1 /* 2 * Copyright (c) 2000, 2011, 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/connode.hpp" 31 #include "opto/divnode.hpp" 32 #include "opto/loopnode.hpp" 33 #include "opto/mulnode.hpp" 34 #include "opto/rootnode.hpp" 35 #include "opto/runtime.hpp" 36 #include "opto/subnode.hpp" 37 38 //------------------------------is_loop_exit----------------------------------- 39 // Given an IfNode, return the loop-exiting projection or NULL if both 40 // arms remain in the loop. 41 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 43 PhaseIdealLoop *phase = _phase; 44 // Test is an IfNode, has 2 projections. If BOTH are in the loop 45 // we need loop unswitching instead of peeling. 46 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 47 return iff->raw_out(0); 48 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 49 return iff->raw_out(1); 50 return NULL; 51 } 52 53 54 //============================================================================= 55 56 57 //------------------------------record_for_igvn---------------------------- 58 // Put loop body on igvn work list 59 void IdealLoopTree::record_for_igvn() { 60 for( uint i = 0; i < _body.size(); i++ ) { 61 Node *n = _body.at(i); 62 _phase->_igvn._worklist.push(n); 63 } 64 } 65 66 //------------------------------compute_exact_trip_count----------------------- 67 // Compute loop exact trip count if possible. Do not recalculate trip count for 68 // split loops (pre-main-post) which have their limits and inits behind Opaque node. 69 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) { 70 if (!_head->as_Loop()->is_valid_counted_loop()) { 71 return; 72 } 73 CountedLoopNode* cl = _head->as_CountedLoop(); 74 // Trip count may become nonexact for iteration split loops since 75 // RCE modifies limits. Note, _trip_count value is not reset since 76 // it is used to limit unrolling of main loop. 77 cl->set_nonexact_trip_count(); 78 79 // Loop's test should be part of loop. 80 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 81 return; // Infinite loop 82 83 #ifdef ASSERT 84 BoolTest::mask bt = cl->loopexit()->test_trip(); 85 assert(bt == BoolTest::lt || bt == BoolTest::gt || 86 bt == BoolTest::ne, "canonical test is expected"); 87 #endif 88 89 Node* init_n = cl->init_trip(); 90 Node* limit_n = cl->limit(); 91 if (init_n != NULL && init_n->is_Con() && 92 limit_n != NULL && limit_n->is_Con()) { 93 // Use longs to avoid integer overflow. 94 int stride_con = cl->stride_con(); 95 long init_con = cl->init_trip()->get_int(); 96 long limit_con = cl->limit()->get_int(); 97 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 98 long trip_count = (limit_con - init_con + stride_m)/stride_con; 99 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { 100 // Set exact trip count. 101 cl->set_exact_trip_count((uint)trip_count); 102 } 103 } 104 } 105 106 //------------------------------compute_profile_trip_cnt---------------------------- 107 // Compute loop trip count from profile data as 108 // (backedge_count + loop_exit_count) / loop_exit_count 109 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 110 if (!_head->is_CountedLoop()) { 111 return; 112 } 113 CountedLoopNode* head = _head->as_CountedLoop(); 114 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 115 return; // Already computed 116 } 117 float trip_cnt = (float)max_jint; // default is big 118 119 Node* back = head->in(LoopNode::LoopBackControl); 120 while (back != head) { 121 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 122 back->in(0) && 123 back->in(0)->is_If() && 124 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 125 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 126 break; 127 } 128 back = phase->idom(back); 129 } 130 if (back != head) { 131 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 132 back->in(0), "if-projection exists"); 133 IfNode* back_if = back->in(0)->as_If(); 134 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 135 136 // Now compute a loop exit count 137 float loop_exit_cnt = 0.0f; 138 for( uint i = 0; i < _body.size(); i++ ) { 139 Node *n = _body[i]; 140 if( n->is_If() ) { 141 IfNode *iff = n->as_If(); 142 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 143 Node *exit = is_loop_exit(iff); 144 if( exit ) { 145 float exit_prob = iff->_prob; 146 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 147 if (exit_prob > PROB_MIN) { 148 float exit_cnt = iff->_fcnt * exit_prob; 149 loop_exit_cnt += exit_cnt; 150 } 151 } 152 } 153 } 154 } 155 if (loop_exit_cnt > 0.0f) { 156 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 157 } else { 158 // No exit count so use 159 trip_cnt = loop_back_cnt; 160 } 161 } 162 #ifndef PRODUCT 163 if (TraceProfileTripCount) { 164 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 165 } 166 #endif 167 head->set_profile_trip_cnt(trip_cnt); 168 } 169 170 //---------------------is_invariant_addition----------------------------- 171 // Return nonzero index of invariant operand for an Add or Sub 172 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 173 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 174 int op = n->Opcode(); 175 if (op == Op_AddI || op == Op_SubI) { 176 bool in1_invar = this->is_invariant(n->in(1)); 177 bool in2_invar = this->is_invariant(n->in(2)); 178 if (in1_invar && !in2_invar) return 1; 179 if (!in1_invar && in2_invar) return 2; 180 } 181 return 0; 182 } 183 184 //---------------------reassociate_add_sub----------------------------- 185 // Reassociate invariant add and subtract expressions: 186 // 187 // inv1 + (x + inv2) => ( inv1 + inv2) + x 188 // (x + inv2) + inv1 => ( inv1 + inv2) + x 189 // inv1 + (x - inv2) => ( inv1 - inv2) + x 190 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 191 // (x + inv2) - inv1 => (-inv1 + inv2) + x 192 // (x - inv2) + inv1 => ( inv1 - inv2) + x 193 // (x - inv2) - inv1 => (-inv1 - inv2) + x 194 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 195 // inv1 - (x - inv2) => ( inv1 + inv2) - x 196 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 197 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 198 // inv1 - (x + inv2) => ( inv1 - inv2) - x 199 // 200 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 201 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 202 if (is_invariant(n1)) return NULL; 203 int inv1_idx = is_invariant_addition(n1, phase); 204 if (!inv1_idx) return NULL; 205 // Don't mess with add of constant (igvn moves them to expression tree root.) 206 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 207 Node* inv1 = n1->in(inv1_idx); 208 Node* n2 = n1->in(3 - inv1_idx); 209 int inv2_idx = is_invariant_addition(n2, phase); 210 if (!inv2_idx) return NULL; 211 Node* x = n2->in(3 - inv2_idx); 212 Node* inv2 = n2->in(inv2_idx); 213 214 bool neg_x = n2->is_Sub() && inv2_idx == 1; 215 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 216 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 217 if (n1->is_Sub() && inv1_idx == 1) { 218 neg_x = !neg_x; 219 neg_inv2 = !neg_inv2; 220 } 221 Node* inv1_c = phase->get_ctrl(inv1); 222 Node* inv2_c = phase->get_ctrl(inv2); 223 Node* n_inv1; 224 if (neg_inv1) { 225 Node *zero = phase->_igvn.intcon(0); 226 phase->set_ctrl(zero, phase->C->root()); 227 n_inv1 = new (phase->C, 3) SubINode(zero, inv1); 228 phase->register_new_node(n_inv1, inv1_c); 229 } else { 230 n_inv1 = inv1; 231 } 232 Node* inv; 233 if (neg_inv2) { 234 inv = new (phase->C, 3) SubINode(n_inv1, inv2); 235 } else { 236 inv = new (phase->C, 3) AddINode(n_inv1, inv2); 237 } 238 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 239 240 Node* addx; 241 if (neg_x) { 242 addx = new (phase->C, 3) SubINode(inv, x); 243 } else { 244 addx = new (phase->C, 3) AddINode(x, inv); 245 } 246 phase->register_new_node(addx, phase->get_ctrl(x)); 247 phase->_igvn.replace_node(n1, addx); 248 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 249 _body.yank(n1); 250 return addx; 251 } 252 253 //---------------------reassociate_invariants----------------------------- 254 // Reassociate invariant expressions: 255 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 256 for (int i = _body.size() - 1; i >= 0; i--) { 257 Node *n = _body.at(i); 258 for (int j = 0; j < 5; j++) { 259 Node* nn = reassociate_add_sub(n, phase); 260 if (nn == NULL) break; 261 n = nn; // again 262 }; 263 } 264 } 265 266 //------------------------------policy_peeling--------------------------------- 267 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 268 // make some loop-invariant test (usually a null-check) happen before the loop. 269 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 270 Node *test = ((IdealLoopTree*)this)->tail(); 271 int body_size = ((IdealLoopTree*)this)->_body.size(); 272 int uniq = phase->C->unique(); 273 // Peeling does loop cloning which can result in O(N^2) node construction 274 if( body_size > 255 /* Prevent overflow for large body_size */ 275 || (body_size * body_size + uniq > MaxNodeLimit) ) { 276 return false; // too large to safely clone 277 } 278 while( test != _head ) { // Scan till run off top of loop 279 if( test->is_If() ) { // Test? 280 Node *ctrl = phase->get_ctrl(test->in(1)); 281 if (ctrl->is_top()) 282 return false; // Found dead test on live IF? No peeling! 283 // Standard IF only has one input value to check for loop invariance 284 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 285 // Condition is not a member of this loop? 286 if( !is_member(phase->get_loop(ctrl)) && 287 is_loop_exit(test) ) 288 return true; // Found reason to peel! 289 } 290 // Walk up dominators to loop _head looking for test which is 291 // executed on every path thru loop. 292 test = phase->idom(test); 293 } 294 return false; 295 } 296 297 //------------------------------peeled_dom_test_elim--------------------------- 298 // If we got the effect of peeling, either by actually peeling or by making 299 // a pre-loop which must execute at least once, we can remove all 300 // loop-invariant dominated tests in the main body. 301 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 302 bool progress = true; 303 while( progress ) { 304 progress = false; // Reset for next iteration 305 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 306 Node *test = prev->in(0); 307 while( test != loop->_head ) { // Scan till run off top of loop 308 309 int p_op = prev->Opcode(); 310 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 311 test->is_If() && // Test? 312 !test->in(1)->is_Con() && // And not already obvious? 313 // Condition is not a member of this loop? 314 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 315 // Walk loop body looking for instances of this test 316 for( uint i = 0; i < loop->_body.size(); i++ ) { 317 Node *n = loop->_body.at(i); 318 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 319 // IfNode was dominated by version in peeled loop body 320 progress = true; 321 dominated_by( old_new[prev->_idx], n ); 322 } 323 } 324 } 325 prev = test; 326 test = idom(test); 327 } // End of scan tests in loop 328 329 } // End of while( progress ) 330 } 331 332 //------------------------------do_peeling------------------------------------- 333 // Peel the first iteration of the given loop. 334 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 335 // The pre-loop illegally has 2 control users (old & new loops). 336 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 337 // Do this by making the old-loop fall-in edges act as if they came 338 // around the loopback from the prior iteration (follow the old-loop 339 // backedges) and then map to the new peeled iteration. This leaves 340 // the pre-loop with only 1 user (the new peeled iteration), but the 341 // peeled-loop backedge has 2 users. 342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 343 // extra backedge user. 344 // 345 // orig 346 // 347 // stmt1 348 // | 349 // v 350 // loop predicate 351 // | 352 // v 353 // loop<----+ 354 // | | 355 // stmt2 | 356 // | | 357 // v | 358 // if ^ 359 // / \ | 360 // / \ | 361 // v v | 362 // false true | 363 // / \ | 364 // / ----+ 365 // | 366 // v 367 // exit 368 // 369 // 370 // after clone loop 371 // 372 // stmt1 373 // | 374 // v 375 // loop predicate 376 // / \ 377 // clone / \ orig 378 // / \ 379 // / \ 380 // v v 381 // +---->loop clone loop<----+ 382 // | | | | 383 // | stmt2 clone stmt2 | 384 // | | | | 385 // | v v | 386 // ^ if clone If ^ 387 // | / \ / \ | 388 // | / \ / \ | 389 // | v v v v | 390 // | true false false true | 391 // | / \ / \ | 392 // +---- \ / ----+ 393 // \ / 394 // 1v v2 395 // region 396 // | 397 // v 398 // exit 399 // 400 // 401 // after peel and predicate move 402 // 403 // stmt1 404 // / 405 // / 406 // clone / orig 407 // / 408 // / +----------+ 409 // / | | 410 // / loop predicate | 411 // / | | 412 // v v | 413 // TOP-->loop clone loop<----+ | 414 // | | | | 415 // stmt2 clone stmt2 | | 416 // | | | ^ 417 // v v | | 418 // if clone If ^ | 419 // / \ / \ | | 420 // / \ / \ | | 421 // v v v v | | 422 // true false false true | | 423 // | \ / \ | | 424 // | \ / ----+ ^ 425 // | \ / | 426 // | 1v v2 | 427 // v region | 428 // | | | 429 // | v | 430 // | exit | 431 // | | 432 // +--------------->-----------------+ 433 // 434 // 435 // final graph 436 // 437 // stmt1 438 // | 439 // v 440 // stmt2 clone 441 // | 442 // v 443 // if clone 444 // / | 445 // / | 446 // v v 447 // false true 448 // | | 449 // | v 450 // | loop predicate 451 // | | 452 // | v 453 // | loop<----+ 454 // | | | 455 // | stmt2 | 456 // | | | 457 // | v | 458 // v if ^ 459 // | / \ | 460 // | / \ | 461 // | v v | 462 // | false true | 463 // | | \ | 464 // v v --+ 465 // region 466 // | 467 // v 468 // exit 469 // 470 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 471 472 C->set_major_progress(); 473 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 474 // 'pre' loop from the main and the 'pre' can no longer have it's 475 // iterations adjusted. Therefore, we need to declare this loop as 476 // no longer a 'main' loop; it will need new pre and post loops before 477 // we can do further RCE. 478 #ifndef PRODUCT 479 if (TraceLoopOpts) { 480 tty->print("Peel "); 481 loop->dump_head(); 482 } 483 #endif 484 Node* head = loop->_head; 485 bool counted_loop = head->is_CountedLoop(); 486 if (counted_loop) { 487 CountedLoopNode *cl = head->as_CountedLoop(); 488 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 489 cl->set_trip_count(cl->trip_count() - 1); 490 if (cl->is_main_loop()) { 491 cl->set_normal_loop(); 492 #ifndef PRODUCT 493 if (PrintOpto && VerifyLoopOptimizations) { 494 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 495 loop->dump_head(); 496 } 497 #endif 498 } 499 } 500 Node* entry = head->in(LoopNode::EntryControl); 501 502 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 503 // The pre-loop illegally has 2 control users (old & new loops). 504 clone_loop( loop, old_new, dom_depth(head) ); 505 506 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 507 // Do this by making the old-loop fall-in edges act as if they came 508 // around the loopback from the prior iteration (follow the old-loop 509 // backedges) and then map to the new peeled iteration. This leaves 510 // the pre-loop with only 1 user (the new peeled iteration), but the 511 // peeled-loop backedge has 2 users. 512 Node* new_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 513 new_exit_value = move_loop_predicates(entry, new_exit_value); 514 _igvn.hash_delete(head); 515 head->set_req(LoopNode::EntryControl, new_exit_value); 516 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 517 Node* old = head->fast_out(j); 518 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 519 new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 520 if (!new_exit_value ) // Backedge value is ALSO loop invariant? 521 // Then loop body backedge value remains the same. 522 new_exit_value = old->in(LoopNode::LoopBackControl); 523 _igvn.hash_delete(old); 524 old->set_req(LoopNode::EntryControl, new_exit_value); 525 } 526 } 527 528 529 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 530 // extra backedge user. 531 Node* new_head = old_new[head->_idx]; 532 _igvn.hash_delete(new_head); 533 new_head->set_req(LoopNode::LoopBackControl, C->top()); 534 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 535 Node* use = new_head->fast_out(j2); 536 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 537 _igvn.hash_delete(use); 538 use->set_req(LoopNode::LoopBackControl, C->top()); 539 } 540 } 541 542 543 // Step 4: Correct dom-depth info. Set to loop-head depth. 544 int dd = dom_depth(head); 545 set_idom(head, head->in(1), dd); 546 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 547 Node *old = loop->_body.at(j3); 548 Node *nnn = old_new[old->_idx]; 549 if (!has_ctrl(nnn)) 550 set_idom(nnn, idom(nnn), dd-1); 551 // While we're at it, remove any SafePoints from the peeled code 552 if (old->Opcode() == Op_SafePoint) { 553 Node *nnn = old_new[old->_idx]; 554 lazy_replace(nnn,nnn->in(TypeFunc::Control)); 555 } 556 } 557 558 // Now force out all loop-invariant dominating tests. The optimizer 559 // finds some, but we _know_ they are all useless. 560 peeled_dom_test_elim(loop,old_new); 561 562 loop->record_for_igvn(); 563 } 564 565 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 566 567 //------------------------------policy_maximally_unroll------------------------ 568 // Calculate exact loop trip count and return true if loop can be maximally 569 // unrolled. 570 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 571 CountedLoopNode *cl = _head->as_CountedLoop(); 572 assert(cl->is_normal_loop(), ""); 573 if (!cl->is_valid_counted_loop()) 574 return false; // Malformed counted loop 575 576 if (!cl->has_exact_trip_count()) { 577 // Trip count is not exact. 578 return false; 579 } 580 581 uint trip_count = cl->trip_count(); 582 // Note, max_juint is used to indicate unknown trip count. 583 assert(trip_count > 1, "one iteration loop should be optimized out already"); 584 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 585 586 // Real policy: if we maximally unroll, does it get too big? 587 // Allow the unrolled mess to get larger than standard loop 588 // size. After all, it will no longer be a loop. 589 uint body_size = _body.size(); 590 uint unroll_limit = (uint)LoopUnrollLimit * 4; 591 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 592 if (trip_count > unroll_limit || body_size > unroll_limit) { 593 return false; 594 } 595 596 // Take into account that after unroll conjoined heads and tails will fold, 597 // otherwise policy_unroll() may allow more unrolling than max unrolling. 598 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; 599 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; 600 if (body_size != tst_body_size) // Check for int overflow 601 return false; 602 if (new_body_size > unroll_limit || 603 // Unrolling can result in a large amount of node construction 604 new_body_size >= MaxNodeLimit - phase->C->unique()) { 605 return false; 606 } 607 608 // Currently we don't have policy to optimize one iteration loops. 609 // Maximally unrolling transformation is used for that: 610 // it is peeled and the original loop become non reachable (dead). 611 // Also fully unroll a loop with few iterations regardless next 612 // conditions since following loop optimizations will split 613 // such loop anyway (pre-main-post). 614 if (trip_count <= 3) 615 return true; 616 617 // Do not unroll a loop with String intrinsics code. 618 // String intrinsics are large and have loops. 619 for (uint k = 0; k < _body.size(); k++) { 620 Node* n = _body.at(k); 621 switch (n->Opcode()) { 622 case Op_StrComp: 623 case Op_StrEquals: 624 case Op_StrIndexOf: 625 case Op_AryEq: { 626 return false; 627 } 628 } // switch 629 } 630 631 return true; // Do maximally unroll 632 } 633 634 635 //------------------------------policy_unroll---------------------------------- 636 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 637 // the loop is a CountedLoop and the body is small enough. 638 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { 639 640 CountedLoopNode *cl = _head->as_CountedLoop(); 641 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 642 643 if (!cl->is_valid_counted_loop()) 644 return false; // Malformed counted loop 645 646 // protect against over-unrolling 647 if (cl->trip_count() <= 1) return false; 648 649 // Check for stride being a small enough constant 650 if (abs(cl->stride_con()) > (1<<3)) return false; 651 652 int future_unroll_ct = cl->unrolled_count() * 2; 653 654 // Don't unroll if the next round of unrolling would push us 655 // over the expected trip count of the loop. One is subtracted 656 // from the expected trip count because the pre-loop normally 657 // executes 1 iteration. 658 if (UnrollLimitForProfileCheck > 0 && 659 cl->profile_trip_cnt() != COUNT_UNKNOWN && 660 future_unroll_ct > UnrollLimitForProfileCheck && 661 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 662 return false; 663 } 664 665 // When unroll count is greater than LoopUnrollMin, don't unroll if: 666 // the residual iterations are more than 10% of the trip count 667 // and rounds of "unroll,optimize" are not making significant progress 668 // Progress defined as current size less than 20% larger than previous size. 669 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 670 future_unroll_ct > LoopUnrollMin && 671 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 672 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 673 return false; 674 } 675 676 Node *init_n = cl->init_trip(); 677 Node *limit_n = cl->limit(); 678 // Non-constant bounds. 679 // Protect against over-unrolling when init or/and limit are not constant 680 // (so that trip_count's init value is maxint) but iv range is known. 681 if (init_n == NULL || !init_n->is_Con() || 682 limit_n == NULL || !limit_n->is_Con()) { 683 Node* phi = cl->phi(); 684 if (phi != NULL) { 685 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 686 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 687 int next_stride = cl->stride_con() * 2; // stride after this unroll 688 if (next_stride > 0) { 689 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 690 iv_type->_lo + next_stride > iv_type->_hi) { 691 return false; // over-unrolling 692 } 693 } else if (next_stride < 0) { 694 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 695 iv_type->_hi + next_stride < iv_type->_lo) { 696 return false; // over-unrolling 697 } 698 } 699 } 700 } 701 702 // Adjust body_size to determine if we unroll or not 703 uint body_size = _body.size(); 704 // Key test to unroll CaffeineMark's Logic test 705 int xors_in_loop = 0; 706 // Also count ModL, DivL and MulL which expand mightly 707 for (uint k = 0; k < _body.size(); k++) { 708 Node* n = _body.at(k); 709 switch (n->Opcode()) { 710 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test 711 case Op_ModL: body_size += 30; break; 712 case Op_DivL: body_size += 30; break; 713 case Op_MulL: body_size += 10; break; 714 case Op_StrComp: 715 case Op_StrEquals: 716 case Op_StrIndexOf: 717 case Op_AryEq: { 718 // Do not unroll a loop with String intrinsics code. 719 // String intrinsics are large and have loops. 720 return false; 721 } 722 } // switch 723 } 724 725 // Check for being too big 726 if (body_size > (uint)LoopUnrollLimit) { 727 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 728 // Normal case: loop too big 729 return false; 730 } 731 732 // Unroll once! (Each trip will soon do double iterations) 733 return true; 734 } 735 736 //------------------------------policy_align----------------------------------- 737 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 738 // expression that does the alignment. Note that only one array base can be 739 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 740 // if we vectorize short memory ops into longer memory ops, we may want to 741 // increase alignment. 742 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 743 return false; 744 } 745 746 //------------------------------policy_range_check----------------------------- 747 // Return TRUE or FALSE if the loop should be range-check-eliminated. 748 // Actually we do iteration-splitting, a more powerful form of RCE. 749 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 750 if( !RangeCheckElimination ) return false; 751 752 CountedLoopNode *cl = _head->as_CountedLoop(); 753 // If we unrolled with no intention of doing RCE and we later 754 // changed our minds, we got no pre-loop. Either we need to 755 // make a new pre-loop, or we gotta disallow RCE. 756 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now. 757 Node *trip_counter = cl->phi(); 758 759 // Check loop body for tests of trip-counter plus loop-invariant vs 760 // loop-invariant. 761 for( uint i = 0; i < _body.size(); i++ ) { 762 Node *iff = _body[i]; 763 if( iff->Opcode() == Op_If ) { // Test? 764 765 // Comparing trip+off vs limit 766 Node *bol = iff->in(1); 767 if( bol->req() != 2 ) continue; // dead constant test 768 if (!bol->is_Bool()) { 769 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 770 continue; 771 } 772 Node *cmp = bol->in(1); 773 774 Node *rc_exp = cmp->in(1); 775 Node *limit = cmp->in(2); 776 777 Node *limit_c = phase->get_ctrl(limit); 778 if( limit_c == phase->C->top() ) 779 return false; // Found dead test on live IF? No RCE! 780 if( is_member(phase->get_loop(limit_c) ) ) { 781 // Compare might have operands swapped; commute them 782 rc_exp = cmp->in(2); 783 limit = cmp->in(1); 784 limit_c = phase->get_ctrl(limit); 785 if( is_member(phase->get_loop(limit_c) ) ) 786 continue; // Both inputs are loop varying; cannot RCE 787 } 788 789 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 790 continue; 791 } 792 // Yeah! Found a test like 'trip+off vs limit' 793 // Test is an IfNode, has 2 projections. If BOTH are in the loop 794 // we need loop unswitching instead of iteration splitting. 795 if( is_loop_exit(iff) ) 796 return true; // Found reason to split iterations 797 } // End of is IF 798 } 799 800 return false; 801 } 802 803 //------------------------------policy_peel_only------------------------------- 804 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 805 // for unrolling loops with NO array accesses. 806 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 807 808 for( uint i = 0; i < _body.size(); i++ ) 809 if( _body[i]->is_Mem() ) 810 return false; 811 812 // No memory accesses at all! 813 return true; 814 } 815 816 //------------------------------clone_up_backedge_goo-------------------------- 817 // If Node n lives in the back_ctrl block and cannot float, we clone a private 818 // version of n in preheader_ctrl block and return that, otherwise return n. 819 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) { 820 if( get_ctrl(n) != back_ctrl ) return n; 821 822 Node *x = NULL; // If required, a clone of 'n' 823 // Check for 'n' being pinned in the backedge. 824 if( n->in(0) && n->in(0) == back_ctrl ) { 825 x = n->clone(); // Clone a copy of 'n' to preheader 826 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 827 } 828 829 // Recursive fixup any other input edges into x. 830 // If there are no changes we can just return 'n', otherwise 831 // we need to clone a private copy and change it. 832 for( uint i = 1; i < n->req(); i++ ) { 833 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) ); 834 if( g != n->in(i) ) { 835 if( !x ) 836 x = n->clone(); 837 x->set_req(i, g); 838 } 839 } 840 if( x ) { // x can legally float to pre-header location 841 register_new_node( x, preheader_ctrl ); 842 return x; 843 } else { // raise n to cover LCA of uses 844 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 845 } 846 return n; 847 } 848 849 //------------------------------insert_pre_post_loops-------------------------- 850 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 851 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 852 // alignment. Useful to unroll loops that do no array accesses. 853 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 854 855 #ifndef PRODUCT 856 if (TraceLoopOpts) { 857 if (peel_only) 858 tty->print("PeelMainPost "); 859 else 860 tty->print("PreMainPost "); 861 loop->dump_head(); 862 } 863 #endif 864 C->set_major_progress(); 865 866 // Find common pieces of the loop being guarded with pre & post loops 867 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 868 assert( main_head->is_normal_loop(), "" ); 869 CountedLoopEndNode *main_end = main_head->loopexit(); 870 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 871 uint dd_main_head = dom_depth(main_head); 872 uint max = main_head->outcnt(); 873 874 Node *pre_header= main_head->in(LoopNode::EntryControl); 875 Node *init = main_head->init_trip(); 876 Node *incr = main_end ->incr(); 877 Node *limit = main_end ->limit(); 878 Node *stride = main_end ->stride(); 879 Node *cmp = main_end ->cmp_node(); 880 BoolTest::mask b_test = main_end->test_trip(); 881 882 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 883 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 884 if( bol->outcnt() != 1 ) { 885 bol = bol->clone(); 886 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 887 _igvn.hash_delete(main_end); 888 main_end->set_req(CountedLoopEndNode::TestValue, bol); 889 } 890 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 891 if( cmp->outcnt() != 1 ) { 892 cmp = cmp->clone(); 893 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 894 _igvn.hash_delete(bol); 895 bol->set_req(1, cmp); 896 } 897 898 //------------------------------ 899 // Step A: Create Post-Loop. 900 Node* main_exit = main_end->proj_out(false); 901 assert( main_exit->Opcode() == Op_IfFalse, "" ); 902 int dd_main_exit = dom_depth(main_exit); 903 904 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 905 // loop pre-header illegally has 2 control users (old & new loops). 906 clone_loop( loop, old_new, dd_main_exit ); 907 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 908 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 909 post_head->set_post_loop(main_head); 910 911 // Reduce the post-loop trip count. 912 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 913 post_end->_prob = PROB_FAIR; 914 915 // Build the main-loop normal exit. 916 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end); 917 _igvn.register_new_node_with_optimizer( new_main_exit ); 918 set_idom(new_main_exit, main_end, dd_main_exit ); 919 set_loop(new_main_exit, loop->_parent); 920 921 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 922 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 923 // (the main-loop trip-counter exit value) because we will be changing 924 // the exit value (via unrolling) so we cannot constant-fold away the zero 925 // trip guard until all unrolling is done. 926 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr); 927 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit ); 928 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test ); 929 register_new_node( zer_opaq, new_main_exit ); 930 register_new_node( zer_cmp , new_main_exit ); 931 register_new_node( zer_bol , new_main_exit ); 932 933 // Build the IfNode 934 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 935 _igvn.register_new_node_with_optimizer( zer_iff ); 936 set_idom(zer_iff, new_main_exit, dd_main_exit); 937 set_loop(zer_iff, loop->_parent); 938 939 // Plug in the false-path, taken if we need to skip post-loop 940 _igvn.hash_delete( main_exit ); 941 main_exit->set_req(0, zer_iff); 942 _igvn._worklist.push(main_exit); 943 set_idom(main_exit, zer_iff, dd_main_exit); 944 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 945 // Make the true-path, must enter the post loop 946 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff ); 947 _igvn.register_new_node_with_optimizer( zer_taken ); 948 set_idom(zer_taken, zer_iff, dd_main_exit); 949 set_loop(zer_taken, loop->_parent); 950 // Plug in the true path 951 _igvn.hash_delete( post_head ); 952 post_head->set_req(LoopNode::EntryControl, zer_taken); 953 set_idom(post_head, zer_taken, dd_main_exit); 954 955 // Step A3: Make the fall-in values to the post-loop come from the 956 // fall-out values of the main-loop. 957 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 958 Node* main_phi = main_head->fast_out(i); 959 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 960 Node *post_phi = old_new[main_phi->_idx]; 961 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 962 post_head->init_control(), 963 main_phi->in(LoopNode::LoopBackControl)); 964 _igvn.hash_delete(post_phi); 965 post_phi->set_req( LoopNode::EntryControl, fallmain ); 966 } 967 } 968 969 // Update local caches for next stanza 970 main_exit = new_main_exit; 971 972 973 //------------------------------ 974 // Step B: Create Pre-Loop. 975 976 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 977 // loop pre-header illegally has 2 control users (old & new loops). 978 clone_loop( loop, old_new, dd_main_head ); 979 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 980 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 981 pre_head->set_pre_loop(main_head); 982 Node *pre_incr = old_new[incr->_idx]; 983 984 // Reduce the pre-loop trip count. 985 pre_end->_prob = PROB_FAIR; 986 987 // Find the pre-loop normal exit. 988 Node* pre_exit = pre_end->proj_out(false); 989 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 990 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end); 991 _igvn.register_new_node_with_optimizer( new_pre_exit ); 992 set_idom(new_pre_exit, pre_end, dd_main_head); 993 set_loop(new_pre_exit, loop->_parent); 994 995 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 996 // pre-loop, the main-loop may not execute at all. Later in life this 997 // zero-trip guard will become the minimum-trip guard when we unroll 998 // the main-loop. 999 Node *min_opaq = new (C, 2) Opaque1Node(C, limit); 1000 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq ); 1001 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test ); 1002 register_new_node( min_opaq, new_pre_exit ); 1003 register_new_node( min_cmp , new_pre_exit ); 1004 register_new_node( min_bol , new_pre_exit ); 1005 1006 // Build the IfNode (assume the main-loop is executed always). 1007 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1008 _igvn.register_new_node_with_optimizer( min_iff ); 1009 set_idom(min_iff, new_pre_exit, dd_main_head); 1010 set_loop(min_iff, loop->_parent); 1011 1012 // Plug in the false-path, taken if we need to skip main-loop 1013 _igvn.hash_delete( pre_exit ); 1014 pre_exit->set_req(0, min_iff); 1015 set_idom(pre_exit, min_iff, dd_main_head); 1016 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1017 // Make the true-path, must enter the main loop 1018 Node *min_taken = new (C, 1) IfTrueNode( min_iff ); 1019 _igvn.register_new_node_with_optimizer( min_taken ); 1020 set_idom(min_taken, min_iff, dd_main_head); 1021 set_loop(min_taken, loop->_parent); 1022 // Plug in the true path 1023 _igvn.hash_delete( main_head ); 1024 main_head->set_req(LoopNode::EntryControl, min_taken); 1025 set_idom(main_head, min_taken, dd_main_head); 1026 1027 // Step B3: Make the fall-in values to the main-loop come from the 1028 // fall-out values of the pre-loop. 1029 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1030 Node* main_phi = main_head->fast_out(i2); 1031 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1032 Node *pre_phi = old_new[main_phi->_idx]; 1033 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1034 main_head->init_control(), 1035 pre_phi->in(LoopNode::LoopBackControl)); 1036 _igvn.hash_delete(main_phi); 1037 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1038 } 1039 } 1040 1041 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1042 // RCE and alignment may change this later. 1043 Node *cmp_end = pre_end->cmp_node(); 1044 assert( cmp_end->in(2) == limit, "" ); 1045 Node *pre_limit = new (C, 3) AddINode( init, stride ); 1046 1047 // Save the original loop limit in this Opaque1 node for 1048 // use by range check elimination. 1049 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit); 1050 1051 register_new_node( pre_limit, pre_head->in(0) ); 1052 register_new_node( pre_opaq , pre_head->in(0) ); 1053 1054 // Since no other users of pre-loop compare, I can hack limit directly 1055 assert( cmp_end->outcnt() == 1, "no other users" ); 1056 _igvn.hash_delete(cmp_end); 1057 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1058 1059 // Special case for not-equal loop bounds: 1060 // Change pre loop test, main loop test, and the 1061 // main loop guard test to use lt or gt depending on stride 1062 // direction: 1063 // positive stride use < 1064 // negative stride use > 1065 1066 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1067 1068 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1069 // Modify pre loop end condition 1070 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1071 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test); 1072 register_new_node( new_bol0, pre_head->in(0) ); 1073 _igvn.hash_delete(pre_end); 1074 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); 1075 // Modify main loop guard condition 1076 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1077 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test); 1078 register_new_node( new_bol1, new_pre_exit ); 1079 _igvn.hash_delete(min_iff); 1080 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1081 // Modify main loop end condition 1082 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1083 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test); 1084 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1085 _igvn.hash_delete(main_end); 1086 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); 1087 } 1088 1089 // Flag main loop 1090 main_head->set_main_loop(); 1091 if( peel_only ) main_head->set_main_no_pre_loop(); 1092 1093 // It's difficult to be precise about the trip-counts 1094 // for the pre/post loops. They are usually very short, 1095 // so guess that 4 trips is a reasonable value. 1096 post_head->set_profile_trip_cnt(4.0); 1097 pre_head->set_profile_trip_cnt(4.0); 1098 1099 // Now force out all loop-invariant dominating tests. The optimizer 1100 // finds some, but we _know_ they are all useless. 1101 peeled_dom_test_elim(loop,old_new); 1102 } 1103 1104 //------------------------------is_invariant----------------------------- 1105 // Return true if n is invariant 1106 bool IdealLoopTree::is_invariant(Node* n) const { 1107 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1108 if (n_c->is_top()) return false; 1109 return !is_member(_phase->get_loop(n_c)); 1110 } 1111 1112 1113 //------------------------------do_unroll-------------------------------------- 1114 // Unroll the loop body one step - make each trip do 2 iterations. 1115 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1116 assert(LoopUnrollLimit, ""); 1117 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1118 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1119 assert(loop_end, ""); 1120 #ifndef PRODUCT 1121 if (PrintOpto && VerifyLoopOptimizations) { 1122 tty->print("Unrolling "); 1123 loop->dump_head(); 1124 } else if (TraceLoopOpts) { 1125 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1126 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1127 } else { 1128 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1129 } 1130 loop->dump_head(); 1131 } 1132 #endif 1133 1134 // Remember loop node count before unrolling to detect 1135 // if rounds of unroll,optimize are making progress 1136 loop_head->set_node_count_before_unroll(loop->_body.size()); 1137 1138 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1139 Node *limit = loop_head->limit(); 1140 Node *init = loop_head->init_trip(); 1141 Node *stride = loop_head->stride(); 1142 1143 Node *opaq = NULL; 1144 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment 1145 assert( loop_head->is_main_loop(), "" ); 1146 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1147 Node *iff = ctrl->in(0); 1148 assert( iff->Opcode() == Op_If, "" ); 1149 Node *bol = iff->in(1); 1150 assert( bol->Opcode() == Op_Bool, "" ); 1151 Node *cmp = bol->in(1); 1152 assert( cmp->Opcode() == Op_CmpI, "" ); 1153 opaq = cmp->in(2); 1154 // Occasionally it's possible for a pre-loop Opaque1 node to be 1155 // optimized away and then another round of loop opts attempted. 1156 // We can not optimize this particular loop in that case. 1157 if( opaq->Opcode() != Op_Opaque1 ) 1158 return; // Cannot find pre-loop! Bail out! 1159 } 1160 1161 C->set_major_progress(); 1162 1163 // Adjust max trip count. The trip count is intentionally rounded 1164 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1165 // the main, unrolled, part of the loop will never execute as it is protected 1166 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1167 // and later determined that part of the unrolled loop was dead. 1168 loop_head->set_trip_count(loop_head->trip_count() / 2); 1169 1170 // Double the count of original iterations in the unrolled loop body. 1171 loop_head->double_unrolled_count(); 1172 1173 // ----------- 1174 // Step 2: Cut back the trip counter for an unroll amount of 2. 1175 // Loop will normally trip (limit - init)/stride_con. Since it's a 1176 // CountedLoop this is exact (stride divides limit-init exactly). 1177 // We are going to double the loop body, so we want to knock off any 1178 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 1179 Node *span = new (C, 3) SubINode( limit, init ); 1180 register_new_node( span, ctrl ); 1181 Node *trip = new (C, 3) DivINode( 0, span, stride ); 1182 register_new_node( trip, ctrl ); 1183 Node *mtwo = _igvn.intcon(-2); 1184 set_ctrl(mtwo, C->root()); 1185 Node *rond = new (C, 3) AndINode( trip, mtwo ); 1186 register_new_node( rond, ctrl ); 1187 Node *spn2 = new (C, 3) MulINode( rond, stride ); 1188 register_new_node( spn2, ctrl ); 1189 Node *lim2 = new (C, 3) AddINode( spn2, init ); 1190 register_new_node( lim2, ctrl ); 1191 1192 // Hammer in the new limit 1193 Node *ctrl2 = loop_end->in(0); 1194 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 ); 1195 register_new_node( cmp2, ctrl2 ); 1196 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() ); 1197 register_new_node( bol2, ctrl2 ); 1198 _igvn.hash_delete(loop_end); 1199 loop_end->set_req(CountedLoopEndNode::TestValue, bol2); 1200 1201 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1202 // Make it a 1-trip test (means at least 2 trips). 1203 if( adjust_min_trip ) { 1204 // Guard test uses an 'opaque' node which is not shared. Hence I 1205 // can edit it's inputs directly. Hammer in the new limit for the 1206 // minimum-trip guard. 1207 assert( opaq->outcnt() == 1, "" ); 1208 _igvn.hash_delete(opaq); 1209 opaq->set_req(1, lim2); 1210 } 1211 1212 // --------- 1213 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1214 // represents the odd iterations; since the loop trips an even number of 1215 // times its backedge is never taken. Kill the backedge. 1216 uint dd = dom_depth(loop_head); 1217 clone_loop( loop, old_new, dd ); 1218 1219 // Make backedges of the clone equal to backedges of the original. 1220 // Make the fall-in from the original come from the fall-out of the clone. 1221 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1222 Node* phi = loop_head->fast_out(j); 1223 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1224 Node *newphi = old_new[phi->_idx]; 1225 _igvn.hash_delete( phi ); 1226 _igvn.hash_delete( newphi ); 1227 1228 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1229 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1230 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1231 } 1232 } 1233 Node *clone_head = old_new[loop_head->_idx]; 1234 _igvn.hash_delete( clone_head ); 1235 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1236 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1237 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1238 loop->_head = clone_head; // New loop header 1239 1240 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1241 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1242 1243 // Kill the clone's backedge 1244 Node *newcle = old_new[loop_end->_idx]; 1245 _igvn.hash_delete( newcle ); 1246 Node *one = _igvn.intcon(1); 1247 set_ctrl(one, C->root()); 1248 newcle->set_req(1, one); 1249 // Force clone into same loop body 1250 uint max = loop->_body.size(); 1251 for( uint k = 0; k < max; k++ ) { 1252 Node *old = loop->_body.at(k); 1253 Node *nnn = old_new[old->_idx]; 1254 loop->_body.push(nnn); 1255 if (!has_ctrl(old)) 1256 set_loop(nnn, loop); 1257 } 1258 1259 loop->record_for_igvn(); 1260 } 1261 1262 //------------------------------do_maximally_unroll---------------------------- 1263 1264 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1265 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1266 assert(cl->trip_count() > 0, ""); 1267 #ifndef PRODUCT 1268 if (TraceLoopOpts) { 1269 tty->print("MaxUnroll %d ", cl->trip_count()); 1270 loop->dump_head(); 1271 } 1272 #endif 1273 1274 // If loop is tripping an odd number of times, peel odd iteration 1275 if ((cl->trip_count() & 1) == 1) { 1276 do_peeling(loop, old_new); 1277 } 1278 1279 // Now its tripping an even number of times remaining. Double loop body. 1280 // Do not adjust pre-guards; they are not needed and do not exist. 1281 if (cl->trip_count() > 0) { 1282 do_unroll(loop, old_new, false); 1283 } 1284 } 1285 1286 //------------------------------dominates_backedge--------------------------------- 1287 // Returns true if ctrl is executed on every complete iteration 1288 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1289 assert(ctrl->is_CFG(), "must be control"); 1290 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1291 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1292 } 1293 1294 //------------------------------add_constraint--------------------------------- 1295 // Constrain the main loop iterations so the condition: 1296 // scale_con * I + offset < limit 1297 // always holds true. That is, either increase the number of iterations in 1298 // the pre-loop or the post-loop until the condition holds true in the main 1299 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1300 // stride and scale are constants (offset and limit often are). 1301 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { 1302 1303 // Compute "I :: (limit-offset)/scale_con" 1304 Node *con = new (C, 3) SubINode( limit, offset ); 1305 register_new_node( con, pre_ctrl ); 1306 Node *scale = _igvn.intcon(scale_con); 1307 set_ctrl(scale, C->root()); 1308 Node *X = new (C, 3) DivINode( 0, con, scale ); 1309 register_new_node( X, pre_ctrl ); 1310 1311 // For positive stride, the pre-loop limit always uses a MAX function 1312 // and the main loop a MIN function. For negative stride these are 1313 // reversed. 1314 1315 // Also for positive stride*scale the affine function is increasing, so the 1316 // pre-loop must check for underflow and the post-loop for overflow. 1317 // Negative stride*scale reverses this; pre-loop checks for overflow and 1318 // post-loop for underflow. 1319 if( stride_con*scale_con > 0 ) { 1320 // Compute I < (limit-offset)/scale_con 1321 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X) 1322 *main_limit = (stride_con > 0) 1323 ? (Node*)(new (C, 3) MinINode( *main_limit, X )) 1324 : (Node*)(new (C, 3) MaxINode( *main_limit, X )); 1325 register_new_node( *main_limit, pre_ctrl ); 1326 1327 } else { 1328 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I 1329 // Add the negation of the main-loop constraint to the pre-loop. 1330 // See footnote [++] below for a derivation of the limit expression. 1331 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1); 1332 set_ctrl(incr, C->root()); 1333 Node *adj = new (C, 3) AddINode( X, incr ); 1334 register_new_node( adj, pre_ctrl ); 1335 *pre_limit = (scale_con > 0) 1336 ? (Node*)new (C, 3) MinINode( *pre_limit, adj ) 1337 : (Node*)new (C, 3) MaxINode( *pre_limit, adj ); 1338 register_new_node( *pre_limit, pre_ctrl ); 1339 1340 // [++] Here's the algebra that justifies the pre-loop limit expression: 1341 // 1342 // NOT( scale_con * I + offset < limit ) 1343 // == 1344 // scale_con * I + offset >= limit 1345 // == 1346 // SGN(scale_con) * I >= (limit-offset)/|scale_con| 1347 // == 1348 // (limit-offset)/|scale_con| <= I * SGN(scale_con) 1349 // == 1350 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con) 1351 // == 1352 // ( if (scale_con > 0) /*common case*/ 1353 // (limit-offset)/scale_con - 1 < I 1354 // else 1355 // (limit-offset)/scale_con + 1 > I 1356 // ) 1357 // ( if (scale_con > 0) /*common case*/ 1358 // (limit-offset)/scale_con + SGN(-scale_con) < I 1359 // else 1360 // (limit-offset)/scale_con + SGN(-scale_con) > I 1361 } 1362 } 1363 1364 1365 //------------------------------is_scaled_iv--------------------------------- 1366 // Return true if exp is a constant times an induction var 1367 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1368 if (exp == iv) { 1369 if (p_scale != NULL) { 1370 *p_scale = 1; 1371 } 1372 return true; 1373 } 1374 int opc = exp->Opcode(); 1375 if (opc == Op_MulI) { 1376 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1377 if (p_scale != NULL) { 1378 *p_scale = exp->in(2)->get_int(); 1379 } 1380 return true; 1381 } 1382 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1383 if (p_scale != NULL) { 1384 *p_scale = exp->in(1)->get_int(); 1385 } 1386 return true; 1387 } 1388 } else if (opc == Op_LShiftI) { 1389 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1390 if (p_scale != NULL) { 1391 *p_scale = 1 << exp->in(2)->get_int(); 1392 } 1393 return true; 1394 } 1395 } 1396 return false; 1397 } 1398 1399 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1400 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1401 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1402 if (is_scaled_iv(exp, iv, p_scale)) { 1403 if (p_offset != NULL) { 1404 Node *zero = _igvn.intcon(0); 1405 set_ctrl(zero, C->root()); 1406 *p_offset = zero; 1407 } 1408 return true; 1409 } 1410 int opc = exp->Opcode(); 1411 if (opc == Op_AddI) { 1412 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1413 if (p_offset != NULL) { 1414 *p_offset = exp->in(2); 1415 } 1416 return true; 1417 } 1418 if (exp->in(2)->is_Con()) { 1419 Node* offset2 = NULL; 1420 if (depth < 2 && 1421 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1422 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1423 if (p_offset != NULL) { 1424 Node *ctrl_off2 = get_ctrl(offset2); 1425 Node* offset = new (C, 3) AddINode(offset2, exp->in(2)); 1426 register_new_node(offset, ctrl_off2); 1427 *p_offset = offset; 1428 } 1429 return true; 1430 } 1431 } 1432 } else if (opc == Op_SubI) { 1433 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1434 if (p_offset != NULL) { 1435 Node *zero = _igvn.intcon(0); 1436 set_ctrl(zero, C->root()); 1437 Node *ctrl_off = get_ctrl(exp->in(2)); 1438 Node* offset = new (C, 3) SubINode(zero, exp->in(2)); 1439 register_new_node(offset, ctrl_off); 1440 *p_offset = offset; 1441 } 1442 return true; 1443 } 1444 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1445 if (p_offset != NULL) { 1446 *p_scale *= -1; 1447 *p_offset = exp->in(1); 1448 } 1449 return true; 1450 } 1451 } 1452 return false; 1453 } 1454 1455 //------------------------------do_range_check--------------------------------- 1456 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1457 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1458 #ifndef PRODUCT 1459 if (PrintOpto && VerifyLoopOptimizations) { 1460 tty->print("Range Check Elimination "); 1461 loop->dump_head(); 1462 } else if (TraceLoopOpts) { 1463 tty->print("RangeCheck "); 1464 loop->dump_head(); 1465 } 1466 #endif 1467 assert(RangeCheckElimination, ""); 1468 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1469 assert(cl->is_main_loop(), ""); 1470 1471 // protect against stride not being a constant 1472 if (!cl->stride_is_con()) 1473 return; 1474 1475 // Find the trip counter; we are iteration splitting based on it 1476 Node *trip_counter = cl->phi(); 1477 // Find the main loop limit; we will trim it's iterations 1478 // to not ever trip end tests 1479 Node *main_limit = cl->limit(); 1480 1481 // Need to find the main-loop zero-trip guard 1482 Node *ctrl = cl->in(LoopNode::EntryControl); 1483 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 1484 Node *iffm = ctrl->in(0); 1485 assert(iffm->Opcode() == Op_If, ""); 1486 Node *bolzm = iffm->in(1); 1487 assert(bolzm->Opcode() == Op_Bool, ""); 1488 Node *cmpzm = bolzm->in(1); 1489 assert(cmpzm->is_Cmp(), ""); 1490 Node *opqzm = cmpzm->in(2); 1491 // Can not optimize a loop if pre-loop Opaque1 node is optimized 1492 // away and then another round of loop opts attempted. 1493 if (opqzm->Opcode() != Op_Opaque1) 1494 return; 1495 assert(opqzm->in(1) == main_limit, "do not understand situation"); 1496 1497 // Find the pre-loop limit; we will expand it's iterations to 1498 // not ever trip low tests. 1499 Node *p_f = iffm->in(0); 1500 assert(p_f->Opcode() == Op_IfFalse, ""); 1501 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 1502 assert(pre_end->loopnode()->is_pre_loop(), ""); 1503 Node *pre_opaq1 = pre_end->limit(); 1504 // Occasionally it's possible for a pre-loop Opaque1 node to be 1505 // optimized away and then another round of loop opts attempted. 1506 // We can not optimize this particular loop in that case. 1507 if (pre_opaq1->Opcode() != Op_Opaque1) 1508 return; 1509 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 1510 Node *pre_limit = pre_opaq->in(1); 1511 1512 // Where do we put new limit calculations 1513 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 1514 1515 // Ensure the original loop limit is available from the 1516 // pre-loop Opaque1 node. 1517 Node *orig_limit = pre_opaq->original_loop_limit(); 1518 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 1519 return; 1520 1521 // Must know if its a count-up or count-down loop 1522 1523 int stride_con = cl->stride_con(); 1524 Node *zero = _igvn.intcon(0); 1525 Node *one = _igvn.intcon(1); 1526 set_ctrl(zero, C->root()); 1527 set_ctrl(one, C->root()); 1528 1529 // Range checks that do not dominate the loop backedge (ie. 1530 // conditionally executed) can lengthen the pre loop limit beyond 1531 // the original loop limit. To prevent this, the pre limit is 1532 // (for stride > 0) MINed with the original loop limit (MAXed 1533 // stride < 0) when some range_check (rc) is conditionally 1534 // executed. 1535 bool conditional_rc = false; 1536 1537 // Check loop body for tests of trip-counter plus loop-invariant vs 1538 // loop-invariant. 1539 for( uint i = 0; i < loop->_body.size(); i++ ) { 1540 Node *iff = loop->_body[i]; 1541 if( iff->Opcode() == Op_If ) { // Test? 1542 1543 // Test is an IfNode, has 2 projections. If BOTH are in the loop 1544 // we need loop unswitching instead of iteration splitting. 1545 Node *exit = loop->is_loop_exit(iff); 1546 if( !exit ) continue; 1547 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 1548 1549 // Get boolean condition to test 1550 Node *i1 = iff->in(1); 1551 if( !i1->is_Bool() ) continue; 1552 BoolNode *bol = i1->as_Bool(); 1553 BoolTest b_test = bol->_test; 1554 // Flip sense of test if exit condition is flipped 1555 if( flip ) 1556 b_test = b_test.negate(); 1557 1558 // Get compare 1559 Node *cmp = bol->in(1); 1560 1561 // Look for trip_counter + offset vs limit 1562 Node *rc_exp = cmp->in(1); 1563 Node *limit = cmp->in(2); 1564 jint scale_con= 1; // Assume trip counter not scaled 1565 1566 Node *limit_c = get_ctrl(limit); 1567 if( loop->is_member(get_loop(limit_c) ) ) { 1568 // Compare might have operands swapped; commute them 1569 b_test = b_test.commute(); 1570 rc_exp = cmp->in(2); 1571 limit = cmp->in(1); 1572 limit_c = get_ctrl(limit); 1573 if( loop->is_member(get_loop(limit_c) ) ) 1574 continue; // Both inputs are loop varying; cannot RCE 1575 } 1576 // Here we know 'limit' is loop invariant 1577 1578 // 'limit' maybe pinned below the zero trip test (probably from a 1579 // previous round of rce), in which case, it can't be used in the 1580 // zero trip test expression which must occur before the zero test's if. 1581 if( limit_c == ctrl ) { 1582 continue; // Don't rce this check but continue looking for other candidates. 1583 } 1584 1585 // Check for scaled induction variable plus an offset 1586 Node *offset = NULL; 1587 1588 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 1589 continue; 1590 } 1591 1592 Node *offset_c = get_ctrl(offset); 1593 if( loop->is_member( get_loop(offset_c) ) ) 1594 continue; // Offset is not really loop invariant 1595 // Here we know 'offset' is loop invariant. 1596 1597 // As above for the 'limit', the 'offset' maybe pinned below the 1598 // zero trip test. 1599 if( offset_c == ctrl ) { 1600 continue; // Don't rce this check but continue looking for other candidates. 1601 } 1602 1603 // At this point we have the expression as: 1604 // scale_con * trip_counter + offset :: limit 1605 // where scale_con, offset and limit are loop invariant. Trip_counter 1606 // monotonically increases by stride_con, a constant. Both (or either) 1607 // stride_con and scale_con can be negative which will flip about the 1608 // sense of the test. 1609 1610 // Adjust pre and main loop limits to guard the correct iteration set 1611 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 1612 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 1613 // The overflow limit: scale*I+offset < limit 1614 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1615 // The underflow limit: 0 <= scale*I+offset. 1616 // Some math yields: -scale*I-(offset+1) < 0 1617 Node *plus_one = new (C, 3) AddINode( offset, one ); 1618 register_new_node( plus_one, pre_ctrl ); 1619 Node *neg_offset = new (C, 3) SubINode( zero, plus_one ); 1620 register_new_node( neg_offset, pre_ctrl ); 1621 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit ); 1622 if (!conditional_rc) { 1623 conditional_rc = !loop->dominates_backedge(iff); 1624 } 1625 } else { 1626 #ifndef PRODUCT 1627 if( PrintOpto ) 1628 tty->print_cr("missed RCE opportunity"); 1629 #endif 1630 continue; // In release mode, ignore it 1631 } 1632 } else { // Otherwise work on normal compares 1633 switch( b_test._test ) { 1634 case BoolTest::ge: // Convert X >= Y to -X <= -Y 1635 scale_con = -scale_con; 1636 offset = new (C, 3) SubINode( zero, offset ); 1637 register_new_node( offset, pre_ctrl ); 1638 limit = new (C, 3) SubINode( zero, limit ); 1639 register_new_node( limit, pre_ctrl ); 1640 // Fall into LE case 1641 case BoolTest::le: // Convert X <= Y to X < Y+1 1642 limit = new (C, 3) AddINode( limit, one ); 1643 register_new_node( limit, pre_ctrl ); 1644 // Fall into LT case 1645 case BoolTest::lt: 1646 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1647 if (!conditional_rc) { 1648 conditional_rc = !loop->dominates_backedge(iff); 1649 } 1650 break; 1651 default: 1652 #ifndef PRODUCT 1653 if( PrintOpto ) 1654 tty->print_cr("missed RCE opportunity"); 1655 #endif 1656 continue; // Unhandled case 1657 } 1658 } 1659 1660 // Kill the eliminated test 1661 C->set_major_progress(); 1662 Node *kill_con = _igvn.intcon( 1-flip ); 1663 set_ctrl(kill_con, C->root()); 1664 _igvn.hash_delete(iff); 1665 iff->set_req(1, kill_con); 1666 _igvn._worklist.push(iff); 1667 // Find surviving projection 1668 assert(iff->is_If(), ""); 1669 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 1670 // Find loads off the surviving projection; remove their control edge 1671 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 1672 Node* cd = dp->fast_out(i); // Control-dependent node 1673 if( cd->is_Load() ) { // Loads can now float around in the loop 1674 _igvn.hash_delete(cd); 1675 // Allow the load to float around in the loop, or before it 1676 // but NOT before the pre-loop. 1677 cd->set_req(0, ctrl); // ctrl, not NULL 1678 _igvn._worklist.push(cd); 1679 --i; 1680 --imax; 1681 } 1682 } 1683 1684 } // End of is IF 1685 1686 } 1687 1688 // Update loop limits 1689 if (conditional_rc) { 1690 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit) 1691 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit); 1692 register_new_node(pre_limit, pre_ctrl); 1693 } 1694 _igvn.hash_delete(pre_opaq); 1695 pre_opaq->set_req(1, pre_limit); 1696 1697 // Note:: we are making the main loop limit no longer precise; 1698 // need to round up based on stride. 1699 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case 1700 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 1701 // Hopefully, compiler will optimize for powers of 2. 1702 Node *ctrl = get_ctrl(main_limit); 1703 Node *stride = cl->stride(); 1704 Node *init = cl->init_trip(); 1705 Node *span = new (C, 3) SubINode(main_limit,init); 1706 register_new_node(span,ctrl); 1707 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 1708 Node *add = new (C, 3) AddINode(span,rndup); 1709 register_new_node(add,ctrl); 1710 Node *div = new (C, 3) DivINode(0,add,stride); 1711 register_new_node(div,ctrl); 1712 Node *mul = new (C, 3) MulINode(div,stride); 1713 register_new_node(mul,ctrl); 1714 Node *newlim = new (C, 3) AddINode(mul,init); 1715 register_new_node(newlim,ctrl); 1716 main_limit = newlim; 1717 } 1718 1719 Node *main_cle = cl->loopexit(); 1720 Node *main_bol = main_cle->in(1); 1721 // Hacking loop bounds; need private copies of exit test 1722 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 1723 _igvn.hash_delete(main_cle); 1724 main_bol = main_bol->clone();// Clone a private BoolNode 1725 register_new_node( main_bol, main_cle->in(0) ); 1726 main_cle->set_req(1,main_bol); 1727 } 1728 Node *main_cmp = main_bol->in(1); 1729 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 1730 _igvn.hash_delete(main_bol); 1731 main_cmp = main_cmp->clone();// Clone a private CmpNode 1732 register_new_node( main_cmp, main_cle->in(0) ); 1733 main_bol->set_req(1,main_cmp); 1734 } 1735 // Hack the now-private loop bounds 1736 _igvn.hash_delete(main_cmp); 1737 main_cmp->set_req(2, main_limit); 1738 _igvn._worklist.push(main_cmp); 1739 // The OpaqueNode is unshared by design 1740 _igvn.hash_delete(opqzm); 1741 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 1742 opqzm->set_req(1,main_limit); 1743 _igvn._worklist.push(opqzm); 1744 } 1745 1746 //------------------------------DCE_loop_body---------------------------------- 1747 // Remove simplistic dead code from loop body 1748 void IdealLoopTree::DCE_loop_body() { 1749 for( uint i = 0; i < _body.size(); i++ ) 1750 if( _body.at(i)->outcnt() == 0 ) 1751 _body.map( i--, _body.pop() ); 1752 } 1753 1754 1755 //------------------------------adjust_loop_exit_prob-------------------------- 1756 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 1757 // Replace with a 1-in-10 exit guess. 1758 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 1759 Node *test = tail(); 1760 while( test != _head ) { 1761 uint top = test->Opcode(); 1762 if( top == Op_IfTrue || top == Op_IfFalse ) { 1763 int test_con = ((ProjNode*)test)->_con; 1764 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 1765 IfNode *iff = test->in(0)->as_If(); 1766 if( iff->outcnt() == 2 ) { // Ignore dead tests 1767 Node *bol = iff->in(1); 1768 if( bol && bol->req() > 1 && bol->in(1) && 1769 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 1770 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 1771 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 1772 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 1773 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 1774 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 1775 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 1776 return; // Allocation loops RARELY take backedge 1777 // Find the OTHER exit path from the IF 1778 Node* ex = iff->proj_out(1-test_con); 1779 float p = iff->_prob; 1780 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 1781 if( top == Op_IfTrue ) { 1782 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 1783 iff->_prob = PROB_STATIC_FREQUENT; 1784 } 1785 } else { 1786 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 1787 iff->_prob = PROB_STATIC_INFREQUENT; 1788 } 1789 } 1790 } 1791 } 1792 } 1793 test = phase->idom(test); 1794 } 1795 } 1796 1797 1798 //------------------------------policy_do_remove_empty_loop-------------------- 1799 // Micro-benchmark spamming. Policy is to always remove empty loops. 1800 // The 'DO' part is to replace the trip counter with the value it will 1801 // have on the last iteration. This will break the loop. 1802 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 1803 // Minimum size must be empty loop 1804 if (_body.size() > EMPTY_LOOP_SIZE) 1805 return false; 1806 1807 if (!_head->is_CountedLoop()) 1808 return false; // Dead loop 1809 CountedLoopNode *cl = _head->as_CountedLoop(); 1810 if (!cl->loopexit()) 1811 return false; // Malformed loop 1812 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 1813 return false; // Infinite loop 1814 1815 #ifdef ASSERT 1816 // Ensure only one phi which is the iv. 1817 Node* iv = NULL; 1818 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 1819 Node* n = cl->fast_out(i); 1820 if (n->Opcode() == Op_Phi) { 1821 assert(iv == NULL, "Too many phis" ); 1822 iv = n; 1823 } 1824 } 1825 assert(iv == cl->phi(), "Wrong phi" ); 1826 #endif 1827 1828 // main and post loops have explicitly created zero trip guard 1829 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 1830 if (needs_guard) { 1831 // Skip guard if values not overlap. 1832 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 1833 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 1834 int stride_con = cl->stride_con(); 1835 if (stride_con > 0) { 1836 needs_guard = (init_t->_hi >= limit_t->_lo); 1837 } else { 1838 needs_guard = (init_t->_lo <= limit_t->_hi); 1839 } 1840 } 1841 if (needs_guard) { 1842 // Check for an obvious zero trip guard. 1843 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 1844 if (inctrl->Opcode() == Op_IfTrue) { 1845 // The test should look like just the backedge of a CountedLoop 1846 Node* iff = inctrl->in(0); 1847 if (iff->is_If()) { 1848 Node* bol = iff->in(1); 1849 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { 1850 Node* cmp = bol->in(1); 1851 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { 1852 needs_guard = false; 1853 } 1854 } 1855 } 1856 } 1857 } 1858 1859 #ifndef PRODUCT 1860 if (PrintOpto) { 1861 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 1862 this->dump_head(); 1863 } else if (TraceLoopOpts) { 1864 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 1865 this->dump_head(); 1866 } 1867 #endif 1868 1869 if (needs_guard) { 1870 // Peel the loop to ensure there's a zero trip guard 1871 Node_List old_new; 1872 phase->do_peeling(this, old_new); 1873 } 1874 1875 // Replace the phi at loop head with the final value of the last 1876 // iteration. Then the CountedLoopEnd will collapse (backedge never 1877 // taken) and all loop-invariant uses of the exit values will be correct. 1878 Node *phi = cl->phi(); 1879 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() ); 1880 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 1881 phase->_igvn.replace_node(phi,final); 1882 phase->C->set_major_progress(); 1883 return true; 1884 } 1885 1886 //------------------------------policy_do_one_iteration_loop------------------- 1887 // Convert one iteration loop into normal code. 1888 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 1889 if (!_head->as_Loop()->is_valid_counted_loop()) 1890 return false; // Only for counted loop 1891 1892 CountedLoopNode *cl = _head->as_CountedLoop(); 1893 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 1894 return false; 1895 } 1896 1897 #ifndef PRODUCT 1898 if(TraceLoopOpts) { 1899 tty->print("OneIteration "); 1900 this->dump_head(); 1901 } 1902 #endif 1903 1904 Node *init_n = cl->init_trip(); 1905 #ifdef ASSERT 1906 // Loop boundaries should be constant since trip count is exact. 1907 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 1908 #endif 1909 // Replace the phi at loop head with the value of the init_trip. 1910 // Then the CountedLoopEnd will collapse (backedge will not be taken) 1911 // and all loop-invariant uses of the exit values will be correct. 1912 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 1913 phase->C->set_major_progress(); 1914 return true; 1915 } 1916 1917 //============================================================================= 1918 //------------------------------iteration_split_impl--------------------------- 1919 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 1920 // Compute exact loop trip count if possible. 1921 compute_exact_trip_count(phase); 1922 1923 // Convert one iteration loop into normal code. 1924 if (policy_do_one_iteration_loop(phase)) 1925 return true; 1926 1927 // Check and remove empty loops (spam micro-benchmarks) 1928 if (policy_do_remove_empty_loop(phase)) 1929 return true; // Here we removed an empty loop 1930 1931 bool should_peel = policy_peeling(phase); // Should we peel? 1932 1933 bool should_unswitch = policy_unswitching(phase); 1934 1935 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 1936 // This removes loop-invariant tests (usually null checks). 1937 if (!_head->is_CountedLoop()) { // Non-counted loop 1938 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 1939 // Partial peel succeeded so terminate this round of loop opts 1940 return false; 1941 } 1942 if (should_peel) { // Should we peel? 1943 #ifndef PRODUCT 1944 if (PrintOpto) tty->print_cr("should_peel"); 1945 #endif 1946 phase->do_peeling(this,old_new); 1947 } else if (should_unswitch) { 1948 phase->do_unswitching(this, old_new); 1949 } 1950 return true; 1951 } 1952 CountedLoopNode *cl = _head->as_CountedLoop(); 1953 1954 if (!cl->loopexit()) return true; // Ignore various kinds of broken loops 1955 1956 // Do nothing special to pre- and post- loops 1957 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 1958 1959 // Compute loop trip count from profile data 1960 compute_profile_trip_cnt(phase); 1961 1962 // Before attempting fancy unrolling, RCE or alignment, see if we want 1963 // to completely unroll this loop or do loop unswitching. 1964 if (cl->is_normal_loop()) { 1965 if (should_unswitch) { 1966 phase->do_unswitching(this, old_new); 1967 return true; 1968 } 1969 bool should_maximally_unroll = policy_maximally_unroll(phase); 1970 if (should_maximally_unroll) { 1971 // Here we did some unrolling and peeling. Eventually we will 1972 // completely unroll this loop and it will no longer be a loop. 1973 phase->do_maximally_unroll(this,old_new); 1974 return true; 1975 } 1976 } 1977 1978 // Skip next optimizations if running low on nodes. Note that 1979 // policy_unswitching and policy_maximally_unroll have this check. 1980 uint nodes_left = MaxNodeLimit - phase->C->unique(); 1981 if ((2 * _body.size()) > nodes_left) { 1982 return true; 1983 } 1984 1985 // Counted loops may be peeled, may need some iterations run up 1986 // front for RCE, and may want to align loop refs to a cache 1987 // line. Thus we clone a full loop up front whose trip count is 1988 // at least 1 (if peeling), but may be several more. 1989 1990 // The main loop will start cache-line aligned with at least 1 1991 // iteration of the unrolled body (zero-trip test required) and 1992 // will have some range checks removed. 1993 1994 // A post-loop will finish any odd iterations (leftover after 1995 // unrolling), plus any needed for RCE purposes. 1996 1997 bool should_unroll = policy_unroll(phase); 1998 1999 bool should_rce = policy_range_check(phase); 2000 2001 bool should_align = policy_align(phase); 2002 2003 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2004 // need a pre-loop. We may still need to peel an initial iteration but 2005 // we will not be needing an unknown number of pre-iterations. 2006 // 2007 // Basically, if may_rce_align reports FALSE first time through, 2008 // we will not be able to later do RCE or Aligning on this loop. 2009 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2010 2011 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2012 // we switch to the pre-/main-/post-loop model. This model also covers 2013 // peeling. 2014 if (should_rce || should_align || should_unroll) { 2015 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2016 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2017 2018 // Adjust the pre- and main-loop limits to let the pre and post loops run 2019 // with full checks, but the main-loop with no checks. Remove said 2020 // checks from the main body. 2021 if (should_rce) 2022 phase->do_range_check(this,old_new); 2023 2024 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2025 // twice as many iterations as before) and the main body limit (only do 2026 // an even number of trips). If we are peeling, we might enable some RCE 2027 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2028 // peeling. 2029 if (should_unroll && !should_peel) 2030 phase->do_unroll(this,old_new, true); 2031 2032 // Adjust the pre-loop limits to align the main body 2033 // iterations. 2034 if (should_align) 2035 Unimplemented(); 2036 2037 } else { // Else we have an unchanged counted loop 2038 if (should_peel) // Might want to peel but do nothing else 2039 phase->do_peeling(this,old_new); 2040 } 2041 return true; 2042 } 2043 2044 2045 //============================================================================= 2046 //------------------------------iteration_split-------------------------------- 2047 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2048 // Recursively iteration split nested loops 2049 if (_child && !_child->iteration_split(phase, old_new)) 2050 return false; 2051 2052 // Clean out prior deadwood 2053 DCE_loop_body(); 2054 2055 2056 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2057 // Replace with a 1-in-10 exit guess. 2058 if (_parent /*not the root loop*/ && 2059 !_irreducible && 2060 // Also ignore the occasional dead backedge 2061 !tail()->is_top()) { 2062 adjust_loop_exit_prob(phase); 2063 } 2064 2065 // Gate unrolling, RCE and peeling efforts. 2066 if (!_child && // If not an inner loop, do not split 2067 !_irreducible && 2068 _allow_optimizations && 2069 !tail()->is_top()) { // Also ignore the occasional dead backedge 2070 if (!_has_call) { 2071 if (!iteration_split_impl(phase, old_new)) { 2072 return false; 2073 } 2074 } else if (policy_unswitching(phase)) { 2075 phase->do_unswitching(this, old_new); 2076 } 2077 } 2078 2079 // Minor offset re-organization to remove loop-fallout uses of 2080 // trip counter when there was no major reshaping. 2081 phase->reorg_offsets(this); 2082 2083 if (_next && !_next->iteration_split(phase, old_new)) 2084 return false; 2085 return true; 2086 } 2087 2088 2089 //============================================================================= 2090 // Process all the loops in the loop tree and replace any fill 2091 // patterns with an intrisc version. 2092 bool PhaseIdealLoop::do_intrinsify_fill() { 2093 bool changed = false; 2094 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2095 IdealLoopTree* lpt = iter.current(); 2096 changed |= intrinsify_fill(lpt); 2097 } 2098 return changed; 2099 } 2100 2101 2102 // Examine an inner loop looking for a a single store of an invariant 2103 // value in a unit stride loop, 2104 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2105 Node*& shift, Node*& con) { 2106 const char* msg = NULL; 2107 Node* msg_node = NULL; 2108 2109 store_value = NULL; 2110 con = NULL; 2111 shift = NULL; 2112 2113 // Process the loop looking for stores. If there are multiple 2114 // stores or extra control flow give at this point. 2115 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2116 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2117 Node* n = lpt->_body.at(i); 2118 if (n->outcnt() == 0) continue; // Ignore dead 2119 if (n->is_Store()) { 2120 if (store != NULL) { 2121 msg = "multiple stores"; 2122 break; 2123 } 2124 int opc = n->Opcode(); 2125 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) { 2126 msg = "oop fills not handled"; 2127 break; 2128 } 2129 Node* value = n->in(MemNode::ValueIn); 2130 if (!lpt->is_invariant(value)) { 2131 msg = "variant store value"; 2132 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2133 msg = "not array address"; 2134 } 2135 store = n; 2136 store_value = value; 2137 } else if (n->is_If() && n != head->loopexit()) { 2138 msg = "extra control flow"; 2139 msg_node = n; 2140 } 2141 } 2142 2143 if (store == NULL) { 2144 // No store in loop 2145 return false; 2146 } 2147 2148 if (msg == NULL && head->stride_con() != 1) { 2149 // could handle negative strides too 2150 if (head->stride_con() < 0) { 2151 msg = "negative stride"; 2152 } else { 2153 msg = "non-unit stride"; 2154 } 2155 } 2156 2157 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2158 msg = "can't handle store address"; 2159 msg_node = store->in(MemNode::Address); 2160 } 2161 2162 if (msg == NULL && 2163 (!store->in(MemNode::Memory)->is_Phi() || 2164 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2165 msg = "store memory isn't proper phi"; 2166 msg_node = store->in(MemNode::Memory); 2167 } 2168 2169 // Make sure there is an appropriate fill routine 2170 BasicType t = store->as_Mem()->memory_type(); 2171 const char* fill_name; 2172 if (msg == NULL && 2173 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2174 msg = "unsupported store"; 2175 msg_node = store; 2176 } 2177 2178 if (msg != NULL) { 2179 #ifndef PRODUCT 2180 if (TraceOptimizeFill) { 2181 tty->print_cr("not fill intrinsic candidate: %s", msg); 2182 if (msg_node != NULL) msg_node->dump(); 2183 } 2184 #endif 2185 return false; 2186 } 2187 2188 // Make sure the address expression can be handled. It should be 2189 // head->phi * elsize + con. head->phi might have a ConvI2L. 2190 Node* elements[4]; 2191 Node* conv = NULL; 2192 bool found_index = false; 2193 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2194 for (int e = 0; e < count; e++) { 2195 Node* n = elements[e]; 2196 if (n->is_Con() && con == NULL) { 2197 con = n; 2198 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2199 Node* value = n->in(1); 2200 #ifdef _LP64 2201 if (value->Opcode() == Op_ConvI2L) { 2202 conv = value; 2203 value = value->in(1); 2204 } 2205 #endif 2206 if (value != head->phi()) { 2207 msg = "unhandled shift in address"; 2208 } else { 2209 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 2210 msg = "scale doesn't match"; 2211 } else { 2212 found_index = true; 2213 shift = n; 2214 } 2215 } 2216 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2217 if (n->in(1) == head->phi()) { 2218 found_index = true; 2219 conv = n; 2220 } else { 2221 msg = "unhandled input to ConvI2L"; 2222 } 2223 } else if (n == head->phi()) { 2224 // no shift, check below for allowed cases 2225 found_index = true; 2226 } else { 2227 msg = "unhandled node in address"; 2228 msg_node = n; 2229 } 2230 } 2231 2232 if (count == -1) { 2233 msg = "malformed address expression"; 2234 msg_node = store; 2235 } 2236 2237 if (!found_index) { 2238 msg = "missing use of index"; 2239 } 2240 2241 // byte sized items won't have a shift 2242 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2243 msg = "can't find shift"; 2244 msg_node = store; 2245 } 2246 2247 if (msg != NULL) { 2248 #ifndef PRODUCT 2249 if (TraceOptimizeFill) { 2250 tty->print_cr("not fill intrinsic: %s", msg); 2251 if (msg_node != NULL) msg_node->dump(); 2252 } 2253 #endif 2254 return false; 2255 } 2256 2257 // No make sure all the other nodes in the loop can be handled 2258 VectorSet ok(Thread::current()->resource_area()); 2259 2260 // store related values are ok 2261 ok.set(store->_idx); 2262 ok.set(store->in(MemNode::Memory)->_idx); 2263 2264 // Loop structure is ok 2265 ok.set(head->_idx); 2266 ok.set(head->loopexit()->_idx); 2267 ok.set(head->phi()->_idx); 2268 ok.set(head->incr()->_idx); 2269 ok.set(head->loopexit()->cmp_node()->_idx); 2270 ok.set(head->loopexit()->in(1)->_idx); 2271 2272 // Address elements are ok 2273 if (con) ok.set(con->_idx); 2274 if (shift) ok.set(shift->_idx); 2275 if (conv) ok.set(conv->_idx); 2276 2277 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2278 Node* n = lpt->_body.at(i); 2279 if (n->outcnt() == 0) continue; // Ignore dead 2280 if (ok.test(n->_idx)) continue; 2281 // Backedge projection is ok 2282 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue; 2283 if (!n->is_AddP()) { 2284 msg = "unhandled node"; 2285 msg_node = n; 2286 break; 2287 } 2288 } 2289 2290 // Make sure no unexpected values are used outside the loop 2291 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2292 Node* n = lpt->_body.at(i); 2293 // These values can be replaced with other nodes if they are used 2294 // outside the loop. 2295 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2296 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2297 Node* use = iter.get(); 2298 if (!lpt->_body.contains(use)) { 2299 msg = "node is used outside loop"; 2300 // lpt->_body.dump(); 2301 msg_node = n; 2302 break; 2303 } 2304 } 2305 } 2306 2307 #ifdef ASSERT 2308 if (TraceOptimizeFill) { 2309 if (msg != NULL) { 2310 tty->print_cr("no fill intrinsic: %s", msg); 2311 if (msg_node != NULL) msg_node->dump(); 2312 } else { 2313 tty->print_cr("fill intrinsic for:"); 2314 } 2315 store->dump(); 2316 if (Verbose) { 2317 lpt->_body.dump(); 2318 } 2319 } 2320 #endif 2321 2322 return msg == NULL; 2323 } 2324 2325 2326 2327 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2328 // Only for counted inner loops 2329 if (!lpt->is_counted() || !lpt->is_inner()) { 2330 return false; 2331 } 2332 2333 // Must have constant stride 2334 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2335 if (!head->stride_is_con() || !head->is_normal_loop()) { 2336 return false; 2337 } 2338 2339 // Check that the body only contains a store of a loop invariant 2340 // value that is indexed by the loop phi. 2341 Node* store = NULL; 2342 Node* store_value = NULL; 2343 Node* shift = NULL; 2344 Node* offset = NULL; 2345 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2346 return false; 2347 } 2348 2349 #ifndef PRODUCT 2350 if (TraceLoopOpts) { 2351 tty->print("ArrayFill "); 2352 lpt->dump_head(); 2353 } 2354 #endif 2355 2356 // Now replace the whole loop body by a call to a fill routine that 2357 // covers the same region as the loop. 2358 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2359 2360 // Build an expression for the beginning of the copy region 2361 Node* index = head->init_trip(); 2362 #ifdef _LP64 2363 index = new (C, 2) ConvI2LNode(index); 2364 _igvn.register_new_node_with_optimizer(index); 2365 #endif 2366 if (shift != NULL) { 2367 // byte arrays don't require a shift but others do. 2368 index = new (C, 3) LShiftXNode(index, shift->in(2)); 2369 _igvn.register_new_node_with_optimizer(index); 2370 } 2371 index = new (C, 4) AddPNode(base, base, index); 2372 _igvn.register_new_node_with_optimizer(index); 2373 Node* from = new (C, 4) AddPNode(base, index, offset); 2374 _igvn.register_new_node_with_optimizer(from); 2375 // Compute the number of elements to copy 2376 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip()); 2377 _igvn.register_new_node_with_optimizer(len); 2378 2379 BasicType t = store->as_Mem()->memory_type(); 2380 bool aligned = false; 2381 if (offset != NULL && head->init_trip()->is_Con()) { 2382 int element_size = type2aelembytes(t); 2383 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2384 } 2385 2386 // Build a call to the fill routine 2387 const char* fill_name; 2388 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2389 assert(fill != NULL, "what?"); 2390 2391 // Convert float/double to int/long for fill routines 2392 if (t == T_FLOAT) { 2393 store_value = new (C, 2) MoveF2INode(store_value); 2394 _igvn.register_new_node_with_optimizer(store_value); 2395 } else if (t == T_DOUBLE) { 2396 store_value = new (C, 2) MoveD2LNode(store_value); 2397 _igvn.register_new_node_with_optimizer(store_value); 2398 } 2399 2400 Node* mem_phi = store->in(MemNode::Memory); 2401 Node* result_ctrl; 2402 Node* result_mem; 2403 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2404 int size = call_type->domain()->cnt(); 2405 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill, 2406 fill_name, TypeAryPtr::get_array_body_type(t)); 2407 call->init_req(TypeFunc::Parms+0, from); 2408 call->init_req(TypeFunc::Parms+1, store_value); 2409 #ifdef _LP64 2410 len = new (C, 2) ConvI2LNode(len); 2411 _igvn.register_new_node_with_optimizer(len); 2412 #endif 2413 call->init_req(TypeFunc::Parms+2, len); 2414 #ifdef _LP64 2415 call->init_req(TypeFunc::Parms+3, C->top()); 2416 #endif 2417 call->init_req( TypeFunc::Control, head->init_control()); 2418 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o 2419 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) ); 2420 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) ); 2421 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) ); 2422 _igvn.register_new_node_with_optimizer(call); 2423 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); 2424 _igvn.register_new_node_with_optimizer(result_ctrl); 2425 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory); 2426 _igvn.register_new_node_with_optimizer(result_mem); 2427 2428 // If this fill is tightly coupled to an allocation and overwrites 2429 // the whole body, allow it to take over the zeroing. 2430 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 2431 if (alloc != NULL && alloc->is_AllocateArray()) { 2432 Node* length = alloc->as_AllocateArray()->Ideal_length(); 2433 if (head->limit() == length && 2434 head->init_trip() == _igvn.intcon(0)) { 2435 if (TraceOptimizeFill) { 2436 tty->print_cr("Eliminated zeroing in allocation"); 2437 } 2438 alloc->maybe_set_complete(&_igvn); 2439 } else { 2440 #ifdef ASSERT 2441 if (TraceOptimizeFill) { 2442 tty->print_cr("filling array but bounds don't match"); 2443 alloc->dump(); 2444 head->init_trip()->dump(); 2445 head->limit()->dump(); 2446 length->dump(); 2447 } 2448 #endif 2449 } 2450 } 2451 2452 // Redirect the old control and memory edges that are outside the loop. 2453 Node* exit = head->loopexit()->proj_out(0); 2454 // Sometimes the memory phi of the head is used as the outgoing 2455 // state of the loop. It's safe in this case to replace it with the 2456 // result_mem. 2457 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 2458 _igvn.replace_node(exit, result_ctrl); 2459 _igvn.replace_node(store, result_mem); 2460 // Any uses the increment outside of the loop become the loop limit. 2461 _igvn.replace_node(head->incr(), head->limit()); 2462 2463 // Disconnect the head from the loop. 2464 for (uint i = 0; i < lpt->_body.size(); i++) { 2465 Node* n = lpt->_body.at(i); 2466 _igvn.replace_node(n, C->top()); 2467 } 2468 2469 return true; 2470 }