1 /* 2 * Copyright (c) 2000, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/convertnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opaquenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 #include "opto/superword.hpp" 42 #include "opto/vectornode.hpp" 43 44 //------------------------------is_loop_exit----------------------------------- 45 // Given an IfNode, return the loop-exiting projection or NULL if both 46 // arms remain in the loop. 47 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 48 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 49 PhaseIdealLoop *phase = _phase; 50 // Test is an IfNode, has 2 projections. If BOTH are in the loop 51 // we need loop unswitching instead of peeling. 52 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 53 return iff->raw_out(0); 54 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 55 return iff->raw_out(1); 56 return NULL; 57 } 58 59 60 //============================================================================= 61 62 63 //------------------------------record_for_igvn---------------------------- 64 // Put loop body on igvn work list 65 void IdealLoopTree::record_for_igvn() { 66 for( uint i = 0; i < _body.size(); i++ ) { 67 Node *n = _body.at(i); 68 _phase->_igvn._worklist.push(n); 69 } 70 } 71 72 //------------------------------compute_exact_trip_count----------------------- 73 // Compute loop exact trip count if possible. Do not recalculate trip count for 74 // split loops (pre-main-post) which have their limits and inits behind Opaque node. 75 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) { 76 if (!_head->as_Loop()->is_valid_counted_loop()) { 77 return; 78 } 79 CountedLoopNode* cl = _head->as_CountedLoop(); 80 // Trip count may become nonexact for iteration split loops since 81 // RCE modifies limits. Note, _trip_count value is not reset since 82 // it is used to limit unrolling of main loop. 83 cl->set_nonexact_trip_count(); 84 85 // Loop's test should be part of loop. 86 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 87 return; // Infinite loop 88 89 #ifdef ASSERT 90 BoolTest::mask bt = cl->loopexit()->test_trip(); 91 assert(bt == BoolTest::lt || bt == BoolTest::gt || 92 bt == BoolTest::ne, "canonical test is expected"); 93 #endif 94 95 Node* init_n = cl->init_trip(); 96 Node* limit_n = cl->limit(); 97 if (init_n != NULL && init_n->is_Con() && 98 limit_n != NULL && limit_n->is_Con()) { 99 // Use longs to avoid integer overflow. 100 int stride_con = cl->stride_con(); 101 jlong init_con = cl->init_trip()->get_int(); 102 jlong limit_con = cl->limit()->get_int(); 103 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 104 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 105 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { 106 // Set exact trip count. 107 cl->set_exact_trip_count((uint)trip_count); 108 } 109 } 110 } 111 112 //------------------------------compute_profile_trip_cnt---------------------------- 113 // Compute loop trip count from profile data as 114 // (backedge_count + loop_exit_count) / loop_exit_count 115 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 116 if (!_head->is_CountedLoop()) { 117 return; 118 } 119 CountedLoopNode* head = _head->as_CountedLoop(); 120 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 121 return; // Already computed 122 } 123 float trip_cnt = (float)max_jint; // default is big 124 125 Node* back = head->in(LoopNode::LoopBackControl); 126 while (back != head) { 127 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 128 back->in(0) && 129 back->in(0)->is_If() && 130 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 131 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 132 break; 133 } 134 back = phase->idom(back); 135 } 136 if (back != head) { 137 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 138 back->in(0), "if-projection exists"); 139 IfNode* back_if = back->in(0)->as_If(); 140 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 141 142 // Now compute a loop exit count 143 float loop_exit_cnt = 0.0f; 144 for( uint i = 0; i < _body.size(); i++ ) { 145 Node *n = _body[i]; 146 if( n->is_If() ) { 147 IfNode *iff = n->as_If(); 148 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 149 Node *exit = is_loop_exit(iff); 150 if( exit ) { 151 float exit_prob = iff->_prob; 152 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 153 if (exit_prob > PROB_MIN) { 154 float exit_cnt = iff->_fcnt * exit_prob; 155 loop_exit_cnt += exit_cnt; 156 } 157 } 158 } 159 } 160 } 161 if (loop_exit_cnt > 0.0f) { 162 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 163 } else { 164 // No exit count so use 165 trip_cnt = loop_back_cnt; 166 } 167 } 168 #ifndef PRODUCT 169 if (TraceProfileTripCount) { 170 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 171 } 172 #endif 173 head->set_profile_trip_cnt(trip_cnt); 174 } 175 176 //---------------------is_invariant_addition----------------------------- 177 // Return nonzero index of invariant operand for an Add or Sub 178 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 179 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 180 int op = n->Opcode(); 181 if (op == Op_AddI || op == Op_SubI) { 182 bool in1_invar = this->is_invariant(n->in(1)); 183 bool in2_invar = this->is_invariant(n->in(2)); 184 if (in1_invar && !in2_invar) return 1; 185 if (!in1_invar && in2_invar) return 2; 186 } 187 return 0; 188 } 189 190 //---------------------reassociate_add_sub----------------------------- 191 // Reassociate invariant add and subtract expressions: 192 // 193 // inv1 + (x + inv2) => ( inv1 + inv2) + x 194 // (x + inv2) + inv1 => ( inv1 + inv2) + x 195 // inv1 + (x - inv2) => ( inv1 - inv2) + x 196 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 197 // (x + inv2) - inv1 => (-inv1 + inv2) + x 198 // (x - inv2) + inv1 => ( inv1 - inv2) + x 199 // (x - inv2) - inv1 => (-inv1 - inv2) + x 200 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 201 // inv1 - (x - inv2) => ( inv1 + inv2) - x 202 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 203 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 204 // inv1 - (x + inv2) => ( inv1 - inv2) - x 205 // 206 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 207 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 208 if (is_invariant(n1)) return NULL; 209 int inv1_idx = is_invariant_addition(n1, phase); 210 if (!inv1_idx) return NULL; 211 // Don't mess with add of constant (igvn moves them to expression tree root.) 212 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 213 Node* inv1 = n1->in(inv1_idx); 214 Node* n2 = n1->in(3 - inv1_idx); 215 int inv2_idx = is_invariant_addition(n2, phase); 216 if (!inv2_idx) return NULL; 217 Node* x = n2->in(3 - inv2_idx); 218 Node* inv2 = n2->in(inv2_idx); 219 220 bool neg_x = n2->is_Sub() && inv2_idx == 1; 221 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 222 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 223 if (n1->is_Sub() && inv1_idx == 1) { 224 neg_x = !neg_x; 225 neg_inv2 = !neg_inv2; 226 } 227 Node* inv1_c = phase->get_ctrl(inv1); 228 Node* inv2_c = phase->get_ctrl(inv2); 229 Node* n_inv1; 230 if (neg_inv1) { 231 Node *zero = phase->_igvn.intcon(0); 232 phase->set_ctrl(zero, phase->C->root()); 233 n_inv1 = new SubINode(zero, inv1); 234 phase->register_new_node(n_inv1, inv1_c); 235 } else { 236 n_inv1 = inv1; 237 } 238 Node* inv; 239 if (neg_inv2) { 240 inv = new SubINode(n_inv1, inv2); 241 } else { 242 inv = new AddINode(n_inv1, inv2); 243 } 244 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 245 246 Node* addx; 247 if (neg_x) { 248 addx = new SubINode(inv, x); 249 } else { 250 addx = new AddINode(x, inv); 251 } 252 phase->register_new_node(addx, phase->get_ctrl(x)); 253 phase->_igvn.replace_node(n1, addx); 254 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 255 _body.yank(n1); 256 return addx; 257 } 258 259 //---------------------reassociate_invariants----------------------------- 260 // Reassociate invariant expressions: 261 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 262 for (int i = _body.size() - 1; i >= 0; i--) { 263 Node *n = _body.at(i); 264 for (int j = 0; j < 5; j++) { 265 Node* nn = reassociate_add_sub(n, phase); 266 if (nn == NULL) break; 267 n = nn; // again 268 }; 269 } 270 } 271 272 //------------------------------policy_peeling--------------------------------- 273 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 274 // make some loop-invariant test (usually a null-check) happen before the loop. 275 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 276 Node *test = ((IdealLoopTree*)this)->tail(); 277 int body_size = ((IdealLoopTree*)this)->_body.size(); 278 // Peeling does loop cloning which can result in O(N^2) node construction 279 if( body_size > 255 /* Prevent overflow for large body_size */ 280 || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) { 281 return false; // too large to safely clone 282 } 283 while( test != _head ) { // Scan till run off top of loop 284 if( test->is_If() ) { // Test? 285 Node *ctrl = phase->get_ctrl(test->in(1)); 286 if (ctrl->is_top()) 287 return false; // Found dead test on live IF? No peeling! 288 // Standard IF only has one input value to check for loop invariance 289 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 290 // Condition is not a member of this loop? 291 if( !is_member(phase->get_loop(ctrl)) && 292 is_loop_exit(test) ) 293 return true; // Found reason to peel! 294 } 295 // Walk up dominators to loop _head looking for test which is 296 // executed on every path thru loop. 297 test = phase->idom(test); 298 } 299 return false; 300 } 301 302 //------------------------------peeled_dom_test_elim--------------------------- 303 // If we got the effect of peeling, either by actually peeling or by making 304 // a pre-loop which must execute at least once, we can remove all 305 // loop-invariant dominated tests in the main body. 306 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 307 bool progress = true; 308 while( progress ) { 309 progress = false; // Reset for next iteration 310 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 311 Node *test = prev->in(0); 312 while( test != loop->_head ) { // Scan till run off top of loop 313 314 int p_op = prev->Opcode(); 315 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 316 test->is_If() && // Test? 317 !test->in(1)->is_Con() && // And not already obvious? 318 // Condition is not a member of this loop? 319 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 320 // Walk loop body looking for instances of this test 321 for( uint i = 0; i < loop->_body.size(); i++ ) { 322 Node *n = loop->_body.at(i); 323 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 324 // IfNode was dominated by version in peeled loop body 325 progress = true; 326 dominated_by( old_new[prev->_idx], n ); 327 } 328 } 329 } 330 prev = test; 331 test = idom(test); 332 } // End of scan tests in loop 333 334 } // End of while( progress ) 335 } 336 337 //------------------------------do_peeling------------------------------------- 338 // Peel the first iteration of the given loop. 339 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 340 // The pre-loop illegally has 2 control users (old & new loops). 341 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 342 // Do this by making the old-loop fall-in edges act as if they came 343 // around the loopback from the prior iteration (follow the old-loop 344 // backedges) and then map to the new peeled iteration. This leaves 345 // the pre-loop with only 1 user (the new peeled iteration), but the 346 // peeled-loop backedge has 2 users. 347 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 348 // extra backedge user. 349 // 350 // orig 351 // 352 // stmt1 353 // | 354 // v 355 // loop predicate 356 // | 357 // v 358 // loop<----+ 359 // | | 360 // stmt2 | 361 // | | 362 // v | 363 // if ^ 364 // / \ | 365 // / \ | 366 // v v | 367 // false true | 368 // / \ | 369 // / ----+ 370 // | 371 // v 372 // exit 373 // 374 // 375 // after clone loop 376 // 377 // stmt1 378 // | 379 // v 380 // loop predicate 381 // / \ 382 // clone / \ orig 383 // / \ 384 // / \ 385 // v v 386 // +---->loop clone loop<----+ 387 // | | | | 388 // | stmt2 clone stmt2 | 389 // | | | | 390 // | v v | 391 // ^ if clone If ^ 392 // | / \ / \ | 393 // | / \ / \ | 394 // | v v v v | 395 // | true false false true | 396 // | / \ / \ | 397 // +---- \ / ----+ 398 // \ / 399 // 1v v2 400 // region 401 // | 402 // v 403 // exit 404 // 405 // 406 // after peel and predicate move 407 // 408 // stmt1 409 // / 410 // / 411 // clone / orig 412 // / 413 // / +----------+ 414 // / | | 415 // / loop predicate | 416 // / | | 417 // v v | 418 // TOP-->loop clone loop<----+ | 419 // | | | | 420 // stmt2 clone stmt2 | | 421 // | | | ^ 422 // v v | | 423 // if clone If ^ | 424 // / \ / \ | | 425 // / \ / \ | | 426 // v v v v | | 427 // true false false true | | 428 // | \ / \ | | 429 // | \ / ----+ ^ 430 // | \ / | 431 // | 1v v2 | 432 // v region | 433 // | | | 434 // | v | 435 // | exit | 436 // | | 437 // +--------------->-----------------+ 438 // 439 // 440 // final graph 441 // 442 // stmt1 443 // | 444 // v 445 // stmt2 clone 446 // | 447 // v 448 // if clone 449 // / | 450 // / | 451 // v v 452 // false true 453 // | | 454 // | v 455 // | loop predicate 456 // | | 457 // | v 458 // | loop<----+ 459 // | | | 460 // | stmt2 | 461 // | | | 462 // | v | 463 // v if ^ 464 // | / \ | 465 // | / \ | 466 // | v v | 467 // | false true | 468 // | | \ | 469 // v v --+ 470 // region 471 // | 472 // v 473 // exit 474 // 475 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 476 477 C->set_major_progress(); 478 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 479 // 'pre' loop from the main and the 'pre' can no longer have it's 480 // iterations adjusted. Therefore, we need to declare this loop as 481 // no longer a 'main' loop; it will need new pre and post loops before 482 // we can do further RCE. 483 #ifndef PRODUCT 484 if (TraceLoopOpts) { 485 tty->print("Peel "); 486 loop->dump_head(); 487 } 488 #endif 489 Node* head = loop->_head; 490 bool counted_loop = head->is_CountedLoop(); 491 if (counted_loop) { 492 CountedLoopNode *cl = head->as_CountedLoop(); 493 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 494 cl->set_trip_count(cl->trip_count() - 1); 495 if (cl->is_main_loop()) { 496 cl->set_normal_loop(); 497 #ifndef PRODUCT 498 if (PrintOpto && VerifyLoopOptimizations) { 499 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 500 loop->dump_head(); 501 } 502 #endif 503 } 504 } 505 Node* entry = head->in(LoopNode::EntryControl); 506 507 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 508 // The pre-loop illegally has 2 control users (old & new loops). 509 clone_loop( loop, old_new, dom_depth(head) ); 510 511 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 512 // Do this by making the old-loop fall-in edges act as if they came 513 // around the loopback from the prior iteration (follow the old-loop 514 // backedges) and then map to the new peeled iteration. This leaves 515 // the pre-loop with only 1 user (the new peeled iteration), but the 516 // peeled-loop backedge has 2 users. 517 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 518 _igvn.hash_delete(head); 519 head->set_req(LoopNode::EntryControl, new_entry); 520 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 521 Node* old = head->fast_out(j); 522 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 523 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 524 if (!new_exit_value ) // Backedge value is ALSO loop invariant? 525 // Then loop body backedge value remains the same. 526 new_exit_value = old->in(LoopNode::LoopBackControl); 527 _igvn.hash_delete(old); 528 old->set_req(LoopNode::EntryControl, new_exit_value); 529 } 530 } 531 532 533 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 534 // extra backedge user. 535 Node* new_head = old_new[head->_idx]; 536 _igvn.hash_delete(new_head); 537 new_head->set_req(LoopNode::LoopBackControl, C->top()); 538 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 539 Node* use = new_head->fast_out(j2); 540 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 541 _igvn.hash_delete(use); 542 use->set_req(LoopNode::LoopBackControl, C->top()); 543 } 544 } 545 546 547 // Step 4: Correct dom-depth info. Set to loop-head depth. 548 int dd = dom_depth(head); 549 set_idom(head, head->in(1), dd); 550 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 551 Node *old = loop->_body.at(j3); 552 Node *nnn = old_new[old->_idx]; 553 if (!has_ctrl(nnn)) 554 set_idom(nnn, idom(nnn), dd-1); 555 } 556 557 // Now force out all loop-invariant dominating tests. The optimizer 558 // finds some, but we _know_ they are all useless. 559 peeled_dom_test_elim(loop,old_new); 560 561 loop->record_for_igvn(); 562 } 563 564 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 565 566 //------------------------------policy_maximally_unroll------------------------ 567 // Calculate exact loop trip count and return true if loop can be maximally 568 // unrolled. 569 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 570 CountedLoopNode *cl = _head->as_CountedLoop(); 571 assert(cl->is_normal_loop(), ""); 572 if (!cl->is_valid_counted_loop()) 573 return false; // Malformed counted loop 574 575 if (!cl->has_exact_trip_count()) { 576 // Trip count is not exact. 577 return false; 578 } 579 580 uint trip_count = cl->trip_count(); 581 // Note, max_juint is used to indicate unknown trip count. 582 assert(trip_count > 1, "one iteration loop should be optimized out already"); 583 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 584 585 // Real policy: if we maximally unroll, does it get too big? 586 // Allow the unrolled mess to get larger than standard loop 587 // size. After all, it will no longer be a loop. 588 uint body_size = _body.size(); 589 uint unroll_limit = (uint)LoopUnrollLimit * 4; 590 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 591 if (trip_count > unroll_limit || body_size > unroll_limit) { 592 return false; 593 } 594 595 // Fully unroll a loop with few iterations regardless next 596 // conditions since following loop optimizations will split 597 // such loop anyway (pre-main-post). 598 if (trip_count <= 3) 599 return true; 600 601 // Take into account that after unroll conjoined heads and tails will fold, 602 // otherwise policy_unroll() may allow more unrolling than max unrolling. 603 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; 604 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; 605 if (body_size != tst_body_size) // Check for int overflow 606 return false; 607 if (new_body_size > unroll_limit || 608 // Unrolling can result in a large amount of node construction 609 new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) { 610 return false; 611 } 612 613 // Do not unroll a loop with String intrinsics code. 614 // String intrinsics are large and have loops. 615 for (uint k = 0; k < _body.size(); k++) { 616 Node* n = _body.at(k); 617 switch (n->Opcode()) { 618 case Op_StrComp: 619 case Op_StrEquals: 620 case Op_StrIndexOf: 621 case Op_EncodeISOArray: 622 case Op_AryEq: { 623 return false; 624 } 625 #if INCLUDE_RTM_OPT 626 case Op_FastLock: 627 case Op_FastUnlock: { 628 // Don't unroll RTM locking code because it is large. 629 if (UseRTMLocking) { 630 return false; 631 } 632 } 633 #endif 634 } // switch 635 } 636 637 return true; // Do maximally unroll 638 } 639 640 641 //------------------------------policy_unroll---------------------------------- 642 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 643 // the loop is a CountedLoop and the body is small enough. 644 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) { 645 646 CountedLoopNode *cl = _head->as_CountedLoop(); 647 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 648 649 if (!cl->is_valid_counted_loop()) 650 return false; // Malformed counted loop 651 652 // Protect against over-unrolling. 653 // After split at least one iteration will be executed in pre-loop. 654 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; 655 656 _local_loop_unroll_limit = LoopUnrollLimit; 657 _local_loop_unroll_factor = 4; 658 int future_unroll_ct = cl->unrolled_count() * 2; 659 if (future_unroll_ct > LoopMaxUnroll) return false; 660 661 if (UseSuperWord) { 662 if (cl->is_reduction_loop() == false) phase->mark_reductions(this); 663 664 // Only attempt slp analysis when user controls do not prohibit it 665 if (LoopMaxUnroll > _local_loop_unroll_factor) { 666 // Once policy_slp_analysis succeeds, mark the loop with the 667 // maximal unroll factor so that we minimize analysis passes 668 if (cl->has_passed_slp() == false) { 669 if (policy_slp_analysis(cl, phase)) { 670 if (_local_loop_unroll_factor > 4) { 671 cl->mark_passed_slp(); 672 cl->set_slp_max_unroll(_local_loop_unroll_factor); 673 } 674 } 675 } 676 677 if (cl->has_passed_slp()) { 678 int slp_max_unroll_factor = cl->slp_max_unroll(); 679 if ((slp_max_unroll_factor > 4) && 680 (slp_max_unroll_factor >= future_unroll_ct)) { 681 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor; 682 if (new_limit > LoopUnrollLimit) { 683 #ifndef PRODUCT 684 if (TraceSuperWordLoopUnrollAnalysis) { 685 tty->print_cr("slp analysis is applying unroll limit %d, the original limit was %d\n", 686 new_limit, _local_loop_unroll_limit); 687 } 688 #endif 689 _local_loop_unroll_limit = new_limit; 690 } 691 } 692 } 693 } 694 } 695 696 // Check for initial stride being a small enough constant 697 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; 698 699 // Don't unroll if the next round of unrolling would push us 700 // over the expected trip count of the loop. One is subtracted 701 // from the expected trip count because the pre-loop normally 702 // executes 1 iteration. 703 if (UnrollLimitForProfileCheck > 0 && 704 cl->profile_trip_cnt() != COUNT_UNKNOWN && 705 future_unroll_ct > UnrollLimitForProfileCheck && 706 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 707 return false; 708 } 709 710 // When unroll count is greater than LoopUnrollMin, don't unroll if: 711 // the residual iterations are more than 10% of the trip count 712 // and rounds of "unroll,optimize" are not making significant progress 713 // Progress defined as current size less than 20% larger than previous size. 714 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 715 future_unroll_ct > LoopUnrollMin && 716 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 717 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 718 return false; 719 } 720 721 Node *init_n = cl->init_trip(); 722 Node *limit_n = cl->limit(); 723 int stride_con = cl->stride_con(); 724 // Non-constant bounds. 725 // Protect against over-unrolling when init or/and limit are not constant 726 // (so that trip_count's init value is maxint) but iv range is known. 727 if (init_n == NULL || !init_n->is_Con() || 728 limit_n == NULL || !limit_n->is_Con()) { 729 Node* phi = cl->phi(); 730 if (phi != NULL) { 731 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 732 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 733 int next_stride = stride_con * 2; // stride after this unroll 734 if (next_stride > 0) { 735 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 736 iv_type->_lo + next_stride > iv_type->_hi) { 737 return false; // over-unrolling 738 } 739 } else if (next_stride < 0) { 740 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 741 iv_type->_hi + next_stride < iv_type->_lo) { 742 return false; // over-unrolling 743 } 744 } 745 } 746 } 747 748 // After unroll limit will be adjusted: new_limit = limit-stride. 749 // Bailout if adjustment overflow. 750 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 751 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || 752 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) 753 return false; // overflow 754 755 // Adjust body_size to determine if we unroll or not 756 uint body_size = _body.size(); 757 // Key test to unroll loop in CRC32 java code 758 int xors_in_loop = 0; 759 // Also count ModL, DivL and MulL which expand mightly 760 for (uint k = 0; k < _body.size(); k++) { 761 Node* n = _body.at(k); 762 switch (n->Opcode()) { 763 case Op_XorI: xors_in_loop++; break; // CRC32 java code 764 case Op_ModL: body_size += 30; break; 765 case Op_DivL: body_size += 30; break; 766 case Op_MulL: body_size += 10; break; 767 case Op_StrComp: 768 case Op_StrEquals: 769 case Op_StrIndexOf: 770 case Op_EncodeISOArray: 771 case Op_AryEq: { 772 // Do not unroll a loop with String intrinsics code. 773 // String intrinsics are large and have loops. 774 return false; 775 } 776 #if INCLUDE_RTM_OPT 777 case Op_FastLock: 778 case Op_FastUnlock: { 779 // Don't unroll RTM locking code because it is large. 780 if (UseRTMLocking) { 781 return false; 782 } 783 } 784 #endif 785 } // switch 786 } 787 788 // Check for being too big 789 if (body_size > (uint)_local_loop_unroll_limit) { 790 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 791 // Normal case: loop too big 792 return false; 793 } 794 795 // Unroll once! (Each trip will soon do double iterations) 796 return true; 797 } 798 799 bool IdealLoopTree::policy_slp_analysis( CountedLoopNode *cl, PhaseIdealLoop *phase ) { 800 // SLP analysis 801 bool not_slp = false; 802 803 // Enable this functionality target by target as needed 804 if (SuperWordLoopUnrollAnalysis) { 805 SuperWord sw(phase); 806 sw.transform_loop(this, false); 807 808 // If the loop is slp canonical analyze it 809 if (sw.early_return() == false) { 810 Arena *a = Thread::current()->resource_area(); 811 int max_vector = Matcher::max_vector_size(T_INT); 812 size_t ignored_size = _body.size()*sizeof(int*); 813 int *ignored_loop_nodes = (int*)a->Amalloc_D(ignored_size); 814 Node_Stack nstack((int)ignored_size); 815 Node *cl_exit = cl->loopexit(); 816 817 // First clear the entries 818 for (uint i = 0; i < _body.size(); i++) { 819 ignored_loop_nodes[i] = -1; 820 } 821 822 // Process the loop, some/all of the stack entries will not be in order, ergo 823 // need to preprocess the ignored initial state before we process the loop 824 for (uint i = 0; i < _body.size(); i++) { 825 Node* n = _body.at(i); 826 if (n == cl->incr() || 827 n->is_reduction() || 828 n->is_AddP() || 829 n->is_Cmp() || 830 n->is_IfTrue() || 831 n->is_CountedLoop() || 832 (n == cl_exit)) { 833 ignored_loop_nodes[i] = n->_idx; 834 continue; 835 } 836 837 if (n->is_If()) { 838 IfNode *iff = n->as_If(); 839 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) { 840 if (is_loop_exit(iff)) { 841 ignored_loop_nodes[i] = n->_idx; 842 continue; 843 } 844 } 845 } 846 847 if (n->is_Phi() && (n->bottom_type() == Type::MEMORY)) { 848 Node* n_tail = n->in(LoopNode::LoopBackControl); 849 if (n_tail != n->in(LoopNode::EntryControl)) { 850 if (!n_tail->is_Mem()) { 851 not_slp = true; 852 break; 853 } 854 } 855 } 856 857 // This must happen after check of phi/if 858 if (n->is_Phi() || n->is_If()) { 859 ignored_loop_nodes[i] = n->_idx; 860 continue; 861 } 862 863 if (n->is_LoadStore() || n->is_MergeMem() || 864 (n->is_Proj() && !n->as_Proj()->is_CFG())) { 865 not_slp = true; 866 break; 867 } 868 869 if (n->is_Mem()) { 870 Node* adr = n->in(MemNode::Address); 871 Node* n_ctrl = phase->get_ctrl(adr); 872 873 // save a queue of post process nodes 874 if (n_ctrl != NULL && is_member(phase->get_loop(n_ctrl))) { 875 MemNode* current = n->as_Mem(); 876 BasicType bt = current->memory_type(); 877 if (is_java_primitive(bt) == false) { 878 ignored_loop_nodes[i] = n->_idx; 879 continue; 880 } 881 882 // Process the memory expression 883 int stack_idx = 0; 884 bool have_side_effects = true; 885 if (adr->is_AddP() == false) { 886 nstack.push(adr, stack_idx++); 887 } else { 888 // Mark the components of the memory operation in nstack 889 SWPointer p1(current, &sw, &nstack, true); 890 have_side_effects = p1.node_stack()->is_nonempty(); 891 } 892 893 // Process the pointer stack 894 while (have_side_effects) { 895 Node* pointer_node = nstack.node(); 896 for (uint j = 0; j < _body.size(); j++) { 897 Node* cur_node = _body.at(j); 898 if (cur_node == pointer_node) { 899 ignored_loop_nodes[j] = cur_node->_idx; 900 break; 901 } 902 } 903 nstack.pop(); 904 have_side_effects = nstack.is_nonempty(); 905 } 906 907 // Cleanup 908 nstack.clear(); 909 } 910 } 911 } 912 913 if (not_slp == false) { 914 // Now we try to find the maximum supported consistent vector which the machine 915 // description can use 916 for (uint i = 0; i < _body.size(); i++) { 917 if (ignored_loop_nodes[i] != -1) continue; 918 919 BasicType bt; 920 Node* n = _body.at(i); 921 if (n->is_Store()) { 922 bt = n->as_Mem()->memory_type(); 923 } else { 924 bt = n->bottom_type()->basic_type(); 925 } 926 927 int cur_max_vector = Matcher::max_vector_size(bt); 928 929 // If a max vector exists which is not larger than _local_loop_unroll_factor 930 // stop looking, we already have the max vector to map to. 931 if (cur_max_vector <= _local_loop_unroll_factor) { 932 not_slp = true; 933 #ifndef PRODUCT 934 if (TraceSuperWordLoopUnrollAnalysis) { 935 tty->print_cr("slp analysis fails: unroll limit equals max vector\n"); 936 } 937 #endif 938 break; 939 } 940 941 // Map the maximal common vector 942 if (VectorNode::implemented(n->Opcode(), cur_max_vector, bt)) { 943 if (cur_max_vector < max_vector) { 944 max_vector = cur_max_vector; 945 } 946 } 947 } 948 if (not_slp == false) _local_loop_unroll_factor = max_vector; 949 } 950 951 if (not_slp) { 952 // Mark the loop as processed so that we do not try again 953 cl->mark_passed_slp(); 954 cl->set_slp_max_unroll(_local_loop_unroll_factor); 955 } 956 957 // Now clean things up 958 a->Afree(ignored_loop_nodes, ignored_size); 959 } 960 } 961 962 return (not_slp == false); 963 } 964 965 //------------------------------policy_align----------------------------------- 966 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 967 // expression that does the alignment. Note that only one array base can be 968 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 969 // if we vectorize short memory ops into longer memory ops, we may want to 970 // increase alignment. 971 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 972 return false; 973 } 974 975 //------------------------------policy_range_check----------------------------- 976 // Return TRUE or FALSE if the loop should be range-check-eliminated. 977 // Actually we do iteration-splitting, a more powerful form of RCE. 978 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 979 if (!RangeCheckElimination) return false; 980 981 CountedLoopNode *cl = _head->as_CountedLoop(); 982 // If we unrolled with no intention of doing RCE and we later 983 // changed our minds, we got no pre-loop. Either we need to 984 // make a new pre-loop, or we gotta disallow RCE. 985 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 986 Node *trip_counter = cl->phi(); 987 988 // Check loop body for tests of trip-counter plus loop-invariant vs 989 // loop-invariant. 990 for (uint i = 0; i < _body.size(); i++) { 991 Node *iff = _body[i]; 992 if (iff->Opcode() == Op_If) { // Test? 993 994 // Comparing trip+off vs limit 995 Node *bol = iff->in(1); 996 if (bol->req() != 2) continue; // dead constant test 997 if (!bol->is_Bool()) { 998 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 999 continue; 1000 } 1001 if (bol->as_Bool()->_test._test == BoolTest::ne) 1002 continue; // not RC 1003 1004 Node *cmp = bol->in(1); 1005 Node *rc_exp = cmp->in(1); 1006 Node *limit = cmp->in(2); 1007 1008 Node *limit_c = phase->get_ctrl(limit); 1009 if( limit_c == phase->C->top() ) 1010 return false; // Found dead test on live IF? No RCE! 1011 if( is_member(phase->get_loop(limit_c) ) ) { 1012 // Compare might have operands swapped; commute them 1013 rc_exp = cmp->in(2); 1014 limit = cmp->in(1); 1015 limit_c = phase->get_ctrl(limit); 1016 if( is_member(phase->get_loop(limit_c) ) ) 1017 continue; // Both inputs are loop varying; cannot RCE 1018 } 1019 1020 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 1021 continue; 1022 } 1023 // Yeah! Found a test like 'trip+off vs limit' 1024 // Test is an IfNode, has 2 projections. If BOTH are in the loop 1025 // we need loop unswitching instead of iteration splitting. 1026 if( is_loop_exit(iff) ) 1027 return true; // Found reason to split iterations 1028 } // End of is IF 1029 } 1030 1031 return false; 1032 } 1033 1034 //------------------------------policy_peel_only------------------------------- 1035 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 1036 // for unrolling loops with NO array accesses. 1037 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 1038 1039 for( uint i = 0; i < _body.size(); i++ ) 1040 if( _body[i]->is_Mem() ) 1041 return false; 1042 1043 // No memory accesses at all! 1044 return true; 1045 } 1046 1047 //------------------------------clone_up_backedge_goo-------------------------- 1048 // If Node n lives in the back_ctrl block and cannot float, we clone a private 1049 // version of n in preheader_ctrl block and return that, otherwise return n. 1050 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { 1051 if( get_ctrl(n) != back_ctrl ) return n; 1052 1053 // Only visit once 1054 if (visited.test_set(n->_idx)) { 1055 Node *x = clones.find(n->_idx); 1056 if (x != NULL) 1057 return x; 1058 return n; 1059 } 1060 1061 Node *x = NULL; // If required, a clone of 'n' 1062 // Check for 'n' being pinned in the backedge. 1063 if( n->in(0) && n->in(0) == back_ctrl ) { 1064 assert(clones.find(n->_idx) == NULL, "dead loop"); 1065 x = n->clone(); // Clone a copy of 'n' to preheader 1066 clones.push(x, n->_idx); 1067 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 1068 } 1069 1070 // Recursive fixup any other input edges into x. 1071 // If there are no changes we can just return 'n', otherwise 1072 // we need to clone a private copy and change it. 1073 for( uint i = 1; i < n->req(); i++ ) { 1074 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); 1075 if( g != n->in(i) ) { 1076 if( !x ) { 1077 assert(clones.find(n->_idx) == NULL, "dead loop"); 1078 x = n->clone(); 1079 clones.push(x, n->_idx); 1080 } 1081 x->set_req(i, g); 1082 } 1083 } 1084 if( x ) { // x can legally float to pre-header location 1085 register_new_node( x, preheader_ctrl ); 1086 return x; 1087 } else { // raise n to cover LCA of uses 1088 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 1089 } 1090 return n; 1091 } 1092 1093 bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { 1094 Node* castii = new CastIINode(incr, TypeInt::INT, true); 1095 castii->set_req(0, ctrl); 1096 register_new_node(castii, ctrl); 1097 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { 1098 Node* n = incr->fast_out(i); 1099 if (n->is_Phi() && n->in(0) == loop) { 1100 int nrep = n->replace_edge(incr, castii); 1101 return true; 1102 } 1103 } 1104 return false; 1105 } 1106 1107 //------------------------------insert_pre_post_loops-------------------------- 1108 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 1109 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 1110 // alignment. Useful to unroll loops that do no array accesses. 1111 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 1112 1113 #ifndef PRODUCT 1114 if (TraceLoopOpts) { 1115 if (peel_only) 1116 tty->print("PeelMainPost "); 1117 else 1118 tty->print("PreMainPost "); 1119 loop->dump_head(); 1120 } 1121 #endif 1122 C->set_major_progress(); 1123 1124 // Find common pieces of the loop being guarded with pre & post loops 1125 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1126 assert( main_head->is_normal_loop(), "" ); 1127 CountedLoopEndNode *main_end = main_head->loopexit(); 1128 guarantee(main_end != NULL, "no loop exit node"); 1129 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 1130 uint dd_main_head = dom_depth(main_head); 1131 uint max = main_head->outcnt(); 1132 1133 Node *pre_header= main_head->in(LoopNode::EntryControl); 1134 Node *init = main_head->init_trip(); 1135 Node *incr = main_end ->incr(); 1136 Node *limit = main_end ->limit(); 1137 Node *stride = main_end ->stride(); 1138 Node *cmp = main_end ->cmp_node(); 1139 BoolTest::mask b_test = main_end->test_trip(); 1140 1141 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 1142 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 1143 if( bol->outcnt() != 1 ) { 1144 bol = bol->clone(); 1145 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 1146 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol); 1147 } 1148 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 1149 if( cmp->outcnt() != 1 ) { 1150 cmp = cmp->clone(); 1151 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 1152 _igvn.replace_input_of(bol, 1, cmp); 1153 } 1154 1155 //------------------------------ 1156 // Step A: Create Post-Loop. 1157 Node* main_exit = main_end->proj_out(false); 1158 assert( main_exit->Opcode() == Op_IfFalse, "" ); 1159 int dd_main_exit = dom_depth(main_exit); 1160 1161 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 1162 // loop pre-header illegally has 2 control users (old & new loops). 1163 clone_loop( loop, old_new, dd_main_exit ); 1164 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 1165 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 1166 post_head->set_post_loop(main_head); 1167 1168 // Reduce the post-loop trip count. 1169 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1170 post_end->_prob = PROB_FAIR; 1171 1172 // Build the main-loop normal exit. 1173 IfFalseNode *new_main_exit = new IfFalseNode(main_end); 1174 _igvn.register_new_node_with_optimizer( new_main_exit ); 1175 set_idom(new_main_exit, main_end, dd_main_exit ); 1176 set_loop(new_main_exit, loop->_parent); 1177 1178 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 1179 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 1180 // (the main-loop trip-counter exit value) because we will be changing 1181 // the exit value (via unrolling) so we cannot constant-fold away the zero 1182 // trip guard until all unrolling is done. 1183 Node *zer_opaq = new Opaque1Node(C, incr); 1184 Node *zer_cmp = new CmpINode( zer_opaq, limit ); 1185 Node *zer_bol = new BoolNode( zer_cmp, b_test ); 1186 register_new_node( zer_opaq, new_main_exit ); 1187 register_new_node( zer_cmp , new_main_exit ); 1188 register_new_node( zer_bol , new_main_exit ); 1189 1190 // Build the IfNode 1191 IfNode *zer_iff = new IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 1192 _igvn.register_new_node_with_optimizer( zer_iff ); 1193 set_idom(zer_iff, new_main_exit, dd_main_exit); 1194 set_loop(zer_iff, loop->_parent); 1195 1196 // Plug in the false-path, taken if we need to skip post-loop 1197 _igvn.replace_input_of(main_exit, 0, zer_iff); 1198 set_idom(main_exit, zer_iff, dd_main_exit); 1199 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 1200 // Make the true-path, must enter the post loop 1201 Node *zer_taken = new IfTrueNode( zer_iff ); 1202 _igvn.register_new_node_with_optimizer( zer_taken ); 1203 set_idom(zer_taken, zer_iff, dd_main_exit); 1204 set_loop(zer_taken, loop->_parent); 1205 // Plug in the true path 1206 _igvn.hash_delete( post_head ); 1207 post_head->set_req(LoopNode::EntryControl, zer_taken); 1208 set_idom(post_head, zer_taken, dd_main_exit); 1209 1210 Arena *a = Thread::current()->resource_area(); 1211 VectorSet visited(a); 1212 Node_Stack clones(a, main_head->back_control()->outcnt()); 1213 // Step A3: Make the fall-in values to the post-loop come from the 1214 // fall-out values of the main-loop. 1215 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 1216 Node* main_phi = main_head->fast_out(i); 1217 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 1218 Node *post_phi = old_new[main_phi->_idx]; 1219 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 1220 post_head->init_control(), 1221 main_phi->in(LoopNode::LoopBackControl), 1222 visited, clones); 1223 _igvn.hash_delete(post_phi); 1224 post_phi->set_req( LoopNode::EntryControl, fallmain ); 1225 } 1226 } 1227 1228 // Update local caches for next stanza 1229 main_exit = new_main_exit; 1230 1231 1232 //------------------------------ 1233 // Step B: Create Pre-Loop. 1234 1235 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 1236 // loop pre-header illegally has 2 control users (old & new loops). 1237 clone_loop( loop, old_new, dd_main_head ); 1238 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1239 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1240 pre_head->set_pre_loop(main_head); 1241 Node *pre_incr = old_new[incr->_idx]; 1242 1243 // Reduce the pre-loop trip count. 1244 pre_end->_prob = PROB_FAIR; 1245 1246 // Find the pre-loop normal exit. 1247 Node* pre_exit = pre_end->proj_out(false); 1248 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 1249 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end); 1250 _igvn.register_new_node_with_optimizer( new_pre_exit ); 1251 set_idom(new_pre_exit, pre_end, dd_main_head); 1252 set_loop(new_pre_exit, loop->_parent); 1253 1254 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1255 // pre-loop, the main-loop may not execute at all. Later in life this 1256 // zero-trip guard will become the minimum-trip guard when we unroll 1257 // the main-loop. 1258 Node *min_opaq = new Opaque1Node(C, limit); 1259 Node *min_cmp = new CmpINode( pre_incr, min_opaq ); 1260 Node *min_bol = new BoolNode( min_cmp, b_test ); 1261 register_new_node( min_opaq, new_pre_exit ); 1262 register_new_node( min_cmp , new_pre_exit ); 1263 register_new_node( min_bol , new_pre_exit ); 1264 1265 // Build the IfNode (assume the main-loop is executed always). 1266 IfNode *min_iff = new IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1267 _igvn.register_new_node_with_optimizer( min_iff ); 1268 set_idom(min_iff, new_pre_exit, dd_main_head); 1269 set_loop(min_iff, loop->_parent); 1270 1271 // Plug in the false-path, taken if we need to skip main-loop 1272 _igvn.hash_delete( pre_exit ); 1273 pre_exit->set_req(0, min_iff); 1274 set_idom(pre_exit, min_iff, dd_main_head); 1275 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1276 // Make the true-path, must enter the main loop 1277 Node *min_taken = new IfTrueNode( min_iff ); 1278 _igvn.register_new_node_with_optimizer( min_taken ); 1279 set_idom(min_taken, min_iff, dd_main_head); 1280 set_loop(min_taken, loop->_parent); 1281 // Plug in the true path 1282 _igvn.hash_delete( main_head ); 1283 main_head->set_req(LoopNode::EntryControl, min_taken); 1284 set_idom(main_head, min_taken, dd_main_head); 1285 1286 visited.Clear(); 1287 clones.clear(); 1288 // Step B3: Make the fall-in values to the main-loop come from the 1289 // fall-out values of the pre-loop. 1290 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1291 Node* main_phi = main_head->fast_out(i2); 1292 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1293 Node *pre_phi = old_new[main_phi->_idx]; 1294 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1295 main_head->init_control(), 1296 pre_phi->in(LoopNode::LoopBackControl), 1297 visited, clones); 1298 _igvn.hash_delete(main_phi); 1299 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1300 } 1301 } 1302 1303 // Nodes inside the loop may be control dependent on a predicate 1304 // that was moved before the preloop. If the back branch of the main 1305 // or post loops becomes dead, those nodes won't be dependent on the 1306 // test that guards that loop nest anymore which could lead to an 1307 // incorrect array access because it executes independently of the 1308 // test that was guarding the loop nest. We add a special CastII on 1309 // the if branch that enters the loop, between the input induction 1310 // variable value and the induction variable Phi to preserve correct 1311 // dependencies. 1312 1313 // CastII for the post loop: 1314 bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); 1315 assert(inserted, "no castII inserted"); 1316 1317 // CastII for the main loop: 1318 inserted = cast_incr_before_loop(pre_incr, min_taken, main_head); 1319 assert(inserted, "no castII inserted"); 1320 1321 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1322 // RCE and alignment may change this later. 1323 Node *cmp_end = pre_end->cmp_node(); 1324 assert( cmp_end->in(2) == limit, "" ); 1325 Node *pre_limit = new AddINode( init, stride ); 1326 1327 // Save the original loop limit in this Opaque1 node for 1328 // use by range check elimination. 1329 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit); 1330 1331 register_new_node( pre_limit, pre_head->in(0) ); 1332 register_new_node( pre_opaq , pre_head->in(0) ); 1333 1334 // Since no other users of pre-loop compare, I can hack limit directly 1335 assert( cmp_end->outcnt() == 1, "no other users" ); 1336 _igvn.hash_delete(cmp_end); 1337 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1338 1339 // Special case for not-equal loop bounds: 1340 // Change pre loop test, main loop test, and the 1341 // main loop guard test to use lt or gt depending on stride 1342 // direction: 1343 // positive stride use < 1344 // negative stride use > 1345 // 1346 // not-equal test is kept for post loop to handle case 1347 // when init > limit when stride > 0 (and reverse). 1348 1349 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1350 1351 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1352 // Modify pre loop end condition 1353 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1354 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test); 1355 register_new_node( new_bol0, pre_head->in(0) ); 1356 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0); 1357 // Modify main loop guard condition 1358 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1359 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test); 1360 register_new_node( new_bol1, new_pre_exit ); 1361 _igvn.hash_delete(min_iff); 1362 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1363 // Modify main loop end condition 1364 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1365 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test); 1366 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1367 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2); 1368 } 1369 1370 // Flag main loop 1371 main_head->set_main_loop(); 1372 if( peel_only ) main_head->set_main_no_pre_loop(); 1373 1374 // Subtract a trip count for the pre-loop. 1375 main_head->set_trip_count(main_head->trip_count() - 1); 1376 1377 // It's difficult to be precise about the trip-counts 1378 // for the pre/post loops. They are usually very short, 1379 // so guess that 4 trips is a reasonable value. 1380 post_head->set_profile_trip_cnt(4.0); 1381 pre_head->set_profile_trip_cnt(4.0); 1382 1383 // Now force out all loop-invariant dominating tests. The optimizer 1384 // finds some, but we _know_ they are all useless. 1385 peeled_dom_test_elim(loop,old_new); 1386 loop->record_for_igvn(); 1387 } 1388 1389 //------------------------------is_invariant----------------------------- 1390 // Return true if n is invariant 1391 bool IdealLoopTree::is_invariant(Node* n) const { 1392 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1393 if (n_c->is_top()) return false; 1394 return !is_member(_phase->get_loop(n_c)); 1395 } 1396 1397 1398 //------------------------------do_unroll-------------------------------------- 1399 // Unroll the loop body one step - make each trip do 2 iterations. 1400 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1401 assert(LoopUnrollLimit, ""); 1402 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1403 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1404 assert(loop_end, ""); 1405 #ifndef PRODUCT 1406 if (PrintOpto && VerifyLoopOptimizations) { 1407 tty->print("Unrolling "); 1408 loop->dump_head(); 1409 } else if (TraceLoopOpts) { 1410 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1411 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1412 } else { 1413 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1414 } 1415 loop->dump_head(); 1416 } 1417 #endif 1418 1419 // Remember loop node count before unrolling to detect 1420 // if rounds of unroll,optimize are making progress 1421 loop_head->set_node_count_before_unroll(loop->_body.size()); 1422 1423 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1424 Node *limit = loop_head->limit(); 1425 Node *init = loop_head->init_trip(); 1426 Node *stride = loop_head->stride(); 1427 1428 Node *opaq = NULL; 1429 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1430 // Search for zero-trip guard. 1431 assert( loop_head->is_main_loop(), "" ); 1432 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1433 Node *iff = ctrl->in(0); 1434 assert( iff->Opcode() == Op_If, "" ); 1435 Node *bol = iff->in(1); 1436 assert( bol->Opcode() == Op_Bool, "" ); 1437 Node *cmp = bol->in(1); 1438 assert( cmp->Opcode() == Op_CmpI, "" ); 1439 opaq = cmp->in(2); 1440 // Occasionally it's possible for a zero-trip guard Opaque1 node to be 1441 // optimized away and then another round of loop opts attempted. 1442 // We can not optimize this particular loop in that case. 1443 if (opaq->Opcode() != Op_Opaque1) 1444 return; // Cannot find zero-trip guard! Bail out! 1445 // Zero-trip test uses an 'opaque' node which is not shared. 1446 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1447 } 1448 1449 C->set_major_progress(); 1450 1451 Node* new_limit = NULL; 1452 if (UnrollLimitCheck) { 1453 int stride_con = stride->get_int(); 1454 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1455 uint old_trip_count = loop_head->trip_count(); 1456 // Verify that unroll policy result is still valid. 1457 assert(old_trip_count > 1 && 1458 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1459 1460 // Adjust loop limit to keep valid iterations number after unroll. 1461 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1462 // which may overflow. 1463 if (!adjust_min_trip) { 1464 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1465 "odd trip count for maximally unroll"); 1466 // Don't need to adjust limit for maximally unroll since trip count is even. 1467 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1468 // Loop's limit is constant. Loop's init could be constant when pre-loop 1469 // become peeled iteration. 1470 jlong init_con = init->get_int(); 1471 // We can keep old loop limit if iterations count stays the same: 1472 // old_trip_count == new_trip_count * 2 1473 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1474 // so we also don't need to adjust zero trip test. 1475 jlong limit_con = limit->get_int(); 1476 // (stride_con*2) not overflow since stride_con <= 8. 1477 int new_stride_con = stride_con * 2; 1478 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1479 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1480 // New trip count should satisfy next conditions. 1481 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1482 uint new_trip_count = (uint)trip_count; 1483 adjust_min_trip = (old_trip_count != new_trip_count*2); 1484 } 1485 1486 if (adjust_min_trip) { 1487 // Step 2: Adjust the trip limit if it is called for. 1488 // The adjustment amount is -stride. Need to make sure if the 1489 // adjustment underflows or overflows, then the main loop is skipped. 1490 Node* cmp = loop_end->cmp_node(); 1491 assert(cmp->in(2) == limit, "sanity"); 1492 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1493 1494 // Verify that policy_unroll result is still valid. 1495 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1496 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1497 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); 1498 1499 if (limit->is_Con()) { 1500 // The check in policy_unroll and the assert above guarantee 1501 // no underflow if limit is constant. 1502 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1503 set_ctrl(new_limit, C->root()); 1504 } else { 1505 // Limit is not constant. 1506 if (loop_head->unrolled_count() == 1) { // only for first unroll 1507 // Separate limit by Opaque node in case it is an incremented 1508 // variable from previous loop to avoid using pre-incremented 1509 // value which could increase register pressure. 1510 // Otherwise reorg_offsets() optimization will create a separate 1511 // Opaque node for each use of trip-counter and as result 1512 // zero trip guard limit will be different from loop limit. 1513 assert(has_ctrl(opaq), "should have it"); 1514 Node* opaq_ctrl = get_ctrl(opaq); 1515 limit = new Opaque2Node( C, limit ); 1516 register_new_node( limit, opaq_ctrl ); 1517 } 1518 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) || 1519 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) { 1520 // No underflow. 1521 new_limit = new SubINode(limit, stride); 1522 } else { 1523 // (limit - stride) may underflow. 1524 // Clamp the adjustment value with MININT or MAXINT: 1525 // 1526 // new_limit = limit-stride 1527 // if (stride > 0) 1528 // new_limit = (limit < new_limit) ? MININT : new_limit; 1529 // else 1530 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1531 // 1532 BoolTest::mask bt = loop_end->test_trip(); 1533 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1534 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1535 set_ctrl(adj_max, C->root()); 1536 Node* old_limit = NULL; 1537 Node* adj_limit = NULL; 1538 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1539 if (loop_head->unrolled_count() > 1 && 1540 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1541 limit->in(CMoveNode::IfTrue) == adj_max && 1542 bol->as_Bool()->_test._test == bt && 1543 bol->in(1)->Opcode() == Op_CmpI && 1544 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1545 // Loop was unrolled before. 1546 // Optimize the limit to avoid nested CMove: 1547 // use original limit as old limit. 1548 old_limit = bol->in(1)->in(1); 1549 // Adjust previous adjusted limit. 1550 adj_limit = limit->in(CMoveNode::IfFalse); 1551 adj_limit = new SubINode(adj_limit, stride); 1552 } else { 1553 old_limit = limit; 1554 adj_limit = new SubINode(limit, stride); 1555 } 1556 assert(old_limit != NULL && adj_limit != NULL, ""); 1557 register_new_node( adj_limit, ctrl ); // adjust amount 1558 Node* adj_cmp = new CmpINode(old_limit, adj_limit); 1559 register_new_node( adj_cmp, ctrl ); 1560 Node* adj_bool = new BoolNode(adj_cmp, bt); 1561 register_new_node( adj_bool, ctrl ); 1562 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1563 } 1564 register_new_node(new_limit, ctrl); 1565 } 1566 assert(new_limit != NULL, ""); 1567 // Replace in loop test. 1568 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1569 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1570 // Don't need to create new test since only one user. 1571 _igvn.hash_delete(cmp); 1572 cmp->set_req(2, new_limit); 1573 } else { 1574 // Create new test since it is shared. 1575 Node* ctrl2 = loop_end->in(0); 1576 Node* cmp2 = cmp->clone(); 1577 cmp2->set_req(2, new_limit); 1578 register_new_node(cmp2, ctrl2); 1579 Node* bol2 = loop_end->in(1)->clone(); 1580 bol2->set_req(1, cmp2); 1581 register_new_node(bol2, ctrl2); 1582 _igvn.replace_input_of(loop_end, 1, bol2); 1583 } 1584 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1585 // Make it a 1-trip test (means at least 2 trips). 1586 1587 // Guard test uses an 'opaque' node which is not shared. Hence I 1588 // can edit it's inputs directly. Hammer in the new limit for the 1589 // minimum-trip guard. 1590 assert(opaq->outcnt() == 1, ""); 1591 _igvn.replace_input_of(opaq, 1, new_limit); 1592 } 1593 1594 // Adjust max trip count. The trip count is intentionally rounded 1595 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1596 // the main, unrolled, part of the loop will never execute as it is protected 1597 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1598 // and later determined that part of the unrolled loop was dead. 1599 loop_head->set_trip_count(old_trip_count / 2); 1600 1601 // Double the count of original iterations in the unrolled loop body. 1602 loop_head->double_unrolled_count(); 1603 1604 } else { // LoopLimitCheck 1605 1606 // Adjust max trip count. The trip count is intentionally rounded 1607 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1608 // the main, unrolled, part of the loop will never execute as it is protected 1609 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1610 // and later determined that part of the unrolled loop was dead. 1611 loop_head->set_trip_count(loop_head->trip_count() / 2); 1612 1613 // Double the count of original iterations in the unrolled loop body. 1614 loop_head->double_unrolled_count(); 1615 1616 // ----------- 1617 // Step 2: Cut back the trip counter for an unroll amount of 2. 1618 // Loop will normally trip (limit - init)/stride_con. Since it's a 1619 // CountedLoop this is exact (stride divides limit-init exactly). 1620 // We are going to double the loop body, so we want to knock off any 1621 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 1622 Node *span = new SubINode( limit, init ); 1623 register_new_node( span, ctrl ); 1624 Node *trip = new DivINode( 0, span, stride ); 1625 register_new_node( trip, ctrl ); 1626 Node *mtwo = _igvn.intcon(-2); 1627 set_ctrl(mtwo, C->root()); 1628 Node *rond = new AndINode( trip, mtwo ); 1629 register_new_node( rond, ctrl ); 1630 Node *spn2 = new MulINode( rond, stride ); 1631 register_new_node( spn2, ctrl ); 1632 new_limit = new AddINode( spn2, init ); 1633 register_new_node( new_limit, ctrl ); 1634 1635 // Hammer in the new limit 1636 Node *ctrl2 = loop_end->in(0); 1637 Node *cmp2 = new CmpINode( loop_head->incr(), new_limit ); 1638 register_new_node( cmp2, ctrl2 ); 1639 Node *bol2 = new BoolNode( cmp2, loop_end->test_trip() ); 1640 register_new_node( bol2, ctrl2 ); 1641 _igvn.replace_input_of(loop_end, CountedLoopEndNode::TestValue, bol2); 1642 1643 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1644 // Make it a 1-trip test (means at least 2 trips). 1645 if( adjust_min_trip ) { 1646 assert( new_limit != NULL, "" ); 1647 // Guard test uses an 'opaque' node which is not shared. Hence I 1648 // can edit it's inputs directly. Hammer in the new limit for the 1649 // minimum-trip guard. 1650 assert( opaq->outcnt() == 1, "" ); 1651 _igvn.hash_delete(opaq); 1652 opaq->set_req(1, new_limit); 1653 } 1654 } // LoopLimitCheck 1655 1656 // --------- 1657 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1658 // represents the odd iterations; since the loop trips an even number of 1659 // times its backedge is never taken. Kill the backedge. 1660 uint dd = dom_depth(loop_head); 1661 clone_loop( loop, old_new, dd ); 1662 1663 // Make backedges of the clone equal to backedges of the original. 1664 // Make the fall-in from the original come from the fall-out of the clone. 1665 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1666 Node* phi = loop_head->fast_out(j); 1667 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1668 Node *newphi = old_new[phi->_idx]; 1669 _igvn.hash_delete( phi ); 1670 _igvn.hash_delete( newphi ); 1671 1672 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1673 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1674 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1675 } 1676 } 1677 Node *clone_head = old_new[loop_head->_idx]; 1678 _igvn.hash_delete( clone_head ); 1679 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1680 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1681 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1682 loop->_head = clone_head; // New loop header 1683 1684 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1685 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1686 1687 // Kill the clone's backedge 1688 Node *newcle = old_new[loop_end->_idx]; 1689 _igvn.hash_delete( newcle ); 1690 Node *one = _igvn.intcon(1); 1691 set_ctrl(one, C->root()); 1692 newcle->set_req(1, one); 1693 // Force clone into same loop body 1694 uint max = loop->_body.size(); 1695 for( uint k = 0; k < max; k++ ) { 1696 Node *old = loop->_body.at(k); 1697 Node *nnn = old_new[old->_idx]; 1698 loop->_body.push(nnn); 1699 if (!has_ctrl(old)) 1700 set_loop(nnn, loop); 1701 } 1702 1703 loop->record_for_igvn(); 1704 } 1705 1706 //------------------------------do_maximally_unroll---------------------------- 1707 1708 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1709 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1710 assert(cl->has_exact_trip_count(), "trip count is not exact"); 1711 assert(cl->trip_count() > 0, ""); 1712 #ifndef PRODUCT 1713 if (TraceLoopOpts) { 1714 tty->print("MaxUnroll %d ", cl->trip_count()); 1715 loop->dump_head(); 1716 } 1717 #endif 1718 1719 // If loop is tripping an odd number of times, peel odd iteration 1720 if ((cl->trip_count() & 1) == 1) { 1721 do_peeling(loop, old_new); 1722 } 1723 1724 // Now its tripping an even number of times remaining. Double loop body. 1725 // Do not adjust pre-guards; they are not needed and do not exist. 1726 if (cl->trip_count() > 0) { 1727 assert((cl->trip_count() & 1) == 0, "missed peeling"); 1728 do_unroll(loop, old_new, false); 1729 } 1730 } 1731 1732 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) { 1733 if (SuperWordReductions == false) return; 1734 1735 CountedLoopNode* loop_head = loop->_head->as_CountedLoop(); 1736 if (loop_head->unrolled_count() > 1) { 1737 return; 1738 } 1739 1740 Node* trip_phi = loop_head->phi(); 1741 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) { 1742 Node* phi = loop_head->fast_out(i); 1743 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) { 1744 // For definitions which are loop inclusive and not tripcounts. 1745 Node* def_node = phi->in(LoopNode::LoopBackControl); 1746 1747 if (def_node != NULL) { 1748 Node* n_ctrl = get_ctrl(def_node); 1749 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) { 1750 // Now test it to see if it fits the standard pattern for a reduction operator. 1751 int opc = def_node->Opcode(); 1752 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type())) { 1753 if (!def_node->is_reduction()) { // Not marked yet 1754 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it 1755 for (unsigned j = 1; j < def_node->req(); j++) { 1756 Node* in = def_node->in(j); 1757 if (in == phi) { 1758 loop_head->mark_has_reductions(); 1759 def_node->add_flag(Node::Flag_is_reduction); 1760 break; 1761 } 1762 } 1763 } 1764 } 1765 } 1766 } 1767 } 1768 } 1769 } 1770 1771 //------------------------------dominates_backedge--------------------------------- 1772 // Returns true if ctrl is executed on every complete iteration 1773 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1774 assert(ctrl->is_CFG(), "must be control"); 1775 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1776 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1777 } 1778 1779 //------------------------------adjust_limit----------------------------------- 1780 // Helper function for add_constraint(). 1781 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { 1782 // Compute "I :: (limit-offset)/scale" 1783 Node *con = new SubINode(rc_limit, offset); 1784 register_new_node(con, pre_ctrl); 1785 Node *X = new DivINode(0, con, scale); 1786 register_new_node(X, pre_ctrl); 1787 1788 // Adjust loop limit 1789 loop_limit = (stride_con > 0) 1790 ? (Node*)(new MinINode(loop_limit, X)) 1791 : (Node*)(new MaxINode(loop_limit, X)); 1792 register_new_node(loop_limit, pre_ctrl); 1793 return loop_limit; 1794 } 1795 1796 //------------------------------add_constraint--------------------------------- 1797 // Constrain the main loop iterations so the conditions: 1798 // low_limit <= scale_con * I + offset < upper_limit 1799 // always holds true. That is, either increase the number of iterations in 1800 // the pre-loop or the post-loop until the condition holds true in the main 1801 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1802 // stride and scale are constants (offset and limit often are). 1803 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 ) { 1804 // For positive stride, the pre-loop limit always uses a MAX function 1805 // and the main loop a MIN function. For negative stride these are 1806 // reversed. 1807 1808 // Also for positive stride*scale the affine function is increasing, so the 1809 // pre-loop must check for underflow and the post-loop for overflow. 1810 // Negative stride*scale reverses this; pre-loop checks for overflow and 1811 // post-loop for underflow. 1812 1813 Node *scale = _igvn.intcon(scale_con); 1814 set_ctrl(scale, C->root()); 1815 1816 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 1817 // The overflow limit: scale*I+offset < upper_limit 1818 // For main-loop compute 1819 // ( if (scale > 0) /* and stride > 0 */ 1820 // I < (upper_limit-offset)/scale 1821 // else /* scale < 0 and stride < 0 */ 1822 // I > (upper_limit-offset)/scale 1823 // ) 1824 // 1825 // (upper_limit-offset) may overflow or underflow. 1826 // But it is fine since main loop will either have 1827 // less iterations or will be skipped in such case. 1828 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); 1829 1830 // The underflow limit: low_limit <= scale*I+offset. 1831 // For pre-loop compute 1832 // NOT(scale*I+offset >= low_limit) 1833 // scale*I+offset < low_limit 1834 // ( if (scale > 0) /* and stride > 0 */ 1835 // I < (low_limit-offset)/scale 1836 // else /* scale < 0 and stride < 0 */ 1837 // I > (low_limit-offset)/scale 1838 // ) 1839 1840 if (low_limit->get_int() == -max_jint) { 1841 if (!RangeLimitCheck) return; 1842 // We need this guard when scale*pre_limit+offset >= limit 1843 // due to underflow. So we need execute pre-loop until 1844 // scale*I+offset >= min_int. But (min_int-offset) will 1845 // underflow when offset > 0 and X will be > original_limit 1846 // when stride > 0. To avoid it we replace positive offset with 0. 1847 // 1848 // Also (min_int+1 == -max_int) is used instead of min_int here 1849 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1850 Node* shift = _igvn.intcon(31); 1851 set_ctrl(shift, C->root()); 1852 Node* sign = new RShiftINode(offset, shift); 1853 register_new_node(sign, pre_ctrl); 1854 offset = new AndINode(offset, sign); 1855 register_new_node(offset, pre_ctrl); 1856 } else { 1857 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1858 // The only problem we have here when offset == min_int 1859 // since (0-min_int) == min_int. It may be fine for stride > 0 1860 // but for stride < 0 X will be < original_limit. To avoid it 1861 // max(pre_limit, original_limit) is used in do_range_check(). 1862 } 1863 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1864 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); 1865 1866 } else { // stride_con*scale_con < 0 1867 // For negative stride*scale pre-loop checks for overflow and 1868 // post-loop for underflow. 1869 // 1870 // The overflow limit: scale*I+offset < upper_limit 1871 // For pre-loop compute 1872 // NOT(scale*I+offset < upper_limit) 1873 // scale*I+offset >= upper_limit 1874 // scale*I+offset+1 > upper_limit 1875 // ( if (scale < 0) /* and stride > 0 */ 1876 // I < (upper_limit-(offset+1))/scale 1877 // else /* scale > 0 and stride < 0 */ 1878 // I > (upper_limit-(offset+1))/scale 1879 // ) 1880 // 1881 // (upper_limit-offset-1) may underflow or overflow. 1882 // To avoid it min(pre_limit, original_limit) is used 1883 // in do_range_check() for stride > 0 and max() for < 0. 1884 Node *one = _igvn.intcon(1); 1885 set_ctrl(one, C->root()); 1886 1887 Node *plus_one = new AddINode(offset, one); 1888 register_new_node( plus_one, pre_ctrl ); 1889 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1890 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); 1891 1892 if (low_limit->get_int() == -max_jint) { 1893 if (!RangeLimitCheck) return; 1894 // We need this guard when scale*main_limit+offset >= limit 1895 // due to underflow. So we need execute main-loop while 1896 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 1897 // underflow when (offset+1) > 0 and X will be < main_limit 1898 // when scale < 0 (and stride > 0). To avoid it we replace 1899 // positive (offset+1) with 0. 1900 // 1901 // Also (min_int+1 == -max_int) is used instead of min_int here 1902 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1903 Node* shift = _igvn.intcon(31); 1904 set_ctrl(shift, C->root()); 1905 Node* sign = new RShiftINode(plus_one, shift); 1906 register_new_node(sign, pre_ctrl); 1907 plus_one = new AndINode(plus_one, sign); 1908 register_new_node(plus_one, pre_ctrl); 1909 } else { 1910 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1911 // The only problem we have here when offset == max_int 1912 // since (max_int+1) == min_int and (0-min_int) == min_int. 1913 // But it is fine since main loop will either have 1914 // less iterations or will be skipped in such case. 1915 } 1916 // The underflow limit: low_limit <= scale*I+offset. 1917 // For main-loop compute 1918 // scale*I+offset+1 > low_limit 1919 // ( if (scale < 0) /* and stride > 0 */ 1920 // I < (low_limit-(offset+1))/scale 1921 // else /* scale > 0 and stride < 0 */ 1922 // I > (low_limit-(offset+1))/scale 1923 // ) 1924 1925 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); 1926 } 1927 } 1928 1929 1930 //------------------------------is_scaled_iv--------------------------------- 1931 // Return true if exp is a constant times an induction var 1932 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1933 if (exp == iv) { 1934 if (p_scale != NULL) { 1935 *p_scale = 1; 1936 } 1937 return true; 1938 } 1939 int opc = exp->Opcode(); 1940 if (opc == Op_MulI) { 1941 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1942 if (p_scale != NULL) { 1943 *p_scale = exp->in(2)->get_int(); 1944 } 1945 return true; 1946 } 1947 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1948 if (p_scale != NULL) { 1949 *p_scale = exp->in(1)->get_int(); 1950 } 1951 return true; 1952 } 1953 } else if (opc == Op_LShiftI) { 1954 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1955 if (p_scale != NULL) { 1956 *p_scale = 1 << exp->in(2)->get_int(); 1957 } 1958 return true; 1959 } 1960 } 1961 return false; 1962 } 1963 1964 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1965 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1966 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1967 if (is_scaled_iv(exp, iv, p_scale)) { 1968 if (p_offset != NULL) { 1969 Node *zero = _igvn.intcon(0); 1970 set_ctrl(zero, C->root()); 1971 *p_offset = zero; 1972 } 1973 return true; 1974 } 1975 int opc = exp->Opcode(); 1976 if (opc == Op_AddI) { 1977 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1978 if (p_offset != NULL) { 1979 *p_offset = exp->in(2); 1980 } 1981 return true; 1982 } 1983 if (exp->in(2)->is_Con()) { 1984 Node* offset2 = NULL; 1985 if (depth < 2 && 1986 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1987 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1988 if (p_offset != NULL) { 1989 Node *ctrl_off2 = get_ctrl(offset2); 1990 Node* offset = new AddINode(offset2, exp->in(2)); 1991 register_new_node(offset, ctrl_off2); 1992 *p_offset = offset; 1993 } 1994 return true; 1995 } 1996 } 1997 } else if (opc == Op_SubI) { 1998 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1999 if (p_offset != NULL) { 2000 Node *zero = _igvn.intcon(0); 2001 set_ctrl(zero, C->root()); 2002 Node *ctrl_off = get_ctrl(exp->in(2)); 2003 Node* offset = new SubINode(zero, exp->in(2)); 2004 register_new_node(offset, ctrl_off); 2005 *p_offset = offset; 2006 } 2007 return true; 2008 } 2009 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 2010 if (p_offset != NULL) { 2011 *p_scale *= -1; 2012 *p_offset = exp->in(1); 2013 } 2014 return true; 2015 } 2016 } 2017 return false; 2018 } 2019 2020 //------------------------------do_range_check--------------------------------- 2021 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 2022 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 2023 #ifndef PRODUCT 2024 if (PrintOpto && VerifyLoopOptimizations) { 2025 tty->print("Range Check Elimination "); 2026 loop->dump_head(); 2027 } else if (TraceLoopOpts) { 2028 tty->print("RangeCheck "); 2029 loop->dump_head(); 2030 } 2031 #endif 2032 assert(RangeCheckElimination, ""); 2033 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2034 assert(cl->is_main_loop(), ""); 2035 2036 // protect against stride not being a constant 2037 if (!cl->stride_is_con()) 2038 return; 2039 2040 // Find the trip counter; we are iteration splitting based on it 2041 Node *trip_counter = cl->phi(); 2042 // Find the main loop limit; we will trim it's iterations 2043 // to not ever trip end tests 2044 Node *main_limit = cl->limit(); 2045 2046 // Need to find the main-loop zero-trip guard 2047 Node *ctrl = cl->in(LoopNode::EntryControl); 2048 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 2049 Node *iffm = ctrl->in(0); 2050 assert(iffm->Opcode() == Op_If, ""); 2051 Node *bolzm = iffm->in(1); 2052 assert(bolzm->Opcode() == Op_Bool, ""); 2053 Node *cmpzm = bolzm->in(1); 2054 assert(cmpzm->is_Cmp(), ""); 2055 Node *opqzm = cmpzm->in(2); 2056 // Can not optimize a loop if zero-trip Opaque1 node is optimized 2057 // away and then another round of loop opts attempted. 2058 if (opqzm->Opcode() != Op_Opaque1) 2059 return; 2060 assert(opqzm->in(1) == main_limit, "do not understand situation"); 2061 2062 // Find the pre-loop limit; we will expand it's iterations to 2063 // not ever trip low tests. 2064 Node *p_f = iffm->in(0); 2065 assert(p_f->Opcode() == Op_IfFalse, ""); 2066 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2067 assert(pre_end->loopnode()->is_pre_loop(), ""); 2068 Node *pre_opaq1 = pre_end->limit(); 2069 // Occasionally it's possible for a pre-loop Opaque1 node to be 2070 // optimized away and then another round of loop opts attempted. 2071 // We can not optimize this particular loop in that case. 2072 if (pre_opaq1->Opcode() != Op_Opaque1) 2073 return; 2074 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 2075 Node *pre_limit = pre_opaq->in(1); 2076 2077 // Where do we put new limit calculations 2078 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 2079 2080 // Ensure the original loop limit is available from the 2081 // pre-loop Opaque1 node. 2082 Node *orig_limit = pre_opaq->original_loop_limit(); 2083 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 2084 return; 2085 2086 // Must know if its a count-up or count-down loop 2087 2088 int stride_con = cl->stride_con(); 2089 Node *zero = _igvn.intcon(0); 2090 Node *one = _igvn.intcon(1); 2091 // Use symmetrical int range [-max_jint,max_jint] 2092 Node *mini = _igvn.intcon(-max_jint); 2093 set_ctrl(zero, C->root()); 2094 set_ctrl(one, C->root()); 2095 set_ctrl(mini, C->root()); 2096 2097 // Range checks that do not dominate the loop backedge (ie. 2098 // conditionally executed) can lengthen the pre loop limit beyond 2099 // the original loop limit. To prevent this, the pre limit is 2100 // (for stride > 0) MINed with the original loop limit (MAXed 2101 // stride < 0) when some range_check (rc) is conditionally 2102 // executed. 2103 bool conditional_rc = false; 2104 2105 // Check loop body for tests of trip-counter plus loop-invariant vs 2106 // loop-invariant. 2107 for( uint i = 0; i < loop->_body.size(); i++ ) { 2108 Node *iff = loop->_body[i]; 2109 if( iff->Opcode() == Op_If ) { // Test? 2110 2111 // Test is an IfNode, has 2 projections. If BOTH are in the loop 2112 // we need loop unswitching instead of iteration splitting. 2113 Node *exit = loop->is_loop_exit(iff); 2114 if( !exit ) continue; 2115 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 2116 2117 // Get boolean condition to test 2118 Node *i1 = iff->in(1); 2119 if( !i1->is_Bool() ) continue; 2120 BoolNode *bol = i1->as_Bool(); 2121 BoolTest b_test = bol->_test; 2122 // Flip sense of test if exit condition is flipped 2123 if( flip ) 2124 b_test = b_test.negate(); 2125 2126 // Get compare 2127 Node *cmp = bol->in(1); 2128 2129 // Look for trip_counter + offset vs limit 2130 Node *rc_exp = cmp->in(1); 2131 Node *limit = cmp->in(2); 2132 jint scale_con= 1; // Assume trip counter not scaled 2133 2134 Node *limit_c = get_ctrl(limit); 2135 if( loop->is_member(get_loop(limit_c) ) ) { 2136 // Compare might have operands swapped; commute them 2137 b_test = b_test.commute(); 2138 rc_exp = cmp->in(2); 2139 limit = cmp->in(1); 2140 limit_c = get_ctrl(limit); 2141 if( loop->is_member(get_loop(limit_c) ) ) 2142 continue; // Both inputs are loop varying; cannot RCE 2143 } 2144 // Here we know 'limit' is loop invariant 2145 2146 // 'limit' maybe pinned below the zero trip test (probably from a 2147 // previous round of rce), in which case, it can't be used in the 2148 // zero trip test expression which must occur before the zero test's if. 2149 if( limit_c == ctrl ) { 2150 continue; // Don't rce this check but continue looking for other candidates. 2151 } 2152 2153 // Check for scaled induction variable plus an offset 2154 Node *offset = NULL; 2155 2156 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 2157 continue; 2158 } 2159 2160 Node *offset_c = get_ctrl(offset); 2161 if( loop->is_member( get_loop(offset_c) ) ) 2162 continue; // Offset is not really loop invariant 2163 // Here we know 'offset' is loop invariant. 2164 2165 // As above for the 'limit', the 'offset' maybe pinned below the 2166 // zero trip test. 2167 if( offset_c == ctrl ) { 2168 continue; // Don't rce this check but continue looking for other candidates. 2169 } 2170 #ifdef ASSERT 2171 if (TraceRangeLimitCheck) { 2172 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 2173 bol->dump(2); 2174 } 2175 #endif 2176 // At this point we have the expression as: 2177 // scale_con * trip_counter + offset :: limit 2178 // where scale_con, offset and limit are loop invariant. Trip_counter 2179 // monotonically increases by stride_con, a constant. Both (or either) 2180 // stride_con and scale_con can be negative which will flip about the 2181 // sense of the test. 2182 2183 // Adjust pre and main loop limits to guard the correct iteration set 2184 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 2185 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 2186 // The underflow and overflow limits: 0 <= scale*I+offset < limit 2187 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); 2188 if (!conditional_rc) { 2189 // (0-offset)/scale could be outside of loop iterations range. 2190 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2191 } 2192 } else { 2193 #ifndef PRODUCT 2194 if( PrintOpto ) 2195 tty->print_cr("missed RCE opportunity"); 2196 #endif 2197 continue; // In release mode, ignore it 2198 } 2199 } else { // Otherwise work on normal compares 2200 switch( b_test._test ) { 2201 case BoolTest::gt: 2202 // Fall into GE case 2203 case BoolTest::ge: 2204 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 2205 scale_con = -scale_con; 2206 offset = new SubINode( zero, offset ); 2207 register_new_node( offset, pre_ctrl ); 2208 limit = new SubINode( zero, limit ); 2209 register_new_node( limit, pre_ctrl ); 2210 // Fall into LE case 2211 case BoolTest::le: 2212 if (b_test._test != BoolTest::gt) { 2213 // Convert X <= Y to X < Y+1 2214 limit = new AddINode( limit, one ); 2215 register_new_node( limit, pre_ctrl ); 2216 } 2217 // Fall into LT case 2218 case BoolTest::lt: 2219 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 2220 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 2221 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 2222 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); 2223 if (!conditional_rc) { 2224 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 2225 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 2226 // still be outside of loop range. 2227 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 2228 } 2229 break; 2230 default: 2231 #ifndef PRODUCT 2232 if( PrintOpto ) 2233 tty->print_cr("missed RCE opportunity"); 2234 #endif 2235 continue; // Unhandled case 2236 } 2237 } 2238 2239 // Kill the eliminated test 2240 C->set_major_progress(); 2241 Node *kill_con = _igvn.intcon( 1-flip ); 2242 set_ctrl(kill_con, C->root()); 2243 _igvn.replace_input_of(iff, 1, kill_con); 2244 // Find surviving projection 2245 assert(iff->is_If(), ""); 2246 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 2247 // Find loads off the surviving projection; remove their control edge 2248 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2249 Node* cd = dp->fast_out(i); // Control-dependent node 2250 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop 2251 // Allow the load to float around in the loop, or before it 2252 // but NOT before the pre-loop. 2253 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 2254 --i; 2255 --imax; 2256 } 2257 } 2258 2259 } // End of is IF 2260 2261 } 2262 2263 // Update loop limits 2264 if (conditional_rc) { 2265 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit) 2266 : (Node*)new MaxINode(pre_limit, orig_limit); 2267 register_new_node(pre_limit, pre_ctrl); 2268 } 2269 _igvn.replace_input_of(pre_opaq, 1, pre_limit); 2270 2271 // Note:: we are making the main loop limit no longer precise; 2272 // need to round up based on stride. 2273 cl->set_nonexact_trip_count(); 2274 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case 2275 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 2276 // Hopefully, compiler will optimize for powers of 2. 2277 Node *ctrl = get_ctrl(main_limit); 2278 Node *stride = cl->stride(); 2279 Node *init = cl->init_trip()->uncast(); 2280 Node *span = new SubINode(main_limit,init); 2281 register_new_node(span,ctrl); 2282 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 2283 Node *add = new AddINode(span,rndup); 2284 register_new_node(add,ctrl); 2285 Node *div = new DivINode(0,add,stride); 2286 register_new_node(div,ctrl); 2287 Node *mul = new MulINode(div,stride); 2288 register_new_node(mul,ctrl); 2289 Node *newlim = new AddINode(mul,init); 2290 register_new_node(newlim,ctrl); 2291 main_limit = newlim; 2292 } 2293 2294 Node *main_cle = cl->loopexit(); 2295 Node *main_bol = main_cle->in(1); 2296 // Hacking loop bounds; need private copies of exit test 2297 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 2298 main_bol = main_bol->clone();// Clone a private BoolNode 2299 register_new_node( main_bol, main_cle->in(0) ); 2300 _igvn.replace_input_of(main_cle, 1, main_bol); 2301 } 2302 Node *main_cmp = main_bol->in(1); 2303 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 2304 main_cmp = main_cmp->clone();// Clone a private CmpNode 2305 register_new_node( main_cmp, main_cle->in(0) ); 2306 _igvn.replace_input_of(main_bol, 1, main_cmp); 2307 } 2308 // Hack the now-private loop bounds 2309 _igvn.replace_input_of(main_cmp, 2, main_limit); 2310 // The OpaqueNode is unshared by design 2311 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 2312 _igvn.replace_input_of(opqzm, 1, main_limit); 2313 } 2314 2315 //------------------------------DCE_loop_body---------------------------------- 2316 // Remove simplistic dead code from loop body 2317 void IdealLoopTree::DCE_loop_body() { 2318 for( uint i = 0; i < _body.size(); i++ ) 2319 if( _body.at(i)->outcnt() == 0 ) 2320 _body.map( i--, _body.pop() ); 2321 } 2322 2323 2324 //------------------------------adjust_loop_exit_prob-------------------------- 2325 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2326 // Replace with a 1-in-10 exit guess. 2327 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 2328 Node *test = tail(); 2329 while( test != _head ) { 2330 uint top = test->Opcode(); 2331 if( top == Op_IfTrue || top == Op_IfFalse ) { 2332 int test_con = ((ProjNode*)test)->_con; 2333 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2334 IfNode *iff = test->in(0)->as_If(); 2335 if( iff->outcnt() == 2 ) { // Ignore dead tests 2336 Node *bol = iff->in(1); 2337 if( bol && bol->req() > 1 && bol->in(1) && 2338 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 2339 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 2340 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 2341 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 2342 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 2343 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 2344 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 2345 return; // Allocation loops RARELY take backedge 2346 // Find the OTHER exit path from the IF 2347 Node* ex = iff->proj_out(1-test_con); 2348 float p = iff->_prob; 2349 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 2350 if( top == Op_IfTrue ) { 2351 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2352 iff->_prob = PROB_STATIC_FREQUENT; 2353 } 2354 } else { 2355 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2356 iff->_prob = PROB_STATIC_INFREQUENT; 2357 } 2358 } 2359 } 2360 } 2361 } 2362 test = phase->idom(test); 2363 } 2364 } 2365 2366 2367 //------------------------------policy_do_remove_empty_loop-------------------- 2368 // Micro-benchmark spamming. Policy is to always remove empty loops. 2369 // The 'DO' part is to replace the trip counter with the value it will 2370 // have on the last iteration. This will break the loop. 2371 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 2372 // Minimum size must be empty loop 2373 if (_body.size() > EMPTY_LOOP_SIZE) 2374 return false; 2375 2376 if (!_head->is_CountedLoop()) 2377 return false; // Dead loop 2378 CountedLoopNode *cl = _head->as_CountedLoop(); 2379 if (!cl->is_valid_counted_loop()) 2380 return false; // Malformed loop 2381 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 2382 return false; // Infinite loop 2383 2384 #ifdef ASSERT 2385 // Ensure only one phi which is the iv. 2386 Node* iv = NULL; 2387 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 2388 Node* n = cl->fast_out(i); 2389 if (n->Opcode() == Op_Phi) { 2390 assert(iv == NULL, "Too many phis" ); 2391 iv = n; 2392 } 2393 } 2394 assert(iv == cl->phi(), "Wrong phi" ); 2395 #endif 2396 2397 // main and post loops have explicitly created zero trip guard 2398 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 2399 if (needs_guard) { 2400 // Skip guard if values not overlap. 2401 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 2402 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 2403 int stride_con = cl->stride_con(); 2404 if (stride_con > 0) { 2405 needs_guard = (init_t->_hi >= limit_t->_lo); 2406 } else { 2407 needs_guard = (init_t->_lo <= limit_t->_hi); 2408 } 2409 } 2410 if (needs_guard) { 2411 // Check for an obvious zero trip guard. 2412 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 2413 if (inctrl->Opcode() == Op_IfTrue) { 2414 // The test should look like just the backedge of a CountedLoop 2415 Node* iff = inctrl->in(0); 2416 if (iff->is_If()) { 2417 Node* bol = iff->in(1); 2418 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { 2419 Node* cmp = bol->in(1); 2420 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { 2421 needs_guard = false; 2422 } 2423 } 2424 } 2425 } 2426 } 2427 2428 #ifndef PRODUCT 2429 if (PrintOpto) { 2430 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 2431 this->dump_head(); 2432 } else if (TraceLoopOpts) { 2433 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 2434 this->dump_head(); 2435 } 2436 #endif 2437 2438 if (needs_guard) { 2439 // Peel the loop to ensure there's a zero trip guard 2440 Node_List old_new; 2441 phase->do_peeling(this, old_new); 2442 } 2443 2444 // Replace the phi at loop head with the final value of the last 2445 // iteration. Then the CountedLoopEnd will collapse (backedge never 2446 // taken) and all loop-invariant uses of the exit values will be correct. 2447 Node *phi = cl->phi(); 2448 Node *exact_limit = phase->exact_limit(this); 2449 if (exact_limit != cl->limit()) { 2450 // We also need to replace the original limit to collapse loop exit. 2451 Node* cmp = cl->loopexit()->cmp_node(); 2452 assert(cl->limit() == cmp->in(2), "sanity"); 2453 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 2454 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 2455 } 2456 // Note: the final value after increment should not overflow since 2457 // counted loop has limit check predicate. 2458 Node *final = new SubINode( exact_limit, cl->stride() ); 2459 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 2460 phase->_igvn.replace_node(phi,final); 2461 phase->C->set_major_progress(); 2462 return true; 2463 } 2464 2465 //------------------------------policy_do_one_iteration_loop------------------- 2466 // Convert one iteration loop into normal code. 2467 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 2468 if (!_head->as_Loop()->is_valid_counted_loop()) 2469 return false; // Only for counted loop 2470 2471 CountedLoopNode *cl = _head->as_CountedLoop(); 2472 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 2473 return false; 2474 } 2475 2476 #ifndef PRODUCT 2477 if(TraceLoopOpts) { 2478 tty->print("OneIteration "); 2479 this->dump_head(); 2480 } 2481 #endif 2482 2483 Node *init_n = cl->init_trip(); 2484 #ifdef ASSERT 2485 // Loop boundaries should be constant since trip count is exact. 2486 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 2487 #endif 2488 // Replace the phi at loop head with the value of the init_trip. 2489 // Then the CountedLoopEnd will collapse (backedge will not be taken) 2490 // and all loop-invariant uses of the exit values will be correct. 2491 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 2492 phase->C->set_major_progress(); 2493 return true; 2494 } 2495 2496 //============================================================================= 2497 //------------------------------iteration_split_impl--------------------------- 2498 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 2499 // Compute exact loop trip count if possible. 2500 compute_exact_trip_count(phase); 2501 2502 // Convert one iteration loop into normal code. 2503 if (policy_do_one_iteration_loop(phase)) 2504 return true; 2505 2506 // Check and remove empty loops (spam micro-benchmarks) 2507 if (policy_do_remove_empty_loop(phase)) 2508 return true; // Here we removed an empty loop 2509 2510 bool should_peel = policy_peeling(phase); // Should we peel? 2511 2512 bool should_unswitch = policy_unswitching(phase); 2513 2514 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 2515 // This removes loop-invariant tests (usually null checks). 2516 if (!_head->is_CountedLoop()) { // Non-counted loop 2517 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 2518 // Partial peel succeeded so terminate this round of loop opts 2519 return false; 2520 } 2521 if (should_peel) { // Should we peel? 2522 #ifndef PRODUCT 2523 if (PrintOpto) tty->print_cr("should_peel"); 2524 #endif 2525 phase->do_peeling(this,old_new); 2526 } else if (should_unswitch) { 2527 phase->do_unswitching(this, old_new); 2528 } 2529 return true; 2530 } 2531 CountedLoopNode *cl = _head->as_CountedLoop(); 2532 2533 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 2534 2535 // Do nothing special to pre- and post- loops 2536 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 2537 2538 // Compute loop trip count from profile data 2539 compute_profile_trip_cnt(phase); 2540 2541 // Before attempting fancy unrolling, RCE or alignment, see if we want 2542 // to completely unroll this loop or do loop unswitching. 2543 if (cl->is_normal_loop()) { 2544 if (should_unswitch) { 2545 phase->do_unswitching(this, old_new); 2546 return true; 2547 } 2548 bool should_maximally_unroll = policy_maximally_unroll(phase); 2549 if (should_maximally_unroll) { 2550 // Here we did some unrolling and peeling. Eventually we will 2551 // completely unroll this loop and it will no longer be a loop. 2552 phase->do_maximally_unroll(this,old_new); 2553 return true; 2554 } 2555 } 2556 2557 // Skip next optimizations if running low on nodes. Note that 2558 // policy_unswitching and policy_maximally_unroll have this check. 2559 int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes(); 2560 if ((int)(2 * _body.size()) > nodes_left) { 2561 return true; 2562 } 2563 2564 // Counted loops may be peeled, may need some iterations run up 2565 // front for RCE, and may want to align loop refs to a cache 2566 // line. Thus we clone a full loop up front whose trip count is 2567 // at least 1 (if peeling), but may be several more. 2568 2569 // The main loop will start cache-line aligned with at least 1 2570 // iteration of the unrolled body (zero-trip test required) and 2571 // will have some range checks removed. 2572 2573 // A post-loop will finish any odd iterations (leftover after 2574 // unrolling), plus any needed for RCE purposes. 2575 2576 bool should_unroll = policy_unroll(phase); 2577 2578 bool should_rce = policy_range_check(phase); 2579 2580 bool should_align = policy_align(phase); 2581 2582 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2583 // need a pre-loop. We may still need to peel an initial iteration but 2584 // we will not be needing an unknown number of pre-iterations. 2585 // 2586 // Basically, if may_rce_align reports FALSE first time through, 2587 // we will not be able to later do RCE or Aligning on this loop. 2588 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2589 2590 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2591 // we switch to the pre-/main-/post-loop model. This model also covers 2592 // peeling. 2593 if (should_rce || should_align || should_unroll) { 2594 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2595 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2596 2597 // Adjust the pre- and main-loop limits to let the pre and post loops run 2598 // with full checks, but the main-loop with no checks. Remove said 2599 // checks from the main body. 2600 if (should_rce) 2601 phase->do_range_check(this,old_new); 2602 2603 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2604 // twice as many iterations as before) and the main body limit (only do 2605 // an even number of trips). If we are peeling, we might enable some RCE 2606 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2607 // peeling. 2608 if (should_unroll && !should_peel) { 2609 phase->do_unroll(this, old_new, true); 2610 } 2611 2612 // Adjust the pre-loop limits to align the main body 2613 // iterations. 2614 if (should_align) 2615 Unimplemented(); 2616 2617 } else { // Else we have an unchanged counted loop 2618 if (should_peel) // Might want to peel but do nothing else 2619 phase->do_peeling(this,old_new); 2620 } 2621 return true; 2622 } 2623 2624 2625 //============================================================================= 2626 //------------------------------iteration_split-------------------------------- 2627 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2628 // Recursively iteration split nested loops 2629 if (_child && !_child->iteration_split(phase, old_new)) 2630 return false; 2631 2632 // Clean out prior deadwood 2633 DCE_loop_body(); 2634 2635 2636 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2637 // Replace with a 1-in-10 exit guess. 2638 if (_parent /*not the root loop*/ && 2639 !_irreducible && 2640 // Also ignore the occasional dead backedge 2641 !tail()->is_top()) { 2642 adjust_loop_exit_prob(phase); 2643 } 2644 2645 // Gate unrolling, RCE and peeling efforts. 2646 if (!_child && // If not an inner loop, do not split 2647 !_irreducible && 2648 _allow_optimizations && 2649 !tail()->is_top()) { // Also ignore the occasional dead backedge 2650 if (!_has_call) { 2651 if (!iteration_split_impl(phase, old_new)) { 2652 return false; 2653 } 2654 } else if (policy_unswitching(phase)) { 2655 phase->do_unswitching(this, old_new); 2656 } 2657 } 2658 2659 // Minor offset re-organization to remove loop-fallout uses of 2660 // trip counter when there was no major reshaping. 2661 phase->reorg_offsets(this); 2662 2663 if (_next && !_next->iteration_split(phase, old_new)) 2664 return false; 2665 return true; 2666 } 2667 2668 2669 //============================================================================= 2670 // Process all the loops in the loop tree and replace any fill 2671 // patterns with an intrisc version. 2672 bool PhaseIdealLoop::do_intrinsify_fill() { 2673 bool changed = false; 2674 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2675 IdealLoopTree* lpt = iter.current(); 2676 changed |= intrinsify_fill(lpt); 2677 } 2678 return changed; 2679 } 2680 2681 2682 // Examine an inner loop looking for a a single store of an invariant 2683 // value in a unit stride loop, 2684 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2685 Node*& shift, Node*& con) { 2686 const char* msg = NULL; 2687 Node* msg_node = NULL; 2688 2689 store_value = NULL; 2690 con = NULL; 2691 shift = NULL; 2692 2693 // Process the loop looking for stores. If there are multiple 2694 // stores or extra control flow give at this point. 2695 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2696 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2697 Node* n = lpt->_body.at(i); 2698 if (n->outcnt() == 0) continue; // Ignore dead 2699 if (n->is_Store()) { 2700 if (store != NULL) { 2701 msg = "multiple stores"; 2702 break; 2703 } 2704 int opc = n->Opcode(); 2705 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 2706 msg = "oop fills not handled"; 2707 break; 2708 } 2709 Node* value = n->in(MemNode::ValueIn); 2710 if (!lpt->is_invariant(value)) { 2711 msg = "variant store value"; 2712 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2713 msg = "not array address"; 2714 } 2715 store = n; 2716 store_value = value; 2717 } else if (n->is_If() && n != head->loopexit()) { 2718 msg = "extra control flow"; 2719 msg_node = n; 2720 } 2721 } 2722 2723 if (store == NULL) { 2724 // No store in loop 2725 return false; 2726 } 2727 2728 if (msg == NULL && head->stride_con() != 1) { 2729 // could handle negative strides too 2730 if (head->stride_con() < 0) { 2731 msg = "negative stride"; 2732 } else { 2733 msg = "non-unit stride"; 2734 } 2735 } 2736 2737 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2738 msg = "can't handle store address"; 2739 msg_node = store->in(MemNode::Address); 2740 } 2741 2742 if (msg == NULL && 2743 (!store->in(MemNode::Memory)->is_Phi() || 2744 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2745 msg = "store memory isn't proper phi"; 2746 msg_node = store->in(MemNode::Memory); 2747 } 2748 2749 // Make sure there is an appropriate fill routine 2750 BasicType t = store->as_Mem()->memory_type(); 2751 const char* fill_name; 2752 if (msg == NULL && 2753 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2754 msg = "unsupported store"; 2755 msg_node = store; 2756 } 2757 2758 if (msg != NULL) { 2759 #ifndef PRODUCT 2760 if (TraceOptimizeFill) { 2761 tty->print_cr("not fill intrinsic candidate: %s", msg); 2762 if (msg_node != NULL) msg_node->dump(); 2763 } 2764 #endif 2765 return false; 2766 } 2767 2768 // Make sure the address expression can be handled. It should be 2769 // head->phi * elsize + con. head->phi might have a ConvI2L. 2770 Node* elements[4]; 2771 Node* conv = NULL; 2772 bool found_index = false; 2773 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2774 for (int e = 0; e < count; e++) { 2775 Node* n = elements[e]; 2776 if (n->is_Con() && con == NULL) { 2777 con = n; 2778 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2779 Node* value = n->in(1); 2780 #ifdef _LP64 2781 if (value->Opcode() == Op_ConvI2L) { 2782 conv = value; 2783 value = value->in(1); 2784 } 2785 #endif 2786 if (value != head->phi()) { 2787 msg = "unhandled shift in address"; 2788 } else { 2789 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 2790 msg = "scale doesn't match"; 2791 } else { 2792 found_index = true; 2793 shift = n; 2794 } 2795 } 2796 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2797 if (n->in(1) == head->phi()) { 2798 found_index = true; 2799 conv = n; 2800 } else { 2801 msg = "unhandled input to ConvI2L"; 2802 } 2803 } else if (n == head->phi()) { 2804 // no shift, check below for allowed cases 2805 found_index = true; 2806 } else { 2807 msg = "unhandled node in address"; 2808 msg_node = n; 2809 } 2810 } 2811 2812 if (count == -1) { 2813 msg = "malformed address expression"; 2814 msg_node = store; 2815 } 2816 2817 if (!found_index) { 2818 msg = "missing use of index"; 2819 } 2820 2821 // byte sized items won't have a shift 2822 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2823 msg = "can't find shift"; 2824 msg_node = store; 2825 } 2826 2827 if (msg != NULL) { 2828 #ifndef PRODUCT 2829 if (TraceOptimizeFill) { 2830 tty->print_cr("not fill intrinsic: %s", msg); 2831 if (msg_node != NULL) msg_node->dump(); 2832 } 2833 #endif 2834 return false; 2835 } 2836 2837 // No make sure all the other nodes in the loop can be handled 2838 VectorSet ok(Thread::current()->resource_area()); 2839 2840 // store related values are ok 2841 ok.set(store->_idx); 2842 ok.set(store->in(MemNode::Memory)->_idx); 2843 2844 CountedLoopEndNode* loop_exit = head->loopexit(); 2845 guarantee(loop_exit != NULL, "no loop exit node"); 2846 2847 // Loop structure is ok 2848 ok.set(head->_idx); 2849 ok.set(loop_exit->_idx); 2850 ok.set(head->phi()->_idx); 2851 ok.set(head->incr()->_idx); 2852 ok.set(loop_exit->cmp_node()->_idx); 2853 ok.set(loop_exit->in(1)->_idx); 2854 2855 // Address elements are ok 2856 if (con) ok.set(con->_idx); 2857 if (shift) ok.set(shift->_idx); 2858 if (conv) ok.set(conv->_idx); 2859 2860 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2861 Node* n = lpt->_body.at(i); 2862 if (n->outcnt() == 0) continue; // Ignore dead 2863 if (ok.test(n->_idx)) continue; 2864 // Backedge projection is ok 2865 if (n->is_IfTrue() && n->in(0) == loop_exit) continue; 2866 if (!n->is_AddP()) { 2867 msg = "unhandled node"; 2868 msg_node = n; 2869 break; 2870 } 2871 } 2872 2873 // Make sure no unexpected values are used outside the loop 2874 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2875 Node* n = lpt->_body.at(i); 2876 // These values can be replaced with other nodes if they are used 2877 // outside the loop. 2878 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2879 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2880 Node* use = iter.get(); 2881 if (!lpt->_body.contains(use)) { 2882 msg = "node is used outside loop"; 2883 // lpt->_body.dump(); 2884 msg_node = n; 2885 break; 2886 } 2887 } 2888 } 2889 2890 #ifdef ASSERT 2891 if (TraceOptimizeFill) { 2892 if (msg != NULL) { 2893 tty->print_cr("no fill intrinsic: %s", msg); 2894 if (msg_node != NULL) msg_node->dump(); 2895 } else { 2896 tty->print_cr("fill intrinsic for:"); 2897 } 2898 store->dump(); 2899 if (Verbose) { 2900 lpt->_body.dump(); 2901 } 2902 } 2903 #endif 2904 2905 return msg == NULL; 2906 } 2907 2908 2909 2910 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2911 // Only for counted inner loops 2912 if (!lpt->is_counted() || !lpt->is_inner()) { 2913 return false; 2914 } 2915 2916 // Must have constant stride 2917 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2918 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 2919 return false; 2920 } 2921 2922 // Check that the body only contains a store of a loop invariant 2923 // value that is indexed by the loop phi. 2924 Node* store = NULL; 2925 Node* store_value = NULL; 2926 Node* shift = NULL; 2927 Node* offset = NULL; 2928 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2929 return false; 2930 } 2931 2932 #ifndef PRODUCT 2933 if (TraceLoopOpts) { 2934 tty->print("ArrayFill "); 2935 lpt->dump_head(); 2936 } 2937 #endif 2938 2939 // Now replace the whole loop body by a call to a fill routine that 2940 // covers the same region as the loop. 2941 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2942 2943 // Build an expression for the beginning of the copy region 2944 Node* index = head->init_trip(); 2945 #ifdef _LP64 2946 index = new ConvI2LNode(index); 2947 _igvn.register_new_node_with_optimizer(index); 2948 #endif 2949 if (shift != NULL) { 2950 // byte arrays don't require a shift but others do. 2951 index = new LShiftXNode(index, shift->in(2)); 2952 _igvn.register_new_node_with_optimizer(index); 2953 } 2954 index = new AddPNode(base, base, index); 2955 _igvn.register_new_node_with_optimizer(index); 2956 Node* from = new AddPNode(base, index, offset); 2957 _igvn.register_new_node_with_optimizer(from); 2958 // Compute the number of elements to copy 2959 Node* len = new SubINode(head->limit(), head->init_trip()); 2960 _igvn.register_new_node_with_optimizer(len); 2961 2962 BasicType t = store->as_Mem()->memory_type(); 2963 bool aligned = false; 2964 if (offset != NULL && head->init_trip()->is_Con()) { 2965 int element_size = type2aelembytes(t); 2966 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2967 } 2968 2969 // Build a call to the fill routine 2970 const char* fill_name; 2971 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2972 assert(fill != NULL, "what?"); 2973 2974 // Convert float/double to int/long for fill routines 2975 if (t == T_FLOAT) { 2976 store_value = new MoveF2INode(store_value); 2977 _igvn.register_new_node_with_optimizer(store_value); 2978 } else if (t == T_DOUBLE) { 2979 store_value = new MoveD2LNode(store_value); 2980 _igvn.register_new_node_with_optimizer(store_value); 2981 } 2982 2983 if (CCallingConventionRequiresIntsAsLongs && 2984 // See StubRoutines::select_fill_function for types. FLOAT has been converted to INT. 2985 (t == T_FLOAT || t == T_INT || is_subword_type(t))) { 2986 store_value = new ConvI2LNode(store_value); 2987 _igvn.register_new_node_with_optimizer(store_value); 2988 } 2989 2990 Node* mem_phi = store->in(MemNode::Memory); 2991 Node* result_ctrl; 2992 Node* result_mem; 2993 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2994 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill, 2995 fill_name, TypeAryPtr::get_array_body_type(t)); 2996 uint cnt = 0; 2997 call->init_req(TypeFunc::Parms + cnt++, from); 2998 call->init_req(TypeFunc::Parms + cnt++, store_value); 2999 if (CCallingConventionRequiresIntsAsLongs) { 3000 call->init_req(TypeFunc::Parms + cnt++, C->top()); 3001 } 3002 #ifdef _LP64 3003 len = new ConvI2LNode(len); 3004 _igvn.register_new_node_with_optimizer(len); 3005 #endif 3006 call->init_req(TypeFunc::Parms + cnt++, len); 3007 #ifdef _LP64 3008 call->init_req(TypeFunc::Parms + cnt++, C->top()); 3009 #endif 3010 call->init_req(TypeFunc::Control, head->init_control()); 3011 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. 3012 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); 3013 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr)); 3014 call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr)); 3015 _igvn.register_new_node_with_optimizer(call); 3016 result_ctrl = new ProjNode(call,TypeFunc::Control); 3017 _igvn.register_new_node_with_optimizer(result_ctrl); 3018 result_mem = new ProjNode(call,TypeFunc::Memory); 3019 _igvn.register_new_node_with_optimizer(result_mem); 3020 3021 /* Disable following optimization until proper fix (add missing checks). 3022 3023 // If this fill is tightly coupled to an allocation and overwrites 3024 // the whole body, allow it to take over the zeroing. 3025 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 3026 if (alloc != NULL && alloc->is_AllocateArray()) { 3027 Node* length = alloc->as_AllocateArray()->Ideal_length(); 3028 if (head->limit() == length && 3029 head->init_trip() == _igvn.intcon(0)) { 3030 if (TraceOptimizeFill) { 3031 tty->print_cr("Eliminated zeroing in allocation"); 3032 } 3033 alloc->maybe_set_complete(&_igvn); 3034 } else { 3035 #ifdef ASSERT 3036 if (TraceOptimizeFill) { 3037 tty->print_cr("filling array but bounds don't match"); 3038 alloc->dump(); 3039 head->init_trip()->dump(); 3040 head->limit()->dump(); 3041 length->dump(); 3042 } 3043 #endif 3044 } 3045 } 3046 */ 3047 3048 // Redirect the old control and memory edges that are outside the loop. 3049 Node* exit = head->loopexit()->proj_out(0); 3050 // Sometimes the memory phi of the head is used as the outgoing 3051 // state of the loop. It's safe in this case to replace it with the 3052 // result_mem. 3053 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 3054 _igvn.replace_node(exit, result_ctrl); 3055 _igvn.replace_node(store, result_mem); 3056 // Any uses the increment outside of the loop become the loop limit. 3057 _igvn.replace_node(head->incr(), head->limit()); 3058 3059 // Disconnect the head from the loop. 3060 for (uint i = 0; i < lpt->_body.size(); i++) { 3061 Node* n = lpt->_body.at(i); 3062 _igvn.replace_node(n, C->top()); 3063 } 3064 3065 return true; 3066 }