1 /* 2 * Copyright (c) 2000, 2016, 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 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_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 && limit_n != NULL) { 98 // Use longs to avoid integer overflow. 99 int stride_con = cl->stride_con(); 100 jlong init_con = phase->_igvn.type(init_n)->is_int()->_lo; 101 jlong limit_con = phase->_igvn.type(limit_n)->is_int()->_hi; 102 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 103 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 104 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { 105 if (init_n->is_Con() && limit_n->is_Con()) { 106 // Set exact trip count. 107 cl->set_exact_trip_count((uint)trip_count); 108 } else if (cl->unrolled_count() == 1) { 109 // Set maximum trip count before unrolling. 110 cl->set_trip_count((uint)trip_count); 111 } 112 } 113 } 114 } 115 116 //------------------------------compute_profile_trip_cnt---------------------------- 117 // Compute loop trip count from profile data as 118 // (backedge_count + loop_exit_count) / loop_exit_count 119 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 120 if (!_head->is_CountedLoop()) { 121 return; 122 } 123 CountedLoopNode* head = _head->as_CountedLoop(); 124 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 125 return; // Already computed 126 } 127 float trip_cnt = (float)max_jint; // default is big 128 129 Node* back = head->in(LoopNode::LoopBackControl); 130 while (back != head) { 131 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 132 back->in(0) && 133 back->in(0)->is_If() && 134 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 135 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 136 break; 137 } 138 back = phase->idom(back); 139 } 140 if (back != head) { 141 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 142 back->in(0), "if-projection exists"); 143 IfNode* back_if = back->in(0)->as_If(); 144 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 145 146 // Now compute a loop exit count 147 float loop_exit_cnt = 0.0f; 148 for( uint i = 0; i < _body.size(); i++ ) { 149 Node *n = _body[i]; 150 if( n->is_If() ) { 151 IfNode *iff = n->as_If(); 152 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 153 Node *exit = is_loop_exit(iff); 154 if( exit ) { 155 float exit_prob = iff->_prob; 156 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 157 if (exit_prob > PROB_MIN) { 158 float exit_cnt = iff->_fcnt * exit_prob; 159 loop_exit_cnt += exit_cnt; 160 } 161 } 162 } 163 } 164 } 165 if (loop_exit_cnt > 0.0f) { 166 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 167 } else { 168 // No exit count so use 169 trip_cnt = loop_back_cnt; 170 } 171 } 172 #ifndef PRODUCT 173 if (TraceProfileTripCount) { 174 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 175 } 176 #endif 177 head->set_profile_trip_cnt(trip_cnt); 178 } 179 180 //---------------------is_invariant_addition----------------------------- 181 // Return nonzero index of invariant operand for an Add or Sub 182 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 183 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 184 int op = n->Opcode(); 185 if (op == Op_AddI || op == Op_SubI) { 186 bool in1_invar = this->is_invariant(n->in(1)); 187 bool in2_invar = this->is_invariant(n->in(2)); 188 if (in1_invar && !in2_invar) return 1; 189 if (!in1_invar && in2_invar) return 2; 190 } 191 return 0; 192 } 193 194 //---------------------reassociate_add_sub----------------------------- 195 // Reassociate invariant add and subtract expressions: 196 // 197 // inv1 + (x + inv2) => ( inv1 + inv2) + x 198 // (x + inv2) + inv1 => ( inv1 + inv2) + x 199 // inv1 + (x - inv2) => ( inv1 - inv2) + x 200 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 201 // (x + inv2) - inv1 => (-inv1 + inv2) + x 202 // (x - inv2) + inv1 => ( inv1 - inv2) + x 203 // (x - inv2) - inv1 => (-inv1 - inv2) + x 204 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 205 // inv1 - (x - inv2) => ( inv1 + inv2) - x 206 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 207 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 208 // inv1 - (x + inv2) => ( inv1 - inv2) - x 209 // 210 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 211 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 212 if (is_invariant(n1)) return NULL; 213 int inv1_idx = is_invariant_addition(n1, phase); 214 if (!inv1_idx) return NULL; 215 // Don't mess with add of constant (igvn moves them to expression tree root.) 216 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 217 Node* inv1 = n1->in(inv1_idx); 218 Node* n2 = n1->in(3 - inv1_idx); 219 int inv2_idx = is_invariant_addition(n2, phase); 220 if (!inv2_idx) return NULL; 221 Node* x = n2->in(3 - inv2_idx); 222 Node* inv2 = n2->in(inv2_idx); 223 224 bool neg_x = n2->is_Sub() && inv2_idx == 1; 225 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 226 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 227 if (n1->is_Sub() && inv1_idx == 1) { 228 neg_x = !neg_x; 229 neg_inv2 = !neg_inv2; 230 } 231 Node* inv1_c = phase->get_ctrl(inv1); 232 Node* inv2_c = phase->get_ctrl(inv2); 233 Node* n_inv1; 234 if (neg_inv1) { 235 Node *zero = phase->_igvn.intcon(0); 236 phase->set_ctrl(zero, phase->C->root()); 237 n_inv1 = new SubINode(zero, inv1); 238 phase->register_new_node(n_inv1, inv1_c); 239 } else { 240 n_inv1 = inv1; 241 } 242 Node* inv; 243 if (neg_inv2) { 244 inv = new SubINode(n_inv1, inv2); 245 } else { 246 inv = new AddINode(n_inv1, inv2); 247 } 248 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 249 250 Node* addx; 251 if (neg_x) { 252 addx = new SubINode(inv, x); 253 } else { 254 addx = new AddINode(x, inv); 255 } 256 phase->register_new_node(addx, phase->get_ctrl(x)); 257 phase->_igvn.replace_node(n1, addx); 258 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 259 _body.yank(n1); 260 return addx; 261 } 262 263 //---------------------reassociate_invariants----------------------------- 264 // Reassociate invariant expressions: 265 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 266 for (int i = _body.size() - 1; i >= 0; i--) { 267 Node *n = _body.at(i); 268 for (int j = 0; j < 5; j++) { 269 Node* nn = reassociate_add_sub(n, phase); 270 if (nn == NULL) break; 271 n = nn; // again 272 }; 273 } 274 } 275 276 //------------------------------policy_peeling--------------------------------- 277 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 278 // make some loop-invariant test (usually a null-check) happen before the loop. 279 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 280 Node *test = ((IdealLoopTree*)this)->tail(); 281 int body_size = ((IdealLoopTree*)this)->_body.size(); 282 // Peeling does loop cloning which can result in O(N^2) node construction 283 if( body_size > 255 /* Prevent overflow for large body_size */ 284 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) { 285 return false; // too large to safely clone 286 } 287 288 // check for vectorized loops, any peeling done was already applied 289 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false; 290 291 while( test != _head ) { // Scan till run off top of loop 292 if( test->is_If() ) { // Test? 293 Node *ctrl = phase->get_ctrl(test->in(1)); 294 if (ctrl->is_top()) 295 return false; // Found dead test on live IF? No peeling! 296 // Standard IF only has one input value to check for loop invariance 297 assert(test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd || test->Opcode() == Op_RangeCheck, "Check this code when new subtype is added"); 298 // Condition is not a member of this loop? 299 if( !is_member(phase->get_loop(ctrl)) && 300 is_loop_exit(test) ) 301 return true; // Found reason to peel! 302 } 303 // Walk up dominators to loop _head looking for test which is 304 // executed on every path thru loop. 305 test = phase->idom(test); 306 } 307 return false; 308 } 309 310 //------------------------------peeled_dom_test_elim--------------------------- 311 // If we got the effect of peeling, either by actually peeling or by making 312 // a pre-loop which must execute at least once, we can remove all 313 // loop-invariant dominated tests in the main body. 314 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 315 bool progress = true; 316 while( progress ) { 317 progress = false; // Reset for next iteration 318 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 319 Node *test = prev->in(0); 320 while( test != loop->_head ) { // Scan till run off top of loop 321 322 int p_op = prev->Opcode(); 323 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 324 test->is_If() && // Test? 325 !test->in(1)->is_Con() && // And not already obvious? 326 // Condition is not a member of this loop? 327 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 328 // Walk loop body looking for instances of this test 329 for( uint i = 0; i < loop->_body.size(); i++ ) { 330 Node *n = loop->_body.at(i); 331 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 332 // IfNode was dominated by version in peeled loop body 333 progress = true; 334 dominated_by( old_new[prev->_idx], n ); 335 } 336 } 337 } 338 prev = test; 339 test = idom(test); 340 } // End of scan tests in loop 341 342 } // End of while( progress ) 343 } 344 345 //------------------------------do_peeling------------------------------------- 346 // Peel the first iteration of the given loop. 347 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 348 // The pre-loop illegally has 2 control users (old & new loops). 349 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 350 // Do this by making the old-loop fall-in edges act as if they came 351 // around the loopback from the prior iteration (follow the old-loop 352 // backedges) and then map to the new peeled iteration. This leaves 353 // the pre-loop with only 1 user (the new peeled iteration), but the 354 // peeled-loop backedge has 2 users. 355 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 356 // extra backedge user. 357 // 358 // orig 359 // 360 // stmt1 361 // | 362 // v 363 // loop predicate 364 // | 365 // v 366 // loop<----+ 367 // | | 368 // stmt2 | 369 // | | 370 // v | 371 // if ^ 372 // / \ | 373 // / \ | 374 // v v | 375 // false true | 376 // / \ | 377 // / ----+ 378 // | 379 // v 380 // exit 381 // 382 // 383 // after clone loop 384 // 385 // stmt1 386 // | 387 // v 388 // loop predicate 389 // / \ 390 // clone / \ orig 391 // / \ 392 // / \ 393 // v v 394 // +---->loop clone loop<----+ 395 // | | | | 396 // | stmt2 clone stmt2 | 397 // | | | | 398 // | v v | 399 // ^ if clone If ^ 400 // | / \ / \ | 401 // | / \ / \ | 402 // | v v v v | 403 // | true false false true | 404 // | / \ / \ | 405 // +---- \ / ----+ 406 // \ / 407 // 1v v2 408 // region 409 // | 410 // v 411 // exit 412 // 413 // 414 // after peel and predicate move 415 // 416 // stmt1 417 // / 418 // / 419 // clone / orig 420 // / 421 // / +----------+ 422 // / | | 423 // / loop predicate | 424 // / | | 425 // v v | 426 // TOP-->loop clone loop<----+ | 427 // | | | | 428 // stmt2 clone stmt2 | | 429 // | | | ^ 430 // v v | | 431 // if clone If ^ | 432 // / \ / \ | | 433 // / \ / \ | | 434 // v v v v | | 435 // true false false true | | 436 // | \ / \ | | 437 // | \ / ----+ ^ 438 // | \ / | 439 // | 1v v2 | 440 // v region | 441 // | | | 442 // | v | 443 // | exit | 444 // | | 445 // +--------------->-----------------+ 446 // 447 // 448 // final graph 449 // 450 // stmt1 451 // | 452 // v 453 // stmt2 clone 454 // | 455 // v 456 // if clone 457 // / | 458 // / | 459 // v v 460 // false true 461 // | | 462 // | v 463 // | loop predicate 464 // | | 465 // | v 466 // | loop<----+ 467 // | | | 468 // | stmt2 | 469 // | | | 470 // | v | 471 // v if ^ 472 // | / \ | 473 // | / \ | 474 // | v v | 475 // | false true | 476 // | | \ | 477 // v v --+ 478 // region 479 // | 480 // v 481 // exit 482 // 483 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 484 485 C->set_major_progress(); 486 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 487 // 'pre' loop from the main and the 'pre' can no longer have its 488 // iterations adjusted. Therefore, we need to declare this loop as 489 // no longer a 'main' loop; it will need new pre and post loops before 490 // we can do further RCE. 491 #ifndef PRODUCT 492 if (TraceLoopOpts) { 493 tty->print("Peel "); 494 loop->dump_head(); 495 } 496 #endif 497 Node* head = loop->_head; 498 bool counted_loop = head->is_CountedLoop(); 499 if (counted_loop) { 500 CountedLoopNode *cl = head->as_CountedLoop(); 501 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 502 cl->set_trip_count(cl->trip_count() - 1); 503 if (cl->is_main_loop()) { 504 cl->set_normal_loop(); 505 #ifndef PRODUCT 506 if (PrintOpto && VerifyLoopOptimizations) { 507 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 508 loop->dump_head(); 509 } 510 #endif 511 } 512 } 513 Node* entry = head->in(LoopNode::EntryControl); 514 515 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 516 // The pre-loop illegally has 2 control users (old & new loops). 517 clone_loop( loop, old_new, dom_depth(head) ); 518 519 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 520 // Do this by making the old-loop fall-in edges act as if they came 521 // around the loopback from the prior iteration (follow the old-loop 522 // backedges) and then map to the new peeled iteration. This leaves 523 // the pre-loop with only 1 user (the new peeled iteration), but the 524 // peeled-loop backedge has 2 users. 525 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 526 _igvn.hash_delete(head); 527 head->set_req(LoopNode::EntryControl, new_entry); 528 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 529 Node* old = head->fast_out(j); 530 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 531 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 532 if (!new_exit_value ) // Backedge value is ALSO loop invariant? 533 // Then loop body backedge value remains the same. 534 new_exit_value = old->in(LoopNode::LoopBackControl); 535 _igvn.hash_delete(old); 536 old->set_req(LoopNode::EntryControl, new_exit_value); 537 } 538 } 539 540 541 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 542 // extra backedge user. 543 Node* new_head = old_new[head->_idx]; 544 _igvn.hash_delete(new_head); 545 new_head->set_req(LoopNode::LoopBackControl, C->top()); 546 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 547 Node* use = new_head->fast_out(j2); 548 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 549 _igvn.hash_delete(use); 550 use->set_req(LoopNode::LoopBackControl, C->top()); 551 } 552 } 553 554 555 // Step 4: Correct dom-depth info. Set to loop-head depth. 556 int dd = dom_depth(head); 557 set_idom(head, head->in(1), dd); 558 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 559 Node *old = loop->_body.at(j3); 560 Node *nnn = old_new[old->_idx]; 561 if (!has_ctrl(nnn)) 562 set_idom(nnn, idom(nnn), dd-1); 563 } 564 565 // Now force out all loop-invariant dominating tests. The optimizer 566 // finds some, but we _know_ they are all useless. 567 peeled_dom_test_elim(loop,old_new); 568 569 loop->record_for_igvn(); 570 } 571 572 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 573 574 //------------------------------policy_maximally_unroll------------------------ 575 // Calculate exact loop trip count and return true if loop can be maximally 576 // unrolled. 577 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 578 CountedLoopNode *cl = _head->as_CountedLoop(); 579 assert(cl->is_normal_loop(), ""); 580 if (!cl->is_valid_counted_loop()) 581 return false; // Malformed counted loop 582 583 if (!cl->has_exact_trip_count()) { 584 // Trip count is not exact. 585 return false; 586 } 587 588 uint trip_count = cl->trip_count(); 589 // Note, max_juint is used to indicate unknown trip count. 590 assert(trip_count > 1, "one iteration loop should be optimized out already"); 591 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 592 593 // Real policy: if we maximally unroll, does it get too big? 594 // Allow the unrolled mess to get larger than standard loop 595 // size. After all, it will no longer be a loop. 596 uint body_size = _body.size(); 597 uint unroll_limit = (uint)LoopUnrollLimit * 4; 598 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 599 if (trip_count > unroll_limit || body_size > unroll_limit) { 600 return false; 601 } 602 603 // Fully unroll a loop with few iterations regardless next 604 // conditions since following loop optimizations will split 605 // such loop anyway (pre-main-post). 606 if (trip_count <= 3) 607 return true; 608 609 // Take into account that after unroll conjoined heads and tails will fold, 610 // otherwise policy_unroll() may allow more unrolling than max unrolling. 611 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; 612 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; 613 if (body_size != tst_body_size) // Check for int overflow 614 return false; 615 if (new_body_size > unroll_limit || 616 // Unrolling can result in a large amount of node construction 617 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) { 618 return false; 619 } 620 621 // Do not unroll a loop with String intrinsics code. 622 // String intrinsics are large and have loops. 623 for (uint k = 0; k < _body.size(); k++) { 624 Node* n = _body.at(k); 625 switch (n->Opcode()) { 626 case Op_StrComp: 627 case Op_StrEquals: 628 case Op_StrIndexOf: 629 case Op_StrIndexOfChar: 630 case Op_EncodeISOArray: 631 case Op_AryEq: 632 case Op_HasNegatives: { 633 return false; 634 } 635 #if INCLUDE_RTM_OPT 636 case Op_FastLock: 637 case Op_FastUnlock: { 638 // Don't unroll RTM locking code because it is large. 639 if (UseRTMLocking) { 640 return false; 641 } 642 } 643 #endif 644 } // switch 645 } 646 647 return true; // Do maximally unroll 648 } 649 650 651 //------------------------------policy_unroll---------------------------------- 652 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 653 // the loop is a CountedLoop and the body is small enough. 654 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) { 655 656 CountedLoopNode *cl = _head->as_CountedLoop(); 657 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 658 659 if (!cl->is_valid_counted_loop()) 660 return false; // Malformed counted loop 661 662 // Protect against over-unrolling. 663 // After split at least one iteration will be executed in pre-loop. 664 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; 665 666 _local_loop_unroll_limit = LoopUnrollLimit; 667 _local_loop_unroll_factor = 4; 668 int future_unroll_ct = cl->unrolled_count() * 2; 669 if (!cl->do_unroll_only()) { 670 if (future_unroll_ct > LoopMaxUnroll) return false; 671 } else { 672 // obey user constraints on vector mapped loops with additional unrolling applied 673 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1; 674 if ((future_unroll_ct / unroll_constraint) > LoopMaxUnroll) return false; 675 } 676 677 // Check for initial stride being a small enough constant 678 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; 679 680 // Don't unroll if the next round of unrolling would push us 681 // over the expected trip count of the loop. One is subtracted 682 // from the expected trip count because the pre-loop normally 683 // executes 1 iteration. 684 if (UnrollLimitForProfileCheck > 0 && 685 cl->profile_trip_cnt() != COUNT_UNKNOWN && 686 future_unroll_ct > UnrollLimitForProfileCheck && 687 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 688 return false; 689 } 690 691 // When unroll count is greater than LoopUnrollMin, don't unroll if: 692 // the residual iterations are more than 10% of the trip count 693 // and rounds of "unroll,optimize" are not making significant progress 694 // Progress defined as current size less than 20% larger than previous size. 695 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 696 future_unroll_ct > LoopUnrollMin && 697 (future_unroll_ct - 1) * (100 / LoopPercentProfileLimit) > cl->profile_trip_cnt() && 698 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 699 return false; 700 } 701 702 Node *init_n = cl->init_trip(); 703 Node *limit_n = cl->limit(); 704 int stride_con = cl->stride_con(); 705 // Non-constant bounds. 706 // Protect against over-unrolling when init or/and limit are not constant 707 // (so that trip_count's init value is maxint) but iv range is known. 708 if (init_n == NULL || !init_n->is_Con() || 709 limit_n == NULL || !limit_n->is_Con()) { 710 Node* phi = cl->phi(); 711 if (phi != NULL) { 712 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 713 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 714 int next_stride = stride_con * 2; // stride after this unroll 715 if (next_stride > 0) { 716 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 717 iv_type->_lo + next_stride > iv_type->_hi) { 718 return false; // over-unrolling 719 } 720 } else if (next_stride < 0) { 721 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 722 iv_type->_hi + next_stride < iv_type->_lo) { 723 return false; // over-unrolling 724 } 725 } 726 } 727 } 728 729 // After unroll limit will be adjusted: new_limit = limit-stride. 730 // Bailout if adjustment overflow. 731 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 732 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || 733 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) 734 return false; // overflow 735 736 // Adjust body_size to determine if we unroll or not 737 uint body_size = _body.size(); 738 // Key test to unroll loop in CRC32 java code 739 int xors_in_loop = 0; 740 // Also count ModL, DivL and MulL which expand mightly 741 for (uint k = 0; k < _body.size(); k++) { 742 Node* n = _body.at(k); 743 switch (n->Opcode()) { 744 case Op_XorI: xors_in_loop++; break; // CRC32 java code 745 case Op_ModL: body_size += 30; break; 746 case Op_DivL: body_size += 30; break; 747 case Op_MulL: body_size += 10; break; 748 case Op_StrComp: 749 case Op_StrEquals: 750 case Op_StrIndexOf: 751 case Op_StrIndexOfChar: 752 case Op_EncodeISOArray: 753 case Op_AryEq: 754 case Op_HasNegatives: { 755 // Do not unroll a loop with String intrinsics code. 756 // String intrinsics are large and have loops. 757 return false; 758 } 759 #if INCLUDE_RTM_OPT 760 case Op_FastLock: 761 case Op_FastUnlock: { 762 // Don't unroll RTM locking code because it is large. 763 if (UseRTMLocking) { 764 return false; 765 } 766 } 767 #endif 768 } // switch 769 } 770 771 if (UseSuperWord) { 772 if (!cl->is_reduction_loop()) { 773 phase->mark_reductions(this); 774 } 775 776 // Only attempt slp analysis when user controls do not prohibit it 777 if (LoopMaxUnroll > _local_loop_unroll_factor) { 778 // Once policy_slp_analysis succeeds, mark the loop with the 779 // maximal unroll factor so that we minimize analysis passes 780 if (future_unroll_ct >= _local_loop_unroll_factor) { 781 policy_unroll_slp_analysis(cl, phase, future_unroll_ct); 782 } 783 } 784 } 785 786 int slp_max_unroll_factor = cl->slp_max_unroll(); 787 if (cl->has_passed_slp()) { 788 if (slp_max_unroll_factor >= future_unroll_ct) return true; 789 // Normal case: loop too big 790 return false; 791 } 792 793 // Check for being too big 794 if (body_size > (uint)_local_loop_unroll_limit) { 795 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 796 // Normal case: loop too big 797 return false; 798 } 799 800 if (cl->do_unroll_only()) { 801 if (TraceSuperWordLoopUnrollAnalysis) { 802 tty->print_cr("policy_unroll passed vector loop(vlen=%d,factor = %d)\n", slp_max_unroll_factor, future_unroll_ct); 803 } 804 } 805 806 // Unroll once! (Each trip will soon do double iterations) 807 return true; 808 } 809 810 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct) { 811 // Enable this functionality target by target as needed 812 if (SuperWordLoopUnrollAnalysis) { 813 if (!cl->was_slp_analyzed()) { 814 SuperWord sw(phase); 815 sw.transform_loop(this, false); 816 817 // If the loop is slp canonical analyze it 818 if (sw.early_return() == false) { 819 sw.unrolling_analysis(_local_loop_unroll_factor); 820 } 821 } 822 823 if (cl->has_passed_slp()) { 824 int slp_max_unroll_factor = cl->slp_max_unroll(); 825 if (slp_max_unroll_factor >= future_unroll_ct) { 826 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor; 827 if (new_limit > LoopUnrollLimit) { 828 if (TraceSuperWordLoopUnrollAnalysis) { 829 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit); 830 } 831 _local_loop_unroll_limit = new_limit; 832 } 833 } 834 } 835 } 836 } 837 838 //------------------------------policy_align----------------------------------- 839 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 840 // expression that does the alignment. Note that only one array base can be 841 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 842 // if we vectorize short memory ops into longer memory ops, we may want to 843 // increase alignment. 844 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 845 return false; 846 } 847 848 //------------------------------policy_range_check----------------------------- 849 // Return TRUE or FALSE if the loop should be range-check-eliminated. 850 // Actually we do iteration-splitting, a more powerful form of RCE. 851 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 852 if (!RangeCheckElimination) return false; 853 854 CountedLoopNode *cl = _head->as_CountedLoop(); 855 // If we unrolled with no intention of doing RCE and we later 856 // changed our minds, we got no pre-loop. Either we need to 857 // make a new pre-loop, or we gotta disallow RCE. 858 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 859 Node *trip_counter = cl->phi(); 860 861 // check for vectorized loops, some opts are no longer needed 862 if (cl->do_unroll_only()) return false; 863 864 // Check loop body for tests of trip-counter plus loop-invariant vs 865 // loop-invariant. 866 for (uint i = 0; i < _body.size(); i++) { 867 Node *iff = _body[i]; 868 if (iff->Opcode() == Op_If || 869 iff->Opcode() == Op_RangeCheck) { // Test? 870 871 // Comparing trip+off vs limit 872 Node *bol = iff->in(1); 873 if (bol->req() != 2) continue; // dead constant test 874 if (!bol->is_Bool()) { 875 assert(bol->Opcode() == Op_Conv2B, "predicate check only"); 876 continue; 877 } 878 if (bol->as_Bool()->_test._test == BoolTest::ne) 879 continue; // not RC 880 881 Node *cmp = bol->in(1); 882 Node *rc_exp = cmp->in(1); 883 Node *limit = cmp->in(2); 884 885 Node *limit_c = phase->get_ctrl(limit); 886 if( limit_c == phase->C->top() ) 887 return false; // Found dead test on live IF? No RCE! 888 if( is_member(phase->get_loop(limit_c) ) ) { 889 // Compare might have operands swapped; commute them 890 rc_exp = cmp->in(2); 891 limit = cmp->in(1); 892 limit_c = phase->get_ctrl(limit); 893 if( is_member(phase->get_loop(limit_c) ) ) 894 continue; // Both inputs are loop varying; cannot RCE 895 } 896 897 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 898 continue; 899 } 900 // Yeah! Found a test like 'trip+off vs limit' 901 // Test is an IfNode, has 2 projections. If BOTH are in the loop 902 // we need loop unswitching instead of iteration splitting. 903 if( is_loop_exit(iff) ) 904 return true; // Found reason to split iterations 905 } // End of is IF 906 } 907 908 return false; 909 } 910 911 //------------------------------policy_peel_only------------------------------- 912 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 913 // for unrolling loops with NO array accesses. 914 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 915 // check for vectorized loops, any peeling done was already applied 916 if (_head->is_CountedLoop() && _head->as_CountedLoop()->do_unroll_only()) return false; 917 918 for( uint i = 0; i < _body.size(); i++ ) 919 if( _body[i]->is_Mem() ) 920 return false; 921 922 // No memory accesses at all! 923 return true; 924 } 925 926 //------------------------------clone_up_backedge_goo-------------------------- 927 // If Node n lives in the back_ctrl block and cannot float, we clone a private 928 // version of n in preheader_ctrl block and return that, otherwise return n. 929 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { 930 if( get_ctrl(n) != back_ctrl ) return n; 931 932 // Only visit once 933 if (visited.test_set(n->_idx)) { 934 Node *x = clones.find(n->_idx); 935 if (x != NULL) 936 return x; 937 return n; 938 } 939 940 Node *x = NULL; // If required, a clone of 'n' 941 // Check for 'n' being pinned in the backedge. 942 if( n->in(0) && n->in(0) == back_ctrl ) { 943 assert(clones.find(n->_idx) == NULL, "dead loop"); 944 x = n->clone(); // Clone a copy of 'n' to preheader 945 clones.push(x, n->_idx); 946 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 947 } 948 949 // Recursive fixup any other input edges into x. 950 // If there are no changes we can just return 'n', otherwise 951 // we need to clone a private copy and change it. 952 for( uint i = 1; i < n->req(); i++ ) { 953 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); 954 if( g != n->in(i) ) { 955 if( !x ) { 956 assert(clones.find(n->_idx) == NULL, "dead loop"); 957 x = n->clone(); 958 clones.push(x, n->_idx); 959 } 960 x->set_req(i, g); 961 } 962 } 963 if( x ) { // x can legally float to pre-header location 964 register_new_node( x, preheader_ctrl ); 965 return x; 966 } else { // raise n to cover LCA of uses 967 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 968 } 969 return n; 970 } 971 972 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { 973 Node* castii = new CastIINode(incr, TypeInt::INT, true); 974 castii->set_req(0, ctrl); 975 register_new_node(castii, ctrl); 976 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { 977 Node* n = incr->fast_out(i); 978 if (n->is_Phi() && n->in(0) == loop) { 979 int nrep = n->replace_edge(incr, castii); 980 return true; 981 } 982 } 983 return false; 984 } 985 986 //------------------------------insert_pre_post_loops-------------------------- 987 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 988 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 989 // alignment. Useful to unroll loops that do no array accesses. 990 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 991 992 #ifndef PRODUCT 993 if (TraceLoopOpts) { 994 if (peel_only) 995 tty->print("PeelMainPost "); 996 else 997 tty->print("PreMainPost "); 998 loop->dump_head(); 999 } 1000 #endif 1001 C->set_major_progress(); 1002 1003 // Find common pieces of the loop being guarded with pre & post loops 1004 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1005 assert( main_head->is_normal_loop(), "" ); 1006 CountedLoopEndNode *main_end = main_head->loopexit(); 1007 guarantee(main_end != NULL, "no loop exit node"); 1008 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 1009 uint dd_main_head = dom_depth(main_head); 1010 uint max = main_head->outcnt(); 1011 1012 Node *pre_header= main_head->in(LoopNode::EntryControl); 1013 Node *init = main_head->init_trip(); 1014 Node *incr = main_end ->incr(); 1015 Node *limit = main_end ->limit(); 1016 Node *stride = main_end ->stride(); 1017 Node *cmp = main_end ->cmp_node(); 1018 BoolTest::mask b_test = main_end->test_trip(); 1019 1020 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 1021 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 1022 if( bol->outcnt() != 1 ) { 1023 bol = bol->clone(); 1024 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 1025 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol); 1026 } 1027 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 1028 if( cmp->outcnt() != 1 ) { 1029 cmp = cmp->clone(); 1030 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 1031 _igvn.replace_input_of(bol, 1, cmp); 1032 } 1033 1034 // Add the post loop 1035 CountedLoopNode *post_head = NULL; 1036 Node *main_exit = insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1037 1038 //------------------------------ 1039 // Step B: Create Pre-Loop. 1040 1041 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 1042 // loop pre-header illegally has 2 control users (old & new loops). 1043 clone_loop( loop, old_new, dd_main_head ); 1044 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1045 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1046 pre_head->set_pre_loop(main_head); 1047 Node *pre_incr = old_new[incr->_idx]; 1048 1049 // Reduce the pre-loop trip count. 1050 pre_end->_prob = PROB_FAIR; 1051 1052 // Find the pre-loop normal exit. 1053 Node* pre_exit = pre_end->proj_out(false); 1054 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 1055 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end); 1056 _igvn.register_new_node_with_optimizer( new_pre_exit ); 1057 set_idom(new_pre_exit, pre_end, dd_main_head); 1058 set_loop(new_pre_exit, loop->_parent); 1059 1060 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1061 // pre-loop, the main-loop may not execute at all. Later in life this 1062 // zero-trip guard will become the minimum-trip guard when we unroll 1063 // the main-loop. 1064 Node *min_opaq = new Opaque1Node(C, limit); 1065 Node *min_cmp = new CmpINode( pre_incr, min_opaq ); 1066 Node *min_bol = new BoolNode( min_cmp, b_test ); 1067 register_new_node( min_opaq, new_pre_exit ); 1068 register_new_node( min_cmp , new_pre_exit ); 1069 register_new_node( min_bol , new_pre_exit ); 1070 1071 // Build the IfNode (assume the main-loop is executed always). 1072 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1073 _igvn.register_new_node_with_optimizer( min_iff ); 1074 set_idom(min_iff, new_pre_exit, dd_main_head); 1075 set_loop(min_iff, loop->_parent); 1076 1077 // Plug in the false-path, taken if we need to skip main-loop 1078 _igvn.hash_delete( pre_exit ); 1079 pre_exit->set_req(0, min_iff); 1080 set_idom(pre_exit, min_iff, dd_main_head); 1081 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1082 // Make the true-path, must enter the main loop 1083 Node *min_taken = new IfTrueNode( min_iff ); 1084 _igvn.register_new_node_with_optimizer( min_taken ); 1085 set_idom(min_taken, min_iff, dd_main_head); 1086 set_loop(min_taken, loop->_parent); 1087 // Plug in the true path 1088 _igvn.hash_delete( main_head ); 1089 main_head->set_req(LoopNode::EntryControl, min_taken); 1090 set_idom(main_head, min_taken, dd_main_head); 1091 1092 Arena *a = Thread::current()->resource_area(); 1093 VectorSet visited(a); 1094 Node_Stack clones(a, main_head->back_control()->outcnt()); 1095 // Step B3: Make the fall-in values to the main-loop come from the 1096 // fall-out values of the pre-loop. 1097 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1098 Node* main_phi = main_head->fast_out(i2); 1099 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1100 Node *pre_phi = old_new[main_phi->_idx]; 1101 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1102 main_head->init_control(), 1103 pre_phi->in(LoopNode::LoopBackControl), 1104 visited, clones); 1105 _igvn.hash_delete(main_phi); 1106 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1107 } 1108 } 1109 1110 // Nodes inside the loop may be control dependent on a predicate 1111 // that was moved before the preloop. If the back branch of the main 1112 // or post loops becomes dead, those nodes won't be dependent on the 1113 // test that guards that loop nest anymore which could lead to an 1114 // incorrect array access because it executes independently of the 1115 // test that was guarding the loop nest. We add a special CastII on 1116 // the if branch that enters the loop, between the input induction 1117 // variable value and the induction variable Phi to preserve correct 1118 // dependencies. 1119 1120 // CastII for the main loop: 1121 bool inserted = cast_incr_before_loop( pre_incr, min_taken, main_head ); 1122 assert(inserted, "no castII inserted"); 1123 1124 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1125 // RCE and alignment may change this later. 1126 Node *cmp_end = pre_end->cmp_node(); 1127 assert( cmp_end->in(2) == limit, "" ); 1128 Node *pre_limit = new AddINode( init, stride ); 1129 1130 // Save the original loop limit in this Opaque1 node for 1131 // use by range check elimination. 1132 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit); 1133 1134 register_new_node( pre_limit, pre_head->in(0) ); 1135 register_new_node( pre_opaq , pre_head->in(0) ); 1136 1137 // Since no other users of pre-loop compare, I can hack limit directly 1138 assert( cmp_end->outcnt() == 1, "no other users" ); 1139 _igvn.hash_delete(cmp_end); 1140 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1141 1142 // Special case for not-equal loop bounds: 1143 // Change pre loop test, main loop test, and the 1144 // main loop guard test to use lt or gt depending on stride 1145 // direction: 1146 // positive stride use < 1147 // negative stride use > 1148 // 1149 // not-equal test is kept for post loop to handle case 1150 // when init > limit when stride > 0 (and reverse). 1151 1152 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1153 1154 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1155 // Modify pre loop end condition 1156 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1157 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test); 1158 register_new_node( new_bol0, pre_head->in(0) ); 1159 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0); 1160 // Modify main loop guard condition 1161 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1162 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test); 1163 register_new_node( new_bol1, new_pre_exit ); 1164 _igvn.hash_delete(min_iff); 1165 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1166 // Modify main loop end condition 1167 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1168 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test); 1169 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1170 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2); 1171 } 1172 1173 // Flag main loop 1174 main_head->set_main_loop(); 1175 if( peel_only ) main_head->set_main_no_pre_loop(); 1176 1177 // Subtract a trip count for the pre-loop. 1178 main_head->set_trip_count(main_head->trip_count() - 1); 1179 1180 // It's difficult to be precise about the trip-counts 1181 // for the pre/post loops. They are usually very short, 1182 // so guess that 4 trips is a reasonable value. 1183 post_head->set_profile_trip_cnt(4.0); 1184 pre_head->set_profile_trip_cnt(4.0); 1185 1186 // Now force out all loop-invariant dominating tests. The optimizer 1187 // finds some, but we _know_ they are all useless. 1188 peeled_dom_test_elim(loop,old_new); 1189 loop->record_for_igvn(); 1190 } 1191 1192 //------------------------------insert_vector_post_loop------------------------ 1193 // Insert a copy of the atomic unrolled vectorized main loop as a post loop, 1194 // unroll_policy has already informed us that more unrolling is about to happen to 1195 // the main loop. The resultant post loop will serve as a vectorized drain loop. 1196 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) { 1197 if (!loop->_head->is_CountedLoop()) return; 1198 1199 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1200 1201 // only process vectorized main loops 1202 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return; 1203 1204 int slp_max_unroll_factor = cl->slp_max_unroll(); 1205 int cur_unroll = cl->unrolled_count(); 1206 1207 if (slp_max_unroll_factor == 0) return; 1208 1209 // only process atomic unroll vector loops (not super unrolled after vectorization) 1210 if (cur_unroll != slp_max_unroll_factor) return; 1211 1212 // we only ever process this one time 1213 if (cl->has_atomic_post_loop()) return; 1214 1215 #ifndef PRODUCT 1216 if (TraceLoopOpts) { 1217 tty->print("PostVector "); 1218 loop->dump_head(); 1219 } 1220 #endif 1221 C->set_major_progress(); 1222 1223 // Find common pieces of the loop being guarded with pre & post loops 1224 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1225 CountedLoopEndNode *main_end = main_head->loopexit(); 1226 guarantee(main_end != NULL, "no loop exit node"); 1227 // diagnostic to show loop end is not properly formed 1228 assert(main_end->outcnt() == 2, "1 true, 1 false path only"); 1229 1230 // mark this loop as processed 1231 main_head->mark_has_atomic_post_loop(); 1232 1233 Node *incr = main_end->incr(); 1234 Node *limit = main_end->limit(); 1235 1236 // In this case we throw away the result as we are not using it to connect anything else. 1237 CountedLoopNode *post_head = NULL; 1238 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1239 1240 // It's difficult to be precise about the trip-counts 1241 // for post loops. They are usually very short, 1242 // so guess that unit vector trips is a reasonable value. 1243 post_head->set_profile_trip_cnt(cur_unroll); 1244 1245 // Now force out all loop-invariant dominating tests. The optimizer 1246 // finds some, but we _know_ they are all useless. 1247 peeled_dom_test_elim(loop, old_new); 1248 loop->record_for_igvn(); 1249 } 1250 1251 1252 //-------------------------insert_scalar_rced_post_loop------------------------ 1253 // Insert a copy of the rce'd main loop as a post loop, 1254 // We have not unrolled the main loop, so this is the right time to inject this. 1255 // Later we will examine the partner of this post loop pair which still has range checks 1256 // to see inject code which tests at runtime if the range checks are applicable. 1257 void PhaseIdealLoop::insert_scalar_rced_post_loop(IdealLoopTree *loop, Node_List &old_new) { 1258 if (!loop->_head->is_CountedLoop()) return; 1259 1260 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1261 1262 // only process RCE'd main loops 1263 if (!cl->is_main_loop() || cl->range_checks_present()) return; 1264 1265 #ifndef PRODUCT 1266 if (TraceLoopOpts) { 1267 tty->print("PostScalarRce "); 1268 loop->dump_head(); 1269 } 1270 #endif 1271 C->set_major_progress(); 1272 1273 // Find common pieces of the loop being guarded with pre & post loops 1274 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1275 CountedLoopEndNode *main_end = main_head->loopexit(); 1276 guarantee(main_end != NULL, "no loop exit node"); 1277 // diagnostic to show loop end is not properly formed 1278 assert(main_end->outcnt() == 2, "1 true, 1 false path only"); 1279 1280 Node *incr = main_end->incr(); 1281 Node *limit = main_end->limit(); 1282 1283 // In this case we throw away the result as we are not using it to connect anything else. 1284 CountedLoopNode *post_head = NULL; 1285 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1286 1287 // It's difficult to be precise about the trip-counts 1288 // for post loops. They are usually very short, 1289 // so guess that unit vector trips is a reasonable value. 1290 post_head->set_profile_trip_cnt(4.0); 1291 post_head->set_is_rce_post_loop(); 1292 1293 // Now force out all loop-invariant dominating tests. The optimizer 1294 // finds some, but we _know_ they are all useless. 1295 peeled_dom_test_elim(loop, old_new); 1296 loop->record_for_igvn(); 1297 } 1298 1299 1300 //------------------------------insert_post_loop------------------------------- 1301 // Insert post loops. Add a post loop to the given loop passed. 1302 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree *loop, Node_List &old_new, 1303 CountedLoopNode *main_head, CountedLoopEndNode *main_end, 1304 Node *incr, Node *limit, CountedLoopNode *&post_head) { 1305 1306 //------------------------------ 1307 // Step A: Create a new post-Loop. 1308 Node* main_exit = main_end->proj_out(false); 1309 assert(main_exit->Opcode() == Op_IfFalse, ""); 1310 int dd_main_exit = dom_depth(main_exit); 1311 1312 // Step A1: Clone the loop body of main. The clone becomes the post-loop. 1313 // The main loop pre-header illegally has 2 control users (old & new loops). 1314 clone_loop(loop, old_new, dd_main_exit); 1315 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, ""); 1316 post_head = old_new[main_head->_idx]->as_CountedLoop(); 1317 post_head->set_normal_loop(); 1318 post_head->set_post_loop(main_head); 1319 1320 // Reduce the post-loop trip count. 1321 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd(); 1322 post_end->_prob = PROB_FAIR; 1323 1324 // Build the main-loop normal exit. 1325 IfFalseNode *new_main_exit = new IfFalseNode(main_end); 1326 _igvn.register_new_node_with_optimizer(new_main_exit); 1327 set_idom(new_main_exit, main_end, dd_main_exit); 1328 set_loop(new_main_exit, loop->_parent); 1329 1330 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 1331 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 1332 // (the previous loop trip-counter exit value) because we will be changing 1333 // the exit value (via additional unrolling) so we cannot constant-fold away the zero 1334 // trip guard until all unrolling is done. 1335 Node *zer_opaq = new Opaque1Node(C, incr); 1336 Node *zer_cmp = new CmpINode(zer_opaq, limit); 1337 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip()); 1338 register_new_node(zer_opaq, new_main_exit); 1339 register_new_node(zer_cmp, new_main_exit); 1340 register_new_node(zer_bol, new_main_exit); 1341 1342 // Build the IfNode 1343 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN); 1344 _igvn.register_new_node_with_optimizer(zer_iff); 1345 set_idom(zer_iff, new_main_exit, dd_main_exit); 1346 set_loop(zer_iff, loop->_parent); 1347 1348 // Plug in the false-path, taken if we need to skip this post-loop 1349 _igvn.replace_input_of(main_exit, 0, zer_iff); 1350 set_idom(main_exit, zer_iff, dd_main_exit); 1351 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 1352 // Make the true-path, must enter this post loop 1353 Node *zer_taken = new IfTrueNode(zer_iff); 1354 _igvn.register_new_node_with_optimizer(zer_taken); 1355 set_idom(zer_taken, zer_iff, dd_main_exit); 1356 set_loop(zer_taken, loop->_parent); 1357 // Plug in the true path 1358 _igvn.hash_delete(post_head); 1359 post_head->set_req(LoopNode::EntryControl, zer_taken); 1360 set_idom(post_head, zer_taken, dd_main_exit); 1361 1362 Arena *a = Thread::current()->resource_area(); 1363 VectorSet visited(a); 1364 Node_Stack clones(a, main_head->back_control()->outcnt()); 1365 // Step A3: Make the fall-in values to the post-loop come from the 1366 // fall-out values of the main-loop. 1367 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 1368 Node* main_phi = main_head->fast_out(i); 1369 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0) { 1370 Node *cur_phi = old_new[main_phi->_idx]; 1371 Node *fallnew = clone_up_backedge_goo(main_head->back_control(), 1372 post_head->init_control(), 1373 main_phi->in(LoopNode::LoopBackControl), 1374 visited, clones); 1375 _igvn.hash_delete(cur_phi); 1376 cur_phi->set_req(LoopNode::EntryControl, fallnew); 1377 } 1378 } 1379 1380 // CastII for the new post loop: 1381 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); 1382 assert(inserted, "no castII inserted"); 1383 1384 return new_main_exit; 1385 } 1386 1387 //------------------------------is_invariant----------------------------- 1388 // Return true if n is invariant 1389 bool IdealLoopTree::is_invariant(Node* n) const { 1390 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1391 if (n_c->is_top()) return false; 1392 return !is_member(_phase->get_loop(n_c)); 1393 } 1394 1395 1396 //------------------------------do_unroll-------------------------------------- 1397 // Unroll the loop body one step - make each trip do 2 iterations. 1398 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1399 assert(LoopUnrollLimit, ""); 1400 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1401 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1402 assert(loop_end, ""); 1403 #ifndef PRODUCT 1404 if (PrintOpto && VerifyLoopOptimizations) { 1405 tty->print("Unrolling "); 1406 loop->dump_head(); 1407 } else if (TraceLoopOpts) { 1408 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1409 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1410 } else { 1411 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1412 } 1413 loop->dump_head(); 1414 } 1415 1416 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1417 Arena* arena = Thread::current()->resource_area(); 1418 Node_Stack stack(arena, C->live_nodes() >> 2); 1419 Node_List rpo_list; 1420 VectorSet visited(arena); 1421 visited.set(loop_head->_idx); 1422 rpo( loop_head, stack, visited, rpo_list ); 1423 dump(loop, rpo_list.size(), rpo_list ); 1424 } 1425 #endif 1426 1427 // Remember loop node count before unrolling to detect 1428 // if rounds of unroll,optimize are making progress 1429 loop_head->set_node_count_before_unroll(loop->_body.size()); 1430 1431 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1432 Node *limit = loop_head->limit(); 1433 Node *init = loop_head->init_trip(); 1434 Node *stride = loop_head->stride(); 1435 1436 Node *opaq = NULL; 1437 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1438 // Search for zero-trip guard. 1439 1440 // Check the shape of the graph at the loop entry. If an inappropriate 1441 // graph shape is encountered, the compiler bails out loop unrolling; 1442 // compilation of the method will still succeed. 1443 if (!is_canonical_loop_entry(loop_head)) { 1444 return; 1445 } 1446 opaq = ctrl->in(0)->in(1)->in(1)->in(2); 1447 // Zero-trip test uses an 'opaque' node which is not shared. 1448 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1449 } 1450 1451 C->set_major_progress(); 1452 1453 Node* new_limit = NULL; 1454 int stride_con = stride->get_int(); 1455 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1456 uint old_trip_count = loop_head->trip_count(); 1457 // Verify that unroll policy result is still valid. 1458 assert(old_trip_count > 1 && 1459 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1460 1461 // Adjust loop limit to keep valid iterations number after unroll. 1462 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1463 // which may overflow. 1464 if (!adjust_min_trip) { 1465 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1466 "odd trip count for maximally unroll"); 1467 // Don't need to adjust limit for maximally unroll since trip count is even. 1468 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1469 // Loop's limit is constant. Loop's init could be constant when pre-loop 1470 // become peeled iteration. 1471 jlong init_con = init->get_int(); 1472 // We can keep old loop limit if iterations count stays the same: 1473 // old_trip_count == new_trip_count * 2 1474 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1475 // so we also don't need to adjust zero trip test. 1476 jlong limit_con = limit->get_int(); 1477 // (stride_con*2) not overflow since stride_con <= 8. 1478 int new_stride_con = stride_con * 2; 1479 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1480 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1481 // New trip count should satisfy next conditions. 1482 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1483 uint new_trip_count = (uint)trip_count; 1484 adjust_min_trip = (old_trip_count != new_trip_count*2); 1485 } 1486 1487 if (adjust_min_trip) { 1488 // Step 2: Adjust the trip limit if it is called for. 1489 // The adjustment amount is -stride. Need to make sure if the 1490 // adjustment underflows or overflows, then the main loop is skipped. 1491 Node* cmp = loop_end->cmp_node(); 1492 assert(cmp->in(2) == limit, "sanity"); 1493 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1494 1495 // Verify that policy_unroll result is still valid. 1496 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1497 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1498 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); 1499 1500 if (limit->is_Con()) { 1501 // The check in policy_unroll and the assert above guarantee 1502 // no underflow if limit is constant. 1503 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1504 set_ctrl(new_limit, C->root()); 1505 } else { 1506 // Limit is not constant. 1507 if (loop_head->unrolled_count() == 1) { // only for first unroll 1508 // Separate limit by Opaque node in case it is an incremented 1509 // variable from previous loop to avoid using pre-incremented 1510 // value which could increase register pressure. 1511 // Otherwise reorg_offsets() optimization will create a separate 1512 // Opaque node for each use of trip-counter and as result 1513 // zero trip guard limit will be different from loop limit. 1514 assert(has_ctrl(opaq), "should have it"); 1515 Node* opaq_ctrl = get_ctrl(opaq); 1516 limit = new Opaque2Node( C, limit ); 1517 register_new_node( limit, opaq_ctrl ); 1518 } 1519 if (stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo) || 1520 stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi)) { 1521 // No underflow. 1522 new_limit = new SubINode(limit, stride); 1523 } else { 1524 // (limit - stride) may underflow. 1525 // Clamp the adjustment value with MININT or MAXINT: 1526 // 1527 // new_limit = limit-stride 1528 // if (stride > 0) 1529 // new_limit = (limit < new_limit) ? MININT : new_limit; 1530 // else 1531 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1532 // 1533 BoolTest::mask bt = loop_end->test_trip(); 1534 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1535 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1536 set_ctrl(adj_max, C->root()); 1537 Node* old_limit = NULL; 1538 Node* adj_limit = NULL; 1539 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1540 if (loop_head->unrolled_count() > 1 && 1541 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1542 limit->in(CMoveNode::IfTrue) == adj_max && 1543 bol->as_Bool()->_test._test == bt && 1544 bol->in(1)->Opcode() == Op_CmpI && 1545 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1546 // Loop was unrolled before. 1547 // Optimize the limit to avoid nested CMove: 1548 // use original limit as old limit. 1549 old_limit = bol->in(1)->in(1); 1550 // Adjust previous adjusted limit. 1551 adj_limit = limit->in(CMoveNode::IfFalse); 1552 adj_limit = new SubINode(adj_limit, stride); 1553 } else { 1554 old_limit = limit; 1555 adj_limit = new SubINode(limit, stride); 1556 } 1557 assert(old_limit != NULL && adj_limit != NULL, ""); 1558 register_new_node( adj_limit, ctrl ); // adjust amount 1559 Node* adj_cmp = new CmpINode(old_limit, adj_limit); 1560 register_new_node( adj_cmp, ctrl ); 1561 Node* adj_bool = new BoolNode(adj_cmp, bt); 1562 register_new_node( adj_bool, ctrl ); 1563 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1564 } 1565 register_new_node(new_limit, ctrl); 1566 } 1567 assert(new_limit != NULL, ""); 1568 // Replace in loop test. 1569 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1570 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1571 // Don't need to create new test since only one user. 1572 _igvn.hash_delete(cmp); 1573 cmp->set_req(2, new_limit); 1574 } else { 1575 // Create new test since it is shared. 1576 Node* ctrl2 = loop_end->in(0); 1577 Node* cmp2 = cmp->clone(); 1578 cmp2->set_req(2, new_limit); 1579 register_new_node(cmp2, ctrl2); 1580 Node* bol2 = loop_end->in(1)->clone(); 1581 bol2->set_req(1, cmp2); 1582 register_new_node(bol2, ctrl2); 1583 _igvn.replace_input_of(loop_end, 1, bol2); 1584 } 1585 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1586 // Make it a 1-trip test (means at least 2 trips). 1587 1588 // Guard test uses an 'opaque' node which is not shared. Hence I 1589 // can edit it's inputs directly. Hammer in the new limit for the 1590 // minimum-trip guard. 1591 assert(opaq->outcnt() == 1, ""); 1592 _igvn.replace_input_of(opaq, 1, new_limit); 1593 } 1594 1595 // Adjust max trip count. The trip count is intentionally rounded 1596 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1597 // the main, unrolled, part of the loop will never execute as it is protected 1598 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1599 // and later determined that part of the unrolled loop was dead. 1600 loop_head->set_trip_count(old_trip_count / 2); 1601 1602 // Double the count of original iterations in the unrolled loop body. 1603 loop_head->double_unrolled_count(); 1604 1605 // --------- 1606 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1607 // represents the odd iterations; since the loop trips an even number of 1608 // times its backedge is never taken. Kill the backedge. 1609 uint dd = dom_depth(loop_head); 1610 clone_loop( loop, old_new, dd ); 1611 1612 // Make backedges of the clone equal to backedges of the original. 1613 // Make the fall-in from the original come from the fall-out of the clone. 1614 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1615 Node* phi = loop_head->fast_out(j); 1616 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1617 Node *newphi = old_new[phi->_idx]; 1618 _igvn.hash_delete( phi ); 1619 _igvn.hash_delete( newphi ); 1620 1621 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1622 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1623 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1624 } 1625 } 1626 Node *clone_head = old_new[loop_head->_idx]; 1627 _igvn.hash_delete( clone_head ); 1628 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1629 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1630 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1631 loop->_head = clone_head; // New loop header 1632 1633 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1634 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1635 1636 // Kill the clone's backedge 1637 Node *newcle = old_new[loop_end->_idx]; 1638 _igvn.hash_delete( newcle ); 1639 Node *one = _igvn.intcon(1); 1640 set_ctrl(one, C->root()); 1641 newcle->set_req(1, one); 1642 // Force clone into same loop body 1643 uint max = loop->_body.size(); 1644 for( uint k = 0; k < max; k++ ) { 1645 Node *old = loop->_body.at(k); 1646 Node *nnn = old_new[old->_idx]; 1647 loop->_body.push(nnn); 1648 if (!has_ctrl(old)) 1649 set_loop(nnn, loop); 1650 } 1651 1652 loop->record_for_igvn(); 1653 1654 #ifndef PRODUCT 1655 if (C->do_vector_loop() && (PrintOpto && VerifyLoopOptimizations || TraceLoopOpts)) { 1656 tty->print("\nnew loop after unroll\n"); loop->dump_head(); 1657 for (uint i = 0; i < loop->_body.size(); i++) { 1658 loop->_body.at(i)->dump(); 1659 } 1660 if(C->clone_map().is_debug()) { 1661 tty->print("\nCloneMap\n"); 1662 Dict* dict = C->clone_map().dict(); 1663 DictI i(dict); 1664 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size()); 1665 for (int ii = 0; i.test(); ++i, ++ii) { 1666 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key)); 1667 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen()); 1668 if (ii % 10 == 9) { 1669 tty->print_cr(" "); 1670 } 1671 } 1672 tty->print_cr(" "); 1673 } 1674 } 1675 #endif 1676 1677 } 1678 1679 //------------------------------do_maximally_unroll---------------------------- 1680 1681 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1682 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1683 assert(cl->has_exact_trip_count(), "trip count is not exact"); 1684 assert(cl->trip_count() > 0, ""); 1685 #ifndef PRODUCT 1686 if (TraceLoopOpts) { 1687 tty->print("MaxUnroll %d ", cl->trip_count()); 1688 loop->dump_head(); 1689 } 1690 #endif 1691 1692 // If loop is tripping an odd number of times, peel odd iteration 1693 if ((cl->trip_count() & 1) == 1) { 1694 do_peeling(loop, old_new); 1695 } 1696 1697 // Now its tripping an even number of times remaining. Double loop body. 1698 // Do not adjust pre-guards; they are not needed and do not exist. 1699 if (cl->trip_count() > 0) { 1700 assert((cl->trip_count() & 1) == 0, "missed peeling"); 1701 do_unroll(loop, old_new, false); 1702 } 1703 } 1704 1705 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) { 1706 if (SuperWordReductions == false) return; 1707 1708 CountedLoopNode* loop_head = loop->_head->as_CountedLoop(); 1709 if (loop_head->unrolled_count() > 1) { 1710 return; 1711 } 1712 1713 Node* trip_phi = loop_head->phi(); 1714 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) { 1715 Node* phi = loop_head->fast_out(i); 1716 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) { 1717 // For definitions which are loop inclusive and not tripcounts. 1718 Node* def_node = phi->in(LoopNode::LoopBackControl); 1719 1720 if (def_node != NULL) { 1721 Node* n_ctrl = get_ctrl(def_node); 1722 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) { 1723 // Now test it to see if it fits the standard pattern for a reduction operator. 1724 int opc = def_node->Opcode(); 1725 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) { 1726 if (!def_node->is_reduction()) { // Not marked yet 1727 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it 1728 bool ok = false; 1729 for (unsigned j = 1; j < def_node->req(); j++) { 1730 Node* in = def_node->in(j); 1731 if (in == phi) { 1732 ok = true; 1733 break; 1734 } 1735 } 1736 1737 // do nothing if we did not match the initial criteria 1738 if (ok == false) { 1739 continue; 1740 } 1741 1742 // The result of the reduction must not be used in the loop 1743 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) { 1744 Node* u = def_node->fast_out(i); 1745 if (!loop->is_member(get_loop(ctrl_or_self(u)))) { 1746 continue; 1747 } 1748 if (u == phi) { 1749 continue; 1750 } 1751 ok = false; 1752 } 1753 1754 // iff the uses conform 1755 if (ok) { 1756 def_node->add_flag(Node::Flag_is_reduction); 1757 loop_head->mark_has_reductions(); 1758 } 1759 } 1760 } 1761 } 1762 } 1763 } 1764 } 1765 } 1766 1767 //------------------------------adjust_limit----------------------------------- 1768 // Helper function for add_constraint(). 1769 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { 1770 // Compute "I :: (limit-offset)/scale" 1771 Node *con = new SubINode(rc_limit, offset); 1772 register_new_node(con, pre_ctrl); 1773 Node *X = new DivINode(0, con, scale); 1774 register_new_node(X, pre_ctrl); 1775 1776 // Adjust loop limit 1777 loop_limit = (stride_con > 0) 1778 ? (Node*)(new MinINode(loop_limit, X)) 1779 : (Node*)(new MaxINode(loop_limit, X)); 1780 register_new_node(loop_limit, pre_ctrl); 1781 return loop_limit; 1782 } 1783 1784 //------------------------------add_constraint--------------------------------- 1785 // Constrain the main loop iterations so the conditions: 1786 // low_limit <= scale_con * I + offset < upper_limit 1787 // always holds true. That is, either increase the number of iterations in 1788 // the pre-loop or the post-loop until the condition holds true in the main 1789 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1790 // stride and scale are constants (offset and limit often are). 1791 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 ) { 1792 // For positive stride, the pre-loop limit always uses a MAX function 1793 // and the main loop a MIN function. For negative stride these are 1794 // reversed. 1795 1796 // Also for positive stride*scale the affine function is increasing, so the 1797 // pre-loop must check for underflow and the post-loop for overflow. 1798 // Negative stride*scale reverses this; pre-loop checks for overflow and 1799 // post-loop for underflow. 1800 1801 Node *scale = _igvn.intcon(scale_con); 1802 set_ctrl(scale, C->root()); 1803 1804 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 1805 // The overflow limit: scale*I+offset < upper_limit 1806 // For main-loop compute 1807 // ( if (scale > 0) /* and stride > 0 */ 1808 // I < (upper_limit-offset)/scale 1809 // else /* scale < 0 and stride < 0 */ 1810 // I > (upper_limit-offset)/scale 1811 // ) 1812 // 1813 // (upper_limit-offset) may overflow or underflow. 1814 // But it is fine since main loop will either have 1815 // less iterations or will be skipped in such case. 1816 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); 1817 1818 // The underflow limit: low_limit <= scale*I+offset. 1819 // For pre-loop compute 1820 // NOT(scale*I+offset >= low_limit) 1821 // scale*I+offset < low_limit 1822 // ( if (scale > 0) /* and stride > 0 */ 1823 // I < (low_limit-offset)/scale 1824 // else /* scale < 0 and stride < 0 */ 1825 // I > (low_limit-offset)/scale 1826 // ) 1827 1828 if (low_limit->get_int() == -max_jint) { 1829 // We need this guard when scale*pre_limit+offset >= limit 1830 // due to underflow. So we need execute pre-loop until 1831 // scale*I+offset >= min_int. But (min_int-offset) will 1832 // underflow when offset > 0 and X will be > original_limit 1833 // when stride > 0. To avoid it we replace positive offset with 0. 1834 // 1835 // Also (min_int+1 == -max_int) is used instead of min_int here 1836 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1837 Node* shift = _igvn.intcon(31); 1838 set_ctrl(shift, C->root()); 1839 Node* sign = new RShiftINode(offset, shift); 1840 register_new_node(sign, pre_ctrl); 1841 offset = new AndINode(offset, sign); 1842 register_new_node(offset, pre_ctrl); 1843 } else { 1844 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1845 // The only problem we have here when offset == min_int 1846 // since (0-min_int) == min_int. It may be fine for stride > 0 1847 // but for stride < 0 X will be < original_limit. To avoid it 1848 // max(pre_limit, original_limit) is used in do_range_check(). 1849 } 1850 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1851 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); 1852 1853 } else { // stride_con*scale_con < 0 1854 // For negative stride*scale pre-loop checks for overflow and 1855 // post-loop for underflow. 1856 // 1857 // The overflow limit: scale*I+offset < upper_limit 1858 // For pre-loop compute 1859 // NOT(scale*I+offset < upper_limit) 1860 // scale*I+offset >= upper_limit 1861 // scale*I+offset+1 > upper_limit 1862 // ( if (scale < 0) /* and stride > 0 */ 1863 // I < (upper_limit-(offset+1))/scale 1864 // else /* scale > 0 and stride < 0 */ 1865 // I > (upper_limit-(offset+1))/scale 1866 // ) 1867 // 1868 // (upper_limit-offset-1) may underflow or overflow. 1869 // To avoid it min(pre_limit, original_limit) is used 1870 // in do_range_check() for stride > 0 and max() for < 0. 1871 Node *one = _igvn.intcon(1); 1872 set_ctrl(one, C->root()); 1873 1874 Node *plus_one = new AddINode(offset, one); 1875 register_new_node( plus_one, pre_ctrl ); 1876 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1877 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); 1878 1879 if (low_limit->get_int() == -max_jint) { 1880 // We need this guard when scale*main_limit+offset >= limit 1881 // due to underflow. So we need execute main-loop while 1882 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 1883 // underflow when (offset+1) > 0 and X will be < main_limit 1884 // when scale < 0 (and stride > 0). To avoid it we replace 1885 // positive (offset+1) with 0. 1886 // 1887 // Also (min_int+1 == -max_int) is used instead of min_int here 1888 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1889 Node* shift = _igvn.intcon(31); 1890 set_ctrl(shift, C->root()); 1891 Node* sign = new RShiftINode(plus_one, shift); 1892 register_new_node(sign, pre_ctrl); 1893 plus_one = new AndINode(plus_one, sign); 1894 register_new_node(plus_one, pre_ctrl); 1895 } else { 1896 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1897 // The only problem we have here when offset == max_int 1898 // since (max_int+1) == min_int and (0-min_int) == min_int. 1899 // But it is fine since main loop will either have 1900 // less iterations or will be skipped in such case. 1901 } 1902 // The underflow limit: low_limit <= scale*I+offset. 1903 // For main-loop compute 1904 // scale*I+offset+1 > low_limit 1905 // ( if (scale < 0) /* and stride > 0 */ 1906 // I < (low_limit-(offset+1))/scale 1907 // else /* scale > 0 and stride < 0 */ 1908 // I > (low_limit-(offset+1))/scale 1909 // ) 1910 1911 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); 1912 } 1913 } 1914 1915 1916 //------------------------------is_scaled_iv--------------------------------- 1917 // Return true if exp is a constant times an induction var 1918 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1919 if (exp == iv) { 1920 if (p_scale != NULL) { 1921 *p_scale = 1; 1922 } 1923 return true; 1924 } 1925 int opc = exp->Opcode(); 1926 if (opc == Op_MulI) { 1927 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1928 if (p_scale != NULL) { 1929 *p_scale = exp->in(2)->get_int(); 1930 } 1931 return true; 1932 } 1933 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1934 if (p_scale != NULL) { 1935 *p_scale = exp->in(1)->get_int(); 1936 } 1937 return true; 1938 } 1939 } else if (opc == Op_LShiftI) { 1940 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1941 if (p_scale != NULL) { 1942 *p_scale = 1 << exp->in(2)->get_int(); 1943 } 1944 return true; 1945 } 1946 } 1947 return false; 1948 } 1949 1950 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1951 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1952 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1953 if (is_scaled_iv(exp, iv, p_scale)) { 1954 if (p_offset != NULL) { 1955 Node *zero = _igvn.intcon(0); 1956 set_ctrl(zero, C->root()); 1957 *p_offset = zero; 1958 } 1959 return true; 1960 } 1961 int opc = exp->Opcode(); 1962 if (opc == Op_AddI) { 1963 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1964 if (p_offset != NULL) { 1965 *p_offset = exp->in(2); 1966 } 1967 return true; 1968 } 1969 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1970 if (p_offset != NULL) { 1971 *p_offset = exp->in(1); 1972 } 1973 return true; 1974 } 1975 if (exp->in(2)->is_Con()) { 1976 Node* offset2 = NULL; 1977 if (depth < 2 && 1978 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1979 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1980 if (p_offset != NULL) { 1981 Node *ctrl_off2 = get_ctrl(offset2); 1982 Node* offset = new AddINode(offset2, exp->in(2)); 1983 register_new_node(offset, ctrl_off2); 1984 *p_offset = offset; 1985 } 1986 return true; 1987 } 1988 } 1989 } else if (opc == Op_SubI) { 1990 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1991 if (p_offset != NULL) { 1992 Node *zero = _igvn.intcon(0); 1993 set_ctrl(zero, C->root()); 1994 Node *ctrl_off = get_ctrl(exp->in(2)); 1995 Node* offset = new SubINode(zero, exp->in(2)); 1996 register_new_node(offset, ctrl_off); 1997 *p_offset = offset; 1998 } 1999 return true; 2000 } 2001 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 2002 if (p_offset != NULL) { 2003 *p_scale *= -1; 2004 *p_offset = exp->in(1); 2005 } 2006 return true; 2007 } 2008 } 2009 return false; 2010 } 2011 2012 //------------------------------do_range_check--------------------------------- 2013 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 2014 int PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 2015 #ifndef PRODUCT 2016 if (PrintOpto && VerifyLoopOptimizations) { 2017 tty->print("Range Check Elimination "); 2018 loop->dump_head(); 2019 } else if (TraceLoopOpts) { 2020 tty->print("RangeCheck "); 2021 loop->dump_head(); 2022 } 2023 #endif 2024 assert(RangeCheckElimination, ""); 2025 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2026 // If we fail before trying to eliminate range checks, set multiversion state 2027 int closed_range_checks = 1; 2028 2029 // protect against stride not being a constant 2030 if (!cl->stride_is_con()) 2031 return closed_range_checks; 2032 2033 // Find the trip counter; we are iteration splitting based on it 2034 Node *trip_counter = cl->phi(); 2035 // Find the main loop limit; we will trim it's iterations 2036 // to not ever trip end tests 2037 Node *main_limit = cl->limit(); 2038 2039 // Check graph shape. Cannot optimize a loop if zero-trip 2040 // Opaque1 node is optimized away and then another round 2041 // of loop opts attempted. 2042 if (!is_canonical_loop_entry(cl)) { 2043 return closed_range_checks; 2044 } 2045 2046 // Need to find the main-loop zero-trip guard 2047 Node *ctrl = cl->in(LoopNode::EntryControl); 2048 Node *iffm = ctrl->in(0); 2049 Node *opqzm = iffm->in(1)->in(1)->in(2); 2050 assert(opqzm->in(1) == main_limit, "do not understand situation"); 2051 2052 // Find the pre-loop limit; we will expand its iterations to 2053 // not ever trip low tests. 2054 Node *p_f = iffm->in(0); 2055 // pre loop may have been optimized out 2056 if (p_f->Opcode() != Op_IfFalse) { 2057 return closed_range_checks; 2058 } 2059 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2060 assert(pre_end->loopnode()->is_pre_loop(), ""); 2061 Node *pre_opaq1 = pre_end->limit(); 2062 // Occasionally it's possible for a pre-loop Opaque1 node to be 2063 // optimized away and then another round of loop opts attempted. 2064 // We can not optimize this particular loop in that case. 2065 if (pre_opaq1->Opcode() != Op_Opaque1) 2066 return closed_range_checks; 2067 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 2068 Node *pre_limit = pre_opaq->in(1); 2069 2070 // Where do we put new limit calculations 2071 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 2072 2073 // Ensure the original loop limit is available from the 2074 // pre-loop Opaque1 node. 2075 Node *orig_limit = pre_opaq->original_loop_limit(); 2076 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 2077 return closed_range_checks; 2078 2079 // Must know if its a count-up or count-down loop 2080 2081 int stride_con = cl->stride_con(); 2082 Node *zero = _igvn.intcon(0); 2083 Node *one = _igvn.intcon(1); 2084 // Use symmetrical int range [-max_jint,max_jint] 2085 Node *mini = _igvn.intcon(-max_jint); 2086 set_ctrl(zero, C->root()); 2087 set_ctrl(one, C->root()); 2088 set_ctrl(mini, C->root()); 2089 2090 // Range checks that do not dominate the loop backedge (ie. 2091 // conditionally executed) can lengthen the pre loop limit beyond 2092 // the original loop limit. To prevent this, the pre limit is 2093 // (for stride > 0) MINed with the original loop limit (MAXed 2094 // stride < 0) when some range_check (rc) is conditionally 2095 // executed. 2096 bool conditional_rc = false; 2097 2098 // Count number of range checks and reduce by load range limits, if zero, 2099 // the loop is in canonical form to multiversion. 2100 closed_range_checks = 0; 2101 2102 // Check loop body for tests of trip-counter plus loop-invariant vs loop-variant. 2103 for( uint i = 0; i < loop->_body.size(); i++ ) { 2104 Node *iff = loop->_body[i]; 2105 if (iff->Opcode() == Op_If || 2106 iff->Opcode() == Op_RangeCheck) { // Test? 2107 // Test is an IfNode, has 2 projections. If BOTH are in the loop 2108 // we need loop unswitching instead of iteration splitting. 2109 closed_range_checks++; 2110 Node *exit = loop->is_loop_exit(iff); 2111 if( !exit ) continue; 2112 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 2113 2114 // Get boolean condition to test 2115 Node *i1 = iff->in(1); 2116 if( !i1->is_Bool() ) continue; 2117 BoolNode *bol = i1->as_Bool(); 2118 BoolTest b_test = bol->_test; 2119 // Flip sense of test if exit condition is flipped 2120 if( flip ) 2121 b_test = b_test.negate(); 2122 2123 // Get compare 2124 Node *cmp = bol->in(1); 2125 2126 // Look for trip_counter + offset vs limit 2127 Node *rc_exp = cmp->in(1); 2128 Node *limit = cmp->in(2); 2129 jint scale_con= 1; // Assume trip counter not scaled 2130 2131 Node *limit_c = get_ctrl(limit); 2132 if( loop->is_member(get_loop(limit_c) ) ) { 2133 // Compare might have operands swapped; commute them 2134 b_test = b_test.commute(); 2135 rc_exp = cmp->in(2); 2136 limit = cmp->in(1); 2137 limit_c = get_ctrl(limit); 2138 if( loop->is_member(get_loop(limit_c) ) ) 2139 continue; // Both inputs are loop varying; cannot RCE 2140 } 2141 // Here we know 'limit' is loop invariant 2142 2143 // 'limit' maybe pinned below the zero trip test (probably from a 2144 // previous round of rce), in which case, it can't be used in the 2145 // zero trip test expression which must occur before the zero test's if. 2146 if( limit_c == ctrl ) { 2147 continue; // Don't rce this check but continue looking for other candidates. 2148 } 2149 2150 // Check for scaled induction variable plus an offset 2151 Node *offset = NULL; 2152 2153 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 2154 continue; 2155 } 2156 2157 Node *offset_c = get_ctrl(offset); 2158 if( loop->is_member( get_loop(offset_c) ) ) 2159 continue; // Offset is not really loop invariant 2160 // Here we know 'offset' is loop invariant. 2161 2162 // As above for the 'limit', the 'offset' maybe pinned below the 2163 // zero trip test. 2164 if( offset_c == ctrl ) { 2165 continue; // Don't rce this check but continue looking for other candidates. 2166 } 2167 #ifdef ASSERT 2168 if (TraceRangeLimitCheck) { 2169 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 2170 bol->dump(2); 2171 } 2172 #endif 2173 // At this point we have the expression as: 2174 // scale_con * trip_counter + offset :: limit 2175 // where scale_con, offset and limit are loop invariant. Trip_counter 2176 // monotonically increases by stride_con, a constant. Both (or either) 2177 // stride_con and scale_con can be negative which will flip about the 2178 // sense of the test. 2179 2180 // Adjust pre and main loop limits to guard the correct iteration set 2181 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 2182 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 2183 // The underflow and overflow limits: 0 <= scale*I+offset < limit 2184 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); 2185 // (0-offset)/scale could be outside of loop iterations range. 2186 conditional_rc = true; 2187 } else { 2188 if (PrintOpto) { 2189 tty->print_cr("missed RCE opportunity"); 2190 } 2191 continue; // In release mode, ignore it 2192 } 2193 } else { // Otherwise work on normal compares 2194 switch( b_test._test ) { 2195 case BoolTest::gt: 2196 // Fall into GE case 2197 case BoolTest::ge: 2198 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 2199 scale_con = -scale_con; 2200 offset = new SubINode( zero, offset ); 2201 register_new_node( offset, pre_ctrl ); 2202 limit = new SubINode( zero, limit ); 2203 register_new_node( limit, pre_ctrl ); 2204 // Fall into LE case 2205 case BoolTest::le: 2206 if (b_test._test != BoolTest::gt) { 2207 // Convert X <= Y to X < Y+1 2208 limit = new AddINode( limit, one ); 2209 register_new_node( limit, pre_ctrl ); 2210 } 2211 // Fall into LT case 2212 case BoolTest::lt: 2213 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 2214 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 2215 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 2216 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); 2217 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 2218 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 2219 // still be outside of loop range. 2220 conditional_rc = true; 2221 break; 2222 default: 2223 if (PrintOpto) { 2224 tty->print_cr("missed RCE opportunity"); 2225 } 2226 continue; // Unhandled case 2227 } 2228 } 2229 2230 // Kill the eliminated test 2231 C->set_major_progress(); 2232 Node *kill_con = _igvn.intcon( 1-flip ); 2233 set_ctrl(kill_con, C->root()); 2234 _igvn.replace_input_of(iff, 1, kill_con); 2235 // Find surviving projection 2236 assert(iff->is_If(), ""); 2237 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 2238 // Find loads off the surviving projection; remove their control edge 2239 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2240 Node* cd = dp->fast_out(i); // Control-dependent node 2241 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop 2242 // Allow the load to float around in the loop, or before it 2243 // but NOT before the pre-loop. 2244 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 2245 --i; 2246 --imax; 2247 } 2248 } 2249 if (limit->Opcode() == Op_LoadRange) { 2250 closed_range_checks--; 2251 } 2252 2253 } // End of is IF 2254 2255 } 2256 2257 // Update loop limits 2258 if (conditional_rc) { 2259 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit) 2260 : (Node*)new MaxINode(pre_limit, orig_limit); 2261 register_new_node(pre_limit, pre_ctrl); 2262 } 2263 _igvn.replace_input_of(pre_opaq, 1, pre_limit); 2264 2265 // Note:: we are making the main loop limit no longer precise; 2266 // need to round up based on stride. 2267 cl->set_nonexact_trip_count(); 2268 Node *main_cle = cl->loopexit(); 2269 Node *main_bol = main_cle->in(1); 2270 // Hacking loop bounds; need private copies of exit test 2271 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 2272 main_bol = main_bol->clone();// Clone a private BoolNode 2273 register_new_node( main_bol, main_cle->in(0) ); 2274 _igvn.replace_input_of(main_cle, 1, main_bol); 2275 } 2276 Node *main_cmp = main_bol->in(1); 2277 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 2278 main_cmp = main_cmp->clone();// Clone a private CmpNode 2279 register_new_node( main_cmp, main_cle->in(0) ); 2280 _igvn.replace_input_of(main_bol, 1, main_cmp); 2281 } 2282 // Hack the now-private loop bounds 2283 _igvn.replace_input_of(main_cmp, 2, main_limit); 2284 // The OpaqueNode is unshared by design 2285 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 2286 _igvn.replace_input_of(opqzm, 1, main_limit); 2287 2288 return closed_range_checks; 2289 } 2290 2291 //------------------------------has_range_checks------------------------------- 2292 // Check to see if RCE cleaned the current loop of range-checks. 2293 void PhaseIdealLoop::has_range_checks(IdealLoopTree *loop) { 2294 assert(RangeCheckElimination, ""); 2295 2296 // skip if not a counted loop 2297 if (!loop->is_counted()) return; 2298 2299 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2300 2301 // skip this loop if it is already checked 2302 if (cl->has_been_range_checked()) return; 2303 2304 // Now check for existence of range checks 2305 for (uint i = 0; i < loop->_body.size(); i++) { 2306 Node *iff = loop->_body[i]; 2307 int iff_opc = iff->Opcode(); 2308 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) { 2309 cl->mark_has_range_checks(); 2310 break; 2311 } 2312 } 2313 cl->set_has_been_range_checked(); 2314 } 2315 2316 //-------------------------multi_version_post_loops---------------------------- 2317 // Check the range checks that remain, if simple, use the bounds to guard 2318 // which version to a post loop we execute, one with range checks or one without 2319 bool PhaseIdealLoop::multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop) { 2320 bool multi_version_succeeded = false; 2321 assert(RangeCheckElimination, ""); 2322 CountedLoopNode *legacy_cl = legacy_loop->_head->as_CountedLoop(); 2323 assert(legacy_cl->is_post_loop(), ""); 2324 2325 // Check for existence of range checks using the unique instance to make a guard with 2326 Unique_Node_List worklist; 2327 for (uint i = 0; i < legacy_loop->_body.size(); i++) { 2328 Node *iff = legacy_loop->_body[i]; 2329 int iff_opc = iff->Opcode(); 2330 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) { 2331 worklist.push(iff); 2332 } 2333 } 2334 2335 // Find RCE'd post loop so that we can stage its guard. 2336 if (!is_canonical_loop_entry(legacy_cl)) return multi_version_succeeded; 2337 Node* ctrl = legacy_cl->in(LoopNode::EntryControl); 2338 Node* iffm = ctrl->in(0); 2339 2340 // Now we test that both the post loops are connected 2341 Node* post_loop_region = iffm->in(0); 2342 if (post_loop_region == NULL) return multi_version_succeeded; 2343 if (!post_loop_region->is_Region()) return multi_version_succeeded; 2344 Node* covering_region = post_loop_region->in(RegionNode::Control+1); 2345 if (covering_region == NULL) return multi_version_succeeded; 2346 if (!covering_region->is_Region()) return multi_version_succeeded; 2347 Node* p_f = covering_region->in(RegionNode::Control); 2348 if (p_f == NULL) return multi_version_succeeded; 2349 if (!p_f->is_IfFalse()) return multi_version_succeeded; 2350 if (!p_f->in(0)->is_CountedLoopEnd()) return multi_version_succeeded; 2351 CountedLoopEndNode* rce_loop_end = p_f->in(0)->as_CountedLoopEnd(); 2352 if (rce_loop_end == NULL) return multi_version_succeeded; 2353 CountedLoopNode* rce_cl = rce_loop_end->loopnode(); 2354 if (rce_cl == NULL || !rce_cl->is_post_loop()) return multi_version_succeeded; 2355 CountedLoopNode *known_rce_cl = rce_loop->_head->as_CountedLoop(); 2356 if (rce_cl != known_rce_cl) return multi_version_succeeded; 2357 2358 // Then we fetch the cover entry test 2359 ctrl = rce_cl->in(LoopNode::EntryControl); 2360 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return multi_version_succeeded; 2361 2362 #ifndef PRODUCT 2363 if (TraceLoopOpts) { 2364 tty->print("PostMultiVersion\n"); 2365 rce_loop->dump_head(); 2366 legacy_loop->dump_head(); 2367 } 2368 #endif 2369 2370 // Now fetch the limit we want to compare against 2371 Node *limit = rce_cl->limit(); 2372 bool first_time = true; 2373 2374 // If we got this far, we identified the post loop which has been RCE'd and 2375 // we have a work list. Now we will try to transform the if guard to cause 2376 // the loop pair to be multi version executed with the determination left to runtime 2377 // or the optimizer if full information is known about the given arrays at compile time. 2378 Node *last_min = NULL; 2379 multi_version_succeeded = true; 2380 while (worklist.size()) { 2381 Node* rc_iffm = worklist.pop(); 2382 if (rc_iffm->is_If()) { 2383 Node *rc_bolzm = rc_iffm->in(1); 2384 if (rc_bolzm->is_Bool()) { 2385 Node *rc_cmpzm = rc_bolzm->in(1); 2386 if (rc_cmpzm->is_Cmp()) { 2387 Node *rc_left = rc_cmpzm->in(2); 2388 if (rc_left->Opcode() != Op_LoadRange) { 2389 multi_version_succeeded = false; 2390 break; 2391 } 2392 if (first_time) { 2393 last_min = rc_left; 2394 first_time = false; 2395 } else { 2396 Node *cur_min = new MinINode(last_min, rc_left); 2397 last_min = cur_min; 2398 _igvn.register_new_node_with_optimizer(last_min); 2399 } 2400 } 2401 } 2402 } 2403 } 2404 2405 // All we have to do is update the limit of the rce loop 2406 // with the min of our expression and the current limit. 2407 // We will use this expression to replace the current limit. 2408 if (last_min && multi_version_succeeded) { 2409 Node *cur_min = new MinINode(last_min, limit); 2410 _igvn.register_new_node_with_optimizer(cur_min); 2411 Node *cmp_node = rce_loop_end->cmp_node(); 2412 _igvn.replace_input_of(cmp_node, 2, cur_min); 2413 set_idom(cmp_node, cur_min, dom_depth(ctrl)); 2414 set_ctrl(cur_min, ctrl); 2415 set_loop(cur_min, rce_loop->_parent); 2416 2417 legacy_cl->mark_is_multiversioned(); 2418 rce_cl->mark_is_multiversioned(); 2419 multi_version_succeeded = true; 2420 2421 C->set_major_progress(); 2422 } 2423 2424 return multi_version_succeeded; 2425 } 2426 2427 //-------------------------poison_rce_post_loop-------------------------------- 2428 // Causes the rce'd post loop to be optimized away if multiversioning fails 2429 void PhaseIdealLoop::poison_rce_post_loop(IdealLoopTree *rce_loop) { 2430 CountedLoopNode *rce_cl = rce_loop->_head->as_CountedLoop(); 2431 Node* ctrl = rce_cl->in(LoopNode::EntryControl); 2432 if (ctrl->is_IfTrue() || ctrl->is_IfFalse()) { 2433 Node* iffm = ctrl->in(0); 2434 if (iffm->is_If()) { 2435 Node* cur_bool = iffm->in(1); 2436 if (cur_bool->is_Bool()) { 2437 Node* cur_cmp = cur_bool->in(1); 2438 if (cur_cmp->is_Cmp()) { 2439 BoolTest::mask new_test = BoolTest::gt; 2440 BoolNode *new_bool = new BoolNode(cur_cmp, new_test); 2441 _igvn.replace_node(cur_bool, new_bool); 2442 _igvn._worklist.push(new_bool); 2443 Node* left_op = cur_cmp->in(1); 2444 _igvn.replace_input_of(cur_cmp, 2, left_op); 2445 C->set_major_progress(); 2446 } 2447 } 2448 } 2449 } 2450 } 2451 2452 //------------------------------DCE_loop_body---------------------------------- 2453 // Remove simplistic dead code from loop body 2454 void IdealLoopTree::DCE_loop_body() { 2455 for( uint i = 0; i < _body.size(); i++ ) 2456 if( _body.at(i)->outcnt() == 0 ) 2457 _body.map( i--, _body.pop() ); 2458 } 2459 2460 2461 //------------------------------adjust_loop_exit_prob-------------------------- 2462 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2463 // Replace with a 1-in-10 exit guess. 2464 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 2465 Node *test = tail(); 2466 while( test != _head ) { 2467 uint top = test->Opcode(); 2468 if( top == Op_IfTrue || top == Op_IfFalse ) { 2469 int test_con = ((ProjNode*)test)->_con; 2470 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2471 IfNode *iff = test->in(0)->as_If(); 2472 if( iff->outcnt() == 2 ) { // Ignore dead tests 2473 Node *bol = iff->in(1); 2474 if( bol && bol->req() > 1 && bol->in(1) && 2475 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 2476 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 2477 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 2478 (bol->in(1)->Opcode() == Op_CompareAndExchangeB ) || 2479 (bol->in(1)->Opcode() == Op_CompareAndExchangeS ) || 2480 (bol->in(1)->Opcode() == Op_CompareAndExchangeI ) || 2481 (bol->in(1)->Opcode() == Op_CompareAndExchangeL ) || 2482 (bol->in(1)->Opcode() == Op_CompareAndExchangeP ) || 2483 (bol->in(1)->Opcode() == Op_CompareAndExchangeN ) || 2484 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapB ) || 2485 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapS ) || 2486 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI ) || 2487 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL ) || 2488 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP ) || 2489 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN ) || 2490 (bol->in(1)->Opcode() == Op_CompareAndSwapB ) || 2491 (bol->in(1)->Opcode() == Op_CompareAndSwapS ) || 2492 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 2493 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 2494 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 2495 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 2496 return; // Allocation loops RARELY take backedge 2497 // Find the OTHER exit path from the IF 2498 Node* ex = iff->proj_out(1-test_con); 2499 float p = iff->_prob; 2500 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 2501 if( top == Op_IfTrue ) { 2502 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2503 iff->_prob = PROB_STATIC_FREQUENT; 2504 } 2505 } else { 2506 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2507 iff->_prob = PROB_STATIC_INFREQUENT; 2508 } 2509 } 2510 } 2511 } 2512 } 2513 test = phase->idom(test); 2514 } 2515 } 2516 2517 #ifdef ASSERT 2518 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) { 2519 Node *ctrl = cl->in(LoopNode::EntryControl); 2520 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 2521 Node *iffm = ctrl->in(0); 2522 assert(iffm->Opcode() == Op_If, ""); 2523 Node *p_f = iffm->in(0); 2524 assert(p_f->Opcode() == Op_IfFalse, ""); 2525 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2526 assert(pre_end->loopnode()->is_pre_loop(), ""); 2527 return pre_end->loopnode(); 2528 } 2529 #endif 2530 2531 // Remove the main and post loops and make the pre loop execute all 2532 // iterations. Useful when the pre loop is found empty. 2533 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) { 2534 CountedLoopEndNode* pre_end = cl->loopexit(); 2535 Node* pre_cmp = pre_end->cmp_node(); 2536 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) { 2537 // Only safe to remove the main loop if the compiler optimized it 2538 // out based on an unknown number of iterations 2539 return; 2540 } 2541 2542 // Can we find the main loop? 2543 if (_next == NULL) { 2544 return; 2545 } 2546 2547 Node* next_head = _next->_head; 2548 if (!next_head->is_CountedLoop()) { 2549 return; 2550 } 2551 2552 CountedLoopNode* main_head = next_head->as_CountedLoop(); 2553 if (!main_head->is_main_loop()) { 2554 return; 2555 } 2556 2557 assert(locate_pre_from_main(main_head) == cl, "bad main loop"); 2558 Node* main_iff = main_head->in(LoopNode::EntryControl)->in(0); 2559 2560 // Remove the Opaque1Node of the pre loop and make it execute all iterations 2561 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2)); 2562 // Remove the Opaque1Node of the main loop so it can be optimized out 2563 Node* main_cmp = main_iff->in(1)->in(1); 2564 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?"); 2565 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1)); 2566 } 2567 2568 //------------------------------policy_do_remove_empty_loop-------------------- 2569 // Micro-benchmark spamming. Policy is to always remove empty loops. 2570 // The 'DO' part is to replace the trip counter with the value it will 2571 // have on the last iteration. This will break the loop. 2572 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 2573 // Minimum size must be empty loop 2574 if (_body.size() > EMPTY_LOOP_SIZE) 2575 return false; 2576 2577 if (!_head->is_CountedLoop()) 2578 return false; // Dead loop 2579 CountedLoopNode *cl = _head->as_CountedLoop(); 2580 if (!cl->is_valid_counted_loop()) 2581 return false; // Malformed loop 2582 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 2583 return false; // Infinite loop 2584 2585 if (cl->is_pre_loop()) { 2586 // If the loop we are removing is a pre-loop then the main and 2587 // post loop can be removed as well 2588 remove_main_post_loops(cl, phase); 2589 } 2590 2591 #ifdef ASSERT 2592 // Ensure only one phi which is the iv. 2593 Node* iv = NULL; 2594 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 2595 Node* n = cl->fast_out(i); 2596 if (n->Opcode() == Op_Phi) { 2597 assert(iv == NULL, "Too many phis" ); 2598 iv = n; 2599 } 2600 } 2601 assert(iv == cl->phi(), "Wrong phi" ); 2602 #endif 2603 2604 // main and post loops have explicitly created zero trip guard 2605 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 2606 if (needs_guard) { 2607 // Skip guard if values not overlap. 2608 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 2609 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 2610 int stride_con = cl->stride_con(); 2611 if (stride_con > 0) { 2612 needs_guard = (init_t->_hi >= limit_t->_lo); 2613 } else { 2614 needs_guard = (init_t->_lo <= limit_t->_hi); 2615 } 2616 } 2617 if (needs_guard) { 2618 // Check for an obvious zero trip guard. 2619 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 2620 if (inctrl->Opcode() == Op_IfTrue || inctrl->Opcode() == Op_IfFalse) { 2621 bool maybe_swapped = (inctrl->Opcode() == Op_IfFalse); 2622 // The test should look like just the backedge of a CountedLoop 2623 Node* iff = inctrl->in(0); 2624 if (iff->is_If()) { 2625 Node* bol = iff->in(1); 2626 if (bol->is_Bool()) { 2627 BoolTest test = bol->as_Bool()->_test; 2628 if (maybe_swapped) { 2629 test._test = test.commute(); 2630 test._test = test.negate(); 2631 } 2632 if (test._test == cl->loopexit()->test_trip()) { 2633 Node* cmp = bol->in(1); 2634 int init_idx = maybe_swapped ? 2 : 1; 2635 int limit_idx = maybe_swapped ? 1 : 2; 2636 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) { 2637 needs_guard = false; 2638 } 2639 } 2640 } 2641 } 2642 } 2643 } 2644 2645 #ifndef PRODUCT 2646 if (PrintOpto) { 2647 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 2648 this->dump_head(); 2649 } else if (TraceLoopOpts) { 2650 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 2651 this->dump_head(); 2652 } 2653 #endif 2654 2655 if (needs_guard) { 2656 // Peel the loop to ensure there's a zero trip guard 2657 Node_List old_new; 2658 phase->do_peeling(this, old_new); 2659 } 2660 2661 // Replace the phi at loop head with the final value of the last 2662 // iteration. Then the CountedLoopEnd will collapse (backedge never 2663 // taken) and all loop-invariant uses of the exit values will be correct. 2664 Node *phi = cl->phi(); 2665 Node *exact_limit = phase->exact_limit(this); 2666 if (exact_limit != cl->limit()) { 2667 // We also need to replace the original limit to collapse loop exit. 2668 Node* cmp = cl->loopexit()->cmp_node(); 2669 assert(cl->limit() == cmp->in(2), "sanity"); 2670 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 2671 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 2672 } 2673 // Note: the final value after increment should not overflow since 2674 // counted loop has limit check predicate. 2675 Node *final = new SubINode( exact_limit, cl->stride() ); 2676 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 2677 phase->_igvn.replace_node(phi,final); 2678 phase->C->set_major_progress(); 2679 return true; 2680 } 2681 2682 //------------------------------policy_do_one_iteration_loop------------------- 2683 // Convert one iteration loop into normal code. 2684 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 2685 if (!_head->as_Loop()->is_valid_counted_loop()) 2686 return false; // Only for counted loop 2687 2688 CountedLoopNode *cl = _head->as_CountedLoop(); 2689 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 2690 return false; 2691 } 2692 2693 #ifndef PRODUCT 2694 if(TraceLoopOpts) { 2695 tty->print("OneIteration "); 2696 this->dump_head(); 2697 } 2698 #endif 2699 2700 Node *init_n = cl->init_trip(); 2701 #ifdef ASSERT 2702 // Loop boundaries should be constant since trip count is exact. 2703 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 2704 #endif 2705 // Replace the phi at loop head with the value of the init_trip. 2706 // Then the CountedLoopEnd will collapse (backedge will not be taken) 2707 // and all loop-invariant uses of the exit values will be correct. 2708 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 2709 phase->C->set_major_progress(); 2710 return true; 2711 } 2712 2713 //============================================================================= 2714 //------------------------------iteration_split_impl--------------------------- 2715 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 2716 // Compute loop trip count if possible. 2717 compute_trip_count(phase); 2718 2719 // Convert one iteration loop into normal code. 2720 if (policy_do_one_iteration_loop(phase)) 2721 return true; 2722 2723 // Check and remove empty loops (spam micro-benchmarks) 2724 if (policy_do_remove_empty_loop(phase)) 2725 return true; // Here we removed an empty loop 2726 2727 bool should_peel = policy_peeling(phase); // Should we peel? 2728 2729 bool should_unswitch = policy_unswitching(phase); 2730 2731 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 2732 // This removes loop-invariant tests (usually null checks). 2733 if (!_head->is_CountedLoop()) { // Non-counted loop 2734 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 2735 // Partial peel succeeded so terminate this round of loop opts 2736 return false; 2737 } 2738 if (should_peel) { // Should we peel? 2739 if (PrintOpto) { tty->print_cr("should_peel"); } 2740 phase->do_peeling(this,old_new); 2741 } else if (should_unswitch) { 2742 phase->do_unswitching(this, old_new); 2743 } 2744 return true; 2745 } 2746 CountedLoopNode *cl = _head->as_CountedLoop(); 2747 2748 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 2749 2750 // Do nothing special to pre- and post- loops 2751 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 2752 2753 // Compute loop trip count from profile data 2754 compute_profile_trip_cnt(phase); 2755 2756 // Before attempting fancy unrolling, RCE or alignment, see if we want 2757 // to completely unroll this loop or do loop unswitching. 2758 if (cl->is_normal_loop()) { 2759 if (should_unswitch) { 2760 phase->do_unswitching(this, old_new); 2761 return true; 2762 } 2763 bool should_maximally_unroll = policy_maximally_unroll(phase); 2764 if (should_maximally_unroll) { 2765 // Here we did some unrolling and peeling. Eventually we will 2766 // completely unroll this loop and it will no longer be a loop. 2767 phase->do_maximally_unroll(this,old_new); 2768 return true; 2769 } 2770 } 2771 2772 // Skip next optimizations if running low on nodes. Note that 2773 // policy_unswitching and policy_maximally_unroll have this check. 2774 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes(); 2775 if ((int)(2 * _body.size()) > nodes_left) { 2776 return true; 2777 } 2778 2779 // Counted loops may be peeled, may need some iterations run up 2780 // front for RCE, and may want to align loop refs to a cache 2781 // line. Thus we clone a full loop up front whose trip count is 2782 // at least 1 (if peeling), but may be several more. 2783 2784 // The main loop will start cache-line aligned with at least 1 2785 // iteration of the unrolled body (zero-trip test required) and 2786 // will have some range checks removed. 2787 2788 // A post-loop will finish any odd iterations (leftover after 2789 // unrolling), plus any needed for RCE purposes. 2790 2791 bool should_unroll = policy_unroll(phase); 2792 2793 bool should_rce = policy_range_check(phase); 2794 2795 bool should_align = policy_align(phase); 2796 2797 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2798 // need a pre-loop. We may still need to peel an initial iteration but 2799 // we will not be needing an unknown number of pre-iterations. 2800 // 2801 // Basically, if may_rce_align reports FALSE first time through, 2802 // we will not be able to later do RCE or Aligning on this loop. 2803 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2804 2805 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2806 // we switch to the pre-/main-/post-loop model. This model also covers 2807 // peeling. 2808 if (should_rce || should_align || should_unroll) { 2809 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2810 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2811 2812 // Adjust the pre- and main-loop limits to let the pre and post loops run 2813 // with full checks, but the main-loop with no checks. Remove said 2814 // checks from the main body. 2815 if (should_rce) { 2816 if (phase->do_range_check(this, old_new) != 0) { 2817 cl->mark_has_range_checks(); 2818 } 2819 } else if (PostLoopMultiversioning) { 2820 phase->has_range_checks(this); 2821 } 2822 2823 if (should_unroll && !should_peel && PostLoopMultiversioning) { 2824 // Try to setup multiversioning on main loops before they are unrolled 2825 if (cl->is_main_loop() && (cl->unrolled_count() == 1)) { 2826 phase->insert_scalar_rced_post_loop(this, old_new); 2827 } 2828 } 2829 2830 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2831 // twice as many iterations as before) and the main body limit (only do 2832 // an even number of trips). If we are peeling, we might enable some RCE 2833 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2834 // peeling. 2835 if (should_unroll && !should_peel) { 2836 if (SuperWordLoopUnrollAnalysis) { 2837 phase->insert_vector_post_loop(this, old_new); 2838 } 2839 phase->do_unroll(this, old_new, true); 2840 } 2841 2842 // Adjust the pre-loop limits to align the main body 2843 // iterations. 2844 if (should_align) 2845 Unimplemented(); 2846 2847 } else { // Else we have an unchanged counted loop 2848 if (should_peel) // Might want to peel but do nothing else 2849 phase->do_peeling(this,old_new); 2850 } 2851 return true; 2852 } 2853 2854 2855 //============================================================================= 2856 //------------------------------iteration_split-------------------------------- 2857 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2858 // Recursively iteration split nested loops 2859 if (_child && !_child->iteration_split(phase, old_new)) 2860 return false; 2861 2862 // Clean out prior deadwood 2863 DCE_loop_body(); 2864 2865 2866 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2867 // Replace with a 1-in-10 exit guess. 2868 if (_parent /*not the root loop*/ && 2869 !_irreducible && 2870 // Also ignore the occasional dead backedge 2871 !tail()->is_top()) { 2872 adjust_loop_exit_prob(phase); 2873 } 2874 2875 // Gate unrolling, RCE and peeling efforts. 2876 if (!_child && // If not an inner loop, do not split 2877 !_irreducible && 2878 _allow_optimizations && 2879 !tail()->is_top()) { // Also ignore the occasional dead backedge 2880 if (!_has_call) { 2881 if (!iteration_split_impl(phase, old_new)) { 2882 return false; 2883 } 2884 } else if (policy_unswitching(phase)) { 2885 phase->do_unswitching(this, old_new); 2886 } 2887 } 2888 2889 // Minor offset re-organization to remove loop-fallout uses of 2890 // trip counter when there was no major reshaping. 2891 phase->reorg_offsets(this); 2892 2893 if (_next && !_next->iteration_split(phase, old_new)) 2894 return false; 2895 return true; 2896 } 2897 2898 2899 //============================================================================= 2900 // Process all the loops in the loop tree and replace any fill 2901 // patterns with an intrinsic version. 2902 bool PhaseIdealLoop::do_intrinsify_fill() { 2903 bool changed = false; 2904 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2905 IdealLoopTree* lpt = iter.current(); 2906 changed |= intrinsify_fill(lpt); 2907 } 2908 return changed; 2909 } 2910 2911 2912 // Examine an inner loop looking for a a single store of an invariant 2913 // value in a unit stride loop, 2914 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2915 Node*& shift, Node*& con) { 2916 const char* msg = NULL; 2917 Node* msg_node = NULL; 2918 2919 store_value = NULL; 2920 con = NULL; 2921 shift = NULL; 2922 2923 // Process the loop looking for stores. If there are multiple 2924 // stores or extra control flow give at this point. 2925 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2926 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2927 Node* n = lpt->_body.at(i); 2928 if (n->outcnt() == 0) continue; // Ignore dead 2929 if (n->is_Store()) { 2930 if (store != NULL) { 2931 msg = "multiple stores"; 2932 break; 2933 } 2934 int opc = n->Opcode(); 2935 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 2936 msg = "oop fills not handled"; 2937 break; 2938 } 2939 Node* value = n->in(MemNode::ValueIn); 2940 if (!lpt->is_invariant(value)) { 2941 msg = "variant store value"; 2942 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2943 msg = "not array address"; 2944 } 2945 store = n; 2946 store_value = value; 2947 } else if (n->is_If() && n != head->loopexit()) { 2948 msg = "extra control flow"; 2949 msg_node = n; 2950 } 2951 } 2952 2953 if (store == NULL) { 2954 // No store in loop 2955 return false; 2956 } 2957 2958 if (msg == NULL && head->stride_con() != 1) { 2959 // could handle negative strides too 2960 if (head->stride_con() < 0) { 2961 msg = "negative stride"; 2962 } else { 2963 msg = "non-unit stride"; 2964 } 2965 } 2966 2967 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2968 msg = "can't handle store address"; 2969 msg_node = store->in(MemNode::Address); 2970 } 2971 2972 if (msg == NULL && 2973 (!store->in(MemNode::Memory)->is_Phi() || 2974 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2975 msg = "store memory isn't proper phi"; 2976 msg_node = store->in(MemNode::Memory); 2977 } 2978 2979 // Make sure there is an appropriate fill routine 2980 BasicType t = store->as_Mem()->memory_type(); 2981 const char* fill_name; 2982 if (msg == NULL && 2983 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2984 msg = "unsupported store"; 2985 msg_node = store; 2986 } 2987 2988 if (msg != NULL) { 2989 #ifndef PRODUCT 2990 if (TraceOptimizeFill) { 2991 tty->print_cr("not fill intrinsic candidate: %s", msg); 2992 if (msg_node != NULL) msg_node->dump(); 2993 } 2994 #endif 2995 return false; 2996 } 2997 2998 // Make sure the address expression can be handled. It should be 2999 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()). 3000 Node* elements[4]; 3001 Node* cast = NULL; 3002 Node* conv = NULL; 3003 bool found_index = false; 3004 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 3005 for (int e = 0; e < count; e++) { 3006 Node* n = elements[e]; 3007 if (n->is_Con() && con == NULL) { 3008 con = n; 3009 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 3010 Node* value = n->in(1); 3011 #ifdef _LP64 3012 if (value->Opcode() == Op_ConvI2L) { 3013 conv = value; 3014 value = value->in(1); 3015 } 3016 if (value->Opcode() == Op_CastII && 3017 value->as_CastII()->has_range_check()) { 3018 // Skip range check dependent CastII nodes 3019 cast = value; 3020 value = value->in(1); 3021 } 3022 #endif 3023 if (value != head->phi()) { 3024 msg = "unhandled shift in address"; 3025 } else { 3026 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 3027 msg = "scale doesn't match"; 3028 } else { 3029 found_index = true; 3030 shift = n; 3031 } 3032 } 3033 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 3034 conv = n; 3035 n = n->in(1); 3036 if (n->Opcode() == Op_CastII && 3037 n->as_CastII()->has_range_check()) { 3038 // Skip range check dependent CastII nodes 3039 cast = n; 3040 n = n->in(1); 3041 } 3042 if (n == head->phi()) { 3043 found_index = true; 3044 } else { 3045 msg = "unhandled input to ConvI2L"; 3046 } 3047 } else if (n == head->phi()) { 3048 // no shift, check below for allowed cases 3049 found_index = true; 3050 } else { 3051 msg = "unhandled node in address"; 3052 msg_node = n; 3053 } 3054 } 3055 3056 if (count == -1) { 3057 msg = "malformed address expression"; 3058 msg_node = store; 3059 } 3060 3061 if (!found_index) { 3062 msg = "missing use of index"; 3063 } 3064 3065 // byte sized items won't have a shift 3066 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 3067 msg = "can't find shift"; 3068 msg_node = store; 3069 } 3070 3071 if (msg != NULL) { 3072 #ifndef PRODUCT 3073 if (TraceOptimizeFill) { 3074 tty->print_cr("not fill intrinsic: %s", msg); 3075 if (msg_node != NULL) msg_node->dump(); 3076 } 3077 #endif 3078 return false; 3079 } 3080 3081 // No make sure all the other nodes in the loop can be handled 3082 VectorSet ok(Thread::current()->resource_area()); 3083 3084 // store related values are ok 3085 ok.set(store->_idx); 3086 ok.set(store->in(MemNode::Memory)->_idx); 3087 3088 CountedLoopEndNode* loop_exit = head->loopexit(); 3089 guarantee(loop_exit != NULL, "no loop exit node"); 3090 3091 // Loop structure is ok 3092 ok.set(head->_idx); 3093 ok.set(loop_exit->_idx); 3094 ok.set(head->phi()->_idx); 3095 ok.set(head->incr()->_idx); 3096 ok.set(loop_exit->cmp_node()->_idx); 3097 ok.set(loop_exit->in(1)->_idx); 3098 3099 // Address elements are ok 3100 if (con) ok.set(con->_idx); 3101 if (shift) ok.set(shift->_idx); 3102 if (cast) ok.set(cast->_idx); 3103 if (conv) ok.set(conv->_idx); 3104 3105 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 3106 Node* n = lpt->_body.at(i); 3107 if (n->outcnt() == 0) continue; // Ignore dead 3108 if (ok.test(n->_idx)) continue; 3109 // Backedge projection is ok 3110 if (n->is_IfTrue() && n->in(0) == loop_exit) continue; 3111 if (!n->is_AddP()) { 3112 msg = "unhandled node"; 3113 msg_node = n; 3114 break; 3115 } 3116 } 3117 3118 // Make sure no unexpected values are used outside the loop 3119 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 3120 Node* n = lpt->_body.at(i); 3121 // These values can be replaced with other nodes if they are used 3122 // outside the loop. 3123 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue; 3124 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 3125 Node* use = iter.get(); 3126 if (!lpt->_body.contains(use)) { 3127 msg = "node is used outside loop"; 3128 // lpt->_body.dump(); 3129 msg_node = n; 3130 break; 3131 } 3132 } 3133 } 3134 3135 #ifdef ASSERT 3136 if (TraceOptimizeFill) { 3137 if (msg != NULL) { 3138 tty->print_cr("no fill intrinsic: %s", msg); 3139 if (msg_node != NULL) msg_node->dump(); 3140 } else { 3141 tty->print_cr("fill intrinsic for:"); 3142 } 3143 store->dump(); 3144 if (Verbose) { 3145 lpt->_body.dump(); 3146 } 3147 } 3148 #endif 3149 3150 return msg == NULL; 3151 } 3152 3153 3154 3155 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 3156 // Only for counted inner loops 3157 if (!lpt->is_counted() || !lpt->is_inner()) { 3158 return false; 3159 } 3160 3161 // Must have constant stride 3162 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 3163 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 3164 return false; 3165 } 3166 3167 // Check that the body only contains a store of a loop invariant 3168 // value that is indexed by the loop phi. 3169 Node* store = NULL; 3170 Node* store_value = NULL; 3171 Node* shift = NULL; 3172 Node* offset = NULL; 3173 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 3174 return false; 3175 } 3176 3177 Node* exit = head->loopexit()->proj_out(0); 3178 if (exit == NULL) { 3179 return false; 3180 } 3181 3182 #ifndef PRODUCT 3183 if (TraceLoopOpts) { 3184 tty->print("ArrayFill "); 3185 lpt->dump_head(); 3186 } 3187 #endif 3188 3189 // Now replace the whole loop body by a call to a fill routine that 3190 // covers the same region as the loop. 3191 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 3192 3193 // Build an expression for the beginning of the copy region 3194 Node* index = head->init_trip(); 3195 #ifdef _LP64 3196 index = new ConvI2LNode(index); 3197 _igvn.register_new_node_with_optimizer(index); 3198 #endif 3199 if (shift != NULL) { 3200 // byte arrays don't require a shift but others do. 3201 index = new LShiftXNode(index, shift->in(2)); 3202 _igvn.register_new_node_with_optimizer(index); 3203 } 3204 index = new AddPNode(base, base, index); 3205 _igvn.register_new_node_with_optimizer(index); 3206 Node* from = new AddPNode(base, index, offset); 3207 _igvn.register_new_node_with_optimizer(from); 3208 // Compute the number of elements to copy 3209 Node* len = new SubINode(head->limit(), head->init_trip()); 3210 _igvn.register_new_node_with_optimizer(len); 3211 3212 BasicType t = store->as_Mem()->memory_type(); 3213 bool aligned = false; 3214 if (offset != NULL && head->init_trip()->is_Con()) { 3215 int element_size = type2aelembytes(t); 3216 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 3217 } 3218 3219 // Build a call to the fill routine 3220 const char* fill_name; 3221 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 3222 assert(fill != NULL, "what?"); 3223 3224 // Convert float/double to int/long for fill routines 3225 if (t == T_FLOAT) { 3226 store_value = new MoveF2INode(store_value); 3227 _igvn.register_new_node_with_optimizer(store_value); 3228 } else if (t == T_DOUBLE) { 3229 store_value = new MoveD2LNode(store_value); 3230 _igvn.register_new_node_with_optimizer(store_value); 3231 } 3232 3233 Node* mem_phi = store->in(MemNode::Memory); 3234 Node* result_ctrl; 3235 Node* result_mem; 3236 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 3237 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill, 3238 fill_name, TypeAryPtr::get_array_body_type(t)); 3239 uint cnt = 0; 3240 call->init_req(TypeFunc::Parms + cnt++, from); 3241 call->init_req(TypeFunc::Parms + cnt++, store_value); 3242 #ifdef _LP64 3243 len = new ConvI2LNode(len); 3244 _igvn.register_new_node_with_optimizer(len); 3245 #endif 3246 call->init_req(TypeFunc::Parms + cnt++, len); 3247 #ifdef _LP64 3248 call->init_req(TypeFunc::Parms + cnt++, C->top()); 3249 #endif 3250 call->init_req(TypeFunc::Control, head->init_control()); 3251 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. 3252 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); 3253 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr)); 3254 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr)); 3255 _igvn.register_new_node_with_optimizer(call); 3256 result_ctrl = new ProjNode(call,TypeFunc::Control); 3257 _igvn.register_new_node_with_optimizer(result_ctrl); 3258 result_mem = new ProjNode(call,TypeFunc::Memory); 3259 _igvn.register_new_node_with_optimizer(result_mem); 3260 3261 /* Disable following optimization until proper fix (add missing checks). 3262 3263 // If this fill is tightly coupled to an allocation and overwrites 3264 // the whole body, allow it to take over the zeroing. 3265 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 3266 if (alloc != NULL && alloc->is_AllocateArray()) { 3267 Node* length = alloc->as_AllocateArray()->Ideal_length(); 3268 if (head->limit() == length && 3269 head->init_trip() == _igvn.intcon(0)) { 3270 if (TraceOptimizeFill) { 3271 tty->print_cr("Eliminated zeroing in allocation"); 3272 } 3273 alloc->maybe_set_complete(&_igvn); 3274 } else { 3275 #ifdef ASSERT 3276 if (TraceOptimizeFill) { 3277 tty->print_cr("filling array but bounds don't match"); 3278 alloc->dump(); 3279 head->init_trip()->dump(); 3280 head->limit()->dump(); 3281 length->dump(); 3282 } 3283 #endif 3284 } 3285 } 3286 */ 3287 3288 // Redirect the old control and memory edges that are outside the loop. 3289 // Sometimes the memory phi of the head is used as the outgoing 3290 // state of the loop. It's safe in this case to replace it with the 3291 // result_mem. 3292 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 3293 lazy_replace(exit, result_ctrl); 3294 _igvn.replace_node(store, result_mem); 3295 // Any uses the increment outside of the loop become the loop limit. 3296 _igvn.replace_node(head->incr(), head->limit()); 3297 3298 // Disconnect the head from the loop. 3299 for (uint i = 0; i < lpt->_body.size(); i++) { 3300 Node* n = lpt->_body.at(i); 3301 _igvn.replace_node(n, C->top()); 3302 } 3303 3304 return true; 3305 }