (and reverse). 1147 1148 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1149 1150 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1151 // Modify pre loop end condition 1152 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1153 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test); 1154 register_new_node( new_bol0, pre_head->in(0) ); 1155 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0); 1156 // Modify main loop guard condition 1157 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1158 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test); 1159 register_new_node( new_bol1, new_pre_exit ); 1160 _igvn.hash_delete(min_iff); 1161 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1162 // Modify main loop end condition 1163 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1164 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test); 1165 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1166 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2); 1167 } 1168 1169 // Flag main loop 1170 main_head->set_main_loop(); 1171 if( peel_only ) main_head->set_main_no_pre_loop(); 1172 1173 // Subtract a trip count for the pre-loop. 1174 main_head->set_trip_count(main_head->trip_count() - 1); 1175 1176 // It's difficult to be precise about the trip-counts 1177 // for the pre/post loops. They are usually very short, 1178 // so guess that 4 trips is a reasonable value. 1179 post_head->set_profile_trip_cnt(4.0); 1180 pre_head->set_profile_trip_cnt(4.0); 1181 1182 // Now force out all loop-invariant dominating tests. The optimizer 1183 // finds some, but we _know_ they are all useless. 1184 peeled_dom_test_elim(loop,old_new); 1185 loop->record_for_igvn(); 1186 } 1187 1188 //------------------------------insert_vector_post_loop------------------------ 1189 // Insert a copy of the atomic unrolled vectorized main loop as a post loop, 1190 // unroll_policy has already informed us that more unrolling is about to happen to 1191 // the main loop. The resultant post loop will serve as a vectorized drain loop. 1192 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) { 1193 if (!loop->_head->is_CountedLoop()) return; 1194 1195 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1196 1197 // only process vectorized main loops 1198 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return; 1199 1200 int slp_max_unroll_factor = cl->slp_max_unroll(); 1201 int cur_unroll = cl->unrolled_count(); 1202 1203 if (slp_max_unroll_factor == 0) return; 1204 1205 // only process atomic unroll vector loops (not super unrolled after vectorization) 1206 if (cur_unroll != slp_max_unroll_factor) return; 1207 1208 // we only ever process this one time 1209 if (cl->has_atomic_post_loop()) return; 1210 1211 #ifndef PRODUCT 1212 if (TraceLoopOpts) { 1213 tty->print("PostVector "); 1214 loop->dump_head(); 1215 } 1216 #endif 1217 C->set_major_progress(); 1218 1219 // Find common pieces of the loop being guarded with pre & post loops 1220 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1221 CountedLoopEndNode *main_end = main_head->loopexit(); 1222 guarantee(main_end != NULL, "no loop exit node"); 1223 // diagnostic to show loop end is not properly formed 1224 assert(main_end->outcnt() == 2, "1 true, 1 false path only"); 1225 1226 // mark this loop as processed 1227 main_head->mark_has_atomic_post_loop(); 1228 1229 Node *incr = main_end->incr(); 1230 Node *limit = main_end->limit(); 1231 1232 // In this case we throw away the result as we are not using it to connect anything else. 1233 CountedLoopNode *post_head = NULL; 1234 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1235 1236 // It's difficult to be precise about the trip-counts 1237 // for post loops. They are usually very short, 1238 // so guess that unit vector trips is a reasonable value. 1239 post_head->set_profile_trip_cnt(cur_unroll); 1240 1241 // Now force out all loop-invariant dominating tests. The optimizer 1242 // finds some, but we _know_ they are all useless. 1243 peeled_dom_test_elim(loop, old_new); 1244 loop->record_for_igvn(); 1245 } 1246 1247 1248 //-------------------------insert_scalar_rced_post_loop------------------------ 1249 // Insert a copy of the rce'd main loop as a post loop, 1250 // We have not unrolled the main loop, so this is the right time to inject this. 1251 // Later we will examine the partner of this post loop pair which still has range checks 1252 // to see inject code which tests at runtime if the range checks are applicable. 1253 void PhaseIdealLoop::insert_scalar_rced_post_loop(IdealLoopTree *loop, Node_List &old_new) { 1254 if (!loop->_head->is_CountedLoop()) return; 1255 1256 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1257 1258 // only process RCE'd main loops 1259 if (!cl->is_main_loop() || cl->range_checks_present()) return; 1260 1261 #ifndef PRODUCT 1262 if (TraceLoopOpts) { 1263 tty->print("PostScalarRce "); 1264 loop->dump_head(); 1265 } 1266 #endif 1267 C->set_major_progress(); 1268 1269 // Find common pieces of the loop being guarded with pre & post loops 1270 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1271 CountedLoopEndNode *main_end = main_head->loopexit(); 1272 guarantee(main_end != NULL, "no loop exit node"); 1273 // diagnostic to show loop end is not properly formed 1274 assert(main_end->outcnt() == 2, "1 true, 1 false path only"); 1275 1276 Node *incr = main_end->incr(); 1277 Node *limit = main_end->limit(); 1278 1279 // In this case we throw away the result as we are not using it to connect anything else. 1280 CountedLoopNode *post_head = NULL; 1281 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1282 1283 // It's difficult to be precise about the trip-counts 1284 // for post loops. They are usually very short, 1285 // so guess that unit vector trips is a reasonable value. 1286 post_head->set_profile_trip_cnt(4.0); 1287 post_head->set_is_rce_post_loop(); 1288 1289 // Now force out all loop-invariant dominating tests. The optimizer 1290 // finds some, but we _know_ they are all useless. 1291 peeled_dom_test_elim(loop, old_new); 1292 loop->record_for_igvn(); 1293 } 1294 1295 1296 //------------------------------insert_post_loop------------------------------- 1297 // Insert post loops. Add a post loop to the given loop passed. 1298 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree *loop, Node_List &old_new, 1299 CountedLoopNode *main_head, CountedLoopEndNode *main_end, 1300 Node *incr, Node *limit, CountedLoopNode *&post_head) { 1301 1302 //------------------------------ 1303 // Step A: Create a new post-Loop. 1304 Node* main_exit = main_end->proj_out(false); 1305 assert(main_exit->Opcode() == Op_IfFalse, ""); 1306 int dd_main_exit = dom_depth(main_exit); 1307 1308 // Step A1: Clone the loop body of main. The clone becomes the vector post-loop. 1309 // The main loop pre-header illegally has 2 control users (old & new loops). 1310 clone_loop(loop, old_new, dd_main_exit); 1311 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, ""); 1312 post_head = old_new[main_head->_idx]->as_CountedLoop(); 1313 post_head->set_normal_loop(); 1314 post_head->set_post_loop(main_head); 1315 1316 // Reduce the post-loop trip count. 1317 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd(); 1318 post_end->_prob = PROB_FAIR; 1319 1320 // Build the main-loop normal exit. 1321 IfFalseNode *new_main_exit = new IfFalseNode(main_end); 1322 _igvn.register_new_node_with_optimizer(new_main_exit); 1323 set_idom(new_main_exit, main_end, dd_main_exit); 1324 set_loop(new_main_exit, loop->_parent); 1325 1326 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 1327 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 1328 // (the previous loop trip-counter exit value) because we will be changing 1329 // the exit value (via additional unrolling) so we cannot constant-fold away the zero 1330 // trip guard until all unrolling is done. 1331 Node *zer_opaq = new Opaque1Node(C, incr); 1332 Node *zer_cmp = new CmpINode(zer_opaq, limit); 1333 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip()); 1334 register_new_node(zer_opaq, new_main_exit); 1335 register_new_node(zer_cmp, new_main_exit); 1336 register_new_node(zer_bol, new_main_exit); 1337 1338 // Build the IfNode 1339 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN); 1340 _igvn.register_new_node_with_optimizer(zer_iff); 1341 set_idom(zer_iff, new_main_exit, dd_main_exit); 1342 set_loop(zer_iff, loop->_parent); 1343 1344 // Plug in the false-path, taken if we need to skip this post-loop 1345 _igvn.replace_input_of(main_exit, 0, zer_iff); 1346 set_idom(main_exit, zer_iff, dd_main_exit); 1347 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 1348 // Make the true-path, must enter this post loop 1349 Node *zer_taken = new IfTrueNode(zer_iff); 1350 _igvn.register_new_node_with_optimizer(zer_taken); 1351 set_idom(zer_taken, zer_iff, dd_main_exit); 1352 set_loop(zer_taken, loop->_parent); 1353 // Plug in the true path 1354 _igvn.hash_delete(post_head); 1355 post_head->set_req(LoopNode::EntryControl, zer_taken); 1356 set_idom(post_head, zer_taken, dd_main_exit); 1357 1358 Arena *a = Thread::current()->resource_area(); 1359 VectorSet visited(a); 1360 Node_Stack clones(a, main_head->back_control()->outcnt()); 1361 // Step A3: Make the fall-in values to the post-loop come from the 1362 // fall-out values of the main-loop. 1363 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 1364 Node* main_phi = main_head->fast_out(i); 1365 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0) { 1366 Node *cur_phi = old_new[main_phi->_idx]; 1367 Node *fallnew = clone_up_backedge_goo(main_head->back_control(), 1368 post_head->init_control(), 1369 main_phi->in(LoopNode::LoopBackControl), 1370 visited, clones); 1371 _igvn.hash_delete(cur_phi); 1372 cur_phi->set_req(LoopNode::EntryControl, fallnew); 1373 } 1374 } 1375 1376 // CastII for the new post loop: 1377 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); 1378 assert(inserted, "no castII inserted"); 1379 1380 return new_main_exit; 1381 } 1382 1383 //------------------------------is_invariant----------------------------- 1384 // Return true if n is invariant 1385 bool IdealLoopTree::is_invariant(Node* n) const { 1386 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1387 if (n_c->is_top()) return false; 1388 return !is_member(_phase->get_loop(n_c)); 1389 } 1390 1391 1392 //------------------------------do_unroll-------------------------------------- 1393 // Unroll the loop body one step - make each trip do 2 iterations. 1394 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1395 assert(LoopUnrollLimit, ""); 1396 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1397 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1398 assert(loop_end, ""); 1399 #ifndef PRODUCT 1400 if (PrintOpto && VerifyLoopOptimizations) { 1401 tty->print("Unrolling "); 1402 loop->dump_head(); 1403 } else if (TraceLoopOpts) { 1404 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1405 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1406 } else { 1407 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1408 } 1409 loop->dump_head(); 1410 } 1411 1412 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1413 Arena* arena = Thread::current()->resource_area(); 1414 Node_Stack stack(arena, C->live_nodes() >> 2); 1415 Node_List rpo_list; 1416 VectorSet visited(arena); 1417 visited.set(loop_head->_idx); 1418 rpo( loop_head, stack, visited, rpo_list ); 1419 dump(loop, rpo_list.size(), rpo_list ); 1420 } 1421 #endif 1422 1423 // Remember loop node count before unrolling to detect 1424 // if rounds of unroll,optimize are making progress 1425 loop_head->set_node_count_before_unroll(loop->_body.size()); 1426 1427 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1428 Node *limit = loop_head->limit(); 1429 Node *init = loop_head->init_trip(); 1430 Node *stride = loop_head->stride(); 1431 1432 Node *opaq = NULL; 1433 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1434 // Search for zero-trip guard. 1435 1436 // Check the shape of the graph at the loop entry. If an inappropriate 1437 // graph shape is encountered, the compiler bails out loop unrolling; 1438 // compilation of the method will still succeed. 1439 if (!is_canonical_loop_entry(loop_head)) { 1440 return; 1441 } 1442 opaq = ctrl->in(0)->in(1)->in(1)->in(2); 1443 // Zero-trip test uses an 'opaque' node which is not shared. 1444 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1445 } 1446 1447 C->set_major_progress(); 1448 1449 Node* new_limit = NULL; 1450 int stride_con = stride->get_int(); 1451 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1452 uint old_trip_count = loop_head->trip_count(); 1453 // Verify that unroll policy result is still valid. 1454 assert(old_trip_count > 1 && 1455 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1456 1457 // Adjust loop limit to keep valid iterations number after unroll. 1458 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1459 // which may overflow. 1460 if (!adjust_min_trip) { 1461 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1462 "odd trip count for maximally unroll"); 1463 // Don't need to adjust limit for maximally unroll since trip count is even. 1464 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1465 // Loop's limit is constant. Loop's init could be constant when pre-loop 1466 // become peeled iteration. 1467 jlong init_con = init->get_int(); 1468 // We can keep old loop limit if iterations count stays the same: 1469 // old_trip_count == new_trip_count * 2 1470 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1471 // so we also don't need to adjust zero trip test. 1472 jlong limit_con = limit->get_int(); 1473 // (stride_con*2) not overflow since stride_con <= 8. 1474 int new_stride_con = stride_con * 2; 1475 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1476 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1477 // New trip count should satisfy next conditions. 1478 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1479 uint new_trip_count = (uint)trip_count; 1480 adjust_min_trip = (old_trip_count != new_trip_count*2); 1481 } 1482 1483 if (adjust_min_trip) { 1484 // Step 2: Adjust the trip limit if it is called for. 1485 // The adjustment amount is -stride. Need to make sure if the 1486 // adjustment underflows or overflows, then the main loop is skipped. 1487 Node* cmp = loop_end->cmp_node(); 1488 assert(cmp->in(2) == limit, "sanity"); 1489 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1490 1491 // Verify that policy_unroll result is still valid. 1492 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1493 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1494 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); 1495 1496 if (limit->is_Con()) { 1497 // The check in policy_unroll and the assert above guarantee 1498 // no underflow if limit is constant. 1499 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1500 set_ctrl(new_limit, C->root()); 1501 } else { 1502 // Limit is not constant. 1503 if (loop_head->unrolled_count() == 1) { // only for first unroll 1504 // Separate limit by Opaque node in case it is an incremented 1505 // variable from previous loop to avoid using pre-incremented 1506 // value which could increase register pressure. 1507 // Otherwise reorg_offsets() optimization will create a separate 1508 // Opaque node for each use of trip-counter and as result 1509 // zero trip guard limit will be different from loop limit. 1510 assert(has_ctrl(opaq), "should have it"); 1511 Node* opaq_ctrl = get_ctrl(opaq); 1512 limit = new Opaque2Node( C, limit ); 1513 register_new_node( limit, opaq_ctrl ); 1514 } 1515 if (stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo) || 1516 stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi)) { 1517 // No underflow. 1518 new_limit = new SubINode(limit, stride); 1519 } else { 1520 // (limit - stride) may underflow. 1521 // Clamp the adjustment value with MININT or MAXINT: 1522 // 1523 // new_limit = limit-stride 1524 // if (stride > 0) 1525 // new_limit = (limit < new_limit) ? MININT : new_limit; 1526 // else 1527 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1528 // 1529 BoolTest::mask bt = loop_end->test_trip(); 1530 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1531 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1532 set_ctrl(adj_max, C->root()); 1533 Node* old_limit = NULL; 1534 Node* adj_limit = NULL; 1535 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1536 if (loop_head->unrolled_count() > 1 && 1537 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1538 limit->in(CMoveNode::IfTrue) == adj_max && 1539 bol->as_Bool()->_test._test == bt && 1540 bol->in(1)->Opcode() == Op_CmpI && 1541 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1542 // Loop was unrolled before. 1543 // Optimize the limit to avoid nested CMove: 1544 // use original limit as old limit. 1545 old_limit = bol->in(1)->in(1); 1546 // Adjust previous adjusted limit. 1547 adj_limit = limit->in(CMoveNode::IfFalse); 1548 adj_limit = new SubINode(adj_limit, stride); 1549 } else { 1550 old_limit = limit; 1551 adj_limit = new SubINode(limit, stride); 1552 } 1553 assert(old_limit != NULL && adj_limit != NULL, ""); 1554 register_new_node( adj_limit, ctrl ); // adjust amount 1555 Node* adj_cmp = new CmpINode(old_limit, adj_limit); 1556 register_new_node( adj_cmp, ctrl ); 1557 Node* adj_bool = new BoolNode(adj_cmp, bt); 1558 register_new_node( adj_bool, ctrl ); 1559 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1560 } 1561 register_new_node(new_limit, ctrl); 1562 } 1563 assert(new_limit != NULL, ""); 1564 // Replace in loop test. 1565 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1566 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1567 // Don't need to create new test since only one user. 1568 _igvn.hash_delete(cmp); 1569 cmp->set_req(2, new_limit); 1570 } else { 1571 // Create new test since it is shared. 1572 Node* ctrl2 = loop_end->in(0); 1573 Node* cmp2 = cmp->clone(); 1574 cmp2->set_req(2, new_limit); 1575 register_new_node(cmp2, ctrl2); 1576 Node* bol2 = loop_end->in(1)->clone(); 1577 bol2->set_req(1, cmp2); 1578 register_new_node(bol2, ctrl2); 1579 _igvn.replace_input_of(loop_end, 1, bol2); 1580 } 1581 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1582 // Make it a 1-trip test (means at least 2 trips). 1583 1584 // Guard test uses an 'opaque' node which is not shared. Hence I 1585 // can edit it's inputs directly. Hammer in the new limit for the 1586 // minimum-trip guard. 1587 assert(opaq->outcnt() == 1, ""); 1588 _igvn.replace_input_of(opaq, 1, new_limit); 1589 } 1590 1591 // Adjust max trip count. The trip count is intentionally rounded 1592 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1593 // the main, unrolled, part of the loop will never execute as it is protected 1594 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1595 // and later determined that part of the unrolled loop was dead. 1596 loop_head->set_trip_count(old_trip_count / 2); 1597 1598 // Double the count of original iterations in the unrolled loop body. 1599 loop_head->double_unrolled_count(); 1600 1601 // --------- 1602 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1603 // represents the odd iterations; since the loop trips an even number of 1604 // times its backedge is never taken. Kill the backedge. 1605 uint dd = dom_depth(loop_head); 1606 clone_loop( loop, old_new, dd ); 1607 1608 // Make backedges of the clone equal to backedges of the original. 1609 // Make the fall-in from the original come from the fall-out of the clone. 1610 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1611 Node* phi = loop_head->fast_out(j); 1612 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1613 Node *newphi = old_new[phi->_idx]; 1614 _igvn.hash_delete( phi ); 1615 _igvn.hash_delete( newphi ); 1616 1617 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1618 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1619 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1620 } 1621 } 1622 Node *clone_head = old_new[loop_head->_idx]; 1623 _igvn.hash_delete( clone_head ); 1624 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1625 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1626 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1627 loop->_head = clone_head; // New loop header 1628 1629 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1630 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1631 1632 // Kill the clone's backedge 1633 Node *newcle = old_new[loop_end->_idx]; 1634 _igvn.hash_delete( newcle ); 1635 Node *one = _igvn.intcon(1); 1636 set_ctrl(one, C->root()); 1637 newcle->set_req(1, one); 1638 // Force clone into same loop body 1639 uint max = loop->_body.size(); 1640 for( uint k = 0; k < max; k++ ) { 1641 Node *old = loop->_body.at(k); 1642 Node *nnn = old_new[old->_idx]; 1643 loop->_body.push(nnn); 1644 if (!has_ctrl(old)) 1645 set_loop(nnn, loop); 1646 } 1647 1648 loop->record_for_igvn(); 1649 1650 #ifndef PRODUCT 1651 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1652 tty->print("\nnew loop after unroll\n"); loop->dump_head(); 1653 for (uint i = 0; i < loop->_body.size(); i++) { 1654 loop->_body.at(i)->dump(); 1655 } 1656 if(C->clone_map().is_debug()) { 1657 tty->print("\nCloneMap\n"); 1658 Dict* dict = C->clone_map().dict(); 1659 DictI i(dict); 1660 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size()); 1661 for (int ii = 0; i.test(); ++i, ++ii) { 1662 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key)); 1663 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen()); 1664 if (ii % 10 == 9) { 1665 tty->print_cr(" "); 1666 } 1667 } 1668 tty->print_cr(" "); 1669 } 1670 } 1671 #endif 1672 1673 } 1674 1675 //------------------------------do_maximally_unroll---------------------------- 1676 1677 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1678 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1679 assert(cl->has_exact_trip_count(), "trip count is not exact"); 1680 assert(cl->trip_count() > 0, ""); 1681 #ifndef PRODUCT 1682 if (TraceLoopOpts) { 1683 tty->print("MaxUnroll %d ", cl->trip_count()); 1684 loop->dump_head(); 1685 } 1686 #endif 1687 1688 // If loop is tripping an odd number of times, peel odd iteration 1689 if ((cl->trip_count() & 1) == 1) { 1690 do_peeling(loop, old_new); 1691 } 1692 1693 // Now its tripping an even number of times remaining. Double loop body. 1694 // Do not adjust pre-guards; they are not needed and do not exist. 1695 if (cl->trip_count() > 0) { 1696 assert((cl->trip_count() & 1) == 0, "missed peeling"); 1697 do_unroll(loop, old_new, false); 1698 } 1699 } 1700 1701 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) { 1702 if (SuperWordReductions == false) return; 1703 1704 CountedLoopNode* loop_head = loop->_head->as_CountedLoop(); 1705 if (loop_head->unrolled_count() > 1) { 1706 return; 1707 } 1708 1709 Node* trip_phi = loop_head->phi(); 1710 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) { 1711 Node* phi = loop_head->fast_out(i); 1712 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) { 1713 // For definitions which are loop inclusive and not tripcounts. 1714 Node* def_node = phi->in(LoopNode::LoopBackControl); 1715 1716 if (def_node != NULL) { 1717 Node* n_ctrl = get_ctrl(def_node); 1718 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) { 1719 // Now test it to see if it fits the standard pattern for a reduction operator. 1720 int opc = def_node->Opcode(); 1721 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) { 1722 if (!def_node->is_reduction()) { // Not marked yet 1723 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it 1724 bool ok = false; 1725 for (unsigned j = 1; j < def_node->req(); j++) { 1726 Node* in = def_node->in(j); 1727 if (in == phi) { 1728 ok = true; 1729 break; 1730 } 1731 } 1732 1733 // do nothing if we did not match the initial criteria 1734 if (ok == false) { 1735 continue; 1736 } 1737 1738 // The result of the reduction must not be used in the loop 1739 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) { 1740 Node* u = def_node->fast_out(i); 1741 if (has_ctrl(u) && !loop->is_member(get_loop(get_ctrl(u)))) { 1742 continue; 1743 } 1744 if (u == phi) { 1745 continue; 1746 } 1747 ok = false; 1748 } 1749 1750 // iff the uses conform 1751 if (ok) { 1752 def_node->add_flag(Node::Flag_is_reduction); 1753 loop_head->mark_has_reductions(); 1754 } 1755 } 1756 } 1757 } 1758 } 1759 } 1760 } 1761 } 1762 1763 //------------------------------dominates_backedge--------------------------------- 1764 // Returns true if ctrl is executed on every complete iteration 1765 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1766 assert(ctrl->is_CFG(), "must be control"); 1767 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1768 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1769 } 1770 1771 //------------------------------adjust_limit----------------------------------- 1772 // Helper function for add_constraint(). 1773 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { 1774 // Compute "I :: (limit-offset)/scale" 1775 Node *con = new SubINode(rc_limit, offset); 1776 register_new_node(con, pre_ctrl); 1777 Node *X = new DivINode(0, con, scale); 1778 register_new_node(X, pre_ctrl); 1779 1780 // Adjust loop limit 1781 loop_limit = (stride_con > 0) 1782 ? (Node*)(new MinINode(loop_limit, X)) 1783 : (Node*)(new MaxINode(loop_limit, X)); 1784 register_new_node(loop_limit, pre_ctrl); 1785 return loop_limit; 1786 } 1787 1788 //------------------------------add_constraint--------------------------------- 1789 // Constrain the main loop iterations so the conditions: 1790 // low_limit <= scale_con * I + offset < upper_limit 1791 // always holds true. That is, either increase the number of iterations in 1792 // the pre-loop or the post-loop until the condition holds true in the main 1793 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1794 // stride and scale are constants (offset and limit often are). 1795 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 ) { 1796 // For positive stride, the pre-loop limit always uses a MAX function 1797 // and the main loop a MIN function. For negative stride these are 1798 // reversed. 1799 1800 // Also for positive stride*scale the affine function is increasing, so the 1801 // pre-loop must check for underflow and the post-loop for overflow. 1802 // Negative stride*scale reverses this; pre-loop checks for overflow and 1803 // post-loop for underflow. 1804 1805 Node *scale = _igvn.intcon(scale_con); 1806 set_ctrl(scale, C->root()); 1807 1808 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 1809 // The overflow limit: scale*I+offset < upper_limit 1810 // For main-loop compute 1811 // ( if (scale > 0) /* and stride > 0 */ 1812 // I < (upper_limit-offset)/scale 1813 // else /* scale < 0 and stride < 0 */ 1814 // I > (upper_limit-offset)/scale 1815 // ) 1816 // 1817 // (upper_limit-offset) may overflow or underflow. 1818 // But it is fine since main loop will either have 1819 // less iterations or will be skipped in such case. 1820 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); 1821 1822 // The underflow limit: low_limit <= scale*I+offset. 1823 // For pre-loop compute 1824 // NOT(scale*I+offset >= low_limit) 1825 // scale*I+offset < low_limit 1826 // ( if (scale > 0) /* and stride > 0 */ 1827 // I < (low_limit-offset)/scale 1828 // else /* scale < 0 and stride < 0 */ 1829 // I > (low_limit-offset)/scale 1830 // ) 1831 1832 if (low_limit->get_int() == -max_jint) { 1833 // We need this guard when scale*pre_limit+offset >= limit 1834 // due to underflow. So we need execute pre-loop until 1835 // scale*I+offset >= min_int. But (min_int-offset) will 1836 // underflow when offset > 0 and X will be > original_limit 1837 // when stride > 0. To avoid it we replace positive offset with 0. 1838 // 1839 // Also (min_int+1 == -max_int) is used instead of min_int here 1840 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1841 Node* shift = _igvn.intcon(31); 1842 set_ctrl(shift, C->root()); 1843 Node* sign = new RShiftINode(offset, shift); 1844 register_new_node(sign, pre_ctrl); 1845 offset = new AndINode(offset, sign); 1846 register_new_node(offset, pre_ctrl); 1847 } else { 1848 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1849 // The only problem we have here when offset == min_int 1850 // since (0-min_int) == min_int. It may be fine for stride > 0 1851 // but for stride < 0 X will be < original_limit. To avoid it 1852 // max(pre_limit, original_limit) is used in do_range_check(). 1853 } 1854 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1855 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); 1856 1857 } else { // stride_con*scale_con < 0 1858 // For negative stride*scale pre-loop checks for overflow and 1859 // post-loop for underflow. 1860 // 1861 // The overflow limit: scale*I+offset < upper_limit 1862 // For pre-loop compute 1863 // NOT(scale*I+offset < upper_limit) 1864 // scale*I+offset >= upper_limit 1865 // scale*I+offset+1 > upper_limit 1866 // ( if (scale < 0) /* and stride > 0 */ 1867 // I < (upper_limit-(offset+1))/scale 1868 // else /* scale > 0 and stride < 0 */ 1869 // I > (upper_limit-(offset+1))/scale 1870 // ) 1871 // 1872 // (upper_limit-offset-1) may underflow or overflow. 1873 // To avoid it min(pre_limit, original_limit) is used 1874 // in do_range_check() for stride > 0 and max() for < 0. 1875 Node *one = _igvn.intcon(1); 1876 set_ctrl(one, C->root()); 1877 1878 Node *plus_one = new AddINode(offset, one); 1879 register_new_node( plus_one, pre_ctrl ); 1880 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1881 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); 1882 1883 if (low_limit->get_int() == -max_jint) { 1884 // We need this guard when scale*main_limit+offset >= limit 1885 // due to underflow. So we need execute main-loop while 1886 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 1887 // underflow when (offset+1) > 0 and X will be < main_limit 1888 // when scale < 0 (and stride > 0). To avoid it we replace 1889 // positive (offset+1) with 0. 1890 // 1891 // Also (min_int+1 == -max_int) is used instead of min_int here 1892 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1893 Node* shift = _igvn.intcon(31); 1894 set_ctrl(shift, C->root()); 1895 Node* sign = new RShiftINode(plus_one, shift); 1896 register_new_node(sign, pre_ctrl); 1897 plus_one = new AndINode(plus_one, sign); 1898 register_new_node(plus_one, pre_ctrl); 1899 } else { 1900 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1901 // The only problem we have here when offset == max_int 1902 // since (max_int+1) == min_int and (0-min_int) == min_int. 1903 // But it is fine since main loop will either have 1904 // less iterations or will be skipped in such case. 1905 } 1906 // The underflow limit: low_limit <= scale*I+offset. 1907 // For main-loop compute 1908 // scale*I+offset+1 > low_limit 1909 // ( if (scale < 0) /* and stride > 0 */ 1910 // I < (low_limit-(offset+1))/scale 1911 // else /* scale > 0 and stride < 0 */ 1912 // I > (low_limit-(offset+1))/scale 1913 // ) 1914 1915 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); 1916 } 1917 } 1918 1919 1920 //------------------------------is_scaled_iv--------------------------------- 1921 // Return true if exp is a constant times an induction var 1922 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1923 if (exp == iv) { 1924 if (p_scale != NULL) { 1925 *p_scale = 1; 1926 } 1927 return true; 1928 } 1929 int opc = exp->Opcode(); 1930 if (opc == Op_MulI) { 1931 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1932 if (p_scale != NULL) { 1933 *p_scale = exp->in(2)->get_int(); 1934 } 1935 return true; 1936 } 1937 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1938 if (p_scale != NULL) { 1939 *p_scale = exp->in(1)->get_int(); 1940 } 1941 return true; 1942 } 1943 } else if (opc == Op_LShiftI) { 1944 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1945 if (p_scale != NULL) { 1946 *p_scale = 1 << exp->in(2)->get_int(); 1947 } 1948 return true; 1949 } 1950 } 1951 return false; 1952 } 1953 1954 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1955 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1956 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1957 if (is_scaled_iv(exp, iv, p_scale)) { 1958 if (p_offset != NULL) { 1959 Node *zero = _igvn.intcon(0); 1960 set_ctrl(zero, C->root()); 1961 *p_offset = zero; 1962 } 1963 return true; 1964 } 1965 int opc = exp->Opcode(); 1966 if (opc == Op_AddI) { 1967 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1968 if (p_offset != NULL) { 1969 *p_offset = exp->in(2); 1970 } 1971 return true; 1972 } 1973 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1974 if (p_offset != NULL) { 1975 *p_offset = exp->in(1); 1976 } 1977 return true; 1978 } 1979 if (exp->in(2)->is_Con()) { 1980 Node* offset2 = NULL; 1981 if (depth < 2 && 1982 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1983 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1984 if (p_offset != NULL) { 1985 Node *ctrl_off2 = get_ctrl(offset2); 1986 Node* offset = new AddINode(offset2, exp->in(2)); 1987 register_new_node(offset, ctrl_off2); 1988 *p_offset = offset; 1989 } 1990 return true; 1991 } 1992 } 1993 } else if (opc == Op_SubI) { 1994 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1995 if (p_offset != NULL) { 1996 Node *zero = _igvn.intcon(0); 1997 set_ctrl(zero, C->root()); 1998 Node *ctrl_off = get_ctrl(exp->in(2)); 1999 Node* offset = new SubINode(zero, exp->in(2)); 2000 register_new_node(offset, ctrl_off); 2001 *p_offset = offset; 2002 } 2003 return true; 2004 } 2005 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 2006 if (p_offset != NULL) { 2007 *p_scale *= -1; 2008 *p_offset = exp->in(1); 2009 } 2010 return true; 2011 } 2012 } 2013 return false; 2014 } 2015 2016 //------------------------------do_range_check--------------------------------- 2017 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 2018 int PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 2019 #ifndef PRODUCT 2020 if (PrintOpto && VerifyLoopOptimizations) { 2021 tty->print("Range Check Elimination "); 2022 loop->dump_head(); 2023 } else if (TraceLoopOpts) { 2024 tty->print("RangeCheck "); 2025 loop->dump_head(); 2026 } 2027 #endif 2028 assert(RangeCheckElimination, ""); 2029 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2030 // If we fail before trying to eliminate range checks, set multiversion state 2031 int closed_range_checks = 1; 2032 2033 // protect against stride not being a constant 2034 if (!cl->stride_is_con()) 2035 return closed_range_checks; 2036 2037 // Find the trip counter; we are iteration splitting based on it 2038 Node *trip_counter = cl->phi(); 2039 // Find the main loop limit; we will trim it's iterations 2040 // to not ever trip end tests 2041 Node *main_limit = cl->limit(); 2042 2043 // Check graph shape. Cannot optimize a loop if zero-trip 2044 // Opaque1 node is optimized away and then another round 2045 // of loop opts attempted. 2046 if (!is_canonical_loop_entry(cl)) { 2047 return closed_range_checks; 2048 } 2049 2050 // Need to find the main-loop zero-trip guard 2051 Node *ctrl = cl->in(LoopNode::EntryControl); 2052 Node *iffm = ctrl->in(0); 2053 Node *opqzm = iffm->in(1)->in(1)->in(2); 2054 assert(opqzm->in(1) == main_limit, "do not understand situation"); 2055 2056 // Find the pre-loop limit; we will expand its iterations to 2057 // not ever trip low tests. 2058 Node *p_f = iffm->in(0); 2059 // pre loop may have been optimized out 2060 if (p_f->Opcode() != Op_IfFalse) { 2061 return closed_range_checks; 2062 } 2063 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2064 assert(pre_end->loopnode()->is_pre_loop(), ""); 2065 Node *pre_opaq1 = pre_end->limit(); 2066 // Occasionally it's possible for a pre-loop Opaque1 node to be 2067 // optimized away and then another round of loop opts attempted. 2068 // We can not optimize this particular loop in that case. 2069 if (pre_opaq1->Opcode() != Op_Opaque1) 2070 return closed_range_checks; 2071 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 2072 Node *pre_limit = pre_opaq->in(1); 2073 2074 // Where do we put new limit calculations 2075 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 2076 2077 // Ensure the original loop limit is available from the 2078 // pre-loop Opaque1 node. 2079 Node *orig_limit = pre_opaq->original_loop_limit(); 2080 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 2081 return closed_range_checks; 2082 2083 // Must know if its a count-up or count-down loop 2084 2085 int stride_con = cl->stride_con(); 2086 Node *zero = _igvn.intcon(0); 2087 Node *one = _igvn.intcon(1); 2088 // Use symmetrical int range [-max_jint,max_jint] 2089 Node *mini = _igvn.intcon(-max_jint); 2090 set_ctrl(zero, C->root()); 2091 set_ctrl(one, C->root()); 2092 set_ctrl(mini, C->root()); 2093 2094 // Range checks that do not dominate the loop backedge (ie. 2095 // conditionally executed) can lengthen the pre loop limit beyond 2096 // the original loop limit. To prevent this, the pre limit is 2097 // (for stride > 0) MINed with the original loop limit (MAXed 2098 // stride < 0) when some range_check (rc) is conditionally 2099 // executed. 2100 bool conditional_rc = false; 2101 2102 // Count number of range checks and reduce by load range limits, if zero, 2103 // the loop is in canonical form to multiversion. 2104 closed_range_checks = 0; 2105 2106 // Check loop body for tests of trip-counter plus loop-invariant vs 2107 // loop-invariant. 2108 for( uint i = 0; i < loop->_body.size(); i++ ) { 2109 Node *iff = loop->_body[i]; 2110 if (iff->Opcode() == Op_If || 2111 iff->Opcode() == Op_RangeCheck) { // Test? 2112 // Test is an IfNode, has 2 projections. If BOTH are in the loop 2113 // we need loop unswitching instead of iteration splitting. 2114 closed_range_checks++; 2115 Node *exit = loop->is_loop_exit(iff); 2116 if( !exit ) continue; 2117 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 2118 2119 // Get boolean condition to test 2120 Node *i1 = iff->in(1); 2121 if( !i1->is_Bool() ) continue; 2122 BoolNode *bol = i1->as_Bool(); 2123 BoolTest b_test = bol->_test; 2124 // Flip sense of test if exit condition is flipped 2125 if( flip ) 2126 b_test = b_test.negate(); 2127 2128 // Get compare 2129 Node *cmp = bol->in(1); 2130 2131 // Look for trip_counter + offset vs limit 2132 Node *rc_exp = cmp->in(1); 2133 Node *limit = cmp->in(2); 2134 jint scale_con= 1; // Assume trip counter not scaled 2135 2136 Node *limit_c = get_ctrl(limit); 2137 if( loop->is_member(get_loop(limit_c) ) ) { 2138 // Compare might have operands swapped; commute them 2139 b_test = b_test.commute(); 2140 rc_exp = cmp->in(2); 2141 limit = cmp->in(1); 2142 limit_c = get_ctrl(limit); 2143 if( loop->is_member(get_loop(limit_c) ) ) 2144 continue; // Both inputs are loop varying; cannot RCE 2145 } 2146 // Here we know 'limit' is loop invariant 2147 2148 // 'limit' maybe pinned below the zero trip test (probably from a 2149 // previous round of rce), in which case, it can't be used in the 2150 // zero trip test expression which must occur before the zero test's if. 2151 if( limit_c == ctrl ) { 2152 continue; // Don't rce this check but continue looking for other candidates. 2153 } 2154 2155 // Check for scaled induction variable plus an offset 2156 Node *offset = NULL; 2157 2158 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 2159 continue; 2160 } 2161 2162 Node *offset_c = get_ctrl(offset); 2163 if( loop->is_member( get_loop(offset_c) ) ) 2164 continue; // Offset is not really loop invariant 2165 // Here we know 'offset' is loop invariant. 2166 2167 // As above for the 'limit', the 'offset' maybe pinned below the 2168 // zero trip test. 2169 if( offset_c == ctrl ) { 2170 continue; // Don't rce this check but continue looking for other candidates. 2171 } 2172 #ifdef ASSERT 2173 if (TraceRangeLimitCheck) { 2174 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 2175 bol->dump(2); 2176 } 2177 #endif 2178 // At this point we have the expression as: 2179 // scale_con * trip_counter + offset :: limit 2180 // where scale_con, offset and limit are loop invariant. Trip_counter 2181 // monotonically increases by stride_con, a constant. Both (or either) 2182 // stride_con and scale_con can be negative which will flip about the 2183 // sense of the test. 2184 2185 // Adjust pre and main loop limits to guard the correct iteration set 2186 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 2187 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 2188 // The underflow and overflow limits: 0 <= scale*I+offset < limit 2189 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); 2190 if (!conditional_rc) { 2191 // (0-offset)/scale could be outside of loop iterations range. 2192 conditional_rc = !loop->dominates_backedge(iff); 2193 } 2194 } else { 2195 if (PrintOpto) { 2196 tty->print_cr("missed RCE opportunity"); 2197 } 2198 continue; // In release mode, ignore it 2199 } 2200 } else { // Otherwise work on normal compares 2201 switch( b_test._test ) { 2202 case BoolTest::gt: 2203 // Fall into GE case 2204 case BoolTest::ge: 2205 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 2206 scale_con = -scale_con; 2207 offset = new SubINode( zero, offset ); 2208 register_new_node( offset, pre_ctrl ); 2209 limit = new SubINode( zero, limit ); 2210 register_new_node( limit, pre_ctrl ); 2211 // Fall into LE case 2212 case BoolTest::le: 2213 if (b_test._test != BoolTest::gt) { 2214 // Convert X <= Y to X < Y+1 2215 limit = new AddINode( limit, one ); 2216 register_new_node( limit, pre_ctrl ); 2217 } 2218 // Fall into LT case 2219 case BoolTest::lt: 2220 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 2221 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 2222 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 2223 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); 2224 if (!conditional_rc) { 2225 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 2226 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 2227 // still be outside of loop range. 2228 conditional_rc = !loop->dominates_backedge(iff); 2229 } 2230 break; 2231 default: 2232 if (PrintOpto) { 2233 tty->print_cr("missed RCE opportunity"); 2234 } 2235 continue; // Unhandled case 2236 } 2237 } 2238 2239 // Kill the eliminated test 2240 C->set_major_progress(); 2241 Node *kill_con = _igvn.intcon( 1-flip ); 2242 set_ctrl(kill_con, C->root()); 2243 _igvn.replace_input_of(iff, 1, kill_con); 2244 // Find surviving projection 2245 assert(iff->is_If(), ""); 2246 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 2247 // Find loads off the surviving projection; remove their control edge 2248 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2249 Node* cd = dp->fast_out(i); // Control-dependent node 2250 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop 2251 // Allow the load to float around in the loop, or before it 2252 // but NOT before the pre-loop. 2253 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 2254 --i; 2255 --imax; 2256 } 2257 } 2258 if (limit->Opcode() == Op_LoadRange) { 2259 closed_range_checks--; 2260 } 2261 2262 } // End of is IF 2263 2264 } 2265 2266 // Update loop limits 2267 if (conditional_rc) { 2268 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit) 2269 : (Node*)new MaxINode(pre_limit, orig_limit); 2270 register_new_node(pre_limit, pre_ctrl); 2271 } 2272 _igvn.replace_input_of(pre_opaq, 1, pre_limit); 2273 2274 // Note:: we are making the main loop limit no longer precise; 2275 // need to round up based on stride. 2276 cl->set_nonexact_trip_count(); 2277 Node *main_cle = cl->loopexit(); 2278 Node *main_bol = main_cle->in(1); 2279 // Hacking loop bounds; need private copies of exit test 2280 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 2281 main_bol = main_bol->clone();// Clone a private BoolNode 2282 register_new_node( main_bol, main_cle->in(0) ); 2283 _igvn.replace_input_of(main_cle, 1, main_bol); 2284 } 2285 Node *main_cmp = main_bol->in(1); 2286 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 2287 main_cmp = main_cmp->clone();// Clone a private CmpNode 2288 register_new_node( main_cmp, main_cle->in(0) ); 2289 _igvn.replace_input_of(main_bol, 1, main_cmp); 2290 } 2291 // Hack the now-private loop bounds 2292 _igvn.replace_input_of(main_cmp, 2, main_limit); 2293 // The OpaqueNode is unshared by design 2294 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 2295 _igvn.replace_input_of(opqzm, 1, main_limit); 2296 2297 return closed_range_checks; 2298 } 2299 2300 //------------------------------has_range_checks------------------------------- 2301 // Check to see if RCE cleaned the current loop of range-checks. 2302 void PhaseIdealLoop::has_range_checks(IdealLoopTree *loop) { 2303 assert(RangeCheckElimination, ""); 2304 2305 // skip if not a counted loop 2306 if (!loop->is_counted()) return; 2307 2308 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2309 2310 // skip this loop if it is already checked 2311 if (cl->has_been_range_checked()) return; 2312 2313 // Now check for existance of range checks 2314 for (uint i = 0; i < loop->_body.size(); i++) { 2315 Node *iff = loop->_body[i]; 2316 int iff_opc = iff->Opcode(); 2317 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) { 2318 cl->mark_has_range_checks(); 2319 break; 2320 } 2321 } 2322 cl->set_has_been_range_checked(); 2323 } 2324 2325 //-------------------------multi_version_post_loops---------------------------- 2326 // Check the range checks that remain, if simple, use the bounds to guard 2327 // which version to a post loop we execute, one with range checks or one without 2328 bool PhaseIdealLoop::multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop) { 2329 bool multi_version_succeeded = false; 2330 assert(RangeCheckElimination, ""); 2331 CountedLoopNode *legacy_cl = legacy_loop->_head->as_CountedLoop(); 2332 assert(legacy_cl->is_post_loop(), ""); 2333 2334 // Check for existance of range checks using the unique instance to make a guard with 2335 Unique_Node_List worklist; 2336 for (uint i = 0; i < legacy_loop->_body.size(); i++) { 2337 Node *iff = legacy_loop->_body[i]; 2338 int iff_opc = iff->Opcode(); 2339 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) { 2340 worklist.push(iff); 2341 } 2342 } 2343 2344 // Find RCE'd post loop so that we can stage its guard. 2345 if (!is_canonical_loop_entry(legacy_cl)) return multi_version_succeeded; 2346 Node* ctrl = legacy_cl->in(LoopNode::EntryControl); 2347 Node* iffm = ctrl->in(0); 2348 2349 // Now we test that both the post loops are connected 2350 Node* post_loop_region = iffm->in(0); 2351 if (post_loop_region == NULL) return multi_version_succeeded; 2352 if (!post_loop_region->is_Region()) return multi_version_succeeded; 2353 Node* covering_region = post_loop_region->in(RegionNode::Control+1); 2354 if (covering_region == NULL) return multi_version_succeeded; 2355 if (!covering_region->is_Region()) return multi_version_succeeded; 2356 Node* p_f = covering_region->in(RegionNode::Control); 2357 if (p_f == NULL) return multi_version_succeeded; 2358 if (!p_f->is_IfFalse()) return multi_version_succeeded; 2359 if (!p_f->in(0)->is_CountedLoopEnd()) return multi_version_succeeded; 2360 CountedLoopEndNode* rce_loop_end = p_f->in(0)->as_CountedLoopEnd(); 2361 if (rce_loop_end == NULL) return multi_version_succeeded; 2362 CountedLoopNode* rce_cl = rce_loop_end->loopnode(); 2363 if (rce_cl == NULL || !rce_cl->is_post_loop()) return multi_version_succeeded; 2364 CountedLoopNode *known_rce_cl = rce_loop->_head->as_CountedLoop(); 2365 if (rce_cl != known_rce_cl) return multi_version_succeeded; 2366 2367 // Then we fetch the cover entry test 2368 ctrl = rce_cl->in(LoopNode::EntryControl); 2369 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return multi_version_succeeded; 2370 2371 #ifndef PRODUCT 2372 if (TraceLoopOpts) { 2373 tty->print("PostMultiVersion\n"); 2374 rce_loop->dump_head(); 2375 legacy_loop->dump_head(); 2376 } 2377 #endif 2378 2379 // Now fetch the limit we want to compare against 2380 Node *limit = rce_cl->limit(); 2381 bool first_time = true; 2382 2383 // If we got this far, we identified the post loop which has been RCE'd and 2384 // we have a work list. Now we will try to transform the if guard to cause 2385 // the loop pair to be multi version executed with the determination left to runtime 2386 // or the optimizer if full information is known about the given arrays at compile time. 2387 Node *last_min = NULL; 2388 multi_version_succeeded = true; 2389 while (worklist.size()) { 2390 Node* rc_iffm = worklist.pop(); 2391 if (rc_iffm->is_If()) { 2392 Node *rc_bolzm = rc_iffm->in(1); 2393 if (rc_bolzm->is_Bool()) { 2394 Node *rc_cmpzm = rc_bolzm->in(1); 2395 if (rc_cmpzm->is_Cmp()) { 2396 Node *rc_left = rc_cmpzm->in(2); 2397 if (rc_left->Opcode() != Op_LoadRange) { 2398 multi_version_succeeded = false; 2399 break; 2400 } 2401 if (first_time) { 2402 last_min = rc_left; 2403 first_time = false; 2404 } else { 2405 Node *cur_min = new MinINode(last_min, rc_left); 2406 last_min = cur_min; 2407 _igvn.register_new_node_with_optimizer(last_min); 2408 } 2409 } 2410 } 2411 } 2412 } 2413 2414 // All we have to do is update the limit of the rce loop 2415 // with the min of our expression and the current limit. 2416 // We will use this expression to replace the current limit. 2417 if (last_min && multi_version_succeeded) { 2418 Node *cur_min = new MinINode(last_min, limit); 2419 _igvn.register_new_node_with_optimizer(cur_min); 2420 Node *cmp_node = rce_loop_end->cmp_node(); 2421 _igvn.replace_input_of(cmp_node, 2, cur_min); 2422 set_idom(cmp_node, cur_min, dom_depth(ctrl)); 2423 set_ctrl(cur_min, ctrl); 2424 set_loop(cur_min, rce_loop->_parent); 2425 2426 legacy_cl->mark_is_multiversioned(); 2427 rce_cl->mark_is_multiversioned(); 2428 multi_version_succeeded = true; 2429 2430 C->set_major_progress(); 2431 } 2432 2433 return multi_version_succeeded; 2434 } 2435 2436 //-------------------------poison_rce_post_loop-------------------------------- 2437 // Causes the rce'd post loop to be optimized away if multiverioning fails 2438 void PhaseIdealLoop::poison_rce_post_loop(IdealLoopTree *rce_loop) { 2439 CountedLoopNode *rce_cl = rce_loop->_head->as_CountedLoop(); 2440 Node* ctrl = rce_cl->in(LoopNode::EntryControl); 2441 if (ctrl->is_IfTrue() || ctrl->is_IfFalse()) { 2442 Node* iffm = ctrl->in(0); 2443 if (iffm->is_If()) { 2444 Node* cur_bool = iffm->in(1); 2445 if (cur_bool->is_Bool()) { 2446 Node* cur_cmp = cur_bool->in(1); 2447 if (cur_cmp->is_Cmp()) { 2448 BoolTest::mask new_test = BoolTest::gt; 2449 BoolNode *new_bool = new BoolNode(cur_cmp, new_test); 2450 _igvn.replace_node(cur_bool, new_bool); 2451 _igvn._worklist.push(new_bool); 2452 Node* left_op = cur_cmp->in(1); 2453 _igvn.replace_input_of(cur_cmp, 2, left_op); 2454 C->set_major_progress(); 2455 } 2456 } 2457 } 2458 } 2459 } 2460 2461 //------------------------------DCE_loop_body---------------------------------- 2462 // Remove simplistic dead code from loop body 2463 void IdealLoopTree::DCE_loop_body() { 2464 for( uint i = 0; i < _body.size(); i++ ) 2465 if( _body.at(i)->outcnt() == 0 ) 2466 _body.map( i--, _body.pop() ); 2467 } 2468 2469 2470 //------------------------------adjust_loop_exit_prob-------------------------- 2471 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2472 // Replace with a 1-in-10 exit guess. 2473 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 2474 Node *test = tail(); 2475 while( test != _head ) { 2476 uint top = test->Opcode(); 2477 if( top == Op_IfTrue || top == Op_IfFalse ) { 2478 int test_con = ((ProjNode*)test)->_con; 2479 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2480 IfNode *iff = test->in(0)->as_If(); 2481 if( iff->outcnt() == 2 ) { // Ignore dead tests 2482 Node *bol = iff->in(1); 2483 if( bol && bol->req() > 1 && bol->in(1) && 2484 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 2485 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 2486 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 2487 (bol->in(1)->Opcode() == Op_CompareAndExchangeI ) || 2488 (bol->in(1)->Opcode() == Op_CompareAndExchangeL ) || 2489 (bol->in(1)->Opcode() == Op_CompareAndExchangeP ) || 2490 (bol->in(1)->Opcode() == Op_CompareAndExchangeN ) || 2491 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI ) || 2492 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL ) || 2493 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP ) || 2494 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN ) || 2495 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 2496 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 2497 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 2498 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 2499 return; // Allocation loops RARELY take backedge 2500 // Find the OTHER exit path from the IF 2501 Node* ex = iff->proj_out(1-test_con); 2502 float p = iff->_prob; 2503 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 2504 if( top == Op_IfTrue ) { 2505 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2506 iff->_prob = PROB_STATIC_FREQUENT; 2507 } 2508 } else { 2509 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2510 iff->_prob = PROB_STATIC_INFREQUENT; 2511 } 2512 } 2513 } 2514 } 2515 } 2516 test = phase->idom(test); 2517 } 2518 } 2519 2520 #ifdef ASSERT 2521 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) { 2522 Node *ctrl = cl->in(LoopNode::EntryControl); 2523 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 2524 Node *iffm = ctrl->in(0); 2525 assert(iffm->Opcode() == Op_If, ""); 2526 Node *p_f = iffm->in(0); 2527 assert(p_f->Opcode() == Op_IfFalse, ""); 2528 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2529 assert(pre_end->loopnode()->is_pre_loop(), ""); 2530 return pre_end->loopnode(); 2531 } 2532 #endif 2533 2534 // Remove the main and post loops and make the pre loop execute all 2535 // iterations. Useful when the pre loop is found empty. 2536 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) { 2537 CountedLoopEndNode* pre_end = cl->loopexit(); 2538 Node* pre_cmp = pre_end->cmp_node(); 2539 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) { 2540 // Only safe to remove the main loop if the compiler optimized it 2541 // out based on an unknown number of iterations 2542 return; 2543 } 2544 2545 // Can we find the main loop? 2546 if (_next == NULL) { 2547 return; 2548 } 2549 2550 Node* next_head = _next->_head; 2551 if (!next_head->is_CountedLoop()) { 2552 return; 2553 } 2554 2555 CountedLoopNode* main_head = next_head->as_CountedLoop(); 2556 if (!main_head->is_main_loop()) { 2557 return; 2558 } 2559 2560 assert(locate_pre_from_main(main_head) == cl, "bad main loop"); 2561 Node* main_iff = main_head->in(LoopNode::EntryControl)->in(0); 2562 2563 // Remove the Opaque1Node of the pre loop and make it execute all iterations 2564 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2)); 2565 // Remove the Opaque1Node of the main loop so it can be optimized out 2566 Node* main_cmp = main_iff->in(1)->in(1); 2567 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?"); 2568 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1)); 2569 } 2570 2571 //------------------------------policy_do_remove_empty_loop-------------------- 2572 // Micro-benchmark spamming. Policy is to always remove empty loops. 2573 // The 'DO' part is to replace the trip counter with the value it will 2574 // have on the last iteration. This will break the loop. 2575 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 2576 // Minimum size must be empty loop 2577 if (_body.size() > EMPTY_LOOP_SIZE) 2578 return false; 2579 2580 if (!_head->is_CountedLoop()) 2581 return false; // Dead loop 2582 CountedLoopNode *cl = _head->as_CountedLoop(); 2583 if (!cl->is_valid_counted_loop()) 2584 return false; // Malformed loop 2585 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 2586 return false; // Infinite loop 2587 2588 if (cl->is_pre_loop()) { 2589 // If the loop we are removing is a pre-loop then the main and 2590 // post loop can be removed as well 2591 remove_main_post_loops(cl, phase); 2592 } 2593 2594 #ifdef ASSERT 2595 // Ensure only one phi which is the iv. 2596 Node* iv = NULL; 2597 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 2598 Node* n = cl->fast_out(i); 2599 if (n->Opcode() == Op_Phi) { 2600 assert(iv == NULL, "Too many phis" ); 2601 iv = n; 2602 } 2603 } 2604 assert(iv == cl->phi(), "Wrong phi" ); 2605 #endif 2606 2607 // main and post loops have explicitly created zero trip guard 2608 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 2609 if (needs_guard) { 2610 // Skip guard if values not overlap. 2611 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 2612 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 2613 int stride_con = cl->stride_con(); 2614 if (stride_con > 0) { 2615 needs_guard = (init_t->_hi >= limit_t->_lo); 2616 } else { 2617 needs_guard = (init_t->_lo <= limit_t->_hi); 2618 } 2619 } 2620 if (needs_guard) { 2621 // Check for an obvious zero trip guard. 2622 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 2623 if (inctrl->Opcode() == Op_IfTrue || inctrl->Opcode() == Op_IfFalse) { 2624 bool maybe_swapped = (inctrl->Opcode() == Op_IfFalse); 2625 // The test should look like just the backedge of a CountedLoop 2626 Node* iff = inctrl->in(0); 2627 if (iff->is_If()) { 2628 Node* bol = iff->in(1); 2629 if (bol->is_Bool()) { 2630 BoolTest test = bol->as_Bool()->_test; 2631 if (maybe_swapped) { 2632 test._test = test.commute(); 2633 test._test = test.negate(); 2634 } 2635 if (test._test == cl->loopexit()->test_trip()) { 2636 Node* cmp = bol->in(1); 2637 int init_idx = maybe_swapped ? 2 : 1; 2638 int limit_idx = maybe_swapped ? 1 : 2; 2639 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) { 2640 needs_guard = false; 2641 } 2642 } 2643 } 2644 } 2645 } 2646 } 2647 2648 #ifndef PRODUCT 2649 if (PrintOpto) { 2650 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 2651 this->dump_head(); 2652 } else if (TraceLoopOpts) { 2653 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 2654 this->dump_head(); 2655 } 2656 #endif 2657 2658 if (needs_guard) { 2659 // Peel the loop to ensure there's a zero trip guard 2660 Node_List old_new; 2661 phase->do_peeling(this, old_new); 2662 } 2663 2664 // Replace the phi at loop head with the final value of the last 2665 // iteration. Then the CountedLoopEnd will collapse (backedge never 2666 // taken) and all loop-invariant uses of the exit values will be correct. 2667 Node *phi = cl->phi(); 2668 Node *exact_limit = phase->exact_limit(this); 2669 if (exact_limit != cl->limit()) { 2670 // We also need to replace the original limit to collapse loop exit. 2671 Node* cmp = cl->loopexit()->cmp_node(); 2672 assert(cl->limit() == cmp->in(2), "sanity"); 2673 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 2674 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 2675 } 2676 // Note: the final value after increment should not overflow since 2677 // counted loop has limit check predicate. 2678 Node *final = new SubINode( exact_limit, cl->stride() ); 2679 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 2680 phase->_igvn.replace_node(phi,final); 2681 phase->C->set_major_progress(); 2682 return true; 2683 } 2684 2685 //------------------------------policy_do_one_iteration_loop------------------- 2686 // Convert one iteration loop into normal code. 2687 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 2688 if (!_head->as_Loop()->is_valid_counted_loop()) 2689 return false; // Only for counted loop 2690 2691 CountedLoopNode *cl = _head->as_CountedLoop(); 2692 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 2693 return false; 2694 } 2695 2696 #ifndef PRODUCT 2697 if(TraceLoopOpts) { 2698 tty->print("OneIteration "); 2699 this->dump_head(); 2700 } 2701 #endif 2702 2703 Node *init_n = cl->init_trip(); 2704 #ifdef ASSERT 2705 // Loop boundaries should be constant since trip count is exact. 2706 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 2707 #endif 2708 // Replace the phi at loop head with the value of the init_trip. 2709 // Then the CountedLoopEnd will collapse (backedge will not be taken) 2710 // and all loop-invariant uses of the exit values will be correct. 2711 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 2712 phase->C->set_major_progress(); 2713 return true; 2714 } 2715 2716 //============================================================================= 2717 //------------------------------iteration_split_impl--------------------------- 2718 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 2719 // Compute exact loop trip count if possible. 2720 compute_exact_trip_count(phase); 2721 2722 // Convert one iteration loop into normal code. 2723 if (policy_do_one_iteration_loop(phase)) 2724 return true; 2725 2726 // Check and remove empty loops (spam micro-benchmarks) 2727 if (policy_do_remove_empty_loop(phase)) 2728 return true; // Here we removed an empty loop 2729 2730 bool should_peel = policy_peeling(phase); // Should we peel? 2731 2732 bool should_unswitch = policy_unswitching(phase); 2733 2734 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 2735 // This removes loop-invariant tests (usually null checks). 2736 if (!_head->is_CountedLoop()) { // Non-counted loop 2737 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 2738 // Partial peel succeeded so terminate this round of loop opts 2739 return false; 2740 } 2741 if (should_peel) { // Should we peel? 2742 if (PrintOpto) { tty->print_cr("should_peel"); } 2743 phase->do_peeling(this,old_new); 2744 } else if (should_unswitch) { 2745 phase->do_unswitching(this, old_new); 2746 } 2747 return true; 2748 } 2749 CountedLoopNode *cl = _head->as_CountedLoop(); 2750 2751 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 2752 2753 // Do nothing special to pre- and post- loops 2754 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 2755 2756 // Compute loop trip count from profile data 2757 compute_profile_trip_cnt(phase); 2758 2759 // Before attempting fancy unrolling, RCE or alignment, see if we want 2760 // to completely unroll this loop or do loop unswitching. 2761 if (cl->is_normal_loop()) { 2762 if (should_unswitch) { 2763 phase->do_unswitching(this, old_new); 2764 return true; 2765 } 2766 bool should_maximally_unroll = policy_maximally_unroll(phase); 2767 if (should_maximally_unroll) { 2768 // Here we did some unrolling and peeling. Eventually we will 2769 // completely unroll this loop and it will no longer be a loop. 2770 phase->do_maximally_unroll(this,old_new); 2771 return true; 2772 } 2773 } 2774 2775 // Skip next optimizations if running low on nodes. Note that 2776 // policy_unswitching and policy_maximally_unroll have this check. 2777 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes(); 2778 if ((int)(2 * _body.size()) > nodes_left) { 2779 return true; 2780 } 2781 2782 // Counted loops may be peeled, may need some iterations run up 2783 // front for RCE, and may want to align loop refs to a cache 2784 // line. Thus we clone a full loop up front whose trip count is 2785 // at least 1 (if peeling), but may be several more. 2786 2787 // The main loop will start cache-line aligned with at least 1 2788 // iteration of the unrolled body (zero-trip test required) and 2789 // will have some range checks removed. 2790 2791 // A post-loop will finish any odd iterations (leftover after 2792 // unrolling), plus any needed for RCE purposes. 2793 2794 bool should_unroll = policy_unroll(phase); 2795 2796 bool should_rce = policy_range_check(phase); 2797 2798 bool should_align = policy_align(phase); 2799 2800 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2801 // need a pre-loop. We may still need to peel an initial iteration but 2802 // we will not be needing an unknown number of pre-iterations. 2803 // 2804 // Basically, if may_rce_align reports FALSE first time through, 2805 // we will not be able to later do RCE or Aligning on this loop. 2806 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2807 2808 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2809 // we switch to the pre-/main-/post-loop model. This model also covers 2810 // peeling. 2811 if (should_rce || should_align || should_unroll) { 2812 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2813 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2814 2815 // Adjust the pre- and main-loop limits to let the pre and post loops run 2816 // with full checks, but the main-loop with no checks. Remove said 2817 // checks from the main body. 2818 if (should_rce) { 2819 if (phase->do_range_check(this, old_new) != 0) { 2820 cl->mark_has_range_checks(); 2821 } 2822 } else if (PostLoopMultiversioning) { 2823 phase->has_range_checks(this); 2824 } 2825 2826 if (should_unroll && !should_peel && PostLoopMultiversioning) { 2827 // Try to setup multiversioning on main loops before they are unrolled 2828 if (cl->is_main_loop() && (cl->unrolled_count() == 1)) { 2829 phase->insert_scalar_rced_post_loop(this, old_new); 2830 } 2831 } 2832 2833 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2834 // twice as many iterations as before) and the main body limit (only do 2835 // an even number of trips). If we are peeling, we might enable some RCE 2836 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2837 // peeling. 2838 if (should_unroll && !should_peel) { 2839 if (SuperWordLoopUnrollAnalysis) { 2840 phase->insert_vector_post_loop(this, old_new); 2841 } 2842 phase->do_unroll(this, old_new, true); 2843 } 2844 2845 // Adjust the pre-loop limits to align the main body 2846 // iterations. 2847 if (should_align) 2848 Unimplemented(); 2849 2850 } else { // Else we have an unchanged counted loop 2851 if (should_peel) // Might want to peel but do nothing else 2852 phase->do_peeling(this,old_new); 2853 } 2854 return true; 2855 } 2856 2857 2858 //============================================================================= 2859 //------------------------------iteration_split-------------------------------- 2860 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2861 // Recursively iteration split nested loops 2862 if (_child && !_child->iteration_split(phase, old_new)) 2863 return false; 2864 2865 // Clean out prior deadwood 2866 DCE_loop_body(); 2867 2868 2869 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2870 // Replace with a 1-in-10 exit guess. 2871 if (_parent /*not the root loop*/ && 2872 !_irreducible && 2873 // Also ignore the occasional dead backedge 2874 !tail()->is_top()) { 2875 adjust_loop_exit_prob(phase); 2876 } 2877 2878 // Gate unrolling, RCE and peeling efforts. 2879 if (!_child && // If not an inner loop, do not split 2880 !_irreducible && 2881 _allow_optimizations && 2882 !tail()->is_top()) { // Also ignore the occasional dead backedge 2883 if (!_has_call) { 2884 if (!iteration_split_impl(phase, old_new)) { 2885 return false; 2886 } 2887 } else if (policy_unswitching(phase)) { 2888 phase->do_unswitching(this, old_new); 2889 } 2890 } 2891 2892 // Minor offset re-organization to remove loop-fallout uses of 2893 // trip counter when there was no major reshaping. 2894 phase->reorg_offsets(this); 2895 2896 if (_next && !_next->iteration_split(phase, old_new)) 2897 return false; 2898 return true; 2899 } 2900 2901 2902 //============================================================================= 2903 // Process all the loops in the loop tree and replace any fill 2904 // patterns with an intrinsic version. 2905 bool PhaseIdealLoop::do_intrinsify_fill() { 2906 bool changed = false; 2907 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2908 IdealLoopTree* lpt = iter.current(); 2909 changed |= intrinsify_fill(lpt); 2910 } 2911 return changed; 2912 } 2913 2914 2915 // Examine an inner loop looking for a a single store of an invariant 2916 // value in a unit stride loop, 2917 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2918 Node*& shift, Node*& con) { 2919 const char* msg = NULL; 2920 Node* msg_node = NULL; 2921 2922 store_value = NULL; 2923 con = NULL; 2924 shift = NULL; 2925 2926 // Process the loop looking for stores. If there are multiple 2927 // stores or extra control flow give at this point. 2928 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2929 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2930 Node* n = lpt->_body.at(i); 2931 if (n->outcnt() == 0) continue; // Ignore dead 2932 if (n->is_Store()) { 2933 if (store != NULL) { 2934 msg = "multiple stores"; 2935 break; 2936 } 2937 int opc = n->Opcode(); 2938 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 2939 msg = "oop fills not handled"; 2940 break; 2941 } 2942 Node* value = n->in(MemNode::ValueIn); 2943 if (!lpt->is_invariant(value)) { 2944 msg = "variant store value"; 2945 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2946 msg = "not array address"; 2947 } 2948 store = n; 2949 store_value = value; 2950 } else if (n->is_If() && n != head->loopexit()) { 2951 msg = "extra control flow"; 2952 msg_node = n; 2953 } 2954 } 2955 2956 if (store == NULL) { 2957 // No store in loop 2958 return false; 2959 } 2960 2961 if (msg == NULL && head->stride_con() != 1) { 2962 // could handle negative strides too 2963 if (head->stride_con() < 0) { 2964 msg = "negative stride"; 2965 } else { 2966 msg = "non-unit stride"; 2967 } 2968 } 2969 2970 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2971 msg = "can't handle store address"; 2972 msg_node = store->in(MemNode::Address); 2973 } 2974 2975 if (msg == NULL && 2976 (!store->in(MemNode::Memory)->is_Phi() || 2977 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2978 msg = "store memory isn't proper phi"; 2979 msg_node = store->in(MemNode::Memory); 2980 } 2981 2982 // Make sure there is an appropriate fill routine 2983 BasicType t = store->as_Mem()->memory_type(); 2984 const char* fill_name; 2985 if (msg == NULL && 2986 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2987 msg = "unsupported store"; 2988 msg_node = store; 2989 } 2990 2991 if (msg != NULL) { 2992 #ifndef PRODUCT 2993 if (TraceOptimizeFill) { 2994 tty->print_cr("not fill intrinsic candidate: %s", msg); 2995 if (msg_node != NULL) msg_node->dump(); 2996 } 2997 #endif 2998 return false; 2999 } 3000 3001 // Make sure the address expression can be handled. It should be 3002 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()). 3003 Node* elements[4]; 3004 Node* cast = NULL; 3005 Node* conv = NULL; 3006 bool found_index = false; 3007 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 3008 for (int e = 0; e < count; e++) { 3009 Node* n = elements[e]; 3010 if (n->is_Con() && con == NULL) { 3011 con = n; 3012 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 3013 Node* value = n->in(1); 3014 #ifdef _LP64 3015 if (value->Opcode() == Op_ConvI2L) { 3016 conv = value; 3017 value = value->in(1); 3018 } 3019 if (value->Opcode() == Op_CastII && 3020 value->as_CastII()->has_range_check()) { 3021 // Skip range check dependent CastII nodes 3022 cast = value; 3023 value = value->in(1); 3024 } 3025 #endif 3026 if (value != head->phi()) { 3027 msg = "unhandled shift in address"; 3028 } else { 3029 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 3030 msg = "scale doesn't match"; 3031 } else { 3032 found_index = true; 3033 shift = n; 3034 } 3035 } 3036 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 3037 conv = n; 3038 n = n->in(1); 3039 if (n->Opcode() == Op_CastII && 3040 n->as_CastII()->has_range_check()) { 3041 // Skip range check dependent CastII nodes 3042 cast = n; 3043 n = n->in(1); 3044 } 3045 if (n == head->phi()) { 3046 found_index = true; 3047 } else { 3048 msg = "unhandled input to ConvI2L"; 3049 } 3050 } else if (n == head->phi()) { 3051 // no shift, check below for allowed cases 3052 found_index = true; 3053 } else { 3054 msg = "unhandled node in address"; 3055 msg_node = n; 3056 } 3057 } 3058 3059 if (count == -1) { 3060 msg = "malformed address expression"; 3061 msg_node = store; 3062 } 3063 3064 if (!found_index) { 3065 msg = "missing use of index"; 3066 } 3067 3068 // byte sized items won't have a shift 3069 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 3070 msg = "can't find shift"; 3071 msg_node = store; 3072 } 3073 3074 if (msg != NULL) { 3075 #ifndef PRODUCT 3076 if (TraceOptimizeFill) { 3077 tty->print_cr("not fill intrinsic: %s", msg); 3078 if (msg_node != NULL) msg_node->dump(); 3079 } 3080 #endif 3081 return false; 3082 } 3083 3084 // No make sure all the other nodes in the loop can be handled 3085 VectorSet ok(Thread::current()->resource_area()); 3086 3087 // store related values are ok 3088 ok.set(store->_idx); 3089 ok.set(store->in(MemNode::Memory)->_idx); 3090 3091 CountedLoopEndNode* loop_exit = head->loopexit(); 3092 guarantee(loop_exit != NULL, "no loop exit node"); 3093 3094 // Loop structure is ok 3095 ok.set(head->_idx); 3096 ok.set(loop_exit->_idx); 3097 ok.set(head->phi()->_idx); 3098 ok.set(head->incr()->_idx); 3099 ok.set(loop_exit->cmp_node()->_idx); 3100 ok.set(loop_exit->in(1)->_idx); 3101 3102 // Address elements are ok 3103 if (con) ok.set(con->_idx); 3104 if (shift) ok.set(shift->_idx); 3105 if (cast) ok.set(cast->_idx); 3106 if (conv) ok.set(conv->_idx); 3107 3108 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 3109 Node* n = lpt->_body.at(i); 3110 if (n->outcnt() == 0) continue; // Ignore dead 3111 if (ok.test(n->_idx)) continue; 3112 // Backedge projection is ok 3113 if (n->is_IfTrue() && n->in(0) == loop_exit) continue; 3114 if (!n->is_AddP()) { 3115 msg = "unhandled node"; 3116 msg_node = n; 3117 break; 3118 } 3119 } 3120 3121 // Make sure no unexpected values are used outside the loop 3122 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 3123 Node* n = lpt->_body.at(i); 3124 // These values can be replaced with other nodes if they are used 3125 // outside the loop. 3126 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue; 3127 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 3128 Node* use = iter.get(); 3129 if (!lpt->_body.contains(use)) { 3130 msg = "node is used outside loop"; 3131 // lpt->_body.dump(); 3132 msg_node = n; 3133 break; 3134 } 3135 } 3136 } 3137 3138 #ifdef ASSERT 3139 if (TraceOptimizeFill) { 3140 if (msg != NULL) { 3141 tty->print_cr("no fill intrinsic: %s", msg); 3142 if (msg_node != NULL) msg_node->dump(); 3143 } else { 3144 tty->print_cr("fill intrinsic for:"); 3145 } 3146 store->dump(); 3147 if (Verbose) { 3148 lpt->_body.dump(); 3149 } 3150 } 3151 #endif 3152 3153 return msg == NULL; 3154 } 3155 3156 3157 3158 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 3159 // Only for counted inner loops 3160 if (!lpt->is_counted() || !lpt->is_inner()) { 3161 return false; 3162 } 3163 3164 // Must have constant stride 3165 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 3166 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 3167 return false; 3168 } 3169 3170 // Check that the body only contains a store of a loop invariant 3171 // value that is indexed by the loop phi. 3172 Node* store = NULL; 3173 Node* store_value = NULL; 3174 Node* shift = NULL; 3175 Node* offset = NULL; 3176 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 3177 return false; 3178 } 3179 3180 #ifndef PRODUCT 3181 if (TraceLoopOpts) { 3182 tty->print("ArrayFill "); 3183 lpt->dump_head(); 3184 } 3185 #endif 3186 3187 // Now replace the whole loop body by a call to a fill routine that 3188 // covers the same region as the loop. 3189 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 3190 3191 // Build an expression for the beginning of the copy region 3192 Node* index = head->init_trip(); 3193 #ifdef _LP64 3194 index = new ConvI2LNode(index); 3195 _igvn.register_new_node_with_optimizer(index); 3196 #endif 3197 if (shift != NULL) { 3198 // byte arrays don't require a shift but others do. 3199 index = new LShiftXNode(index, shift->in(2)); 3200 _igvn.register_new_node_with_optimizer(index); 3201 } 3202 index = new AddPNode(base, base, index); 3203 _igvn.register_new_node_with_optimizer(index); 3204 Node* from = new AddPNode(base, index, offset); 3205 _igvn.register_new_node_with_optimizer(from); 3206 // Compute the number of elements to copy 3207 Node* len = new SubINode(head->limit(), head->init_trip()); 3208 _igvn.register_new_node_with_optimizer(len); 3209 3210 BasicType t = store->as_Mem()->memory_type(); 3211 bool aligned = false; 3212 if (offset != NULL && head->init_trip()->is_Con()) { 3213 int element_size = type2aelembytes(t); 3214 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 3215 } 3216 3217 // Build a call to the fill routine 3218 const char* fill_name; 3219 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 3220 assert(fill != NULL, "what?"); 3221 3222 // Convert float/double to int/long for fill routines 3223 if (t == T_FLOAT) { 3224 store_value = new MoveF2INode(store_value); 3225 _igvn.register_new_node_with_optimizer(store_value); 3226 } else if (t == T_DOUBLE) { 3227 store_value = new MoveD2LNode(store_value); 3228 _igvn.register_new_node_with_optimizer(store_value); 3229 } 3230 3231 Node* mem_phi = store->in(MemNode::Memory); 3232 Node* result_ctrl; 3233 Node* result_mem; 3234 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 3235 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill, 3236 fill_name, TypeAryPtr::get_array_body_type(t)); 3237 uint cnt = 0; 3238 call->init_req(TypeFunc::Parms + cnt++, from); 3239 call->init_req(TypeFunc::Parms + cnt++, store_value); 3240 #ifdef _LP64 3241 len = new ConvI2LNode(len); 3242 _igvn.register_new_node_with_optimizer(len); 3243 #endif 3244 call->init_req(TypeFunc::Parms + cnt++, len); 3245 #ifdef _LP64 3246 call->init_req(TypeFunc::Parms + cnt++, C->top()); 3247 #endif 3248 call->init_req(TypeFunc::Control, head->init_control()); 3249 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. 3250 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); 3251 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr)); 3252 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr)); 3253 _igvn.register_new_node_with_optimizer(call); 3254 result_ctrl = new ProjNode(call,TypeFunc::Control); 3255 _igvn.register_new_node_with_optimizer(result_ctrl); 3256 result_mem = new ProjNode(call,TypeFunc::Memory); 3257 _igvn.register_new_node_with_optimizer(result_mem); 3258 3259 /* Disable following optimization until proper fix (add missing checks). 3260 3261 // If this fill is tightly coupled to an allocation and overwrites 3262 // the whole body, allow it to take over the zeroing. 3263 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 3264 if (alloc != NULL && alloc->is_AllocateArray()) { 3265 Node* length = alloc->as_AllocateArray()->Ideal_length(); 3266 if (head->limit() == length && 3267 head->init_trip() == _igvn.intcon(0)) { 3268 if (TraceOptimizeFill) { 3269 tty->print_cr("Eliminated zeroing in allocation"); 3270 } 3271 alloc->maybe_set_complete(&_igvn); 3272 } else { 3273 #ifdef ASSERT 3274 if (TraceOptimizeFill) { 3275 tty->print_cr("filling array but bounds don't match"); 3276 alloc->dump(); 3277 head->init_trip()->dump(); 3278 head->limit()->dump(); 3279 length->dump(); 3280 } 3281 #endif 3282 } 3283 } 3284 */ 3285 3286 // Redirect the old control and memory edges that are outside the loop. 3287 Node* exit = head->loopexit()->proj_out(0); 3288 // Sometimes the memory phi of the head is used as the outgoing 3289 // state of the loop. It's safe in this case to replace it with the 3290 // result_mem. 3291 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 3292 lazy_replace(exit, result_ctrl); 3293 _igvn.replace_node(store, result_mem); 3294 // Any uses the increment outside of the loop become the loop limit. 3295 _igvn.replace_node(head->incr(), head->limit()); 3296 3297 // Disconnect the head from the loop. 3298 for (uint i = 0; i < lpt->_body.size(); i++) { 3299 Node* n = lpt->_body.at(i); 3300 _igvn.replace_node(n, C->top()); 3301 } 3302 3303 return true; 3304 }