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) 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) SubINode(n_inv1, inv2); 235 } else { 236 inv = new (phase->C) 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) SubINode(inv, x); 243 } else { 244 addx = new (phase->C) 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_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 513 _igvn.hash_delete(head); 514 head->set_req(LoopNode::EntryControl, new_entry); 515 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 516 Node* old = head->fast_out(j); 517 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 518 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 519 if (!new_exit_value ) // Backedge value is ALSO loop invariant? 520 // Then loop body backedge value remains the same. 521 new_exit_value = old->in(LoopNode::LoopBackControl); 522 _igvn.hash_delete(old); 523 old->set_req(LoopNode::EntryControl, new_exit_value); 524 } 525 } 526 527 528 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 529 // extra backedge user. 530 Node* new_head = old_new[head->_idx]; 531 _igvn.hash_delete(new_head); 532 new_head->set_req(LoopNode::LoopBackControl, C->top()); 533 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 534 Node* use = new_head->fast_out(j2); 535 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 536 _igvn.hash_delete(use); 537 use->set_req(LoopNode::LoopBackControl, C->top()); 538 } 539 } 540 541 542 // Step 4: Correct dom-depth info. Set to loop-head depth. 543 int dd = dom_depth(head); 544 set_idom(head, head->in(1), dd); 545 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 546 Node *old = loop->_body.at(j3); 547 Node *nnn = old_new[old->_idx]; 548 if (!has_ctrl(nnn)) 549 set_idom(nnn, idom(nnn), dd-1); 550 } 551 552 // Now force out all loop-invariant dominating tests. The optimizer 553 // finds some, but we _know_ they are all useless. 554 peeled_dom_test_elim(loop,old_new); 555 556 loop->record_for_igvn(); 557 } 558 559 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 560 561 //------------------------------policy_maximally_unroll------------------------ 562 // Calculate exact loop trip count and return true if loop can be maximally 563 // unrolled. 564 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 565 CountedLoopNode *cl = _head->as_CountedLoop(); 566 assert(cl->is_normal_loop(), ""); 567 if (!cl->is_valid_counted_loop()) 568 return false; // Malformed counted loop 569 570 if (!cl->has_exact_trip_count()) { 571 // Trip count is not exact. 572 return false; 573 } 574 575 uint trip_count = cl->trip_count(); 576 // Note, max_juint is used to indicate unknown trip count. 577 assert(trip_count > 1, "one iteration loop should be optimized out already"); 578 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 579 580 // Real policy: if we maximally unroll, does it get too big? 581 // Allow the unrolled mess to get larger than standard loop 582 // size. After all, it will no longer be a loop. 583 uint body_size = _body.size(); 584 uint unroll_limit = (uint)LoopUnrollLimit * 4; 585 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 586 if (trip_count > unroll_limit || body_size > unroll_limit) { 587 return false; 588 } 589 590 // Fully unroll a loop with few iterations regardless next 591 // conditions since following loop optimizations will split 592 // such loop anyway (pre-main-post). 593 if (trip_count <= 3) 594 return true; 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 // Do not unroll a loop with String intrinsics code. 609 // String intrinsics are large and have loops. 610 for (uint k = 0; k < _body.size(); k++) { 611 Node* n = _body.at(k); 612 switch (n->Opcode()) { 613 case Op_StrComp: 614 case Op_StrEquals: 615 case Op_StrIndexOf: 616 case Op_AryEq: { 617 return false; 618 } 619 } // switch 620 } 621 622 return true; // Do maximally unroll 623 } 624 625 626 #define MAX_UNROLL 16 // maximum number of unrolls for main loop 627 628 //------------------------------policy_unroll---------------------------------- 629 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 630 // the loop is a CountedLoop and the body is small enough. 631 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { 632 633 CountedLoopNode *cl = _head->as_CountedLoop(); 634 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 635 636 if (!cl->is_valid_counted_loop()) 637 return false; // Malformed counted loop 638 639 // Protect against over-unrolling. 640 // After split at least one iteration will be executed in pre-loop. 641 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; 642 643 int future_unroll_ct = cl->unrolled_count() * 2; 644 if (future_unroll_ct > MAX_UNROLL) return false; 645 646 // Check for initial stride being a small enough constant 647 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; 648 649 // Don't unroll if the next round of unrolling would push us 650 // over the expected trip count of the loop. One is subtracted 651 // from the expected trip count because the pre-loop normally 652 // executes 1 iteration. 653 if (UnrollLimitForProfileCheck > 0 && 654 cl->profile_trip_cnt() != COUNT_UNKNOWN && 655 future_unroll_ct > UnrollLimitForProfileCheck && 656 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 657 return false; 658 } 659 660 // When unroll count is greater than LoopUnrollMin, don't unroll if: 661 // the residual iterations are more than 10% of the trip count 662 // and rounds of "unroll,optimize" are not making significant progress 663 // Progress defined as current size less than 20% larger than previous size. 664 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 665 future_unroll_ct > LoopUnrollMin && 666 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 667 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 668 return false; 669 } 670 671 Node *init_n = cl->init_trip(); 672 Node *limit_n = cl->limit(); 673 int stride_con = cl->stride_con(); 674 // Non-constant bounds. 675 // Protect against over-unrolling when init or/and limit are not constant 676 // (so that trip_count's init value is maxint) but iv range is known. 677 if (init_n == NULL || !init_n->is_Con() || 678 limit_n == NULL || !limit_n->is_Con()) { 679 Node* phi = cl->phi(); 680 if (phi != NULL) { 681 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 682 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 683 int next_stride = stride_con * 2; // stride after this unroll 684 if (next_stride > 0) { 685 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 686 iv_type->_lo + next_stride > iv_type->_hi) { 687 return false; // over-unrolling 688 } 689 } else if (next_stride < 0) { 690 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 691 iv_type->_hi + next_stride < iv_type->_lo) { 692 return false; // over-unrolling 693 } 694 } 695 } 696 } 697 698 // After unroll limit will be adjusted: new_limit = limit-stride. 699 // Bailout if adjustment overflow. 700 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 701 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || 702 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) 703 return false; // overflow 704 705 // Adjust body_size to determine if we unroll or not 706 uint body_size = _body.size(); 707 // Key test to unroll loop in CRC32 java code 708 int xors_in_loop = 0; 709 // Also count ModL, DivL and MulL which expand mightly 710 for (uint k = 0; k < _body.size(); k++) { 711 Node* n = _body.at(k); 712 switch (n->Opcode()) { 713 case Op_XorI: xors_in_loop++; break; // CRC32 java code 714 case Op_ModL: body_size += 30; break; 715 case Op_DivL: body_size += 30; break; 716 case Op_MulL: body_size += 10; break; 717 case Op_StrComp: 718 case Op_StrEquals: 719 case Op_StrIndexOf: 720 case Op_AryEq: { 721 // Do not unroll a loop with String intrinsics code. 722 // String intrinsics are large and have loops. 723 return false; 724 } 725 } // switch 726 } 727 728 // Check for being too big 729 if (body_size > (uint)LoopUnrollLimit) { 730 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 731 // Normal case: loop too big 732 return false; 733 } 734 735 // Unroll once! (Each trip will soon do double iterations) 736 return true; 737 } 738 739 //------------------------------policy_align----------------------------------- 740 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 741 // expression that does the alignment. Note that only one array base can be 742 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 743 // if we vectorize short memory ops into longer memory ops, we may want to 744 // increase alignment. 745 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 746 return false; 747 } 748 749 //------------------------------policy_range_check----------------------------- 750 // Return TRUE or FALSE if the loop should be range-check-eliminated. 751 // Actually we do iteration-splitting, a more powerful form of RCE. 752 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 753 if (!RangeCheckElimination) return false; 754 755 CountedLoopNode *cl = _head->as_CountedLoop(); 756 // If we unrolled with no intention of doing RCE and we later 757 // changed our minds, we got no pre-loop. Either we need to 758 // make a new pre-loop, or we gotta disallow RCE. 759 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 760 Node *trip_counter = cl->phi(); 761 762 // Check loop body for tests of trip-counter plus loop-invariant vs 763 // loop-invariant. 764 for (uint i = 0; i < _body.size(); i++) { 765 Node *iff = _body[i]; 766 if (iff->Opcode() == Op_If) { // Test? 767 768 // Comparing trip+off vs limit 769 Node *bol = iff->in(1); 770 if (bol->req() != 2) continue; // dead constant test 771 if (!bol->is_Bool()) { 772 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 773 continue; 774 } 775 if (bol->as_Bool()->_test._test == BoolTest::ne) 776 continue; // not RC 777 778 Node *cmp = bol->in(1); 779 780 Node *rc_exp = cmp->in(1); 781 Node *limit = cmp->in(2); 782 783 Node *limit_c = phase->get_ctrl(limit); 784 if( limit_c == phase->C->top() ) 785 return false; // Found dead test on live IF? No RCE! 786 if( is_member(phase->get_loop(limit_c) ) ) { 787 // Compare might have operands swapped; commute them 788 rc_exp = cmp->in(2); 789 limit = cmp->in(1); 790 limit_c = phase->get_ctrl(limit); 791 if( is_member(phase->get_loop(limit_c) ) ) 792 continue; // Both inputs are loop varying; cannot RCE 793 } 794 795 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 796 continue; 797 } 798 // Yeah! Found a test like 'trip+off vs limit' 799 // Test is an IfNode, has 2 projections. If BOTH are in the loop 800 // we need loop unswitching instead of iteration splitting. 801 if( is_loop_exit(iff) ) 802 return true; // Found reason to split iterations 803 } // End of is IF 804 } 805 806 return false; 807 } 808 809 //------------------------------policy_peel_only------------------------------- 810 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 811 // for unrolling loops with NO array accesses. 812 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 813 814 for( uint i = 0; i < _body.size(); i++ ) 815 if( _body[i]->is_Mem() ) 816 return false; 817 818 // No memory accesses at all! 819 return true; 820 } 821 822 //------------------------------clone_up_backedge_goo-------------------------- 823 // If Node n lives in the back_ctrl block and cannot float, we clone a private 824 // version of n in preheader_ctrl block and return that, otherwise return n. 825 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { 826 if( get_ctrl(n) != back_ctrl ) return n; 827 828 // Only visit once 829 if (visited.test_set(n->_idx)) { 830 Node *x = clones.find(n->_idx); 831 if (x != NULL) 832 return x; 833 return n; 834 } 835 836 Node *x = NULL; // If required, a clone of 'n' 837 // Check for 'n' being pinned in the backedge. 838 if( n->in(0) && n->in(0) == back_ctrl ) { 839 assert(clones.find(n->_idx) == NULL, "dead loop"); 840 x = n->clone(); // Clone a copy of 'n' to preheader 841 clones.push(x, n->_idx); 842 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 843 } 844 845 // Recursive fixup any other input edges into x. 846 // If there are no changes we can just return 'n', otherwise 847 // we need to clone a private copy and change it. 848 for( uint i = 1; i < n->req(); i++ ) { 849 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); 850 if( g != n->in(i) ) { 851 if( !x ) { 852 assert(clones.find(n->_idx) == NULL, "dead loop"); 853 x = n->clone(); 854 clones.push(x, n->_idx); 855 } 856 x->set_req(i, g); 857 } 858 } 859 if( x ) { // x can legally float to pre-header location 860 register_new_node( x, preheader_ctrl ); 861 return x; 862 } else { // raise n to cover LCA of uses 863 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 864 } 865 return n; 866 } 867 868 //------------------------------insert_pre_post_loops-------------------------- 869 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 870 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 871 // alignment. Useful to unroll loops that do no array accesses. 872 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 873 874 #ifndef PRODUCT 875 if (TraceLoopOpts) { 876 if (peel_only) 877 tty->print("PeelMainPost "); 878 else 879 tty->print("PreMainPost "); 880 loop->dump_head(); 881 } 882 #endif 883 C->set_major_progress(); 884 885 // Find common pieces of the loop being guarded with pre & post loops 886 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 887 assert( main_head->is_normal_loop(), "" ); 888 CountedLoopEndNode *main_end = main_head->loopexit(); 889 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 890 uint dd_main_head = dom_depth(main_head); 891 uint max = main_head->outcnt(); 892 893 Node *pre_header= main_head->in(LoopNode::EntryControl); 894 Node *init = main_head->init_trip(); 895 Node *incr = main_end ->incr(); 896 Node *limit = main_end ->limit(); 897 Node *stride = main_end ->stride(); 898 Node *cmp = main_end ->cmp_node(); 899 BoolTest::mask b_test = main_end->test_trip(); 900 901 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 902 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 903 if( bol->outcnt() != 1 ) { 904 bol = bol->clone(); 905 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 906 _igvn.hash_delete(main_end); 907 main_end->set_req(CountedLoopEndNode::TestValue, bol); 908 } 909 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 910 if( cmp->outcnt() != 1 ) { 911 cmp = cmp->clone(); 912 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 913 _igvn.hash_delete(bol); 914 bol->set_req(1, cmp); 915 } 916 917 //------------------------------ 918 // Step A: Create Post-Loop. 919 Node* main_exit = main_end->proj_out(false); 920 assert( main_exit->Opcode() == Op_IfFalse, "" ); 921 int dd_main_exit = dom_depth(main_exit); 922 923 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 924 // loop pre-header illegally has 2 control users (old & new loops). 925 clone_loop( loop, old_new, dd_main_exit ); 926 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 927 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 928 post_head->set_post_loop(main_head); 929 930 // Reduce the post-loop trip count. 931 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 932 post_end->_prob = PROB_FAIR; 933 934 // Build the main-loop normal exit. 935 IfFalseNode *new_main_exit = new (C) IfFalseNode(main_end); 936 _igvn.register_new_node_with_optimizer( new_main_exit ); 937 set_idom(new_main_exit, main_end, dd_main_exit ); 938 set_loop(new_main_exit, loop->_parent); 939 940 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 941 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 942 // (the main-loop trip-counter exit value) because we will be changing 943 // the exit value (via unrolling) so we cannot constant-fold away the zero 944 // trip guard until all unrolling is done. 945 Node *zer_opaq = new (C) Opaque1Node(C, incr); 946 Node *zer_cmp = new (C) CmpINode( zer_opaq, limit ); 947 Node *zer_bol = new (C) BoolNode( zer_cmp, b_test ); 948 register_new_node( zer_opaq, new_main_exit ); 949 register_new_node( zer_cmp , new_main_exit ); 950 register_new_node( zer_bol , new_main_exit ); 951 952 // Build the IfNode 953 IfNode *zer_iff = new (C) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 954 _igvn.register_new_node_with_optimizer( zer_iff ); 955 set_idom(zer_iff, new_main_exit, dd_main_exit); 956 set_loop(zer_iff, loop->_parent); 957 958 // Plug in the false-path, taken if we need to skip post-loop 959 _igvn.replace_input_of(main_exit, 0, zer_iff); 960 set_idom(main_exit, zer_iff, dd_main_exit); 961 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 962 // Make the true-path, must enter the post loop 963 Node *zer_taken = new (C) IfTrueNode( zer_iff ); 964 _igvn.register_new_node_with_optimizer( zer_taken ); 965 set_idom(zer_taken, zer_iff, dd_main_exit); 966 set_loop(zer_taken, loop->_parent); 967 // Plug in the true path 968 _igvn.hash_delete( post_head ); 969 post_head->set_req(LoopNode::EntryControl, zer_taken); 970 set_idom(post_head, zer_taken, dd_main_exit); 971 972 Arena *a = Thread::current()->resource_area(); 973 VectorSet visited(a); 974 Node_Stack clones(a, main_head->back_control()->outcnt()); 975 // Step A3: Make the fall-in values to the post-loop come from the 976 // fall-out values of the main-loop. 977 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 978 Node* main_phi = main_head->fast_out(i); 979 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 980 Node *post_phi = old_new[main_phi->_idx]; 981 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 982 post_head->init_control(), 983 main_phi->in(LoopNode::LoopBackControl), 984 visited, clones); 985 _igvn.hash_delete(post_phi); 986 post_phi->set_req( LoopNode::EntryControl, fallmain ); 987 } 988 } 989 990 // Update local caches for next stanza 991 main_exit = new_main_exit; 992 993 994 //------------------------------ 995 // Step B: Create Pre-Loop. 996 997 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 998 // loop pre-header illegally has 2 control users (old & new loops). 999 clone_loop( loop, old_new, dd_main_head ); 1000 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1001 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1002 pre_head->set_pre_loop(main_head); 1003 Node *pre_incr = old_new[incr->_idx]; 1004 1005 // Reduce the pre-loop trip count. 1006 pre_end->_prob = PROB_FAIR; 1007 1008 // Find the pre-loop normal exit. 1009 Node* pre_exit = pre_end->proj_out(false); 1010 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 1011 IfFalseNode *new_pre_exit = new (C) IfFalseNode(pre_end); 1012 _igvn.register_new_node_with_optimizer( new_pre_exit ); 1013 set_idom(new_pre_exit, pre_end, dd_main_head); 1014 set_loop(new_pre_exit, loop->_parent); 1015 1016 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1017 // pre-loop, the main-loop may not execute at all. Later in life this 1018 // zero-trip guard will become the minimum-trip guard when we unroll 1019 // the main-loop. 1020 Node *min_opaq = new (C) Opaque1Node(C, limit); 1021 Node *min_cmp = new (C) CmpINode( pre_incr, min_opaq ); 1022 Node *min_bol = new (C) BoolNode( min_cmp, b_test ); 1023 register_new_node( min_opaq, new_pre_exit ); 1024 register_new_node( min_cmp , new_pre_exit ); 1025 register_new_node( min_bol , new_pre_exit ); 1026 1027 // Build the IfNode (assume the main-loop is executed always). 1028 IfNode *min_iff = new (C) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 1029 _igvn.register_new_node_with_optimizer( min_iff ); 1030 set_idom(min_iff, new_pre_exit, dd_main_head); 1031 set_loop(min_iff, loop->_parent); 1032 1033 // Plug in the false-path, taken if we need to skip main-loop 1034 _igvn.hash_delete( pre_exit ); 1035 pre_exit->set_req(0, min_iff); 1036 set_idom(pre_exit, min_iff, dd_main_head); 1037 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 1038 // Make the true-path, must enter the main loop 1039 Node *min_taken = new (C) IfTrueNode( min_iff ); 1040 _igvn.register_new_node_with_optimizer( min_taken ); 1041 set_idom(min_taken, min_iff, dd_main_head); 1042 set_loop(min_taken, loop->_parent); 1043 // Plug in the true path 1044 _igvn.hash_delete( main_head ); 1045 main_head->set_req(LoopNode::EntryControl, min_taken); 1046 set_idom(main_head, min_taken, dd_main_head); 1047 1048 visited.Clear(); 1049 clones.clear(); 1050 // Step B3: Make the fall-in values to the main-loop come from the 1051 // fall-out values of the pre-loop. 1052 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1053 Node* main_phi = main_head->fast_out(i2); 1054 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 1055 Node *pre_phi = old_new[main_phi->_idx]; 1056 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1057 main_head->init_control(), 1058 pre_phi->in(LoopNode::LoopBackControl), 1059 visited, clones); 1060 _igvn.hash_delete(main_phi); 1061 main_phi->set_req( LoopNode::EntryControl, fallpre ); 1062 } 1063 } 1064 1065 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1066 // RCE and alignment may change this later. 1067 Node *cmp_end = pre_end->cmp_node(); 1068 assert( cmp_end->in(2) == limit, "" ); 1069 Node *pre_limit = new (C) AddINode( init, stride ); 1070 1071 // Save the original loop limit in this Opaque1 node for 1072 // use by range check elimination. 1073 Node *pre_opaq = new (C) Opaque1Node(C, pre_limit, limit); 1074 1075 register_new_node( pre_limit, pre_head->in(0) ); 1076 register_new_node( pre_opaq , pre_head->in(0) ); 1077 1078 // Since no other users of pre-loop compare, I can hack limit directly 1079 assert( cmp_end->outcnt() == 1, "no other users" ); 1080 _igvn.hash_delete(cmp_end); 1081 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1082 1083 // Special case for not-equal loop bounds: 1084 // Change pre loop test, main loop test, and the 1085 // main loop guard test to use lt or gt depending on stride 1086 // direction: 1087 // positive stride use < 1088 // negative stride use > 1089 // 1090 // not-equal test is kept for post loop to handle case 1091 // when init > limit when stride > 0 (and reverse). 1092 1093 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1094 1095 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1096 // Modify pre loop end condition 1097 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1098 BoolNode* new_bol0 = new (C) BoolNode(pre_bol->in(1), new_test); 1099 register_new_node( new_bol0, pre_head->in(0) ); 1100 _igvn.hash_delete(pre_end); 1101 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); 1102 // Modify main loop guard condition 1103 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1104 BoolNode* new_bol1 = new (C) BoolNode(min_bol->in(1), new_test); 1105 register_new_node( new_bol1, new_pre_exit ); 1106 _igvn.hash_delete(min_iff); 1107 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1108 // Modify main loop end condition 1109 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1110 BoolNode* new_bol2 = new (C) BoolNode(main_bol->in(1), new_test); 1111 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 1112 _igvn.hash_delete(main_end); 1113 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); 1114 } 1115 1116 // Flag main loop 1117 main_head->set_main_loop(); 1118 if( peel_only ) main_head->set_main_no_pre_loop(); 1119 1120 // Subtract a trip count for the pre-loop. 1121 main_head->set_trip_count(main_head->trip_count() - 1); 1122 1123 // It's difficult to be precise about the trip-counts 1124 // for the pre/post loops. They are usually very short, 1125 // so guess that 4 trips is a reasonable value. 1126 post_head->set_profile_trip_cnt(4.0); 1127 pre_head->set_profile_trip_cnt(4.0); 1128 1129 // Now force out all loop-invariant dominating tests. The optimizer 1130 // finds some, but we _know_ they are all useless. 1131 peeled_dom_test_elim(loop,old_new); 1132 } 1133 1134 //------------------------------is_invariant----------------------------- 1135 // Return true if n is invariant 1136 bool IdealLoopTree::is_invariant(Node* n) const { 1137 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1138 if (n_c->is_top()) return false; 1139 return !is_member(_phase->get_loop(n_c)); 1140 } 1141 1142 1143 //------------------------------do_unroll-------------------------------------- 1144 // Unroll the loop body one step - make each trip do 2 iterations. 1145 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 1146 assert(LoopUnrollLimit, ""); 1147 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1148 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1149 assert(loop_end, ""); 1150 #ifndef PRODUCT 1151 if (PrintOpto && VerifyLoopOptimizations) { 1152 tty->print("Unrolling "); 1153 loop->dump_head(); 1154 } else if (TraceLoopOpts) { 1155 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1156 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1157 } else { 1158 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1159 } 1160 loop->dump_head(); 1161 } 1162 #endif 1163 1164 // Remember loop node count before unrolling to detect 1165 // if rounds of unroll,optimize are making progress 1166 loop_head->set_node_count_before_unroll(loop->_body.size()); 1167 1168 Node *ctrl = loop_head->in(LoopNode::EntryControl); 1169 Node *limit = loop_head->limit(); 1170 Node *init = loop_head->init_trip(); 1171 Node *stride = loop_head->stride(); 1172 1173 Node *opaq = NULL; 1174 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1175 // Search for zero-trip guard. 1176 assert( loop_head->is_main_loop(), "" ); 1177 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1178 Node *iff = ctrl->in(0); 1179 assert( iff->Opcode() == Op_If, "" ); 1180 Node *bol = iff->in(1); 1181 assert( bol->Opcode() == Op_Bool, "" ); 1182 Node *cmp = bol->in(1); 1183 assert( cmp->Opcode() == Op_CmpI, "" ); 1184 opaq = cmp->in(2); 1185 // Occasionally it's possible for a zero-trip guard Opaque1 node to be 1186 // optimized away and then another round of loop opts attempted. 1187 // We can not optimize this particular loop in that case. 1188 if (opaq->Opcode() != Op_Opaque1) 1189 return; // Cannot find zero-trip guard! Bail out! 1190 // Zero-trip test uses an 'opaque' node which is not shared. 1191 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1192 } 1193 1194 C->set_major_progress(); 1195 1196 Node* new_limit = NULL; 1197 if (UnrollLimitCheck) { 1198 int stride_con = stride->get_int(); 1199 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1200 uint old_trip_count = loop_head->trip_count(); 1201 // Verify that unroll policy result is still valid. 1202 assert(old_trip_count > 1 && 1203 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1204 1205 // Adjust loop limit to keep valid iterations number after unroll. 1206 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1207 // which may overflow. 1208 if (!adjust_min_trip) { 1209 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1210 "odd trip count for maximally unroll"); 1211 // Don't need to adjust limit for maximally unroll since trip count is even. 1212 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1213 // Loop's limit is constant. Loop's init could be constant when pre-loop 1214 // become peeled iteration. 1215 long init_con = init->get_int(); 1216 // We can keep old loop limit if iterations count stays the same: 1217 // old_trip_count == new_trip_count * 2 1218 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1219 // so we also don't need to adjust zero trip test. 1220 long limit_con = limit->get_int(); 1221 // (stride_con*2) not overflow since stride_con <= 8. 1222 int new_stride_con = stride_con * 2; 1223 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1224 long trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1225 // New trip count should satisfy next conditions. 1226 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1227 uint new_trip_count = (uint)trip_count; 1228 adjust_min_trip = (old_trip_count != new_trip_count*2); 1229 } 1230 1231 if (adjust_min_trip) { 1232 // Step 2: Adjust the trip limit if it is called for. 1233 // The adjustment amount is -stride. Need to make sure if the 1234 // adjustment underflows or overflows, then the main loop is skipped. 1235 Node* cmp = loop_end->cmp_node(); 1236 assert(cmp->in(2) == limit, "sanity"); 1237 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1238 1239 // Verify that policy_unroll result is still valid. 1240 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1241 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1242 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); 1243 1244 if (limit->is_Con()) { 1245 // The check in policy_unroll and the assert above guarantee 1246 // no underflow if limit is constant. 1247 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1248 set_ctrl(new_limit, C->root()); 1249 } else { 1250 // Limit is not constant. 1251 if (loop_head->unrolled_count() == 1) { // only for first unroll 1252 // Separate limit by Opaque node in case it is an incremented 1253 // variable from previous loop to avoid using pre-incremented 1254 // value which could increase register pressure. 1255 // Otherwise reorg_offsets() optimization will create a separate 1256 // Opaque node for each use of trip-counter and as result 1257 // zero trip guard limit will be different from loop limit. 1258 assert(has_ctrl(opaq), "should have it"); 1259 Node* opaq_ctrl = get_ctrl(opaq); 1260 limit = new (C) Opaque2Node( C, limit ); 1261 register_new_node( limit, opaq_ctrl ); 1262 } 1263 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) || 1264 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) { 1265 // No underflow. 1266 new_limit = new (C) SubINode(limit, stride); 1267 } else { 1268 // (limit - stride) may underflow. 1269 // Clamp the adjustment value with MININT or MAXINT: 1270 // 1271 // new_limit = limit-stride 1272 // if (stride > 0) 1273 // new_limit = (limit < new_limit) ? MININT : new_limit; 1274 // else 1275 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1276 // 1277 BoolTest::mask bt = loop_end->test_trip(); 1278 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1279 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1280 set_ctrl(adj_max, C->root()); 1281 Node* old_limit = NULL; 1282 Node* adj_limit = NULL; 1283 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1284 if (loop_head->unrolled_count() > 1 && 1285 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1286 limit->in(CMoveNode::IfTrue) == adj_max && 1287 bol->as_Bool()->_test._test == bt && 1288 bol->in(1)->Opcode() == Op_CmpI && 1289 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1290 // Loop was unrolled before. 1291 // Optimize the limit to avoid nested CMove: 1292 // use original limit as old limit. 1293 old_limit = bol->in(1)->in(1); 1294 // Adjust previous adjusted limit. 1295 adj_limit = limit->in(CMoveNode::IfFalse); 1296 adj_limit = new (C) SubINode(adj_limit, stride); 1297 } else { 1298 old_limit = limit; 1299 adj_limit = new (C) SubINode(limit, stride); 1300 } 1301 assert(old_limit != NULL && adj_limit != NULL, ""); 1302 register_new_node( adj_limit, ctrl ); // adjust amount 1303 Node* adj_cmp = new (C) CmpINode(old_limit, adj_limit); 1304 register_new_node( adj_cmp, ctrl ); 1305 Node* adj_bool = new (C) BoolNode(adj_cmp, bt); 1306 register_new_node( adj_bool, ctrl ); 1307 new_limit = new (C) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1308 } 1309 register_new_node(new_limit, ctrl); 1310 } 1311 assert(new_limit != NULL, ""); 1312 // Replace in loop test. 1313 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1314 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1315 // Don't need to create new test since only one user. 1316 _igvn.hash_delete(cmp); 1317 cmp->set_req(2, new_limit); 1318 } else { 1319 // Create new test since it is shared. 1320 Node* ctrl2 = loop_end->in(0); 1321 Node* cmp2 = cmp->clone(); 1322 cmp2->set_req(2, new_limit); 1323 register_new_node(cmp2, ctrl2); 1324 Node* bol2 = loop_end->in(1)->clone(); 1325 bol2->set_req(1, cmp2); 1326 register_new_node(bol2, ctrl2); 1327 _igvn.hash_delete(loop_end); 1328 loop_end->set_req(1, bol2); 1329 } 1330 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1331 // Make it a 1-trip test (means at least 2 trips). 1332 1333 // Guard test uses an 'opaque' node which is not shared. Hence I 1334 // can edit it's inputs directly. Hammer in the new limit for the 1335 // minimum-trip guard. 1336 assert(opaq->outcnt() == 1, ""); 1337 _igvn.hash_delete(opaq); 1338 opaq->set_req(1, new_limit); 1339 } 1340 1341 // Adjust max trip count. The trip count is intentionally rounded 1342 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1343 // the main, unrolled, part of the loop will never execute as it is protected 1344 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1345 // and later determined that part of the unrolled loop was dead. 1346 loop_head->set_trip_count(old_trip_count / 2); 1347 1348 // Double the count of original iterations in the unrolled loop body. 1349 loop_head->double_unrolled_count(); 1350 1351 } else { // LoopLimitCheck 1352 1353 // Adjust max trip count. The trip count is intentionally rounded 1354 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1355 // the main, unrolled, part of the loop will never execute as it is protected 1356 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1357 // and later determined that part of the unrolled loop was dead. 1358 loop_head->set_trip_count(loop_head->trip_count() / 2); 1359 1360 // Double the count of original iterations in the unrolled loop body. 1361 loop_head->double_unrolled_count(); 1362 1363 // ----------- 1364 // Step 2: Cut back the trip counter for an unroll amount of 2. 1365 // Loop will normally trip (limit - init)/stride_con. Since it's a 1366 // CountedLoop this is exact (stride divides limit-init exactly). 1367 // We are going to double the loop body, so we want to knock off any 1368 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 1369 Node *span = new (C) SubINode( limit, init ); 1370 register_new_node( span, ctrl ); 1371 Node *trip = new (C) DivINode( 0, span, stride ); 1372 register_new_node( trip, ctrl ); 1373 Node *mtwo = _igvn.intcon(-2); 1374 set_ctrl(mtwo, C->root()); 1375 Node *rond = new (C) AndINode( trip, mtwo ); 1376 register_new_node( rond, ctrl ); 1377 Node *spn2 = new (C) MulINode( rond, stride ); 1378 register_new_node( spn2, ctrl ); 1379 new_limit = new (C) AddINode( spn2, init ); 1380 register_new_node( new_limit, ctrl ); 1381 1382 // Hammer in the new limit 1383 Node *ctrl2 = loop_end->in(0); 1384 Node *cmp2 = new (C) CmpINode( loop_head->incr(), new_limit ); 1385 register_new_node( cmp2, ctrl2 ); 1386 Node *bol2 = new (C) BoolNode( cmp2, loop_end->test_trip() ); 1387 register_new_node( bol2, ctrl2 ); 1388 _igvn.hash_delete(loop_end); 1389 loop_end->set_req(CountedLoopEndNode::TestValue, bol2); 1390 1391 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1392 // Make it a 1-trip test (means at least 2 trips). 1393 if( adjust_min_trip ) { 1394 assert( new_limit != NULL, "" ); 1395 // Guard test uses an 'opaque' node which is not shared. Hence I 1396 // can edit it's inputs directly. Hammer in the new limit for the 1397 // minimum-trip guard. 1398 assert( opaq->outcnt() == 1, "" ); 1399 _igvn.hash_delete(opaq); 1400 opaq->set_req(1, new_limit); 1401 } 1402 } // LoopLimitCheck 1403 1404 // --------- 1405 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1406 // represents the odd iterations; since the loop trips an even number of 1407 // times its backedge is never taken. Kill the backedge. 1408 uint dd = dom_depth(loop_head); 1409 clone_loop( loop, old_new, dd ); 1410 1411 // Make backedges of the clone equal to backedges of the original. 1412 // Make the fall-in from the original come from the fall-out of the clone. 1413 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1414 Node* phi = loop_head->fast_out(j); 1415 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1416 Node *newphi = old_new[phi->_idx]; 1417 _igvn.hash_delete( phi ); 1418 _igvn.hash_delete( newphi ); 1419 1420 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1421 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1422 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1423 } 1424 } 1425 Node *clone_head = old_new[loop_head->_idx]; 1426 _igvn.hash_delete( clone_head ); 1427 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1428 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1429 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1430 loop->_head = clone_head; // New loop header 1431 1432 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1433 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1434 1435 // Kill the clone's backedge 1436 Node *newcle = old_new[loop_end->_idx]; 1437 _igvn.hash_delete( newcle ); 1438 Node *one = _igvn.intcon(1); 1439 set_ctrl(one, C->root()); 1440 newcle->set_req(1, one); 1441 // Force clone into same loop body 1442 uint max = loop->_body.size(); 1443 for( uint k = 0; k < max; k++ ) { 1444 Node *old = loop->_body.at(k); 1445 Node *nnn = old_new[old->_idx]; 1446 loop->_body.push(nnn); 1447 if (!has_ctrl(old)) 1448 set_loop(nnn, loop); 1449 } 1450 1451 loop->record_for_igvn(); 1452 } 1453 1454 //------------------------------do_maximally_unroll---------------------------- 1455 1456 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1457 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1458 assert(cl->has_exact_trip_count(), "trip count is not exact"); 1459 assert(cl->trip_count() > 0, ""); 1460 #ifndef PRODUCT 1461 if (TraceLoopOpts) { 1462 tty->print("MaxUnroll %d ", cl->trip_count()); 1463 loop->dump_head(); 1464 } 1465 #endif 1466 1467 // If loop is tripping an odd number of times, peel odd iteration 1468 if ((cl->trip_count() & 1) == 1) { 1469 do_peeling(loop, old_new); 1470 } 1471 1472 // Now its tripping an even number of times remaining. Double loop body. 1473 // Do not adjust pre-guards; they are not needed and do not exist. 1474 if (cl->trip_count() > 0) { 1475 assert((cl->trip_count() & 1) == 0, "missed peeling"); 1476 do_unroll(loop, old_new, false); 1477 } 1478 } 1479 1480 //------------------------------dominates_backedge--------------------------------- 1481 // Returns true if ctrl is executed on every complete iteration 1482 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1483 assert(ctrl->is_CFG(), "must be control"); 1484 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1485 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1486 } 1487 1488 //------------------------------adjust_limit----------------------------------- 1489 // Helper function for add_constraint(). 1490 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) { 1491 // Compute "I :: (limit-offset)/scale" 1492 Node *con = new (C) SubINode(rc_limit, offset); 1493 register_new_node(con, pre_ctrl); 1494 Node *X = new (C) DivINode(0, con, scale); 1495 register_new_node(X, pre_ctrl); 1496 1497 // Adjust loop limit 1498 loop_limit = (stride_con > 0) 1499 ? (Node*)(new (C) MinINode(loop_limit, X)) 1500 : (Node*)(new (C) MaxINode(loop_limit, X)); 1501 register_new_node(loop_limit, pre_ctrl); 1502 return loop_limit; 1503 } 1504 1505 //------------------------------add_constraint--------------------------------- 1506 // Constrain the main loop iterations so the conditions: 1507 // low_limit <= scale_con * I + offset < upper_limit 1508 // always holds true. That is, either increase the number of iterations in 1509 // the pre-loop or the post-loop until the condition holds true in the main 1510 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1511 // stride and scale are constants (offset and limit often are). 1512 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { 1513 // For positive stride, the pre-loop limit always uses a MAX function 1514 // and the main loop a MIN function. For negative stride these are 1515 // reversed. 1516 1517 // Also for positive stride*scale the affine function is increasing, so the 1518 // pre-loop must check for underflow and the post-loop for overflow. 1519 // Negative stride*scale reverses this; pre-loop checks for overflow and 1520 // post-loop for underflow. 1521 1522 Node *scale = _igvn.intcon(scale_con); 1523 set_ctrl(scale, C->root()); 1524 1525 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 1526 // The overflow limit: scale*I+offset < upper_limit 1527 // For main-loop compute 1528 // ( if (scale > 0) /* and stride > 0 */ 1529 // I < (upper_limit-offset)/scale 1530 // else /* scale < 0 and stride < 0 */ 1531 // I > (upper_limit-offset)/scale 1532 // ) 1533 // 1534 // (upper_limit-offset) may overflow or underflow. 1535 // But it is fine since main loop will either have 1536 // less iterations or will be skipped in such case. 1537 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl); 1538 1539 // The underflow limit: low_limit <= scale*I+offset. 1540 // For pre-loop compute 1541 // NOT(scale*I+offset >= low_limit) 1542 // scale*I+offset < low_limit 1543 // ( if (scale > 0) /* and stride > 0 */ 1544 // I < (low_limit-offset)/scale 1545 // else /* scale < 0 and stride < 0 */ 1546 // I > (low_limit-offset)/scale 1547 // ) 1548 1549 if (low_limit->get_int() == -max_jint) { 1550 if (!RangeLimitCheck) return; 1551 // We need this guard when scale*pre_limit+offset >= limit 1552 // due to underflow. So we need execute pre-loop until 1553 // scale*I+offset >= min_int. But (min_int-offset) will 1554 // underflow when offset > 0 and X will be > original_limit 1555 // when stride > 0. To avoid it we replace positive offset with 0. 1556 // 1557 // Also (min_int+1 == -max_int) is used instead of min_int here 1558 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1559 Node* shift = _igvn.intcon(31); 1560 set_ctrl(shift, C->root()); 1561 Node* sign = new (C) RShiftINode(offset, shift); 1562 register_new_node(sign, pre_ctrl); 1563 offset = new (C) AndINode(offset, sign); 1564 register_new_node(offset, pre_ctrl); 1565 } else { 1566 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1567 // The only problem we have here when offset == min_int 1568 // since (0-min_int) == min_int. It may be fine for stride > 0 1569 // but for stride < 0 X will be < original_limit. To avoid it 1570 // max(pre_limit, original_limit) is used in do_range_check(). 1571 } 1572 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1573 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl); 1574 1575 } else { // stride_con*scale_con < 0 1576 // For negative stride*scale pre-loop checks for overflow and 1577 // post-loop for underflow. 1578 // 1579 // The overflow limit: scale*I+offset < upper_limit 1580 // For pre-loop compute 1581 // NOT(scale*I+offset < upper_limit) 1582 // scale*I+offset >= upper_limit 1583 // scale*I+offset+1 > upper_limit 1584 // ( if (scale < 0) /* and stride > 0 */ 1585 // I < (upper_limit-(offset+1))/scale 1586 // else /* scale > 0 and stride < 0 */ 1587 // I > (upper_limit-(offset+1))/scale 1588 // ) 1589 // 1590 // (upper_limit-offset-1) may underflow or overflow. 1591 // To avoid it min(pre_limit, original_limit) is used 1592 // in do_range_check() for stride > 0 and max() for < 0. 1593 Node *one = _igvn.intcon(1); 1594 set_ctrl(one, C->root()); 1595 1596 Node *plus_one = new (C) AddINode(offset, one); 1597 register_new_node( plus_one, pre_ctrl ); 1598 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 1599 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl); 1600 1601 if (low_limit->get_int() == -max_jint) { 1602 if (!RangeLimitCheck) return; 1603 // We need this guard when scale*main_limit+offset >= limit 1604 // due to underflow. So we need execute main-loop while 1605 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 1606 // underflow when (offset+1) > 0 and X will be < main_limit 1607 // when scale < 0 (and stride > 0). To avoid it we replace 1608 // positive (offset+1) with 0. 1609 // 1610 // Also (min_int+1 == -max_int) is used instead of min_int here 1611 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 1612 Node* shift = _igvn.intcon(31); 1613 set_ctrl(shift, C->root()); 1614 Node* sign = new (C) RShiftINode(plus_one, shift); 1615 register_new_node(sign, pre_ctrl); 1616 plus_one = new (C) AndINode(plus_one, sign); 1617 register_new_node(plus_one, pre_ctrl); 1618 } else { 1619 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 1620 // The only problem we have here when offset == max_int 1621 // since (max_int+1) == min_int and (0-min_int) == min_int. 1622 // But it is fine since main loop will either have 1623 // less iterations or will be skipped in such case. 1624 } 1625 // The underflow limit: low_limit <= scale*I+offset. 1626 // For main-loop compute 1627 // scale*I+offset+1 > low_limit 1628 // ( if (scale < 0) /* and stride > 0 */ 1629 // I < (low_limit-(offset+1))/scale 1630 // else /* scale > 0 and stride < 0 */ 1631 // I > (low_limit-(offset+1))/scale 1632 // ) 1633 1634 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl); 1635 } 1636 } 1637 1638 1639 //------------------------------is_scaled_iv--------------------------------- 1640 // Return true if exp is a constant times an induction var 1641 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1642 if (exp == iv) { 1643 if (p_scale != NULL) { 1644 *p_scale = 1; 1645 } 1646 return true; 1647 } 1648 int opc = exp->Opcode(); 1649 if (opc == Op_MulI) { 1650 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1651 if (p_scale != NULL) { 1652 *p_scale = exp->in(2)->get_int(); 1653 } 1654 return true; 1655 } 1656 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1657 if (p_scale != NULL) { 1658 *p_scale = exp->in(1)->get_int(); 1659 } 1660 return true; 1661 } 1662 } else if (opc == Op_LShiftI) { 1663 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1664 if (p_scale != NULL) { 1665 *p_scale = 1 << exp->in(2)->get_int(); 1666 } 1667 return true; 1668 } 1669 } 1670 return false; 1671 } 1672 1673 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1674 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1675 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1676 if (is_scaled_iv(exp, iv, p_scale)) { 1677 if (p_offset != NULL) { 1678 Node *zero = _igvn.intcon(0); 1679 set_ctrl(zero, C->root()); 1680 *p_offset = zero; 1681 } 1682 return true; 1683 } 1684 int opc = exp->Opcode(); 1685 if (opc == Op_AddI) { 1686 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1687 if (p_offset != NULL) { 1688 *p_offset = exp->in(2); 1689 } 1690 return true; 1691 } 1692 if (exp->in(2)->is_Con()) { 1693 Node* offset2 = NULL; 1694 if (depth < 2 && 1695 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1696 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1697 if (p_offset != NULL) { 1698 Node *ctrl_off2 = get_ctrl(offset2); 1699 Node* offset = new (C) AddINode(offset2, exp->in(2)); 1700 register_new_node(offset, ctrl_off2); 1701 *p_offset = offset; 1702 } 1703 return true; 1704 } 1705 } 1706 } else if (opc == Op_SubI) { 1707 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1708 if (p_offset != NULL) { 1709 Node *zero = _igvn.intcon(0); 1710 set_ctrl(zero, C->root()); 1711 Node *ctrl_off = get_ctrl(exp->in(2)); 1712 Node* offset = new (C) SubINode(zero, exp->in(2)); 1713 register_new_node(offset, ctrl_off); 1714 *p_offset = offset; 1715 } 1716 return true; 1717 } 1718 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1719 if (p_offset != NULL) { 1720 *p_scale *= -1; 1721 *p_offset = exp->in(1); 1722 } 1723 return true; 1724 } 1725 } 1726 return false; 1727 } 1728 1729 //------------------------------do_range_check--------------------------------- 1730 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1731 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1732 #ifndef PRODUCT 1733 if (PrintOpto && VerifyLoopOptimizations) { 1734 tty->print("Range Check Elimination "); 1735 loop->dump_head(); 1736 } else if (TraceLoopOpts) { 1737 tty->print("RangeCheck "); 1738 loop->dump_head(); 1739 } 1740 #endif 1741 assert(RangeCheckElimination, ""); 1742 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1743 assert(cl->is_main_loop(), ""); 1744 1745 // protect against stride not being a constant 1746 if (!cl->stride_is_con()) 1747 return; 1748 1749 // Find the trip counter; we are iteration splitting based on it 1750 Node *trip_counter = cl->phi(); 1751 // Find the main loop limit; we will trim it's iterations 1752 // to not ever trip end tests 1753 Node *main_limit = cl->limit(); 1754 1755 // Need to find the main-loop zero-trip guard 1756 Node *ctrl = cl->in(LoopNode::EntryControl); 1757 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 1758 Node *iffm = ctrl->in(0); 1759 assert(iffm->Opcode() == Op_If, ""); 1760 Node *bolzm = iffm->in(1); 1761 assert(bolzm->Opcode() == Op_Bool, ""); 1762 Node *cmpzm = bolzm->in(1); 1763 assert(cmpzm->is_Cmp(), ""); 1764 Node *opqzm = cmpzm->in(2); 1765 // Can not optimize a loop if zero-trip Opaque1 node is optimized 1766 // away and then another round of loop opts attempted. 1767 if (opqzm->Opcode() != Op_Opaque1) 1768 return; 1769 assert(opqzm->in(1) == main_limit, "do not understand situation"); 1770 1771 // Find the pre-loop limit; we will expand it's iterations to 1772 // not ever trip low tests. 1773 Node *p_f = iffm->in(0); 1774 assert(p_f->Opcode() == Op_IfFalse, ""); 1775 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 1776 assert(pre_end->loopnode()->is_pre_loop(), ""); 1777 Node *pre_opaq1 = pre_end->limit(); 1778 // Occasionally it's possible for a pre-loop Opaque1 node to be 1779 // optimized away and then another round of loop opts attempted. 1780 // We can not optimize this particular loop in that case. 1781 if (pre_opaq1->Opcode() != Op_Opaque1) 1782 return; 1783 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 1784 Node *pre_limit = pre_opaq->in(1); 1785 1786 // Where do we put new limit calculations 1787 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 1788 1789 // Ensure the original loop limit is available from the 1790 // pre-loop Opaque1 node. 1791 Node *orig_limit = pre_opaq->original_loop_limit(); 1792 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 1793 return; 1794 1795 // Must know if its a count-up or count-down loop 1796 1797 int stride_con = cl->stride_con(); 1798 Node *zero = _igvn.intcon(0); 1799 Node *one = _igvn.intcon(1); 1800 // Use symmetrical int range [-max_jint,max_jint] 1801 Node *mini = _igvn.intcon(-max_jint); 1802 set_ctrl(zero, C->root()); 1803 set_ctrl(one, C->root()); 1804 set_ctrl(mini, C->root()); 1805 1806 // Range checks that do not dominate the loop backedge (ie. 1807 // conditionally executed) can lengthen the pre loop limit beyond 1808 // the original loop limit. To prevent this, the pre limit is 1809 // (for stride > 0) MINed with the original loop limit (MAXed 1810 // stride < 0) when some range_check (rc) is conditionally 1811 // executed. 1812 bool conditional_rc = false; 1813 1814 // Check loop body for tests of trip-counter plus loop-invariant vs 1815 // loop-invariant. 1816 for( uint i = 0; i < loop->_body.size(); i++ ) { 1817 Node *iff = loop->_body[i]; 1818 if( iff->Opcode() == Op_If ) { // Test? 1819 1820 // Test is an IfNode, has 2 projections. If BOTH are in the loop 1821 // we need loop unswitching instead of iteration splitting. 1822 Node *exit = loop->is_loop_exit(iff); 1823 if( !exit ) continue; 1824 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 1825 1826 // Get boolean condition to test 1827 Node *i1 = iff->in(1); 1828 if( !i1->is_Bool() ) continue; 1829 BoolNode *bol = i1->as_Bool(); 1830 BoolTest b_test = bol->_test; 1831 // Flip sense of test if exit condition is flipped 1832 if( flip ) 1833 b_test = b_test.negate(); 1834 1835 // Get compare 1836 Node *cmp = bol->in(1); 1837 1838 // Look for trip_counter + offset vs limit 1839 Node *rc_exp = cmp->in(1); 1840 Node *limit = cmp->in(2); 1841 jint scale_con= 1; // Assume trip counter not scaled 1842 1843 Node *limit_c = get_ctrl(limit); 1844 if( loop->is_member(get_loop(limit_c) ) ) { 1845 // Compare might have operands swapped; commute them 1846 b_test = b_test.commute(); 1847 rc_exp = cmp->in(2); 1848 limit = cmp->in(1); 1849 limit_c = get_ctrl(limit); 1850 if( loop->is_member(get_loop(limit_c) ) ) 1851 continue; // Both inputs are loop varying; cannot RCE 1852 } 1853 // Here we know 'limit' is loop invariant 1854 1855 // 'limit' maybe pinned below the zero trip test (probably from a 1856 // previous round of rce), in which case, it can't be used in the 1857 // zero trip test expression which must occur before the zero test's if. 1858 if( limit_c == ctrl ) { 1859 continue; // Don't rce this check but continue looking for other candidates. 1860 } 1861 1862 // Check for scaled induction variable plus an offset 1863 Node *offset = NULL; 1864 1865 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 1866 continue; 1867 } 1868 1869 Node *offset_c = get_ctrl(offset); 1870 if( loop->is_member( get_loop(offset_c) ) ) 1871 continue; // Offset is not really loop invariant 1872 // Here we know 'offset' is loop invariant. 1873 1874 // As above for the 'limit', the 'offset' maybe pinned below the 1875 // zero trip test. 1876 if( offset_c == ctrl ) { 1877 continue; // Don't rce this check but continue looking for other candidates. 1878 } 1879 #ifdef ASSERT 1880 if (TraceRangeLimitCheck) { 1881 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 1882 bol->dump(2); 1883 } 1884 #endif 1885 // At this point we have the expression as: 1886 // scale_con * trip_counter + offset :: limit 1887 // where scale_con, offset and limit are loop invariant. Trip_counter 1888 // monotonically increases by stride_con, a constant. Both (or either) 1889 // stride_con and scale_con can be negative which will flip about the 1890 // sense of the test. 1891 1892 // Adjust pre and main loop limits to guard the correct iteration set 1893 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 1894 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 1895 // The underflow and overflow limits: 0 <= scale*I+offset < limit 1896 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); 1897 if (!conditional_rc) { 1898 // (0-offset)/scale could be outside of loop iterations range. 1899 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 1900 } 1901 } else { 1902 #ifndef PRODUCT 1903 if( PrintOpto ) 1904 tty->print_cr("missed RCE opportunity"); 1905 #endif 1906 continue; // In release mode, ignore it 1907 } 1908 } else { // Otherwise work on normal compares 1909 switch( b_test._test ) { 1910 case BoolTest::gt: 1911 // Fall into GE case 1912 case BoolTest::ge: 1913 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 1914 scale_con = -scale_con; 1915 offset = new (C) SubINode( zero, offset ); 1916 register_new_node( offset, pre_ctrl ); 1917 limit = new (C) SubINode( zero, limit ); 1918 register_new_node( limit, pre_ctrl ); 1919 // Fall into LE case 1920 case BoolTest::le: 1921 if (b_test._test != BoolTest::gt) { 1922 // Convert X <= Y to X < Y+1 1923 limit = new (C) AddINode( limit, one ); 1924 register_new_node( limit, pre_ctrl ); 1925 } 1926 // Fall into LT case 1927 case BoolTest::lt: 1928 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 1929 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 1930 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 1931 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); 1932 if (!conditional_rc) { 1933 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 1934 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 1935 // still be outside of loop range. 1936 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 1937 } 1938 break; 1939 default: 1940 #ifndef PRODUCT 1941 if( PrintOpto ) 1942 tty->print_cr("missed RCE opportunity"); 1943 #endif 1944 continue; // Unhandled case 1945 } 1946 } 1947 1948 // Kill the eliminated test 1949 C->set_major_progress(); 1950 Node *kill_con = _igvn.intcon( 1-flip ); 1951 set_ctrl(kill_con, C->root()); 1952 _igvn.replace_input_of(iff, 1, kill_con); 1953 // Find surviving projection 1954 assert(iff->is_If(), ""); 1955 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 1956 // Find loads off the surviving projection; remove their control edge 1957 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 1958 Node* cd = dp->fast_out(i); // Control-dependent node 1959 if( cd->is_Load() ) { // Loads can now float around in the loop 1960 // Allow the load to float around in the loop, or before it 1961 // but NOT before the pre-loop. 1962 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 1963 --i; 1964 --imax; 1965 } 1966 } 1967 1968 } // End of is IF 1969 1970 } 1971 1972 // Update loop limits 1973 if (conditional_rc) { 1974 pre_limit = (stride_con > 0) ? (Node*)new (C) MinINode(pre_limit, orig_limit) 1975 : (Node*)new (C) MaxINode(pre_limit, orig_limit); 1976 register_new_node(pre_limit, pre_ctrl); 1977 } 1978 _igvn.hash_delete(pre_opaq); 1979 pre_opaq->set_req(1, pre_limit); 1980 1981 // Note:: we are making the main loop limit no longer precise; 1982 // need to round up based on stride. 1983 cl->set_nonexact_trip_count(); 1984 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case 1985 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 1986 // Hopefully, compiler will optimize for powers of 2. 1987 Node *ctrl = get_ctrl(main_limit); 1988 Node *stride = cl->stride(); 1989 Node *init = cl->init_trip(); 1990 Node *span = new (C) SubINode(main_limit,init); 1991 register_new_node(span,ctrl); 1992 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 1993 Node *add = new (C) AddINode(span,rndup); 1994 register_new_node(add,ctrl); 1995 Node *div = new (C) DivINode(0,add,stride); 1996 register_new_node(div,ctrl); 1997 Node *mul = new (C) MulINode(div,stride); 1998 register_new_node(mul,ctrl); 1999 Node *newlim = new (C) AddINode(mul,init); 2000 register_new_node(newlim,ctrl); 2001 main_limit = newlim; 2002 } 2003 2004 Node *main_cle = cl->loopexit(); 2005 Node *main_bol = main_cle->in(1); 2006 // Hacking loop bounds; need private copies of exit test 2007 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 2008 _igvn.hash_delete(main_cle); 2009 main_bol = main_bol->clone();// Clone a private BoolNode 2010 register_new_node( main_bol, main_cle->in(0) ); 2011 main_cle->set_req(1,main_bol); 2012 } 2013 Node *main_cmp = main_bol->in(1); 2014 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 2015 _igvn.hash_delete(main_bol); 2016 main_cmp = main_cmp->clone();// Clone a private CmpNode 2017 register_new_node( main_cmp, main_cle->in(0) ); 2018 main_bol->set_req(1,main_cmp); 2019 } 2020 // Hack the now-private loop bounds 2021 _igvn.replace_input_of(main_cmp, 2, main_limit); 2022 // The OpaqueNode is unshared by design 2023 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 2024 _igvn.replace_input_of(opqzm, 1, main_limit); 2025 } 2026 2027 //------------------------------DCE_loop_body---------------------------------- 2028 // Remove simplistic dead code from loop body 2029 void IdealLoopTree::DCE_loop_body() { 2030 for( uint i = 0; i < _body.size(); i++ ) 2031 if( _body.at(i)->outcnt() == 0 ) 2032 _body.map( i--, _body.pop() ); 2033 } 2034 2035 2036 //------------------------------adjust_loop_exit_prob-------------------------- 2037 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2038 // Replace with a 1-in-10 exit guess. 2039 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 2040 Node *test = tail(); 2041 while( test != _head ) { 2042 uint top = test->Opcode(); 2043 if( top == Op_IfTrue || top == Op_IfFalse ) { 2044 int test_con = ((ProjNode*)test)->_con; 2045 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2046 IfNode *iff = test->in(0)->as_If(); 2047 if( iff->outcnt() == 2 ) { // Ignore dead tests 2048 Node *bol = iff->in(1); 2049 if( bol && bol->req() > 1 && bol->in(1) && 2050 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 2051 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 2052 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 2053 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 2054 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 2055 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 2056 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 2057 return; // Allocation loops RARELY take backedge 2058 // Find the OTHER exit path from the IF 2059 Node* ex = iff->proj_out(1-test_con); 2060 float p = iff->_prob; 2061 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 2062 if( top == Op_IfTrue ) { 2063 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2064 iff->_prob = PROB_STATIC_FREQUENT; 2065 } 2066 } else { 2067 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2068 iff->_prob = PROB_STATIC_INFREQUENT; 2069 } 2070 } 2071 } 2072 } 2073 } 2074 test = phase->idom(test); 2075 } 2076 } 2077 2078 2079 //------------------------------policy_do_remove_empty_loop-------------------- 2080 // Micro-benchmark spamming. Policy is to always remove empty loops. 2081 // The 'DO' part is to replace the trip counter with the value it will 2082 // have on the last iteration. This will break the loop. 2083 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 2084 // Minimum size must be empty loop 2085 if (_body.size() > EMPTY_LOOP_SIZE) 2086 return false; 2087 2088 if (!_head->is_CountedLoop()) 2089 return false; // Dead loop 2090 CountedLoopNode *cl = _head->as_CountedLoop(); 2091 if (!cl->is_valid_counted_loop()) 2092 return false; // Malformed loop 2093 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 2094 return false; // Infinite loop 2095 2096 #ifdef ASSERT 2097 // Ensure only one phi which is the iv. 2098 Node* iv = NULL; 2099 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 2100 Node* n = cl->fast_out(i); 2101 if (n->Opcode() == Op_Phi) { 2102 assert(iv == NULL, "Too many phis" ); 2103 iv = n; 2104 } 2105 } 2106 assert(iv == cl->phi(), "Wrong phi" ); 2107 #endif 2108 2109 // main and post loops have explicitly created zero trip guard 2110 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 2111 if (needs_guard) { 2112 // Skip guard if values not overlap. 2113 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 2114 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 2115 int stride_con = cl->stride_con(); 2116 if (stride_con > 0) { 2117 needs_guard = (init_t->_hi >= limit_t->_lo); 2118 } else { 2119 needs_guard = (init_t->_lo <= limit_t->_hi); 2120 } 2121 } 2122 if (needs_guard) { 2123 // Check for an obvious zero trip guard. 2124 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); 2125 if (inctrl->Opcode() == Op_IfTrue) { 2126 // The test should look like just the backedge of a CountedLoop 2127 Node* iff = inctrl->in(0); 2128 if (iff->is_If()) { 2129 Node* bol = iff->in(1); 2130 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { 2131 Node* cmp = bol->in(1); 2132 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { 2133 needs_guard = false; 2134 } 2135 } 2136 } 2137 } 2138 } 2139 2140 #ifndef PRODUCT 2141 if (PrintOpto) { 2142 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 2143 this->dump_head(); 2144 } else if (TraceLoopOpts) { 2145 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 2146 this->dump_head(); 2147 } 2148 #endif 2149 2150 if (needs_guard) { 2151 // Peel the loop to ensure there's a zero trip guard 2152 Node_List old_new; 2153 phase->do_peeling(this, old_new); 2154 } 2155 2156 // Replace the phi at loop head with the final value of the last 2157 // iteration. Then the CountedLoopEnd will collapse (backedge never 2158 // taken) and all loop-invariant uses of the exit values will be correct. 2159 Node *phi = cl->phi(); 2160 Node *exact_limit = phase->exact_limit(this); 2161 if (exact_limit != cl->limit()) { 2162 // We also need to replace the original limit to collapse loop exit. 2163 Node* cmp = cl->loopexit()->cmp_node(); 2164 assert(cl->limit() == cmp->in(2), "sanity"); 2165 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 2166 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 2167 } 2168 // Note: the final value after increment should not overflow since 2169 // counted loop has limit check predicate. 2170 Node *final = new (phase->C) SubINode( exact_limit, cl->stride() ); 2171 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 2172 phase->_igvn.replace_node(phi,final); 2173 phase->C->set_major_progress(); 2174 return true; 2175 } 2176 2177 //------------------------------policy_do_one_iteration_loop------------------- 2178 // Convert one iteration loop into normal code. 2179 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { 2180 if (!_head->as_Loop()->is_valid_counted_loop()) 2181 return false; // Only for counted loop 2182 2183 CountedLoopNode *cl = _head->as_CountedLoop(); 2184 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 2185 return false; 2186 } 2187 2188 #ifndef PRODUCT 2189 if(TraceLoopOpts) { 2190 tty->print("OneIteration "); 2191 this->dump_head(); 2192 } 2193 #endif 2194 2195 Node *init_n = cl->init_trip(); 2196 #ifdef ASSERT 2197 // Loop boundaries should be constant since trip count is exact. 2198 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 2199 #endif 2200 // Replace the phi at loop head with the value of the init_trip. 2201 // Then the CountedLoopEnd will collapse (backedge will not be taken) 2202 // and all loop-invariant uses of the exit values will be correct. 2203 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 2204 phase->C->set_major_progress(); 2205 return true; 2206 } 2207 2208 //============================================================================= 2209 //------------------------------iteration_split_impl--------------------------- 2210 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 2211 // Compute exact loop trip count if possible. 2212 compute_exact_trip_count(phase); 2213 2214 // Convert one iteration loop into normal code. 2215 if (policy_do_one_iteration_loop(phase)) 2216 return true; 2217 2218 // Check and remove empty loops (spam micro-benchmarks) 2219 if (policy_do_remove_empty_loop(phase)) 2220 return true; // Here we removed an empty loop 2221 2222 bool should_peel = policy_peeling(phase); // Should we peel? 2223 2224 bool should_unswitch = policy_unswitching(phase); 2225 2226 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 2227 // This removes loop-invariant tests (usually null checks). 2228 if (!_head->is_CountedLoop()) { // Non-counted loop 2229 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 2230 // Partial peel succeeded so terminate this round of loop opts 2231 return false; 2232 } 2233 if (should_peel) { // Should we peel? 2234 #ifndef PRODUCT 2235 if (PrintOpto) tty->print_cr("should_peel"); 2236 #endif 2237 phase->do_peeling(this,old_new); 2238 } else if (should_unswitch) { 2239 phase->do_unswitching(this, old_new); 2240 } 2241 return true; 2242 } 2243 CountedLoopNode *cl = _head->as_CountedLoop(); 2244 2245 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 2246 2247 // Do nothing special to pre- and post- loops 2248 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 2249 2250 // Compute loop trip count from profile data 2251 compute_profile_trip_cnt(phase); 2252 2253 // Before attempting fancy unrolling, RCE or alignment, see if we want 2254 // to completely unroll this loop or do loop unswitching. 2255 if (cl->is_normal_loop()) { 2256 if (should_unswitch) { 2257 phase->do_unswitching(this, old_new); 2258 return true; 2259 } 2260 bool should_maximally_unroll = policy_maximally_unroll(phase); 2261 if (should_maximally_unroll) { 2262 // Here we did some unrolling and peeling. Eventually we will 2263 // completely unroll this loop and it will no longer be a loop. 2264 phase->do_maximally_unroll(this,old_new); 2265 return true; 2266 } 2267 } 2268 2269 // Skip next optimizations if running low on nodes. Note that 2270 // policy_unswitching and policy_maximally_unroll have this check. 2271 uint nodes_left = MaxNodeLimit - phase->C->unique(); 2272 if ((2 * _body.size()) > nodes_left) { 2273 return true; 2274 } 2275 2276 // Counted loops may be peeled, may need some iterations run up 2277 // front for RCE, and may want to align loop refs to a cache 2278 // line. Thus we clone a full loop up front whose trip count is 2279 // at least 1 (if peeling), but may be several more. 2280 2281 // The main loop will start cache-line aligned with at least 1 2282 // iteration of the unrolled body (zero-trip test required) and 2283 // will have some range checks removed. 2284 2285 // A post-loop will finish any odd iterations (leftover after 2286 // unrolling), plus any needed for RCE purposes. 2287 2288 bool should_unroll = policy_unroll(phase); 2289 2290 bool should_rce = policy_range_check(phase); 2291 2292 bool should_align = policy_align(phase); 2293 2294 // If not RCE'ing (iteration splitting) or Aligning, then we do not 2295 // need a pre-loop. We may still need to peel an initial iteration but 2296 // we will not be needing an unknown number of pre-iterations. 2297 // 2298 // Basically, if may_rce_align reports FALSE first time through, 2299 // we will not be able to later do RCE or Aligning on this loop. 2300 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 2301 2302 // If we have any of these conditions (RCE, alignment, unrolling) met, then 2303 // we switch to the pre-/main-/post-loop model. This model also covers 2304 // peeling. 2305 if (should_rce || should_align || should_unroll) { 2306 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops 2307 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 2308 2309 // Adjust the pre- and main-loop limits to let the pre and post loops run 2310 // with full checks, but the main-loop with no checks. Remove said 2311 // checks from the main body. 2312 if (should_rce) 2313 phase->do_range_check(this,old_new); 2314 2315 // Double loop body for unrolling. Adjust the minimum-trip test (will do 2316 // twice as many iterations as before) and the main body limit (only do 2317 // an even number of trips). If we are peeling, we might enable some RCE 2318 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 2319 // peeling. 2320 if (should_unroll && !should_peel) 2321 phase->do_unroll(this,old_new, true); 2322 2323 // Adjust the pre-loop limits to align the main body 2324 // iterations. 2325 if (should_align) 2326 Unimplemented(); 2327 2328 } else { // Else we have an unchanged counted loop 2329 if (should_peel) // Might want to peel but do nothing else 2330 phase->do_peeling(this,old_new); 2331 } 2332 return true; 2333 } 2334 2335 2336 //============================================================================= 2337 //------------------------------iteration_split-------------------------------- 2338 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 2339 // Recursively iteration split nested loops 2340 if (_child && !_child->iteration_split(phase, old_new)) 2341 return false; 2342 2343 // Clean out prior deadwood 2344 DCE_loop_body(); 2345 2346 2347 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 2348 // Replace with a 1-in-10 exit guess. 2349 if (_parent /*not the root loop*/ && 2350 !_irreducible && 2351 // Also ignore the occasional dead backedge 2352 !tail()->is_top()) { 2353 adjust_loop_exit_prob(phase); 2354 } 2355 2356 // Gate unrolling, RCE and peeling efforts. 2357 if (!_child && // If not an inner loop, do not split 2358 !_irreducible && 2359 _allow_optimizations && 2360 !tail()->is_top()) { // Also ignore the occasional dead backedge 2361 if (!_has_call) { 2362 if (!iteration_split_impl(phase, old_new)) { 2363 return false; 2364 } 2365 } else if (policy_unswitching(phase)) { 2366 phase->do_unswitching(this, old_new); 2367 } 2368 } 2369 2370 // Minor offset re-organization to remove loop-fallout uses of 2371 // trip counter when there was no major reshaping. 2372 phase->reorg_offsets(this); 2373 2374 if (_next && !_next->iteration_split(phase, old_new)) 2375 return false; 2376 return true; 2377 } 2378 2379 2380 //============================================================================= 2381 // Process all the loops in the loop tree and replace any fill 2382 // patterns with an intrisc version. 2383 bool PhaseIdealLoop::do_intrinsify_fill() { 2384 bool changed = false; 2385 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2386 IdealLoopTree* lpt = iter.current(); 2387 changed |= intrinsify_fill(lpt); 2388 } 2389 return changed; 2390 } 2391 2392 2393 // Examine an inner loop looking for a a single store of an invariant 2394 // value in a unit stride loop, 2395 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2396 Node*& shift, Node*& con) { 2397 const char* msg = NULL; 2398 Node* msg_node = NULL; 2399 2400 store_value = NULL; 2401 con = NULL; 2402 shift = NULL; 2403 2404 // Process the loop looking for stores. If there are multiple 2405 // stores or extra control flow give at this point. 2406 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2407 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2408 Node* n = lpt->_body.at(i); 2409 if (n->outcnt() == 0) continue; // Ignore dead 2410 if (n->is_Store()) { 2411 if (store != NULL) { 2412 msg = "multiple stores"; 2413 break; 2414 } 2415 int opc = n->Opcode(); 2416 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 2417 msg = "oop fills not handled"; 2418 break; 2419 } 2420 Node* value = n->in(MemNode::ValueIn); 2421 if (!lpt->is_invariant(value)) { 2422 msg = "variant store value"; 2423 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2424 msg = "not array address"; 2425 } 2426 store = n; 2427 store_value = value; 2428 } else if (n->is_If() && n != head->loopexit()) { 2429 msg = "extra control flow"; 2430 msg_node = n; 2431 } 2432 } 2433 2434 if (store == NULL) { 2435 // No store in loop 2436 return false; 2437 } 2438 2439 if (msg == NULL && head->stride_con() != 1) { 2440 // could handle negative strides too 2441 if (head->stride_con() < 0) { 2442 msg = "negative stride"; 2443 } else { 2444 msg = "non-unit stride"; 2445 } 2446 } 2447 2448 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2449 msg = "can't handle store address"; 2450 msg_node = store->in(MemNode::Address); 2451 } 2452 2453 if (msg == NULL && 2454 (!store->in(MemNode::Memory)->is_Phi() || 2455 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2456 msg = "store memory isn't proper phi"; 2457 msg_node = store->in(MemNode::Memory); 2458 } 2459 2460 // Make sure there is an appropriate fill routine 2461 BasicType t = store->as_Mem()->memory_type(); 2462 const char* fill_name; 2463 if (msg == NULL && 2464 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2465 msg = "unsupported store"; 2466 msg_node = store; 2467 } 2468 2469 if (msg != NULL) { 2470 #ifndef PRODUCT 2471 if (TraceOptimizeFill) { 2472 tty->print_cr("not fill intrinsic candidate: %s", msg); 2473 if (msg_node != NULL) msg_node->dump(); 2474 } 2475 #endif 2476 return false; 2477 } 2478 2479 // Make sure the address expression can be handled. It should be 2480 // head->phi * elsize + con. head->phi might have a ConvI2L. 2481 Node* elements[4]; 2482 Node* conv = NULL; 2483 bool found_index = false; 2484 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2485 for (int e = 0; e < count; e++) { 2486 Node* n = elements[e]; 2487 if (n->is_Con() && con == NULL) { 2488 con = n; 2489 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2490 Node* value = n->in(1); 2491 #ifdef _LP64 2492 if (value->Opcode() == Op_ConvI2L) { 2493 conv = value; 2494 value = value->in(1); 2495 } 2496 #endif 2497 if (value != head->phi()) { 2498 msg = "unhandled shift in address"; 2499 } else { 2500 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 2501 msg = "scale doesn't match"; 2502 } else { 2503 found_index = true; 2504 shift = n; 2505 } 2506 } 2507 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2508 if (n->in(1) == head->phi()) { 2509 found_index = true; 2510 conv = n; 2511 } else { 2512 msg = "unhandled input to ConvI2L"; 2513 } 2514 } else if (n == head->phi()) { 2515 // no shift, check below for allowed cases 2516 found_index = true; 2517 } else { 2518 msg = "unhandled node in address"; 2519 msg_node = n; 2520 } 2521 } 2522 2523 if (count == -1) { 2524 msg = "malformed address expression"; 2525 msg_node = store; 2526 } 2527 2528 if (!found_index) { 2529 msg = "missing use of index"; 2530 } 2531 2532 // byte sized items won't have a shift 2533 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2534 msg = "can't find shift"; 2535 msg_node = store; 2536 } 2537 2538 if (msg != NULL) { 2539 #ifndef PRODUCT 2540 if (TraceOptimizeFill) { 2541 tty->print_cr("not fill intrinsic: %s", msg); 2542 if (msg_node != NULL) msg_node->dump(); 2543 } 2544 #endif 2545 return false; 2546 } 2547 2548 // No make sure all the other nodes in the loop can be handled 2549 VectorSet ok(Thread::current()->resource_area()); 2550 2551 // store related values are ok 2552 ok.set(store->_idx); 2553 ok.set(store->in(MemNode::Memory)->_idx); 2554 2555 // Loop structure is ok 2556 ok.set(head->_idx); 2557 ok.set(head->loopexit()->_idx); 2558 ok.set(head->phi()->_idx); 2559 ok.set(head->incr()->_idx); 2560 ok.set(head->loopexit()->cmp_node()->_idx); 2561 ok.set(head->loopexit()->in(1)->_idx); 2562 2563 // Address elements are ok 2564 if (con) ok.set(con->_idx); 2565 if (shift) ok.set(shift->_idx); 2566 if (conv) ok.set(conv->_idx); 2567 2568 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2569 Node* n = lpt->_body.at(i); 2570 if (n->outcnt() == 0) continue; // Ignore dead 2571 if (ok.test(n->_idx)) continue; 2572 // Backedge projection is ok 2573 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue; 2574 if (!n->is_AddP()) { 2575 msg = "unhandled node"; 2576 msg_node = n; 2577 break; 2578 } 2579 } 2580 2581 // Make sure no unexpected values are used outside the loop 2582 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2583 Node* n = lpt->_body.at(i); 2584 // These values can be replaced with other nodes if they are used 2585 // outside the loop. 2586 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2587 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2588 Node* use = iter.get(); 2589 if (!lpt->_body.contains(use)) { 2590 msg = "node is used outside loop"; 2591 // lpt->_body.dump(); 2592 msg_node = n; 2593 break; 2594 } 2595 } 2596 } 2597 2598 #ifdef ASSERT 2599 if (TraceOptimizeFill) { 2600 if (msg != NULL) { 2601 tty->print_cr("no fill intrinsic: %s", msg); 2602 if (msg_node != NULL) msg_node->dump(); 2603 } else { 2604 tty->print_cr("fill intrinsic for:"); 2605 } 2606 store->dump(); 2607 if (Verbose) { 2608 lpt->_body.dump(); 2609 } 2610 } 2611 #endif 2612 2613 return msg == NULL; 2614 } 2615 2616 2617 2618 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2619 // Only for counted inner loops 2620 if (!lpt->is_counted() || !lpt->is_inner()) { 2621 return false; 2622 } 2623 2624 // Must have constant stride 2625 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2626 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 2627 return false; 2628 } 2629 2630 // Check that the body only contains a store of a loop invariant 2631 // value that is indexed by the loop phi. 2632 Node* store = NULL; 2633 Node* store_value = NULL; 2634 Node* shift = NULL; 2635 Node* offset = NULL; 2636 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2637 return false; 2638 } 2639 2640 #ifndef PRODUCT 2641 if (TraceLoopOpts) { 2642 tty->print("ArrayFill "); 2643 lpt->dump_head(); 2644 } 2645 #endif 2646 2647 // Now replace the whole loop body by a call to a fill routine that 2648 // covers the same region as the loop. 2649 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2650 2651 // Build an expression for the beginning of the copy region 2652 Node* index = head->init_trip(); 2653 #ifdef _LP64 2654 index = new (C) ConvI2LNode(index); 2655 _igvn.register_new_node_with_optimizer(index); 2656 #endif 2657 if (shift != NULL) { 2658 // byte arrays don't require a shift but others do. 2659 index = new (C) LShiftXNode(index, shift->in(2)); 2660 _igvn.register_new_node_with_optimizer(index); 2661 } 2662 index = new (C) AddPNode(base, base, index); 2663 _igvn.register_new_node_with_optimizer(index); 2664 Node* from = new (C) AddPNode(base, index, offset); 2665 _igvn.register_new_node_with_optimizer(from); 2666 // Compute the number of elements to copy 2667 Node* len = new (C) SubINode(head->limit(), head->init_trip()); 2668 _igvn.register_new_node_with_optimizer(len); 2669 2670 BasicType t = store->as_Mem()->memory_type(); 2671 bool aligned = false; 2672 if (offset != NULL && head->init_trip()->is_Con()) { 2673 int element_size = type2aelembytes(t); 2674 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2675 } 2676 2677 // Build a call to the fill routine 2678 const char* fill_name; 2679 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2680 assert(fill != NULL, "what?"); 2681 2682 // Convert float/double to int/long for fill routines 2683 if (t == T_FLOAT) { 2684 store_value = new (C) MoveF2INode(store_value); 2685 _igvn.register_new_node_with_optimizer(store_value); 2686 } else if (t == T_DOUBLE) { 2687 store_value = new (C) MoveD2LNode(store_value); 2688 _igvn.register_new_node_with_optimizer(store_value); 2689 } 2690 2691 Node* mem_phi = store->in(MemNode::Memory); 2692 Node* result_ctrl; 2693 Node* result_mem; 2694 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2695 CallLeafNode *call = new (C) CallLeafNoFPNode(call_type, fill, 2696 fill_name, TypeAryPtr::get_array_body_type(t)); 2697 call->init_req(TypeFunc::Parms+0, from); 2698 call->init_req(TypeFunc::Parms+1, store_value); 2699 #ifdef _LP64 2700 len = new (C) ConvI2LNode(len); 2701 _igvn.register_new_node_with_optimizer(len); 2702 #endif 2703 call->init_req(TypeFunc::Parms+2, len); 2704 #ifdef _LP64 2705 call->init_req(TypeFunc::Parms+3, C->top()); 2706 #endif 2707 call->init_req( TypeFunc::Control, head->init_control()); 2708 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o 2709 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) ); 2710 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) ); 2711 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) ); 2712 _igvn.register_new_node_with_optimizer(call); 2713 result_ctrl = new (C) ProjNode(call,TypeFunc::Control); 2714 _igvn.register_new_node_with_optimizer(result_ctrl); 2715 result_mem = new (C) ProjNode(call,TypeFunc::Memory); 2716 _igvn.register_new_node_with_optimizer(result_mem); 2717 2718 // If this fill is tightly coupled to an allocation and overwrites 2719 // the whole body, allow it to take over the zeroing. 2720 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 2721 if (alloc != NULL && alloc->is_AllocateArray()) { 2722 Node* length = alloc->as_AllocateArray()->Ideal_length(); 2723 if (head->limit() == length && 2724 head->init_trip() == _igvn.intcon(0)) { 2725 if (TraceOptimizeFill) { 2726 tty->print_cr("Eliminated zeroing in allocation"); 2727 } 2728 alloc->maybe_set_complete(&_igvn); 2729 } else { 2730 #ifdef ASSERT 2731 if (TraceOptimizeFill) { 2732 tty->print_cr("filling array but bounds don't match"); 2733 alloc->dump(); 2734 head->init_trip()->dump(); 2735 head->limit()->dump(); 2736 length->dump(); 2737 } 2738 #endif 2739 } 2740 } 2741 2742 // Redirect the old control and memory edges that are outside the loop. 2743 Node* exit = head->loopexit()->proj_out(0); 2744 // Sometimes the memory phi of the head is used as the outgoing 2745 // state of the loop. It's safe in this case to replace it with the 2746 // result_mem. 2747 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 2748 _igvn.replace_node(exit, result_ctrl); 2749 _igvn.replace_node(store, result_mem); 2750 // Any uses the increment outside of the loop become the loop limit. 2751 _igvn.replace_node(head->incr(), head->limit()); 2752 2753 // Disconnect the head from the loop. 2754 for (uint i = 0; i < lpt->_body.size(); i++) { 2755 Node* n = lpt->_body.at(i); 2756 _igvn.replace_node(n, C->top()); 2757 } 2758 2759 return true; 2760 }