1 /* 2 * Copyright (c) 2000, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/castnode.hpp" 31 #include "opto/connode.hpp" 32 #include "opto/convertnode.hpp" 33 #include "opto/divnode.hpp" 34 #include "opto/loopnode.hpp" 35 #include "opto/mulnode.hpp" 36 #include "opto/movenode.hpp" 37 #include "opto/opaquenode.hpp" 38 #include "opto/rootnode.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 #include "opto/superword.hpp" 42 #include "opto/vectornode.hpp" 43 44 //------------------------------is_loop_exit----------------------------------- 45 // Given an IfNode, return the loop-exiting projection or NULL if both 46 // arms remain in the loop. 47 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 48 if (iff->outcnt() != 2) return NULL; // Ignore partially dead tests 49 PhaseIdealLoop *phase = _phase; 50 // Test is an IfNode, has 2 projections. If BOTH are in the loop 51 // we need loop unswitching instead of peeling. 52 if (!is_member(phase->get_loop(iff->raw_out(0)))) 53 return iff->raw_out(0); 54 if (!is_member(phase->get_loop(iff->raw_out(1)))) 55 return iff->raw_out(1); 56 return NULL; 57 } 58 59 60 //============================================================================= 61 62 63 //------------------------------record_for_igvn---------------------------- 64 // Put loop body on igvn work list 65 void IdealLoopTree::record_for_igvn() { 66 for (uint i = 0; i < _body.size(); i++) { 67 Node *n = _body.at(i); 68 _phase->_igvn._worklist.push(n); 69 } 70 // put body of outer strip mined loop on igvn work list as well 71 if (_head->is_CountedLoop() && _head->as_Loop()->is_strip_mined()) { 72 CountedLoopNode* l = _head->as_CountedLoop(); 73 Node* outer_loop = l->outer_loop(); 74 assert(outer_loop != NULL, "missing piece of strip mined loop"); 75 _phase->_igvn._worklist.push(outer_loop); 76 Node* outer_loop_tail = l->outer_loop_tail(); 77 assert(outer_loop_tail != NULL, "missing piece of strip mined loop"); 78 _phase->_igvn._worklist.push(outer_loop_tail); 79 Node* outer_loop_end = l->outer_loop_end(); 80 assert(outer_loop_end != NULL, "missing piece of strip mined loop"); 81 _phase->_igvn._worklist.push(outer_loop_end); 82 Node* outer_safepoint = l->outer_safepoint(); 83 assert(outer_safepoint != NULL, "missing piece of strip mined loop"); 84 _phase->_igvn._worklist.push(outer_safepoint); 85 Node* cle_out = _head->as_CountedLoop()->loopexit()->proj_out(false); 86 assert(cle_out != NULL, "missing piece of strip mined loop"); 87 _phase->_igvn._worklist.push(cle_out); 88 } 89 } 90 91 //------------------------------compute_exact_trip_count----------------------- 92 // Compute loop trip count if possible. Do not recalculate trip count for 93 // split loops (pre-main-post) which have their limits and inits behind Opaque node. 94 void IdealLoopTree::compute_trip_count(PhaseIdealLoop* phase) { 95 if (!_head->as_Loop()->is_valid_counted_loop()) { 96 return; 97 } 98 CountedLoopNode* cl = _head->as_CountedLoop(); 99 // Trip count may become nonexact for iteration split loops since 100 // RCE modifies limits. Note, _trip_count value is not reset since 101 // it is used to limit unrolling of main loop. 102 cl->set_nonexact_trip_count(); 103 104 // Loop's test should be part of loop. 105 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 106 return; // Infinite loop 107 108 #ifdef ASSERT 109 BoolTest::mask bt = cl->loopexit()->test_trip(); 110 assert(bt == BoolTest::lt || bt == BoolTest::gt || 111 bt == BoolTest::ne, "canonical test is expected"); 112 #endif 113 114 Node* init_n = cl->init_trip(); 115 Node* limit_n = cl->limit(); 116 if (init_n != NULL && limit_n != NULL) { 117 // Use longs to avoid integer overflow. 118 int stride_con = cl->stride_con(); 119 const TypeInt* init_type = phase->_igvn.type(init_n)->is_int(); 120 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 121 jlong init_con = (stride_con > 0) ? init_type->_lo : init_type->_hi; 122 jlong limit_con = (stride_con > 0) ? limit_type->_hi : limit_type->_lo; 123 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 124 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 125 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { 126 if (init_n->is_Con() && limit_n->is_Con()) { 127 // Set exact trip count. 128 cl->set_exact_trip_count((uint)trip_count); 129 } else if (cl->unrolled_count() == 1) { 130 // Set maximum trip count before unrolling. 131 cl->set_trip_count((uint)trip_count); 132 } 133 } 134 } 135 } 136 137 //------------------------------compute_profile_trip_cnt---------------------------- 138 // Compute loop trip count from profile data as 139 // (backedge_count + loop_exit_count) / loop_exit_count 140 141 float IdealLoopTree::compute_profile_trip_cnt_helper(Node* n) { 142 if (n->is_If()) { 143 IfNode *iff = n->as_If(); 144 if (iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN) { 145 Node *exit = is_loop_exit(iff); 146 if (exit) { 147 float exit_prob = iff->_prob; 148 if (exit->Opcode() == Op_IfFalse) { 149 exit_prob = 1.0 - exit_prob; 150 } 151 if (exit_prob > PROB_MIN) { 152 float exit_cnt = iff->_fcnt * exit_prob; 153 return exit_cnt; 154 } 155 } 156 } 157 } 158 if (n->is_Jump()) { 159 JumpNode *jmp = n->as_Jump(); 160 if (jmp->_fcnt != COUNT_UNKNOWN) { 161 float* probs = jmp->_probs; 162 float exit_prob = 0; 163 PhaseIdealLoop *phase = _phase; 164 for (DUIterator_Fast imax, i = jmp->fast_outs(imax); i < imax; i++) { 165 JumpProjNode* u = jmp->fast_out(i)->as_JumpProj(); 166 if (!is_member(_phase->get_loop(u))) { 167 exit_prob += probs[u->_con]; 168 } 169 } 170 return exit_prob * jmp->_fcnt; 171 } 172 } 173 return 0; 174 } 175 176 void IdealLoopTree::compute_profile_trip_cnt(PhaseIdealLoop *phase) { 177 if (!_head->is_Loop()) { 178 return; 179 } 180 LoopNode* head = _head->as_Loop(); 181 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 182 return; // Already computed 183 } 184 float trip_cnt = (float)max_jint; // default is big 185 186 Node* back = head->in(LoopNode::LoopBackControl); 187 while (back != head) { 188 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 189 back->in(0) && 190 back->in(0)->is_If() && 191 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 192 back->in(0)->as_If()->_prob != PROB_UNKNOWN && 193 (back->Opcode() == Op_IfTrue ? 1-back->in(0)->as_If()->_prob : back->in(0)->as_If()->_prob) > PROB_MIN) { 194 break; 195 } 196 back = phase->idom(back); 197 } 198 if (back != head) { 199 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 200 back->in(0), "if-projection exists"); 201 IfNode* back_if = back->in(0)->as_If(); 202 float loop_back_cnt = back_if->_fcnt * (back->Opcode() == Op_IfTrue ? back_if->_prob : (1 - back_if->_prob)); 203 204 // Now compute a loop exit count 205 float loop_exit_cnt = 0.0f; 206 if (_child == NULL) { 207 for (uint i = 0; i < _body.size(); i++) { 208 Node *n = _body[i]; 209 loop_exit_cnt += compute_profile_trip_cnt_helper(n); 210 } 211 } else { 212 ResourceMark rm; 213 Unique_Node_List wq; 214 wq.push(back); 215 for (uint i = 0; i < wq.size(); i++) { 216 Node *n = wq.at(i); 217 assert(n->is_CFG(), "only control nodes"); 218 if (n != head) { 219 if (n->is_Region()) { 220 for (uint j = 1; j < n->req(); j++) { 221 wq.push(n->in(j)); 222 } 223 } else { 224 loop_exit_cnt += compute_profile_trip_cnt_helper(n); 225 wq.push(n->in(0)); 226 } 227 } 228 } 229 230 } 231 if (loop_exit_cnt > 0.0f) { 232 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 233 } else { 234 // No exit count so use 235 trip_cnt = loop_back_cnt; 236 } 237 } else { 238 head->mark_profile_trip_failed(); 239 } 240 #ifndef PRODUCT 241 if (TraceProfileTripCount) { 242 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 243 } 244 #endif 245 head->set_profile_trip_cnt(trip_cnt); 246 } 247 248 //---------------------is_invariant_addition----------------------------- 249 // Return nonzero index of invariant operand for an Add or Sub 250 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 251 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 252 int op = n->Opcode(); 253 if (op == Op_AddI || op == Op_SubI) { 254 bool in1_invar = this->is_invariant(n->in(1)); 255 bool in2_invar = this->is_invariant(n->in(2)); 256 if (in1_invar && !in2_invar) return 1; 257 if (!in1_invar && in2_invar) return 2; 258 } 259 return 0; 260 } 261 262 //---------------------reassociate_add_sub----------------------------- 263 // Reassociate invariant add and subtract expressions: 264 // 265 // inv1 + (x + inv2) => ( inv1 + inv2) + x 266 // (x + inv2) + inv1 => ( inv1 + inv2) + x 267 // inv1 + (x - inv2) => ( inv1 - inv2) + x 268 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 269 // (x + inv2) - inv1 => (-inv1 + inv2) + x 270 // (x - inv2) + inv1 => ( inv1 - inv2) + x 271 // (x - inv2) - inv1 => (-inv1 - inv2) + x 272 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 273 // inv1 - (x - inv2) => ( inv1 + inv2) - x 274 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 275 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 276 // inv1 - (x + inv2) => ( inv1 - inv2) - x 277 // 278 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 279 if ((!n1->is_Add() && !n1->is_Sub()) || n1->outcnt() == 0) return NULL; 280 if (is_invariant(n1)) return NULL; 281 int inv1_idx = is_invariant_addition(n1, phase); 282 if (!inv1_idx) return NULL; 283 // Don't mess with add of constant (igvn moves them to expression tree root.) 284 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 285 Node* inv1 = n1->in(inv1_idx); 286 Node* n2 = n1->in(3 - inv1_idx); 287 int inv2_idx = is_invariant_addition(n2, phase); 288 if (!inv2_idx) return NULL; 289 Node* x = n2->in(3 - inv2_idx); 290 Node* inv2 = n2->in(inv2_idx); 291 292 bool neg_x = n2->is_Sub() && inv2_idx == 1; 293 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 294 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 295 if (n1->is_Sub() && inv1_idx == 1) { 296 neg_x = !neg_x; 297 neg_inv2 = !neg_inv2; 298 } 299 Node* inv1_c = phase->get_ctrl(inv1); 300 Node* inv2_c = phase->get_ctrl(inv2); 301 Node* n_inv1; 302 if (neg_inv1) { 303 Node *zero = phase->_igvn.intcon(0); 304 phase->set_ctrl(zero, phase->C->root()); 305 n_inv1 = new SubINode(zero, inv1); 306 phase->register_new_node(n_inv1, inv1_c); 307 } else { 308 n_inv1 = inv1; 309 } 310 Node* inv; 311 if (neg_inv2) { 312 inv = new SubINode(n_inv1, inv2); 313 } else { 314 inv = new AddINode(n_inv1, inv2); 315 } 316 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 317 318 Node* addx; 319 if (neg_x) { 320 addx = new SubINode(inv, x); 321 } else { 322 addx = new AddINode(x, inv); 323 } 324 phase->register_new_node(addx, phase->get_ctrl(x)); 325 phase->_igvn.replace_node(n1, addx); 326 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 327 _body.yank(n1); 328 return addx; 329 } 330 331 //---------------------reassociate_invariants----------------------------- 332 // Reassociate invariant expressions: 333 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 334 for (int i = _body.size() - 1; i >= 0; i--) { 335 Node *n = _body.at(i); 336 for (int j = 0; j < 5; j++) { 337 Node* nn = reassociate_add_sub(n, phase); 338 if (nn == NULL) break; 339 n = nn; // again 340 }; 341 } 342 } 343 344 //------------------------------policy_peeling--------------------------------- 345 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 346 // make some loop-invariant test (usually a null-check) happen before the loop. 347 bool IdealLoopTree::policy_peeling(PhaseIdealLoop *phase) const { 348 IdealLoopTree *loop = (IdealLoopTree*)this; 349 350 // If nodes are depleted, some transform has miscalculated its needs. 351 assert(!phase->exceeding_node_budget(), "sanity"); 352 353 uint body_size = loop->_body.size(); 354 // Peeling does loop cloning which can result in O(N^2) node construction 355 if (body_size > 255) { 356 return false; // Prevent overflow for large body size 357 } 358 uint estimate = body_size * body_size; 359 if (phase->exceeding_node_budget(estimate)) { 360 return false; // Too large to safely clone 361 } 362 363 // check for vectorized loops, any peeling done was already applied 364 if (_head->is_CountedLoop()) { 365 CountedLoopNode* cl = _head->as_CountedLoop(); 366 if (cl->is_unroll_only() || cl->trip_count() == 1) { 367 return false; 368 } 369 } 370 371 Node* test = loop->tail(); 372 373 while (test != _head) { // Scan till run off top of loop 374 if (test->is_If()) { // Test? 375 Node *ctrl = phase->get_ctrl(test->in(1)); 376 if (ctrl->is_top()) { 377 return false; // Found dead test on live IF? No peeling! 378 } 379 // Standard IF only has one input value to check for loop invariance 380 assert(test->Opcode() == Op_If || 381 test->Opcode() == Op_CountedLoopEnd || 382 test->Opcode() == Op_RangeCheck, 383 "Check this code when new subtype is added"); 384 // Condition is not a member of this loop? 385 if (!is_member(phase->get_loop(ctrl)) && is_loop_exit(test)) { 386 // Found reason to peel! 387 return phase->may_require_nodes(estimate); 388 } 389 } 390 // Walk up dominators to loop _head looking for test which is 391 // executed on every path thru loop. 392 test = phase->idom(test); 393 } 394 return false; 395 } 396 397 //------------------------------peeled_dom_test_elim--------------------------- 398 // If we got the effect of peeling, either by actually peeling or by making 399 // a pre-loop which must execute at least once, we can remove all 400 // loop-invariant dominated tests in the main body. 401 void PhaseIdealLoop::peeled_dom_test_elim(IdealLoopTree *loop, Node_List &old_new) { 402 bool progress = true; 403 while (progress) { 404 progress = false; // Reset for next iteration 405 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 406 Node *test = prev->in(0); 407 while (test != loop->_head) { // Scan till run off top of loop 408 409 int p_op = prev->Opcode(); 410 if ((p_op == Op_IfFalse || p_op == Op_IfTrue) && 411 test->is_If() && // Test? 412 !test->in(1)->is_Con() && // And not already obvious? 413 // Condition is not a member of this loop? 414 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 415 // Walk loop body looking for instances of this test 416 for (uint i = 0; i < loop->_body.size(); i++) { 417 Node *n = loop->_body.at(i); 418 if (n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/) { 419 // IfNode was dominated by version in peeled loop body 420 progress = true; 421 dominated_by(old_new[prev->_idx], n); 422 } 423 } 424 } 425 prev = test; 426 test = idom(test); 427 } // End of scan tests in loop 428 429 } // End of while (progress) 430 } 431 432 //------------------------------do_peeling------------------------------------- 433 // Peel the first iteration of the given loop. 434 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 435 // The pre-loop illegally has 2 control users (old & new loops). 436 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 437 // Do this by making the old-loop fall-in edges act as if they came 438 // around the loopback from the prior iteration (follow the old-loop 439 // backedges) and then map to the new peeled iteration. This leaves 440 // the pre-loop with only 1 user (the new peeled iteration), but the 441 // peeled-loop backedge has 2 users. 442 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 443 // extra backedge user. 444 // 445 // orig 446 // 447 // stmt1 448 // | 449 // v 450 // loop predicate 451 // | 452 // v 453 // loop<----+ 454 // | | 455 // stmt2 | 456 // | | 457 // v | 458 // if ^ 459 // / \ | 460 // / \ | 461 // v v | 462 // false true | 463 // / \ | 464 // / ----+ 465 // | 466 // v 467 // exit 468 // 469 // 470 // after clone loop 471 // 472 // stmt1 473 // | 474 // v 475 // loop predicate 476 // / \ 477 // clone / \ orig 478 // / \ 479 // / \ 480 // v v 481 // +---->loop clone loop<----+ 482 // | | | | 483 // | stmt2 clone stmt2 | 484 // | | | | 485 // | v v | 486 // ^ if clone If ^ 487 // | / \ / \ | 488 // | / \ / \ | 489 // | v v v v | 490 // | true false false true | 491 // | / \ / \ | 492 // +---- \ / ----+ 493 // \ / 494 // 1v v2 495 // region 496 // | 497 // v 498 // exit 499 // 500 // 501 // after peel and predicate move 502 // 503 // stmt1 504 // / 505 // / 506 // clone / orig 507 // / 508 // / +----------+ 509 // / | | 510 // / loop predicate | 511 // / | | 512 // v v | 513 // TOP-->loop clone loop<----+ | 514 // | | | | 515 // stmt2 clone stmt2 | | 516 // | | | ^ 517 // v v | | 518 // if clone If ^ | 519 // / \ / \ | | 520 // / \ / \ | | 521 // v v v v | | 522 // true false false true | | 523 // | \ / \ | | 524 // | \ / ----+ ^ 525 // | \ / | 526 // | 1v v2 | 527 // v region | 528 // | | | 529 // | v | 530 // | exit | 531 // | | 532 // +--------------->-----------------+ 533 // 534 // 535 // final graph 536 // 537 // stmt1 538 // | 539 // v 540 // stmt2 clone 541 // | 542 // v 543 // if clone 544 // / | 545 // / | 546 // v v 547 // false true 548 // | | 549 // | v 550 // | loop predicate 551 // | | 552 // | v 553 // | loop<----+ 554 // | | | 555 // | stmt2 | 556 // | | | 557 // | v | 558 // v if ^ 559 // | / \ | 560 // | / \ | 561 // | v v | 562 // | false true | 563 // | | \ | 564 // v v --+ 565 // region 566 // | 567 // v 568 // exit 569 // 570 void PhaseIdealLoop::do_peeling(IdealLoopTree *loop, Node_List &old_new) { 571 572 C->set_major_progress(); 573 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 574 // 'pre' loop from the main and the 'pre' can no longer have its 575 // iterations adjusted. Therefore, we need to declare this loop as 576 // no longer a 'main' loop; it will need new pre and post loops before 577 // we can do further RCE. 578 #ifndef PRODUCT 579 if (TraceLoopOpts) { 580 tty->print("Peel "); 581 loop->dump_head(); 582 } 583 #endif 584 LoopNode* head = loop->_head->as_Loop(); 585 bool counted_loop = head->is_CountedLoop(); 586 if (counted_loop) { 587 CountedLoopNode *cl = head->as_CountedLoop(); 588 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 589 cl->set_trip_count(cl->trip_count() - 1); 590 if (cl->is_main_loop()) { 591 cl->set_normal_loop(); 592 #ifndef PRODUCT 593 if (PrintOpto && VerifyLoopOptimizations) { 594 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 595 loop->dump_head(); 596 } 597 #endif 598 } 599 } 600 Node* entry = head->in(LoopNode::EntryControl); 601 602 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 603 // The pre-loop illegally has 2 control users (old & new loops). 604 clone_loop(loop, old_new, dom_depth(head->skip_strip_mined()), ControlAroundStripMined); 605 606 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 607 // Do this by making the old-loop fall-in edges act as if they came 608 // around the loopback from the prior iteration (follow the old-loop 609 // backedges) and then map to the new peeled iteration. This leaves 610 // the pre-loop with only 1 user (the new peeled iteration), but the 611 // peeled-loop backedge has 2 users. 612 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; 613 _igvn.hash_delete(head->skip_strip_mined()); 614 head->skip_strip_mined()->set_req(LoopNode::EntryControl, new_entry); 615 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { 616 Node* old = head->fast_out(j); 617 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { 618 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 619 if (!new_exit_value) // Backedge value is ALSO loop invariant? 620 // Then loop body backedge value remains the same. 621 new_exit_value = old->in(LoopNode::LoopBackControl); 622 _igvn.hash_delete(old); 623 old->set_req(LoopNode::EntryControl, new_exit_value); 624 } 625 } 626 627 628 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 629 // extra backedge user. 630 Node* new_head = old_new[head->_idx]; 631 _igvn.hash_delete(new_head); 632 new_head->set_req(LoopNode::LoopBackControl, C->top()); 633 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { 634 Node* use = new_head->fast_out(j2); 635 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) { 636 _igvn.hash_delete(use); 637 use->set_req(LoopNode::LoopBackControl, C->top()); 638 } 639 } 640 641 642 // Step 4: Correct dom-depth info. Set to loop-head depth. 643 int dd = dom_depth(head); 644 set_idom(head, head->in(1), dd); 645 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 646 Node *old = loop->_body.at(j3); 647 Node *nnn = old_new[old->_idx]; 648 if (!has_ctrl(nnn)) { 649 set_idom(nnn, idom(nnn), dd-1); 650 } 651 } 652 653 // Now force out all loop-invariant dominating tests. The optimizer 654 // finds some, but we _know_ they are all useless. 655 peeled_dom_test_elim(loop,old_new); 656 657 loop->record_for_igvn(); 658 } 659 660 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop 661 662 //------------------------------policy_maximally_unroll------------------------ 663 // Calculate exact loop trip count and return true if loop can be maximally 664 // unrolled. 665 bool IdealLoopTree::policy_maximally_unroll(PhaseIdealLoop *phase) const { 666 CountedLoopNode *cl = _head->as_CountedLoop(); 667 assert(cl->is_normal_loop(), ""); 668 if (!cl->is_valid_counted_loop()) { 669 return false; // Malformed counted loop 670 } 671 if (!cl->has_exact_trip_count()) { 672 // Trip count is not exact. 673 return false; 674 } 675 676 uint trip_count = cl->trip_count(); 677 // Note, max_juint is used to indicate unknown trip count. 678 assert(trip_count > 1, "one iteration loop should be optimized out already"); 679 assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 680 681 // If nodes are depleted, some transform has miscalculated its needs. 682 assert(!phase->exceeding_node_budget(), "sanity"); 683 684 // Real policy: if we maximally unroll, does it get too big? 685 // Allow the unrolled mess to get larger than standard loop 686 // size. After all, it will no longer be a loop. 687 uint body_size = _body.size(); 688 uint unroll_limit = (uint)LoopUnrollLimit * 4; 689 assert((intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 690 if (trip_count > unroll_limit || body_size > unroll_limit) { 691 return false; 692 } 693 694 // Take into account that after unroll conjoined heads and tails will fold, 695 // otherwise policy_unroll() may allow more unrolling than max unrolling. 696 uint new_body_size = est_loop_clone_sz(trip_count, body_size - EMPTY_LOOP_SIZE); 697 698 if (new_body_size == UINT_MAX) { // Check for bad estimate (overflow). 699 return false; 700 } 701 702 // Fully unroll a loop with few iterations regardless next conditions since 703 // following loop optimizations will split such loop anyway (pre-main-post). 704 if (trip_count <= 3) { 705 return phase->may_require_nodes(new_body_size); 706 } 707 708 if (new_body_size > unroll_limit || 709 // Unrolling can result in a large amount of node construction 710 phase->exceeding_node_budget(new_body_size)) { 711 return false; 712 } 713 714 // Do not unroll a loop with String intrinsics code. 715 // String intrinsics are large and have loops. 716 for (uint k = 0; k < _body.size(); k++) { 717 Node* n = _body.at(k); 718 switch (n->Opcode()) { 719 case Op_StrComp: 720 case Op_StrEquals: 721 case Op_StrIndexOf: 722 case Op_StrIndexOfChar: 723 case Op_EncodeISOArray: 724 case Op_AryEq: 725 case Op_HasNegatives: { 726 return false; 727 } 728 #if INCLUDE_RTM_OPT 729 case Op_FastLock: 730 case Op_FastUnlock: { 731 // Don't unroll RTM locking code because it is large. 732 if (UseRTMLocking) { 733 return false; 734 } 735 } 736 #endif 737 } // switch 738 } 739 740 return phase->may_require_nodes(new_body_size); 741 } 742 743 744 //------------------------------policy_unroll---------------------------------- 745 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if the 746 // loop is a CountedLoop and the body is small enough. 747 bool IdealLoopTree::policy_unroll(PhaseIdealLoop *phase) { 748 749 CountedLoopNode *cl = _head->as_CountedLoop(); 750 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 751 752 if (!cl->is_valid_counted_loop()) { 753 return false; // Malformed counted loop 754 } 755 756 // If nodes are depleted, some transform has miscalculated its needs. 757 assert(!phase->exceeding_node_budget(), "sanity"); 758 759 // Protect against over-unrolling. 760 // After split at least one iteration will be executed in pre-loop. 761 if (cl->trip_count() <= (cl->is_normal_loop() ? 2u : 1u)) { 762 return false; 763 } 764 _local_loop_unroll_limit = LoopUnrollLimit; 765 _local_loop_unroll_factor = 4; 766 int future_unroll_cnt = cl->unrolled_count() * 2; 767 if (!cl->is_vectorized_loop()) { 768 if (future_unroll_cnt > LoopMaxUnroll) return false; 769 } else { 770 // obey user constraints on vector mapped loops with additional unrolling applied 771 int unroll_constraint = (cl->slp_max_unroll()) ? cl->slp_max_unroll() : 1; 772 if ((future_unroll_cnt / unroll_constraint) > LoopMaxUnroll) return false; 773 } 774 775 // Check for initial stride being a small enough constant 776 if (abs(cl->stride_con()) > (1<<2)*future_unroll_cnt) return false; 777 778 // Don't unroll if the next round of unrolling would push us 779 // over the expected trip count of the loop. One is subtracted 780 // from the expected trip count because the pre-loop normally 781 // executes 1 iteration. 782 if (UnrollLimitForProfileCheck > 0 && 783 cl->profile_trip_cnt() != COUNT_UNKNOWN && 784 future_unroll_cnt > UnrollLimitForProfileCheck && 785 (float)future_unroll_cnt > cl->profile_trip_cnt() - 1.0) { 786 return false; 787 } 788 789 // When unroll count is greater than LoopUnrollMin, don't unroll if: 790 // the residual iterations are more than 10% of the trip count 791 // and rounds of "unroll,optimize" are not making significant progress 792 // Progress defined as current size less than 20% larger than previous size. 793 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 794 future_unroll_cnt > LoopUnrollMin && 795 (future_unroll_cnt - 1) * (100 / LoopPercentProfileLimit) > cl->profile_trip_cnt() && 796 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 797 return false; 798 } 799 800 Node *init_n = cl->init_trip(); 801 Node *limit_n = cl->limit(); 802 int stride_con = cl->stride_con(); 803 if (limit_n == NULL) return false; // We will dereference it below. 804 805 // Non-constant bounds. 806 // Protect against over-unrolling when init or/and limit are not constant 807 // (so that trip_count's init value is maxint) but iv range is known. 808 if (init_n == NULL || !init_n->is_Con() || !limit_n->is_Con()) { 809 Node* phi = cl->phi(); 810 if (phi != NULL) { 811 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 812 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 813 int next_stride = stride_con * 2; // stride after this unroll 814 if (next_stride > 0) { 815 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 816 iv_type->_lo + next_stride > iv_type->_hi) { 817 return false; // over-unrolling 818 } 819 } else if (next_stride < 0) { 820 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 821 iv_type->_hi + next_stride < iv_type->_lo) { 822 return false; // over-unrolling 823 } 824 } 825 } 826 } 827 828 // After unroll limit will be adjusted: new_limit = limit-stride. 829 // Bailout if adjustment overflow. 830 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); 831 if ((stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi)) || 832 (stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))) 833 return false; // overflow 834 835 // Adjust body_size to determine if we unroll or not 836 uint body_size = _body.size(); 837 // Key test to unroll loop in CRC32 java code 838 int xors_in_loop = 0; 839 // Also count ModL, DivL and MulL which expand mightly 840 for (uint k = 0; k < _body.size(); k++) { 841 Node* n = _body.at(k); 842 switch (n->Opcode()) { 843 case Op_XorI: xors_in_loop++; break; // CRC32 java code 844 case Op_ModL: body_size += 30; break; 845 case Op_DivL: body_size += 30; break; 846 case Op_MulL: body_size += 10; break; 847 case Op_StrComp: 848 case Op_StrEquals: 849 case Op_StrIndexOf: 850 case Op_StrIndexOfChar: 851 case Op_EncodeISOArray: 852 case Op_AryEq: 853 case Op_HasNegatives: { 854 // Do not unroll a loop with String intrinsics code. 855 // String intrinsics are large and have loops. 856 return false; 857 } 858 #if INCLUDE_RTM_OPT 859 case Op_FastLock: 860 case Op_FastUnlock: { 861 // Don't unroll RTM locking code because it is large. 862 if (UseRTMLocking) { 863 return false; 864 } 865 } 866 #endif 867 } // switch 868 } 869 870 if (UseSuperWord) { 871 if (!cl->is_reduction_loop()) { 872 phase->mark_reductions(this); 873 } 874 875 // Only attempt slp analysis when user controls do not prohibit it 876 if (LoopMaxUnroll > _local_loop_unroll_factor) { 877 // Once policy_slp_analysis succeeds, mark the loop with the 878 // maximal unroll factor so that we minimize analysis passes 879 if (future_unroll_cnt >= _local_loop_unroll_factor) { 880 policy_unroll_slp_analysis(cl, phase, future_unroll_cnt); 881 } 882 } 883 } 884 885 int slp_max_unroll_factor = cl->slp_max_unroll(); 886 if ((LoopMaxUnroll < slp_max_unroll_factor) && FLAG_IS_DEFAULT(LoopMaxUnroll) && UseSubwordForMaxVector) { 887 LoopMaxUnroll = slp_max_unroll_factor; 888 } 889 890 uint estimate = est_loop_clone_sz(2, body_size); 891 892 if (cl->has_passed_slp()) { 893 if (slp_max_unroll_factor >= future_unroll_cnt) { 894 return phase->may_require_nodes(estimate); 895 } 896 return false; // Loop too big. 897 } 898 899 // Check for being too big 900 if (body_size > (uint)_local_loop_unroll_limit) { 901 if ((cl->is_subword_loop() || xors_in_loop >= 4) && body_size < 4u * LoopUnrollLimit) { 902 return phase->may_require_nodes(estimate); 903 } 904 return false; // Loop too big. 905 } 906 907 if (cl->is_unroll_only()) { 908 if (TraceSuperWordLoopUnrollAnalysis) { 909 tty->print_cr("policy_unroll passed vector loop(vlen=%d, factor=%d)\n", 910 slp_max_unroll_factor, future_unroll_cnt); 911 } 912 } 913 914 // Unroll once! (Each trip will soon do double iterations) 915 return phase->may_require_nodes(estimate); 916 } 917 918 void IdealLoopTree::policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_cnt) { 919 920 // If nodes are depleted, some transform has miscalculated its needs. 921 assert(!phase->exceeding_node_budget(), "sanity"); 922 923 // Enable this functionality target by target as needed 924 if (SuperWordLoopUnrollAnalysis) { 925 if (!cl->was_slp_analyzed()) { 926 SuperWord sw(phase); 927 sw.transform_loop(this, false); 928 929 // If the loop is slp canonical analyze it 930 if (sw.early_return() == false) { 931 sw.unrolling_analysis(_local_loop_unroll_factor); 932 } 933 } 934 935 if (cl->has_passed_slp()) { 936 int slp_max_unroll_factor = cl->slp_max_unroll(); 937 if (slp_max_unroll_factor >= future_unroll_cnt) { 938 int new_limit = cl->node_count_before_unroll() * slp_max_unroll_factor; 939 if (new_limit > LoopUnrollLimit) { 940 if (TraceSuperWordLoopUnrollAnalysis) { 941 tty->print_cr("slp analysis unroll=%d, default limit=%d\n", new_limit, _local_loop_unroll_limit); 942 } 943 _local_loop_unroll_limit = new_limit; 944 } 945 } 946 } 947 } 948 } 949 950 //------------------------------policy_align----------------------------------- 951 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 952 // expression that does the alignment. Note that only one array base can be 953 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 954 // if we vectorize short memory ops into longer memory ops, we may want to 955 // increase alignment. 956 bool IdealLoopTree::policy_align(PhaseIdealLoop *phase) const { 957 return false; 958 } 959 960 //------------------------------policy_range_check----------------------------- 961 // Return TRUE or FALSE if the loop should be range-check-eliminated. 962 // Actually we do iteration-splitting, a more powerful form of RCE. 963 bool IdealLoopTree::policy_range_check(PhaseIdealLoop *phase) const { 964 if (!RangeCheckElimination) return false; 965 966 // If nodes are depleted, some transform has miscalculated its needs. 967 assert(!phase->exceeding_node_budget(), "sanity"); 968 969 CountedLoopNode *cl = _head->as_CountedLoop(); 970 // If we unrolled with no intention of doing RCE and we later 971 // changed our minds, we got no pre-loop. Either we need to 972 // make a new pre-loop, or we gotta disallow RCE. 973 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now. 974 Node *trip_counter = cl->phi(); 975 976 // check for vectorized loops, some opts are no longer needed 977 if (cl->is_unroll_only()) return false; 978 979 // Check loop body for tests of trip-counter plus loop-invariant vs 980 // loop-invariant. 981 for (uint i = 0; i < _body.size(); i++) { 982 Node *iff = _body[i]; 983 if (iff->Opcode() == Op_If || 984 iff->Opcode() == Op_RangeCheck) { // Test? 985 986 // Comparing trip+off vs limit 987 Node *bol = iff->in(1); 988 if (bol->req() != 2) { 989 continue; // dead constant test 990 } 991 if (!bol->is_Bool()) { 992 assert(bol->Opcode() == Op_Conv2B, "predicate check only"); 993 continue; 994 } 995 if (bol->as_Bool()->_test._test == BoolTest::ne) { 996 continue; // not RC 997 } 998 Node *cmp = bol->in(1); 999 Node *rc_exp = cmp->in(1); 1000 Node *limit = cmp->in(2); 1001 1002 Node *limit_c = phase->get_ctrl(limit); 1003 if (limit_c == phase->C->top()) { 1004 return false; // Found dead test on live IF? No RCE! 1005 } 1006 if (is_member(phase->get_loop(limit_c))) { 1007 // Compare might have operands swapped; commute them 1008 rc_exp = cmp->in(2); 1009 limit = cmp->in(1); 1010 limit_c = phase->get_ctrl(limit); 1011 if (is_member(phase->get_loop(limit_c))) { 1012 continue; // Both inputs are loop varying; cannot RCE 1013 } 1014 } 1015 1016 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 1017 continue; 1018 } 1019 // Found a test like 'trip+off vs limit'. Test is an IfNode, has two 1020 // (2) projections. If BOTH are in the loop we need loop unswitching 1021 // instead of iteration splitting. 1022 if (is_loop_exit(iff)) { 1023 // Found valid reason to split iterations (if there is room). 1024 // NOTE: Usually a gross overestimate. 1025 return phase->may_require_nodes(est_loop_clone_sz(2, _body.size())); 1026 } 1027 } // End of is IF 1028 } 1029 1030 return false; 1031 } 1032 1033 //------------------------------policy_peel_only------------------------------- 1034 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 1035 // for unrolling loops with NO array accesses. 1036 bool IdealLoopTree::policy_peel_only(PhaseIdealLoop *phase) const { 1037 1038 // If nodes are depleted, some transform has miscalculated its needs. 1039 assert(!phase->exceeding_node_budget(), "sanity"); 1040 1041 // check for vectorized loops, any peeling done was already applied 1042 if (_head->is_CountedLoop() && _head->as_CountedLoop()->is_unroll_only()) { 1043 return false; 1044 } 1045 1046 for (uint i = 0; i < _body.size(); i++) { 1047 if (_body[i]->is_Mem()) { 1048 return false; 1049 } 1050 } 1051 // No memory accesses at all! 1052 return true; 1053 } 1054 1055 //------------------------------clone_up_backedge_goo-------------------------- 1056 // If Node n lives in the back_ctrl block and cannot float, we clone a private 1057 // version of n in preheader_ctrl block and return that, otherwise return n. 1058 Node *PhaseIdealLoop::clone_up_backedge_goo(Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones) { 1059 if (get_ctrl(n) != back_ctrl) return n; 1060 1061 // Only visit once 1062 if (visited.test_set(n->_idx)) { 1063 Node *x = clones.find(n->_idx); 1064 return (x != NULL) ? x : n; 1065 } 1066 1067 Node *x = NULL; // If required, a clone of 'n' 1068 // Check for 'n' being pinned in the backedge. 1069 if (n->in(0) && n->in(0) == back_ctrl) { 1070 assert(clones.find(n->_idx) == NULL, "dead loop"); 1071 x = n->clone(); // Clone a copy of 'n' to preheader 1072 clones.push(x, n->_idx); 1073 x->set_req(0, preheader_ctrl); // Fix x's control input to preheader 1074 } 1075 1076 // Recursive fixup any other input edges into x. 1077 // If there are no changes we can just return 'n', otherwise 1078 // we need to clone a private copy and change it. 1079 for (uint i = 1; i < n->req(); i++) { 1080 Node *g = clone_up_backedge_goo(back_ctrl, preheader_ctrl, n->in(i), visited, clones); 1081 if (g != n->in(i)) { 1082 if (!x) { 1083 assert(clones.find(n->_idx) == NULL, "dead loop"); 1084 x = n->clone(); 1085 clones.push(x, n->_idx); 1086 } 1087 x->set_req(i, g); 1088 } 1089 } 1090 if (x) { // x can legally float to pre-header location 1091 register_new_node(x, preheader_ctrl); 1092 return x; 1093 } else { // raise n to cover LCA of uses 1094 set_ctrl(n, find_non_split_ctrl(back_ctrl->in(0))); 1095 } 1096 return n; 1097 } 1098 1099 Node* PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { 1100 Node* castii = new CastIINode(incr, TypeInt::INT, true); 1101 castii->set_req(0, ctrl); 1102 register_new_node(castii, ctrl); 1103 for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { 1104 Node* n = incr->fast_out(i); 1105 if (n->is_Phi() && n->in(0) == loop) { 1106 int nrep = n->replace_edge(incr, castii); 1107 return castii; 1108 } 1109 } 1110 return NULL; 1111 } 1112 1113 // Make a copy of the skeleton range check predicates before the main 1114 // loop and set the initial value of loop as input. After unrolling, 1115 // the range of values for the induction variable in the main loop can 1116 // fall outside the allowed range of values by the array access (main 1117 // loop is never executed). When that happens, range check 1118 // CastII/ConvI2L nodes cause some data paths to die. For consistency, 1119 // the control paths must die too but the range checks were removed by 1120 // predication. The range checks that we add here guarantee that they do. 1121 void PhaseIdealLoop::duplicate_predicates_helper(Node* predicate, Node* start, Node* end, 1122 IdealLoopTree* outer_loop, LoopNode* outer_main_head, 1123 uint dd_main_head) { 1124 if (predicate != NULL) { 1125 IfNode* iff = predicate->in(0)->as_If(); 1126 ProjNode* uncommon_proj = iff->proj_out(1 - predicate->as_Proj()->_con); 1127 Node* rgn = uncommon_proj->unique_ctrl_out(); 1128 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); 1129 assert(iff->in(1)->in(1)->Opcode() == Op_Opaque1, "unexpected predicate shape"); 1130 predicate = iff->in(0); 1131 Node* current_proj = outer_main_head->in(LoopNode::EntryControl); 1132 Node* prev_proj = current_proj; 1133 while (predicate != NULL && predicate->is_Proj() && predicate->in(0)->is_If()) { 1134 iff = predicate->in(0)->as_If(); 1135 uncommon_proj = iff->proj_out(1 - predicate->as_Proj()->_con); 1136 if (uncommon_proj->unique_ctrl_out() != rgn) 1137 break; 1138 if (iff->in(1)->Opcode() == Op_Opaque4) { 1139 assert(skeleton_predicate_has_opaque(iff), "unexpected"); 1140 // Clone the predicate twice and initialize one with the initial 1141 // value of the loop induction variable. Leave the other predicate 1142 // to be initialized when increasing the stride during loop unrolling. 1143 prev_proj = clone_skeleton_predicate(iff, start, predicate, uncommon_proj, current_proj, outer_loop, prev_proj); 1144 assert(skeleton_predicate_has_opaque(prev_proj->in(0)->as_If()) == (start->Opcode() == Op_Opaque1), ""); 1145 prev_proj = clone_skeleton_predicate(iff, end, predicate, uncommon_proj, current_proj, outer_loop, prev_proj); 1146 assert(skeleton_predicate_has_opaque(prev_proj->in(0)->as_If()) == (end->Opcode() == Op_Opaque1), ""); 1147 // Remove the skeleton predicate from the pre-loop 1148 _igvn.replace_input_of(iff, 1, _igvn.intcon(1)); 1149 } 1150 predicate = predicate->in(0)->in(0); 1151 } 1152 _igvn.replace_input_of(outer_main_head, LoopNode::EntryControl, prev_proj); 1153 set_idom(outer_main_head, prev_proj, dd_main_head); 1154 } 1155 } 1156 1157 static bool skeleton_follow_inputs(Node* n, int op) { 1158 return (n->is_Bool() || 1159 n->is_Cmp() || 1160 op == Op_AndL || 1161 op == Op_OrL || 1162 op == Op_RShiftL || 1163 op == Op_LShiftL || 1164 op == Op_AddL || 1165 op == Op_AddI || 1166 op == Op_MulL || 1167 op == Op_MulI || 1168 op == Op_SubL || 1169 op == Op_SubI || 1170 op == Op_ConvI2L); 1171 } 1172 1173 bool PhaseIdealLoop::skeleton_predicate_has_opaque(IfNode* iff) { 1174 ResourceMark rm; 1175 Unique_Node_List wq; 1176 wq.push(iff->in(1)->in(1)); 1177 for (uint i = 0; i < wq.size(); i++) { 1178 Node* n = wq.at(i); 1179 int op = n->Opcode(); 1180 if (skeleton_follow_inputs(n, op)) { 1181 for (uint j = 1; j < n->req(); j++) { 1182 Node* m = n->in(j); 1183 if (m != NULL) { 1184 wq.push(m); 1185 } 1186 } 1187 continue; 1188 } 1189 if (op == Op_Opaque1) { 1190 return true; 1191 } 1192 } 1193 return false; 1194 } 1195 1196 Node* PhaseIdealLoop::clone_skeleton_predicate(Node* iff, Node* value, Node* predicate, Node* uncommon_proj, 1197 Node* current_proj, IdealLoopTree* outer_loop, Node* prev_proj) { 1198 Node_Stack to_clone(2); 1199 to_clone.push(iff->in(1), 1); 1200 uint current = C->unique(); 1201 Node* result = NULL; 1202 // Look for the opaque node to replace with the new value 1203 // and clone everything in between. We keep the Opaque4 node 1204 // so the duplicated predicates are eliminated once loop 1205 // opts are over: they are here only to keep the IR graph 1206 // consistent. 1207 do { 1208 Node* n = to_clone.node(); 1209 uint i = to_clone.index(); 1210 Node* m = n->in(i); 1211 int op = m->Opcode(); 1212 if (skeleton_follow_inputs(m, op)) { 1213 to_clone.push(m, 1); 1214 continue; 1215 } 1216 if (op == Op_Opaque1) { 1217 if (n->_idx < current) { 1218 n = n->clone(); 1219 } 1220 n->set_req(i, value); 1221 register_new_node(n, current_proj); 1222 to_clone.set_node(n); 1223 } 1224 for (;;) { 1225 Node* cur = to_clone.node(); 1226 uint j = to_clone.index(); 1227 if (j+1 < cur->req()) { 1228 to_clone.set_index(j+1); 1229 break; 1230 } 1231 to_clone.pop(); 1232 if (to_clone.size() == 0) { 1233 result = cur; 1234 break; 1235 } 1236 Node* next = to_clone.node(); 1237 j = to_clone.index(); 1238 if (next->in(j) != cur) { 1239 assert(cur->_idx >= current || next->in(j)->Opcode() == Op_Opaque1, "new node or Opaque1 being replaced"); 1240 if (next->_idx < current) { 1241 next = next->clone(); 1242 register_new_node(next, current_proj); 1243 to_clone.set_node(next); 1244 } 1245 next->set_req(j, cur); 1246 } 1247 } 1248 } while (result == NULL); 1249 assert(result->_idx >= current, "new node expected"); 1250 1251 Node* proj = predicate->clone(); 1252 Node* other_proj = uncommon_proj->clone(); 1253 Node* new_iff = iff->clone(); 1254 new_iff->set_req(1, result); 1255 proj->set_req(0, new_iff); 1256 other_proj->set_req(0, new_iff); 1257 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr); 1258 register_new_node(frame, C->start()); 1259 // It's impossible for the predicate to fail at runtime. Use an Halt node. 1260 Node* halt = new HaltNode(other_proj, frame); 1261 C->root()->add_req(halt); 1262 new_iff->set_req(0, prev_proj); 1263 1264 register_control(new_iff, outer_loop->_parent, prev_proj); 1265 register_control(proj, outer_loop->_parent, new_iff); 1266 register_control(other_proj, _ltree_root, new_iff); 1267 register_control(halt, _ltree_root, other_proj); 1268 return proj; 1269 } 1270 1271 void PhaseIdealLoop::duplicate_predicates(CountedLoopNode* pre_head, Node* start, Node* end, 1272 IdealLoopTree* outer_loop, LoopNode* outer_main_head, 1273 uint dd_main_head) { 1274 if (UseLoopPredicate) { 1275 Node* entry = pre_head->in(LoopNode::EntryControl); 1276 Node* predicate = NULL; 1277 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 1278 if (predicate != NULL) { 1279 entry = skip_loop_predicates(entry); 1280 } 1281 Node* profile_predicate = NULL; 1282 if (UseProfiledLoopPredicate) { 1283 profile_predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); 1284 if (profile_predicate != NULL) { 1285 entry = skip_loop_predicates(entry); 1286 } 1287 } 1288 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 1289 duplicate_predicates_helper(predicate, start, end, outer_loop, outer_main_head, dd_main_head); 1290 duplicate_predicates_helper(profile_predicate, start, end, outer_loop, outer_main_head, dd_main_head); 1291 } 1292 } 1293 1294 //------------------------------insert_pre_post_loops-------------------------- 1295 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 1296 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 1297 // alignment. Useful to unroll loops that do no array accesses. 1298 void PhaseIdealLoop::insert_pre_post_loops(IdealLoopTree *loop, Node_List &old_new, bool peel_only) { 1299 1300 #ifndef PRODUCT 1301 if (TraceLoopOpts) { 1302 if (peel_only) 1303 tty->print("PeelMainPost "); 1304 else 1305 tty->print("PreMainPost "); 1306 loop->dump_head(); 1307 } 1308 #endif 1309 C->set_major_progress(); 1310 1311 // Find common pieces of the loop being guarded with pre & post loops 1312 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1313 assert(main_head->is_normal_loop(), ""); 1314 CountedLoopEndNode *main_end = main_head->loopexit(); 1315 assert(main_end->outcnt() == 2, "1 true, 1 false path only"); 1316 1317 Node *pre_header= main_head->in(LoopNode::EntryControl); 1318 Node *init = main_head->init_trip(); 1319 Node *incr = main_end ->incr(); 1320 Node *limit = main_end ->limit(); 1321 Node *stride = main_end ->stride(); 1322 Node *cmp = main_end ->cmp_node(); 1323 BoolTest::mask b_test = main_end->test_trip(); 1324 1325 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 1326 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 1327 if (bol->outcnt() != 1) { 1328 bol = bol->clone(); 1329 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 1330 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, bol); 1331 } 1332 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 1333 if (cmp->outcnt() != 1) { 1334 cmp = cmp->clone(); 1335 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 1336 _igvn.replace_input_of(bol, 1, cmp); 1337 } 1338 1339 // Add the post loop 1340 CountedLoopNode *post_head = NULL; 1341 Node *main_exit = insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1342 1343 //------------------------------ 1344 // Step B: Create Pre-Loop. 1345 1346 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 1347 // loop pre-header illegally has 2 control users (old & new loops). 1348 LoopNode* outer_main_head = main_head; 1349 IdealLoopTree* outer_loop = loop; 1350 if (main_head->is_strip_mined()) { 1351 main_head->verify_strip_mined(1); 1352 outer_main_head = main_head->outer_loop(); 1353 outer_loop = loop->_parent; 1354 assert(outer_loop->_head == outer_main_head, "broken loop tree"); 1355 } 1356 uint dd_main_head = dom_depth(outer_main_head); 1357 clone_loop(loop, old_new, dd_main_head, ControlAroundStripMined); 1358 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 1359 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 1360 pre_head->set_pre_loop(main_head); 1361 Node *pre_incr = old_new[incr->_idx]; 1362 1363 // Reduce the pre-loop trip count. 1364 pre_end->_prob = PROB_FAIR; 1365 1366 // Find the pre-loop normal exit. 1367 Node* pre_exit = pre_end->proj_out(false); 1368 assert(pre_exit->Opcode() == Op_IfFalse, ""); 1369 IfFalseNode *new_pre_exit = new IfFalseNode(pre_end); 1370 _igvn.register_new_node_with_optimizer(new_pre_exit); 1371 set_idom(new_pre_exit, pre_end, dd_main_head); 1372 set_loop(new_pre_exit, outer_loop->_parent); 1373 1374 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 1375 // pre-loop, the main-loop may not execute at all. Later in life this 1376 // zero-trip guard will become the minimum-trip guard when we unroll 1377 // the main-loop. 1378 Node *min_opaq = new Opaque1Node(C, limit); 1379 Node *min_cmp = new CmpINode(pre_incr, min_opaq); 1380 Node *min_bol = new BoolNode(min_cmp, b_test); 1381 register_new_node(min_opaq, new_pre_exit); 1382 register_new_node(min_cmp , new_pre_exit); 1383 register_new_node(min_bol , new_pre_exit); 1384 1385 // Build the IfNode (assume the main-loop is executed always). 1386 IfNode *min_iff = new IfNode(new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN); 1387 _igvn.register_new_node_with_optimizer(min_iff); 1388 set_idom(min_iff, new_pre_exit, dd_main_head); 1389 set_loop(min_iff, outer_loop->_parent); 1390 1391 // Plug in the false-path, taken if we need to skip main-loop 1392 _igvn.hash_delete(pre_exit); 1393 pre_exit->set_req(0, min_iff); 1394 set_idom(pre_exit, min_iff, dd_main_head); 1395 set_idom(pre_exit->unique_ctrl_out(), min_iff, dd_main_head); 1396 // Make the true-path, must enter the main loop 1397 Node *min_taken = new IfTrueNode(min_iff); 1398 _igvn.register_new_node_with_optimizer(min_taken); 1399 set_idom(min_taken, min_iff, dd_main_head); 1400 set_loop(min_taken, outer_loop->_parent); 1401 // Plug in the true path 1402 _igvn.hash_delete(outer_main_head); 1403 outer_main_head->set_req(LoopNode::EntryControl, min_taken); 1404 set_idom(outer_main_head, min_taken, dd_main_head); 1405 1406 Arena *a = Thread::current()->resource_area(); 1407 VectorSet visited(a); 1408 Node_Stack clones(a, main_head->back_control()->outcnt()); 1409 // Step B3: Make the fall-in values to the main-loop come from the 1410 // fall-out values of the pre-loop. 1411 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 1412 Node* main_phi = main_head->fast_out(i2); 1413 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0) { 1414 Node *pre_phi = old_new[main_phi->_idx]; 1415 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 1416 main_head->skip_strip_mined()->in(LoopNode::EntryControl), 1417 pre_phi->in(LoopNode::LoopBackControl), 1418 visited, clones); 1419 _igvn.hash_delete(main_phi); 1420 main_phi->set_req(LoopNode::EntryControl, fallpre); 1421 } 1422 } 1423 1424 // Nodes inside the loop may be control dependent on a predicate 1425 // that was moved before the preloop. If the back branch of the main 1426 // or post loops becomes dead, those nodes won't be dependent on the 1427 // test that guards that loop nest anymore which could lead to an 1428 // incorrect array access because it executes independently of the 1429 // test that was guarding the loop nest. We add a special CastII on 1430 // the if branch that enters the loop, between the input induction 1431 // variable value and the induction variable Phi to preserve correct 1432 // dependencies. 1433 1434 // CastII for the main loop: 1435 Node* castii = cast_incr_before_loop(pre_incr, min_taken, main_head); 1436 assert(castii != NULL, "no castII inserted"); 1437 Node* opaque_castii = new Opaque1Node(C, castii); 1438 register_new_node(opaque_castii, outer_main_head->in(LoopNode::EntryControl)); 1439 duplicate_predicates(pre_head, castii, opaque_castii, outer_loop, outer_main_head, dd_main_head); 1440 1441 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 1442 // RCE and alignment may change this later. 1443 Node *cmp_end = pre_end->cmp_node(); 1444 assert(cmp_end->in(2) == limit, ""); 1445 Node *pre_limit = new AddINode(init, stride); 1446 1447 // Save the original loop limit in this Opaque1 node for 1448 // use by range check elimination. 1449 Node *pre_opaq = new Opaque1Node(C, pre_limit, limit); 1450 1451 register_new_node(pre_limit, pre_head->in(0)); 1452 register_new_node(pre_opaq , pre_head->in(0)); 1453 1454 // Since no other users of pre-loop compare, I can hack limit directly 1455 assert(cmp_end->outcnt() == 1, "no other users"); 1456 _igvn.hash_delete(cmp_end); 1457 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 1458 1459 // Special case for not-equal loop bounds: 1460 // Change pre loop test, main loop test, and the 1461 // main loop guard test to use lt or gt depending on stride 1462 // direction: 1463 // positive stride use < 1464 // negative stride use > 1465 // 1466 // not-equal test is kept for post loop to handle case 1467 // when init > limit when stride > 0 (and reverse). 1468 1469 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 1470 1471 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 1472 // Modify pre loop end condition 1473 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1474 BoolNode* new_bol0 = new BoolNode(pre_bol->in(1), new_test); 1475 register_new_node(new_bol0, pre_head->in(0)); 1476 _igvn.replace_input_of(pre_end, CountedLoopEndNode::TestValue, new_bol0); 1477 // Modify main loop guard condition 1478 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 1479 BoolNode* new_bol1 = new BoolNode(min_bol->in(1), new_test); 1480 register_new_node(new_bol1, new_pre_exit); 1481 _igvn.hash_delete(min_iff); 1482 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 1483 // Modify main loop end condition 1484 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 1485 BoolNode* new_bol2 = new BoolNode(main_bol->in(1), new_test); 1486 register_new_node(new_bol2, main_end->in(CountedLoopEndNode::TestControl)); 1487 _igvn.replace_input_of(main_end, CountedLoopEndNode::TestValue, new_bol2); 1488 } 1489 1490 // Flag main loop 1491 main_head->set_main_loop(); 1492 if (peel_only) { 1493 main_head->set_main_no_pre_loop(); 1494 } 1495 1496 // Subtract a trip count for the pre-loop. 1497 main_head->set_trip_count(main_head->trip_count() - 1); 1498 1499 // It's difficult to be precise about the trip-counts 1500 // for the pre/post loops. They are usually very short, 1501 // so guess that 4 trips is a reasonable value. 1502 post_head->set_profile_trip_cnt(4.0); 1503 pre_head->set_profile_trip_cnt(4.0); 1504 1505 // Now force out all loop-invariant dominating tests. The optimizer 1506 // finds some, but we _know_ they are all useless. 1507 peeled_dom_test_elim(loop,old_new); 1508 loop->record_for_igvn(); 1509 } 1510 1511 //------------------------------insert_vector_post_loop------------------------ 1512 // Insert a copy of the atomic unrolled vectorized main loop as a post loop, 1513 // unroll_policy has already informed us that more unrolling is about to 1514 // happen to the main loop. The resultant post loop will serve as a 1515 // vectorized drain loop. 1516 void PhaseIdealLoop::insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new) { 1517 if (!loop->_head->is_CountedLoop()) return; 1518 1519 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1520 1521 // only process vectorized main loops 1522 if (!cl->is_vectorized_loop() || !cl->is_main_loop()) return; 1523 1524 if (!may_require_nodes(est_loop_clone_sz(2, loop->_body.size()))) { 1525 return; 1526 } 1527 int slp_max_unroll_factor = cl->slp_max_unroll(); 1528 int cur_unroll = cl->unrolled_count(); 1529 1530 if (slp_max_unroll_factor == 0) return; 1531 1532 // only process atomic unroll vector loops (not super unrolled after vectorization) 1533 if (cur_unroll != slp_max_unroll_factor) return; 1534 1535 // we only ever process this one time 1536 if (cl->has_atomic_post_loop()) return; 1537 1538 #ifndef PRODUCT 1539 if (TraceLoopOpts) { 1540 tty->print("PostVector "); 1541 loop->dump_head(); 1542 } 1543 #endif 1544 C->set_major_progress(); 1545 1546 // Find common pieces of the loop being guarded with pre & post loops 1547 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1548 CountedLoopEndNode *main_end = main_head->loopexit(); 1549 // diagnostic to show loop end is not properly formed 1550 assert(main_end->outcnt() == 2, "1 true, 1 false path only"); 1551 1552 // mark this loop as processed 1553 main_head->mark_has_atomic_post_loop(); 1554 1555 Node *incr = main_end->incr(); 1556 Node *limit = main_end->limit(); 1557 1558 // In this case we throw away the result as we are not using it to connect anything else. 1559 CountedLoopNode *post_head = NULL; 1560 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1561 1562 // It's difficult to be precise about the trip-counts 1563 // for post loops. They are usually very short, 1564 // so guess that unit vector trips is a reasonable value. 1565 post_head->set_profile_trip_cnt(cur_unroll); 1566 1567 // Now force out all loop-invariant dominating tests. The optimizer 1568 // finds some, but we _know_ they are all useless. 1569 peeled_dom_test_elim(loop, old_new); 1570 loop->record_for_igvn(); 1571 } 1572 1573 1574 //-------------------------insert_scalar_rced_post_loop------------------------ 1575 // Insert a copy of the rce'd main loop as a post loop, 1576 // We have not unrolled the main loop, so this is the right time to inject this. 1577 // Later we will examine the partner of this post loop pair which still has range checks 1578 // to see inject code which tests at runtime if the range checks are applicable. 1579 void PhaseIdealLoop::insert_scalar_rced_post_loop(IdealLoopTree *loop, Node_List &old_new) { 1580 if (!loop->_head->is_CountedLoop()) return; 1581 1582 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1583 1584 // only process RCE'd main loops 1585 if (!cl->is_main_loop() || cl->range_checks_present()) return; 1586 1587 #ifndef PRODUCT 1588 if (TraceLoopOpts) { 1589 tty->print("PostScalarRce "); 1590 loop->dump_head(); 1591 } 1592 #endif 1593 C->set_major_progress(); 1594 1595 // Find common pieces of the loop being guarded with pre & post loops 1596 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 1597 CountedLoopEndNode *main_end = main_head->loopexit(); 1598 // diagnostic to show loop end is not properly formed 1599 assert(main_end->outcnt() == 2, "1 true, 1 false path only"); 1600 1601 Node *incr = main_end->incr(); 1602 Node *limit = main_end->limit(); 1603 1604 // In this case we throw away the result as we are not using it to connect anything else. 1605 CountedLoopNode *post_head = NULL; 1606 insert_post_loop(loop, old_new, main_head, main_end, incr, limit, post_head); 1607 1608 // It's difficult to be precise about the trip-counts 1609 // for post loops. They are usually very short, 1610 // so guess that unit vector trips is a reasonable value. 1611 post_head->set_profile_trip_cnt(4.0); 1612 post_head->set_is_rce_post_loop(); 1613 1614 // Now force out all loop-invariant dominating tests. The optimizer 1615 // finds some, but we _know_ they are all useless. 1616 peeled_dom_test_elim(loop, old_new); 1617 loop->record_for_igvn(); 1618 } 1619 1620 1621 //------------------------------insert_post_loop------------------------------- 1622 // Insert post loops. Add a post loop to the given loop passed. 1623 Node *PhaseIdealLoop::insert_post_loop(IdealLoopTree *loop, Node_List &old_new, 1624 CountedLoopNode *main_head, CountedLoopEndNode *main_end, 1625 Node *incr, Node *limit, CountedLoopNode *&post_head) { 1626 IfNode* outer_main_end = main_end; 1627 IdealLoopTree* outer_loop = loop; 1628 if (main_head->is_strip_mined()) { 1629 main_head->verify_strip_mined(1); 1630 outer_main_end = main_head->outer_loop_end(); 1631 outer_loop = loop->_parent; 1632 assert(outer_loop->_head == main_head->in(LoopNode::EntryControl), "broken loop tree"); 1633 } 1634 1635 //------------------------------ 1636 // Step A: Create a new post-Loop. 1637 Node* main_exit = outer_main_end->proj_out(false); 1638 assert(main_exit->Opcode() == Op_IfFalse, ""); 1639 int dd_main_exit = dom_depth(main_exit); 1640 1641 // Step A1: Clone the loop body of main. The clone becomes the post-loop. 1642 // The main loop pre-header illegally has 2 control users (old & new loops). 1643 clone_loop(loop, old_new, dd_main_exit, ControlAroundStripMined); 1644 assert(old_new[main_end->_idx]->Opcode() == Op_CountedLoopEnd, ""); 1645 post_head = old_new[main_head->_idx]->as_CountedLoop(); 1646 post_head->set_normal_loop(); 1647 post_head->set_post_loop(main_head); 1648 1649 // Reduce the post-loop trip count. 1650 CountedLoopEndNode* post_end = old_new[main_end->_idx]->as_CountedLoopEnd(); 1651 post_end->_prob = PROB_FAIR; 1652 1653 // Build the main-loop normal exit. 1654 IfFalseNode *new_main_exit = new IfFalseNode(outer_main_end); 1655 _igvn.register_new_node_with_optimizer(new_main_exit); 1656 set_idom(new_main_exit, outer_main_end, dd_main_exit); 1657 set_loop(new_main_exit, outer_loop->_parent); 1658 1659 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 1660 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 1661 // (the previous loop trip-counter exit value) because we will be changing 1662 // the exit value (via additional unrolling) so we cannot constant-fold away the zero 1663 // trip guard until all unrolling is done. 1664 Node *zer_opaq = new Opaque1Node(C, incr); 1665 Node *zer_cmp = new CmpINode(zer_opaq, limit); 1666 Node *zer_bol = new BoolNode(zer_cmp, main_end->test_trip()); 1667 register_new_node(zer_opaq, new_main_exit); 1668 register_new_node(zer_cmp, new_main_exit); 1669 register_new_node(zer_bol, new_main_exit); 1670 1671 // Build the IfNode 1672 IfNode *zer_iff = new IfNode(new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN); 1673 _igvn.register_new_node_with_optimizer(zer_iff); 1674 set_idom(zer_iff, new_main_exit, dd_main_exit); 1675 set_loop(zer_iff, outer_loop->_parent); 1676 1677 // Plug in the false-path, taken if we need to skip this post-loop 1678 _igvn.replace_input_of(main_exit, 0, zer_iff); 1679 set_idom(main_exit, zer_iff, dd_main_exit); 1680 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 1681 // Make the true-path, must enter this post loop 1682 Node *zer_taken = new IfTrueNode(zer_iff); 1683 _igvn.register_new_node_with_optimizer(zer_taken); 1684 set_idom(zer_taken, zer_iff, dd_main_exit); 1685 set_loop(zer_taken, outer_loop->_parent); 1686 // Plug in the true path 1687 _igvn.hash_delete(post_head); 1688 post_head->set_req(LoopNode::EntryControl, zer_taken); 1689 set_idom(post_head, zer_taken, dd_main_exit); 1690 1691 Arena *a = Thread::current()->resource_area(); 1692 VectorSet visited(a); 1693 Node_Stack clones(a, main_head->back_control()->outcnt()); 1694 // Step A3: Make the fall-in values to the post-loop come from the 1695 // fall-out values of the main-loop. 1696 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 1697 Node* main_phi = main_head->fast_out(i); 1698 if (main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0) { 1699 Node *cur_phi = old_new[main_phi->_idx]; 1700 Node *fallnew = clone_up_backedge_goo(main_head->back_control(), 1701 post_head->init_control(), 1702 main_phi->in(LoopNode::LoopBackControl), 1703 visited, clones); 1704 _igvn.hash_delete(cur_phi); 1705 cur_phi->set_req(LoopNode::EntryControl, fallnew); 1706 } 1707 } 1708 1709 // CastII for the new post loop: 1710 Node* castii = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); 1711 assert(castii != NULL, "no castII inserted"); 1712 1713 return new_main_exit; 1714 } 1715 1716 //------------------------------is_invariant----------------------------- 1717 // Return true if n is invariant 1718 bool IdealLoopTree::is_invariant(Node* n) const { 1719 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 1720 if (n_c->is_top()) return false; 1721 return !is_member(_phase->get_loop(n_c)); 1722 } 1723 1724 void PhaseIdealLoop::update_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con) { 1725 // Search for skeleton predicates and update them according to the new stride 1726 Node* entry = ctrl; 1727 Node* prev_proj = ctrl; 1728 LoopNode* outer_loop_head = loop_head->skip_strip_mined(); 1729 IdealLoopTree* outer_loop = get_loop(outer_loop_head); 1730 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) { 1731 IfNode* iff = entry->in(0)->as_If(); 1732 ProjNode* proj = iff->proj_out(1 - entry->as_Proj()->_con); 1733 if (proj->unique_ctrl_out()->Opcode() != Op_Halt) { 1734 break; 1735 } 1736 if (iff->in(1)->Opcode() == Op_Opaque4) { 1737 // Look for predicate with an Opaque1 node that can be used as a template 1738 if (!skeleton_predicate_has_opaque(iff)) { 1739 // No Opaque1 node? It's either the check for the first value 1740 // of the first iteration or the check for the last value of 1741 // the first iteration of an unrolled loop. We can't 1742 // tell. Kill it in any case. 1743 _igvn.replace_input_of(iff, 1, iff->in(1)->in(2)); 1744 } else { 1745 // Add back the predicate for the value at the beginning of the first entry 1746 prev_proj = clone_skeleton_predicate(iff, init, entry, proj, ctrl, outer_loop, prev_proj); 1747 assert(!skeleton_predicate_has_opaque(prev_proj->in(0)->as_If()), "unexpected"); 1748 // Compute the value of the loop induction variable at the end of the 1749 // first iteration of the unrolled loop: init + new_stride_con - init_inc 1750 int init_inc = stride_con/loop_head->unrolled_count(); 1751 assert(init_inc != 0, "invalid loop increment"); 1752 int new_stride_con = stride_con * 2; 1753 Node* max_value = _igvn.intcon(new_stride_con - init_inc); 1754 max_value = new AddINode(init, max_value); 1755 register_new_node(max_value, get_ctrl(iff->in(1))); 1756 prev_proj = clone_skeleton_predicate(iff, max_value, entry, proj, ctrl, outer_loop, prev_proj); 1757 assert(!skeleton_predicate_has_opaque(prev_proj->in(0)->as_If()), "unexpected"); 1758 } 1759 } 1760 entry = entry->in(0)->in(0); 1761 } 1762 if (prev_proj != ctrl) { 1763 _igvn.replace_input_of(outer_loop_head, LoopNode::EntryControl, prev_proj); 1764 set_idom(outer_loop_head, prev_proj, dom_depth(outer_loop_head)); 1765 } 1766 } 1767 1768 //------------------------------do_unroll-------------------------------------- 1769 // Unroll the loop body one step - make each trip do 2 iterations. 1770 void PhaseIdealLoop::do_unroll(IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip) { 1771 assert(LoopUnrollLimit, ""); 1772 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 1773 CountedLoopEndNode *loop_end = loop_head->loopexit(); 1774 #ifndef PRODUCT 1775 if (PrintOpto && VerifyLoopOptimizations) { 1776 tty->print("Unrolling "); 1777 loop->dump_head(); 1778 } else if (TraceLoopOpts) { 1779 if (loop_head->trip_count() < (uint)LoopUnrollLimit) { 1780 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); 1781 } else { 1782 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 1783 } 1784 loop->dump_head(); 1785 } 1786 1787 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) { 1788 Arena* arena = Thread::current()->resource_area(); 1789 Node_Stack stack(arena, C->live_nodes() >> 2); 1790 Node_List rpo_list; 1791 VectorSet visited(arena); 1792 visited.set(loop_head->_idx); 1793 rpo(loop_head, stack, visited, rpo_list); 1794 dump(loop, rpo_list.size(), rpo_list); 1795 } 1796 #endif 1797 1798 // Remember loop node count before unrolling to detect 1799 // if rounds of unroll,optimize are making progress 1800 loop_head->set_node_count_before_unroll(loop->_body.size()); 1801 1802 Node *ctrl = loop_head->skip_strip_mined()->in(LoopNode::EntryControl); 1803 Node *limit = loop_head->limit(); 1804 Node *init = loop_head->init_trip(); 1805 Node *stride = loop_head->stride(); 1806 1807 Node *opaq = NULL; 1808 if (adjust_min_trip) { // If not maximally unrolling, need adjustment 1809 // Search for zero-trip guard. 1810 1811 // Check the shape of the graph at the loop entry. If an inappropriate 1812 // graph shape is encountered, the compiler bails out loop unrolling; 1813 // compilation of the method will still succeed. 1814 if (!is_canonical_loop_entry(loop_head)) { 1815 return; 1816 } 1817 opaq = loop_head->skip_predicates()->in(0)->in(1)->in(1)->in(2); 1818 // Zero-trip test uses an 'opaque' node which is not shared. 1819 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); 1820 } 1821 1822 C->set_major_progress(); 1823 1824 Node* new_limit = NULL; 1825 int stride_con = stride->get_int(); 1826 int stride_p = (stride_con > 0) ? stride_con : -stride_con; 1827 uint old_trip_count = loop_head->trip_count(); 1828 // Verify that unroll policy result is still valid. 1829 assert(old_trip_count > 1 && 1830 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); 1831 1832 update_skeleton_predicates(ctrl, loop_head, init, stride_con); 1833 1834 // Adjust loop limit to keep valid iterations number after unroll. 1835 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride 1836 // which may overflow. 1837 if (!adjust_min_trip) { 1838 assert(old_trip_count > 1 && (old_trip_count & 1) == 0, 1839 "odd trip count for maximally unroll"); 1840 // Don't need to adjust limit for maximally unroll since trip count is even. 1841 } else if (loop_head->has_exact_trip_count() && init->is_Con()) { 1842 // Loop's limit is constant. Loop's init could be constant when pre-loop 1843 // become peeled iteration. 1844 jlong init_con = init->get_int(); 1845 // We can keep old loop limit if iterations count stays the same: 1846 // old_trip_count == new_trip_count * 2 1847 // Note: since old_trip_count >= 2 then new_trip_count >= 1 1848 // so we also don't need to adjust zero trip test. 1849 jlong limit_con = limit->get_int(); 1850 // (stride_con*2) not overflow since stride_con <= 8. 1851 int new_stride_con = stride_con * 2; 1852 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); 1853 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; 1854 // New trip count should satisfy next conditions. 1855 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); 1856 uint new_trip_count = (uint)trip_count; 1857 adjust_min_trip = (old_trip_count != new_trip_count*2); 1858 } 1859 1860 if (adjust_min_trip) { 1861 // Step 2: Adjust the trip limit if it is called for. 1862 // The adjustment amount is -stride. Need to make sure if the 1863 // adjustment underflows or overflows, then the main loop is skipped. 1864 Node* cmp = loop_end->cmp_node(); 1865 assert(cmp->in(2) == limit, "sanity"); 1866 assert(opaq != NULL && opaq->in(1) == limit, "sanity"); 1867 1868 // Verify that policy_unroll result is still valid. 1869 const TypeInt* limit_type = _igvn.type(limit)->is_int(); 1870 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || 1871 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), 1872 "sanity"); 1873 1874 if (limit->is_Con()) { 1875 // The check in policy_unroll and the assert above guarantee 1876 // no underflow if limit is constant. 1877 new_limit = _igvn.intcon(limit->get_int() - stride_con); 1878 set_ctrl(new_limit, C->root()); 1879 } else { 1880 // Limit is not constant. 1881 if (loop_head->unrolled_count() == 1) { // only for first unroll 1882 // Separate limit by Opaque node in case it is an incremented 1883 // variable from previous loop to avoid using pre-incremented 1884 // value which could increase register pressure. 1885 // Otherwise reorg_offsets() optimization will create a separate 1886 // Opaque node for each use of trip-counter and as result 1887 // zero trip guard limit will be different from loop limit. 1888 assert(has_ctrl(opaq), "should have it"); 1889 Node* opaq_ctrl = get_ctrl(opaq); 1890 limit = new Opaque2Node(C, limit); 1891 register_new_node(limit, opaq_ctrl); 1892 } 1893 if ((stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo)) || 1894 (stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi))) { 1895 // No underflow. 1896 new_limit = new SubINode(limit, stride); 1897 } else { 1898 // (limit - stride) may underflow. 1899 // Clamp the adjustment value with MININT or MAXINT: 1900 // 1901 // new_limit = limit-stride 1902 // if (stride > 0) 1903 // new_limit = (limit < new_limit) ? MININT : new_limit; 1904 // else 1905 // new_limit = (limit > new_limit) ? MAXINT : new_limit; 1906 // 1907 BoolTest::mask bt = loop_end->test_trip(); 1908 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 1909 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); 1910 set_ctrl(adj_max, C->root()); 1911 Node* old_limit = NULL; 1912 Node* adj_limit = NULL; 1913 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; 1914 if (loop_head->unrolled_count() > 1 && 1915 limit->is_CMove() && limit->Opcode() == Op_CMoveI && 1916 limit->in(CMoveNode::IfTrue) == adj_max && 1917 bol->as_Bool()->_test._test == bt && 1918 bol->in(1)->Opcode() == Op_CmpI && 1919 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { 1920 // Loop was unrolled before. 1921 // Optimize the limit to avoid nested CMove: 1922 // use original limit as old limit. 1923 old_limit = bol->in(1)->in(1); 1924 // Adjust previous adjusted limit. 1925 adj_limit = limit->in(CMoveNode::IfFalse); 1926 adj_limit = new SubINode(adj_limit, stride); 1927 } else { 1928 old_limit = limit; 1929 adj_limit = new SubINode(limit, stride); 1930 } 1931 assert(old_limit != NULL && adj_limit != NULL, ""); 1932 register_new_node(adj_limit, ctrl); // adjust amount 1933 Node* adj_cmp = new CmpINode(old_limit, adj_limit); 1934 register_new_node(adj_cmp, ctrl); 1935 Node* adj_bool = new BoolNode(adj_cmp, bt); 1936 register_new_node(adj_bool, ctrl); 1937 new_limit = new CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); 1938 } 1939 register_new_node(new_limit, ctrl); 1940 } 1941 1942 assert(new_limit != NULL, ""); 1943 // Replace in loop test. 1944 assert(loop_end->in(1)->in(1) == cmp, "sanity"); 1945 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { 1946 // Don't need to create new test since only one user. 1947 _igvn.hash_delete(cmp); 1948 cmp->set_req(2, new_limit); 1949 } else { 1950 // Create new test since it is shared. 1951 Node* ctrl2 = loop_end->in(0); 1952 Node* cmp2 = cmp->clone(); 1953 cmp2->set_req(2, new_limit); 1954 register_new_node(cmp2, ctrl2); 1955 Node* bol2 = loop_end->in(1)->clone(); 1956 bol2->set_req(1, cmp2); 1957 register_new_node(bol2, ctrl2); 1958 _igvn.replace_input_of(loop_end, 1, bol2); 1959 } 1960 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1961 // Make it a 1-trip test (means at least 2 trips). 1962 1963 // Guard test uses an 'opaque' node which is not shared. Hence I 1964 // can edit it's inputs directly. Hammer in the new limit for the 1965 // minimum-trip guard. 1966 assert(opaq->outcnt() == 1, ""); 1967 _igvn.replace_input_of(opaq, 1, new_limit); 1968 } 1969 1970 // Adjust max trip count. The trip count is intentionally rounded 1971 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 1972 // the main, unrolled, part of the loop will never execute as it is protected 1973 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 1974 // and later determined that part of the unrolled loop was dead. 1975 loop_head->set_trip_count(old_trip_count / 2); 1976 1977 // Double the count of original iterations in the unrolled loop body. 1978 loop_head->double_unrolled_count(); 1979 1980 // --------- 1981 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1982 // represents the odd iterations; since the loop trips an even number of 1983 // times its backedge is never taken. Kill the backedge. 1984 uint dd = dom_depth(loop_head); 1985 clone_loop(loop, old_new, dd, IgnoreStripMined); 1986 1987 // Make backedges of the clone equal to backedges of the original. 1988 // Make the fall-in from the original come from the fall-out of the clone. 1989 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1990 Node* phi = loop_head->fast_out(j); 1991 if (phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0) { 1992 Node *newphi = old_new[phi->_idx]; 1993 _igvn.hash_delete(phi); 1994 _igvn.hash_delete(newphi); 1995 1996 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1997 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1998 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1999 } 2000 } 2001 Node *clone_head = old_new[loop_head->_idx]; 2002 _igvn.hash_delete(clone_head); 2003 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 2004 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 2005 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 2006 loop->_head = clone_head; // New loop header 2007 2008 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 2009 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 2010 2011 // Kill the clone's backedge 2012 Node *newcle = old_new[loop_end->_idx]; 2013 _igvn.hash_delete(newcle); 2014 Node *one = _igvn.intcon(1); 2015 set_ctrl(one, C->root()); 2016 newcle->set_req(1, one); 2017 // Force clone into same loop body 2018 uint max = loop->_body.size(); 2019 for (uint k = 0; k < max; k++) { 2020 Node *old = loop->_body.at(k); 2021 Node *nnn = old_new[old->_idx]; 2022 loop->_body.push(nnn); 2023 if (!has_ctrl(old)) { 2024 set_loop(nnn, loop); 2025 } 2026 } 2027 2028 loop->record_for_igvn(); 2029 loop_head->clear_strip_mined(); 2030 2031 #ifndef PRODUCT 2032 if (C->do_vector_loop() && (PrintOpto && (VerifyLoopOptimizations || TraceLoopOpts))) { 2033 tty->print("\nnew loop after unroll\n"); loop->dump_head(); 2034 for (uint i = 0; i < loop->_body.size(); i++) { 2035 loop->_body.at(i)->dump(); 2036 } 2037 if (C->clone_map().is_debug()) { 2038 tty->print("\nCloneMap\n"); 2039 Dict* dict = C->clone_map().dict(); 2040 DictI i(dict); 2041 tty->print_cr("Dict@%p[%d] = ", dict, dict->Size()); 2042 for (int ii = 0; i.test(); ++i, ++ii) { 2043 NodeCloneInfo cl((uint64_t)dict->operator[]((void*)i._key)); 2044 tty->print("%d->%d:%d,", (int)(intptr_t)i._key, cl.idx(), cl.gen()); 2045 if (ii % 10 == 9) { 2046 tty->print_cr(" "); 2047 } 2048 } 2049 tty->print_cr(" "); 2050 } 2051 } 2052 #endif 2053 } 2054 2055 //------------------------------do_maximally_unroll---------------------------- 2056 2057 void PhaseIdealLoop::do_maximally_unroll(IdealLoopTree *loop, Node_List &old_new) { 2058 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2059 assert(cl->has_exact_trip_count(), "trip count is not exact"); 2060 assert(cl->trip_count() > 0, ""); 2061 #ifndef PRODUCT 2062 if (TraceLoopOpts) { 2063 tty->print("MaxUnroll %d ", cl->trip_count()); 2064 loop->dump_head(); 2065 } 2066 #endif 2067 2068 // If loop is tripping an odd number of times, peel odd iteration 2069 if ((cl->trip_count() & 1) == 1) { 2070 do_peeling(loop, old_new); 2071 } 2072 2073 // Now its tripping an even number of times remaining. Double loop body. 2074 // Do not adjust pre-guards; they are not needed and do not exist. 2075 if (cl->trip_count() > 0) { 2076 assert((cl->trip_count() & 1) == 0, "missed peeling"); 2077 do_unroll(loop, old_new, false); 2078 } 2079 } 2080 2081 void PhaseIdealLoop::mark_reductions(IdealLoopTree *loop) { 2082 if (SuperWordReductions == false) return; 2083 2084 CountedLoopNode* loop_head = loop->_head->as_CountedLoop(); 2085 if (loop_head->unrolled_count() > 1) { 2086 return; 2087 } 2088 2089 Node* trip_phi = loop_head->phi(); 2090 for (DUIterator_Fast imax, i = loop_head->fast_outs(imax); i < imax; i++) { 2091 Node* phi = loop_head->fast_out(i); 2092 if (phi->is_Phi() && phi->outcnt() > 0 && phi != trip_phi) { 2093 // For definitions which are loop inclusive and not tripcounts. 2094 Node* def_node = phi->in(LoopNode::LoopBackControl); 2095 2096 if (def_node != NULL) { 2097 Node* n_ctrl = get_ctrl(def_node); 2098 if (n_ctrl != NULL && loop->is_member(get_loop(n_ctrl))) { 2099 // Now test it to see if it fits the standard pattern for a reduction operator. 2100 int opc = def_node->Opcode(); 2101 if (opc != ReductionNode::opcode(opc, def_node->bottom_type()->basic_type()) 2102 || opc == Op_MinD || opc == Op_MinF || opc == Op_MaxD || opc == Op_MaxF) { 2103 if (!def_node->is_reduction()) { // Not marked yet 2104 // To be a reduction, the arithmetic node must have the phi as input and provide a def to it 2105 bool ok = false; 2106 for (unsigned j = 1; j < def_node->req(); j++) { 2107 Node* in = def_node->in(j); 2108 if (in == phi) { 2109 ok = true; 2110 break; 2111 } 2112 } 2113 2114 // do nothing if we did not match the initial criteria 2115 if (ok == false) { 2116 continue; 2117 } 2118 2119 // The result of the reduction must not be used in the loop 2120 for (DUIterator_Fast imax, i = def_node->fast_outs(imax); i < imax && ok; i++) { 2121 Node* u = def_node->fast_out(i); 2122 if (!loop->is_member(get_loop(ctrl_or_self(u)))) { 2123 continue; 2124 } 2125 if (u == phi) { 2126 continue; 2127 } 2128 ok = false; 2129 } 2130 2131 // iff the uses conform 2132 if (ok) { 2133 def_node->add_flag(Node::Flag_is_reduction); 2134 loop_head->mark_has_reductions(); 2135 } 2136 } 2137 } 2138 } 2139 } 2140 } 2141 } 2142 } 2143 2144 //------------------------------adjust_limit----------------------------------- 2145 // Helper function for add_constraint(). 2146 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up) { 2147 // Compute "I :: (limit-offset)/scale" 2148 Node *con = new SubINode(rc_limit, offset); 2149 register_new_node(con, pre_ctrl); 2150 Node *X = new DivINode(0, con, scale); 2151 register_new_node(X, pre_ctrl); 2152 2153 // When the absolute value of scale is greater than one, the integer 2154 // division may round limit down so add one to the limit. 2155 if (round_up) { 2156 X = new AddINode(X, _igvn.intcon(1)); 2157 register_new_node(X, pre_ctrl); 2158 } 2159 2160 // Adjust loop limit 2161 loop_limit = (stride_con > 0) 2162 ? (Node*)(new MinINode(loop_limit, X)) 2163 : (Node*)(new MaxINode(loop_limit, X)); 2164 register_new_node(loop_limit, pre_ctrl); 2165 return loop_limit; 2166 } 2167 2168 //------------------------------add_constraint--------------------------------- 2169 // Constrain the main loop iterations so the conditions: 2170 // low_limit <= scale_con * I + offset < upper_limit 2171 // always holds true. That is, either increase the number of iterations in 2172 // the pre-loop or the post-loop until the condition holds true in the main 2173 // loop. Stride, scale, offset and limit are all loop invariant. Further, 2174 // stride and scale are constants (offset and limit often are). 2175 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) { 2176 // For positive stride, the pre-loop limit always uses a MAX function 2177 // and the main loop a MIN function. For negative stride these are 2178 // reversed. 2179 2180 // Also for positive stride*scale the affine function is increasing, so the 2181 // pre-loop must check for underflow and the post-loop for overflow. 2182 // Negative stride*scale reverses this; pre-loop checks for overflow and 2183 // post-loop for underflow. 2184 2185 Node *scale = _igvn.intcon(scale_con); 2186 set_ctrl(scale, C->root()); 2187 2188 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow 2189 // The overflow limit: scale*I+offset < upper_limit 2190 // For main-loop compute 2191 // ( if (scale > 0) /* and stride > 0 */ 2192 // I < (upper_limit-offset)/scale 2193 // else /* scale < 0 and stride < 0 */ 2194 // I > (upper_limit-offset)/scale 2195 // ) 2196 // 2197 // (upper_limit-offset) may overflow or underflow. 2198 // But it is fine since main loop will either have 2199 // less iterations or will be skipped in such case. 2200 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl, false); 2201 2202 // The underflow limit: low_limit <= scale*I+offset. 2203 // For pre-loop compute 2204 // NOT(scale*I+offset >= low_limit) 2205 // scale*I+offset < low_limit 2206 // ( if (scale > 0) /* and stride > 0 */ 2207 // I < (low_limit-offset)/scale 2208 // else /* scale < 0 and stride < 0 */ 2209 // I > (low_limit-offset)/scale 2210 // ) 2211 2212 if (low_limit->get_int() == -max_jint) { 2213 // We need this guard when scale*pre_limit+offset >= limit 2214 // due to underflow. So we need execute pre-loop until 2215 // scale*I+offset >= min_int. But (min_int-offset) will 2216 // underflow when offset > 0 and X will be > original_limit 2217 // when stride > 0. To avoid it we replace positive offset with 0. 2218 // 2219 // Also (min_int+1 == -max_int) is used instead of min_int here 2220 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 2221 Node* shift = _igvn.intcon(31); 2222 set_ctrl(shift, C->root()); 2223 Node* sign = new RShiftINode(offset, shift); 2224 register_new_node(sign, pre_ctrl); 2225 offset = new AndINode(offset, sign); 2226 register_new_node(offset, pre_ctrl); 2227 } else { 2228 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 2229 // The only problem we have here when offset == min_int 2230 // since (0-min_int) == min_int. It may be fine for stride > 0 2231 // but for stride < 0 X will be < original_limit. To avoid it 2232 // max(pre_limit, original_limit) is used in do_range_check(). 2233 } 2234 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 2235 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl, 2236 scale_con > 1 && stride_con > 0); 2237 2238 } else { // stride_con*scale_con < 0 2239 // For negative stride*scale pre-loop checks for overflow and 2240 // post-loop for underflow. 2241 // 2242 // The overflow limit: scale*I+offset < upper_limit 2243 // For pre-loop compute 2244 // NOT(scale*I+offset < upper_limit) 2245 // scale*I+offset >= upper_limit 2246 // scale*I+offset+1 > upper_limit 2247 // ( if (scale < 0) /* and stride > 0 */ 2248 // I < (upper_limit-(offset+1))/scale 2249 // else /* scale > 0 and stride < 0 */ 2250 // I > (upper_limit-(offset+1))/scale 2251 // ) 2252 // 2253 // (upper_limit-offset-1) may underflow or overflow. 2254 // To avoid it min(pre_limit, original_limit) is used 2255 // in do_range_check() for stride > 0 and max() for < 0. 2256 Node *one = _igvn.intcon(1); 2257 set_ctrl(one, C->root()); 2258 2259 Node *plus_one = new AddINode(offset, one); 2260 register_new_node(plus_one, pre_ctrl); 2261 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); 2262 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl, 2263 scale_con < -1 && stride_con > 0); 2264 2265 if (low_limit->get_int() == -max_jint) { 2266 // We need this guard when scale*main_limit+offset >= limit 2267 // due to underflow. So we need execute main-loop while 2268 // scale*I+offset+1 > min_int. But (min_int-offset-1) will 2269 // underflow when (offset+1) > 0 and X will be < main_limit 2270 // when scale < 0 (and stride > 0). To avoid it we replace 2271 // positive (offset+1) with 0. 2272 // 2273 // Also (min_int+1 == -max_int) is used instead of min_int here 2274 // to avoid problem with scale == -1 (min_int/(-1) == min_int). 2275 Node* shift = _igvn.intcon(31); 2276 set_ctrl(shift, C->root()); 2277 Node* sign = new RShiftINode(plus_one, shift); 2278 register_new_node(sign, pre_ctrl); 2279 plus_one = new AndINode(plus_one, sign); 2280 register_new_node(plus_one, pre_ctrl); 2281 } else { 2282 assert(low_limit->get_int() == 0, "wrong low limit for range check"); 2283 // The only problem we have here when offset == max_int 2284 // since (max_int+1) == min_int and (0-min_int) == min_int. 2285 // But it is fine since main loop will either have 2286 // less iterations or will be skipped in such case. 2287 } 2288 // The underflow limit: low_limit <= scale*I+offset. 2289 // For main-loop compute 2290 // scale*I+offset+1 > low_limit 2291 // ( if (scale < 0) /* and stride > 0 */ 2292 // I < (low_limit-(offset+1))/scale 2293 // else /* scale > 0 and stride < 0 */ 2294 // I > (low_limit-(offset+1))/scale 2295 // ) 2296 2297 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl, 2298 false); 2299 } 2300 } 2301 2302 2303 //------------------------------is_scaled_iv--------------------------------- 2304 // Return true if exp is a constant times an induction var 2305 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 2306 if (exp == iv) { 2307 if (p_scale != NULL) { 2308 *p_scale = 1; 2309 } 2310 return true; 2311 } 2312 int opc = exp->Opcode(); 2313 if (opc == Op_MulI) { 2314 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 2315 if (p_scale != NULL) { 2316 *p_scale = exp->in(2)->get_int(); 2317 } 2318 return true; 2319 } 2320 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 2321 if (p_scale != NULL) { 2322 *p_scale = exp->in(1)->get_int(); 2323 } 2324 return true; 2325 } 2326 } else if (opc == Op_LShiftI) { 2327 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 2328 if (p_scale != NULL) { 2329 *p_scale = 1 << exp->in(2)->get_int(); 2330 } 2331 return true; 2332 } 2333 } 2334 return false; 2335 } 2336 2337 //-----------------------------is_scaled_iv_plus_offset------------------------------ 2338 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 2339 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 2340 if (is_scaled_iv(exp, iv, p_scale)) { 2341 if (p_offset != NULL) { 2342 Node *zero = _igvn.intcon(0); 2343 set_ctrl(zero, C->root()); 2344 *p_offset = zero; 2345 } 2346 return true; 2347 } 2348 int opc = exp->Opcode(); 2349 if (opc == Op_AddI) { 2350 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 2351 if (p_offset != NULL) { 2352 *p_offset = exp->in(2); 2353 } 2354 return true; 2355 } 2356 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 2357 if (p_offset != NULL) { 2358 *p_offset = exp->in(1); 2359 } 2360 return true; 2361 } 2362 if (exp->in(2)->is_Con()) { 2363 Node* offset2 = NULL; 2364 if (depth < 2 && 2365 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 2366 p_offset != NULL ? &offset2 : NULL, depth+1)) { 2367 if (p_offset != NULL) { 2368 Node *ctrl_off2 = get_ctrl(offset2); 2369 Node* offset = new AddINode(offset2, exp->in(2)); 2370 register_new_node(offset, ctrl_off2); 2371 *p_offset = offset; 2372 } 2373 return true; 2374 } 2375 } 2376 } else if (opc == Op_SubI) { 2377 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 2378 if (p_offset != NULL) { 2379 Node *zero = _igvn.intcon(0); 2380 set_ctrl(zero, C->root()); 2381 Node *ctrl_off = get_ctrl(exp->in(2)); 2382 Node* offset = new SubINode(zero, exp->in(2)); 2383 register_new_node(offset, ctrl_off); 2384 *p_offset = offset; 2385 } 2386 return true; 2387 } 2388 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 2389 if (p_offset != NULL) { 2390 *p_scale *= -1; 2391 *p_offset = exp->in(1); 2392 } 2393 return true; 2394 } 2395 } 2396 return false; 2397 } 2398 2399 // Same as PhaseIdealLoop::duplicate_predicates() but for range checks 2400 // eliminated by iteration splitting. 2401 Node* PhaseIdealLoop::add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl, 2402 Node* predicate_proj, int scale_con, Node* offset, 2403 Node* limit, jint stride_con, Node* value) { 2404 bool overflow = false; 2405 BoolNode* bol = rc_predicate(loop, predicate_proj, scale_con, offset, value, NULL, stride_con, limit, (stride_con > 0) != (scale_con > 0), overflow); 2406 Node* opaque_bol = new Opaque4Node(C, bol, _igvn.intcon(1)); 2407 register_new_node(opaque_bol, predicate_proj); 2408 IfNode* new_iff = NULL; 2409 if (overflow) { 2410 new_iff = new IfNode(predicate_proj, opaque_bol, PROB_MAX, COUNT_UNKNOWN); 2411 } else { 2412 new_iff = new RangeCheckNode(predicate_proj, opaque_bol, PROB_MAX, COUNT_UNKNOWN); 2413 } 2414 register_control(new_iff, loop->_parent, predicate_proj); 2415 Node* iffalse = new IfFalseNode(new_iff); 2416 register_control(iffalse, _ltree_root, new_iff); 2417 ProjNode* iftrue = new IfTrueNode(new_iff); 2418 register_control(iftrue, loop->_parent, new_iff); 2419 Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr); 2420 register_new_node(frame, C->start()); 2421 Node* halt = new HaltNode(iffalse, frame); 2422 register_control(halt, _ltree_root, iffalse); 2423 C->root()->add_req(halt); 2424 return iftrue; 2425 } 2426 2427 //------------------------------do_range_check--------------------------------- 2428 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 2429 int PhaseIdealLoop::do_range_check(IdealLoopTree *loop, Node_List &old_new) { 2430 #ifndef PRODUCT 2431 if (PrintOpto && VerifyLoopOptimizations) { 2432 tty->print("Range Check Elimination "); 2433 loop->dump_head(); 2434 } else if (TraceLoopOpts) { 2435 tty->print("RangeCheck "); 2436 loop->dump_head(); 2437 } 2438 #endif 2439 2440 assert(RangeCheckElimination, ""); 2441 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2442 // If we fail before trying to eliminate range checks, set multiversion state 2443 int closed_range_checks = 1; 2444 2445 // protect against stride not being a constant 2446 if (!cl->stride_is_con()) { 2447 return closed_range_checks; 2448 } 2449 // Find the trip counter; we are iteration splitting based on it 2450 Node *trip_counter = cl->phi(); 2451 // Find the main loop limit; we will trim it's iterations 2452 // to not ever trip end tests 2453 Node *main_limit = cl->limit(); 2454 2455 // Check graph shape. Cannot optimize a loop if zero-trip 2456 // Opaque1 node is optimized away and then another round 2457 // of loop opts attempted. 2458 if (!is_canonical_loop_entry(cl)) { 2459 return closed_range_checks; 2460 } 2461 2462 // Need to find the main-loop zero-trip guard 2463 Node *ctrl = cl->skip_predicates(); 2464 Node *iffm = ctrl->in(0); 2465 Node *opqzm = iffm->in(1)->in(1)->in(2); 2466 assert(opqzm->in(1) == main_limit, "do not understand situation"); 2467 2468 // Find the pre-loop limit; we will expand its iterations to 2469 // not ever trip low tests. 2470 Node *p_f = iffm->in(0); 2471 // pre loop may have been optimized out 2472 if (p_f->Opcode() != Op_IfFalse) { 2473 return closed_range_checks; 2474 } 2475 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2476 assert(pre_end->loopnode()->is_pre_loop(), ""); 2477 Node *pre_opaq1 = pre_end->limit(); 2478 // Occasionally it's possible for a pre-loop Opaque1 node to be 2479 // optimized away and then another round of loop opts attempted. 2480 // We can not optimize this particular loop in that case. 2481 if (pre_opaq1->Opcode() != Op_Opaque1) { 2482 return closed_range_checks; 2483 } 2484 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 2485 Node *pre_limit = pre_opaq->in(1); 2486 2487 // Where do we put new limit calculations 2488 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 2489 2490 // Ensure the original loop limit is available from the 2491 // pre-loop Opaque1 node. 2492 Node *orig_limit = pre_opaq->original_loop_limit(); 2493 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) { 2494 return closed_range_checks; 2495 } 2496 // Must know if its a count-up or count-down loop 2497 2498 int stride_con = cl->stride_con(); 2499 Node *zero = _igvn.intcon(0); 2500 Node *one = _igvn.intcon(1); 2501 // Use symmetrical int range [-max_jint,max_jint] 2502 Node *mini = _igvn.intcon(-max_jint); 2503 set_ctrl(zero, C->root()); 2504 set_ctrl(one, C->root()); 2505 set_ctrl(mini, C->root()); 2506 2507 // Range checks that do not dominate the loop backedge (ie. 2508 // conditionally executed) can lengthen the pre loop limit beyond 2509 // the original loop limit. To prevent this, the pre limit is 2510 // (for stride > 0) MINed with the original loop limit (MAXed 2511 // stride < 0) when some range_check (rc) is conditionally 2512 // executed. 2513 bool conditional_rc = false; 2514 2515 // Count number of range checks and reduce by load range limits, if zero, 2516 // the loop is in canonical form to multiversion. 2517 closed_range_checks = 0; 2518 2519 Node* predicate_proj = cl->skip_strip_mined()->in(LoopNode::EntryControl); 2520 assert(predicate_proj->is_Proj() && predicate_proj->in(0)->is_If(), "if projection only"); 2521 2522 // Check loop body for tests of trip-counter plus loop-invariant vs loop-variant. 2523 for (uint i = 0; i < loop->_body.size(); i++) { 2524 Node *iff = loop->_body[i]; 2525 if (iff->Opcode() == Op_If || 2526 iff->Opcode() == Op_RangeCheck) { // Test? 2527 // Test is an IfNode, has 2 projections. If BOTH are in the loop 2528 // we need loop unswitching instead of iteration splitting. 2529 closed_range_checks++; 2530 Node *exit = loop->is_loop_exit(iff); 2531 if (!exit) continue; 2532 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 2533 2534 // Get boolean condition to test 2535 Node *i1 = iff->in(1); 2536 if (!i1->is_Bool()) continue; 2537 BoolNode *bol = i1->as_Bool(); 2538 BoolTest b_test = bol->_test; 2539 // Flip sense of test if exit condition is flipped 2540 if (flip) { 2541 b_test = b_test.negate(); 2542 } 2543 // Get compare 2544 Node *cmp = bol->in(1); 2545 2546 // Look for trip_counter + offset vs limit 2547 Node *rc_exp = cmp->in(1); 2548 Node *limit = cmp->in(2); 2549 int scale_con= 1; // Assume trip counter not scaled 2550 2551 Node *limit_c = get_ctrl(limit); 2552 if (loop->is_member(get_loop(limit_c))) { 2553 // Compare might have operands swapped; commute them 2554 b_test = b_test.commute(); 2555 rc_exp = cmp->in(2); 2556 limit = cmp->in(1); 2557 limit_c = get_ctrl(limit); 2558 if (loop->is_member(get_loop(limit_c))) { 2559 continue; // Both inputs are loop varying; cannot RCE 2560 } 2561 } 2562 // Here we know 'limit' is loop invariant 2563 2564 // 'limit' maybe pinned below the zero trip test (probably from a 2565 // previous round of rce), in which case, it can't be used in the 2566 // zero trip test expression which must occur before the zero test's if. 2567 if (is_dominator(ctrl, limit_c)) { 2568 continue; // Don't rce this check but continue looking for other candidates. 2569 } 2570 2571 // Check for scaled induction variable plus an offset 2572 Node *offset = NULL; 2573 2574 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 2575 continue; 2576 } 2577 2578 Node *offset_c = get_ctrl(offset); 2579 if (loop->is_member(get_loop(offset_c))) { 2580 continue; // Offset is not really loop invariant 2581 } 2582 // Here we know 'offset' is loop invariant. 2583 2584 // As above for the 'limit', the 'offset' maybe pinned below the 2585 // zero trip test. 2586 if (is_dominator(ctrl, offset_c)) { 2587 continue; // Don't rce this check but continue looking for other candidates. 2588 } 2589 #ifdef ASSERT 2590 if (TraceRangeLimitCheck) { 2591 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); 2592 bol->dump(2); 2593 } 2594 #endif 2595 // At this point we have the expression as: 2596 // scale_con * trip_counter + offset :: limit 2597 // where scale_con, offset and limit are loop invariant. Trip_counter 2598 // monotonically increases by stride_con, a constant. Both (or either) 2599 // stride_con and scale_con can be negative which will flip about the 2600 // sense of the test. 2601 2602 // Adjust pre and main loop limits to guard the correct iteration set 2603 if (cmp->Opcode() == Op_CmpU) { // Unsigned compare is really 2 tests 2604 if (b_test._test == BoolTest::lt) { // Range checks always use lt 2605 // The underflow and overflow limits: 0 <= scale*I+offset < limit 2606 add_constraint(stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit); 2607 // (0-offset)/scale could be outside of loop iterations range. 2608 conditional_rc = true; 2609 Node* init = cl->init_trip(); 2610 Node* opaque_init = new Opaque1Node(C, init); 2611 register_new_node(opaque_init, predicate_proj); 2612 // template predicate so it can be updated on next unrolling 2613 predicate_proj = add_range_check_predicate(loop, cl, predicate_proj, scale_con, offset, limit, stride_con, opaque_init); 2614 assert(skeleton_predicate_has_opaque(predicate_proj->in(0)->as_If()), "unexpected"); 2615 // predicate on first value of first iteration 2616 predicate_proj = add_range_check_predicate(loop, cl, predicate_proj, scale_con, offset, limit, stride_con, init); 2617 assert(!skeleton_predicate_has_opaque(predicate_proj->in(0)->as_If()), "unexpected"); 2618 int init_inc = stride_con/cl->unrolled_count(); 2619 assert(init_inc != 0, "invalid loop increment"); 2620 Node* max_value = _igvn.intcon(stride_con - init_inc); 2621 max_value = new AddINode(init, max_value); 2622 register_new_node(max_value, predicate_proj); 2623 // predicate on last value of first iteration (in case unrolling has already happened) 2624 predicate_proj = add_range_check_predicate(loop, cl, predicate_proj, scale_con, offset, limit, stride_con, max_value); 2625 assert(!skeleton_predicate_has_opaque(predicate_proj->in(0)->as_If()), "unexpected"); 2626 } else { 2627 if (PrintOpto) { 2628 tty->print_cr("missed RCE opportunity"); 2629 } 2630 continue; // In release mode, ignore it 2631 } 2632 } else { // Otherwise work on normal compares 2633 switch(b_test._test) { 2634 case BoolTest::gt: 2635 // Fall into GE case 2636 case BoolTest::ge: 2637 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit 2638 scale_con = -scale_con; 2639 offset = new SubINode(zero, offset); 2640 register_new_node(offset, pre_ctrl); 2641 limit = new SubINode(zero, limit); 2642 register_new_node(limit, pre_ctrl); 2643 // Fall into LE case 2644 case BoolTest::le: 2645 if (b_test._test != BoolTest::gt) { 2646 // Convert X <= Y to X < Y+1 2647 limit = new AddINode(limit, one); 2648 register_new_node(limit, pre_ctrl); 2649 } 2650 // Fall into LT case 2651 case BoolTest::lt: 2652 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit 2653 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here 2654 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. 2655 add_constraint(stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit); 2656 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. 2657 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could 2658 // still be outside of loop range. 2659 conditional_rc = true; 2660 break; 2661 default: 2662 if (PrintOpto) { 2663 tty->print_cr("missed RCE opportunity"); 2664 } 2665 continue; // Unhandled case 2666 } 2667 } 2668 2669 // Kill the eliminated test 2670 C->set_major_progress(); 2671 Node *kill_con = _igvn.intcon(1-flip); 2672 set_ctrl(kill_con, C->root()); 2673 _igvn.replace_input_of(iff, 1, kill_con); 2674 // Find surviving projection 2675 assert(iff->is_If(), ""); 2676 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 2677 // Find loads off the surviving projection; remove their control edge 2678 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 2679 Node* cd = dp->fast_out(i); // Control-dependent node 2680 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop 2681 // Allow the load to float around in the loop, or before it 2682 // but NOT before the pre-loop. 2683 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 2684 --i; 2685 --imax; 2686 } 2687 } 2688 if (limit->Opcode() == Op_LoadRange) { 2689 closed_range_checks--; 2690 } 2691 } // End of is IF 2692 } 2693 if (predicate_proj != cl->skip_strip_mined()->in(LoopNode::EntryControl)) { 2694 _igvn.replace_input_of(cl->skip_strip_mined(), LoopNode::EntryControl, predicate_proj); 2695 set_idom(cl->skip_strip_mined(), predicate_proj, dom_depth(cl->skip_strip_mined())); 2696 } 2697 2698 // Update loop limits 2699 if (conditional_rc) { 2700 pre_limit = (stride_con > 0) ? (Node*)new MinINode(pre_limit, orig_limit) 2701 : (Node*)new MaxINode(pre_limit, orig_limit); 2702 register_new_node(pre_limit, pre_ctrl); 2703 } 2704 _igvn.replace_input_of(pre_opaq, 1, pre_limit); 2705 2706 // Note:: we are making the main loop limit no longer precise; 2707 // need to round up based on stride. 2708 cl->set_nonexact_trip_count(); 2709 Node *main_cle = cl->loopexit(); 2710 Node *main_bol = main_cle->in(1); 2711 // Hacking loop bounds; need private copies of exit test 2712 if (main_bol->outcnt() > 1) { // BoolNode shared? 2713 main_bol = main_bol->clone(); // Clone a private BoolNode 2714 register_new_node(main_bol, main_cle->in(0)); 2715 _igvn.replace_input_of(main_cle, 1, main_bol); 2716 } 2717 Node *main_cmp = main_bol->in(1); 2718 if (main_cmp->outcnt() > 1) { // CmpNode shared? 2719 main_cmp = main_cmp->clone(); // Clone a private CmpNode 2720 register_new_node(main_cmp, main_cle->in(0)); 2721 _igvn.replace_input_of(main_bol, 1, main_cmp); 2722 } 2723 // Hack the now-private loop bounds 2724 _igvn.replace_input_of(main_cmp, 2, main_limit); 2725 // The OpaqueNode is unshared by design 2726 assert(opqzm->outcnt() == 1, "cannot hack shared node"); 2727 _igvn.replace_input_of(opqzm, 1, main_limit); 2728 2729 return closed_range_checks; 2730 } 2731 2732 //------------------------------has_range_checks------------------------------- 2733 // Check to see if RCE cleaned the current loop of range-checks. 2734 void PhaseIdealLoop::has_range_checks(IdealLoopTree *loop) { 2735 assert(RangeCheckElimination, ""); 2736 2737 // skip if not a counted loop 2738 if (!loop->is_counted()) return; 2739 2740 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2741 2742 // skip this loop if it is already checked 2743 if (cl->has_been_range_checked()) return; 2744 2745 // Now check for existence of range checks 2746 for (uint i = 0; i < loop->_body.size(); i++) { 2747 Node *iff = loop->_body[i]; 2748 int iff_opc = iff->Opcode(); 2749 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) { 2750 cl->mark_has_range_checks(); 2751 break; 2752 } 2753 } 2754 cl->set_has_been_range_checked(); 2755 } 2756 2757 //-------------------------multi_version_post_loops---------------------------- 2758 // Check the range checks that remain, if simple, use the bounds to guard 2759 // which version to a post loop we execute, one with range checks or one without 2760 bool PhaseIdealLoop::multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop) { 2761 bool multi_version_succeeded = false; 2762 assert(RangeCheckElimination, ""); 2763 CountedLoopNode *legacy_cl = legacy_loop->_head->as_CountedLoop(); 2764 assert(legacy_cl->is_post_loop(), ""); 2765 2766 // Check for existence of range checks using the unique instance to make a guard with 2767 Unique_Node_List worklist; 2768 for (uint i = 0; i < legacy_loop->_body.size(); i++) { 2769 Node *iff = legacy_loop->_body[i]; 2770 int iff_opc = iff->Opcode(); 2771 if (iff_opc == Op_If || iff_opc == Op_RangeCheck) { 2772 worklist.push(iff); 2773 } 2774 } 2775 2776 // Find RCE'd post loop so that we can stage its guard. 2777 if (!is_canonical_loop_entry(legacy_cl)) return multi_version_succeeded; 2778 Node* ctrl = legacy_cl->in(LoopNode::EntryControl); 2779 Node* iffm = ctrl->in(0); 2780 2781 // Now we test that both the post loops are connected 2782 Node* post_loop_region = iffm->in(0); 2783 if (post_loop_region == NULL) return multi_version_succeeded; 2784 if (!post_loop_region->is_Region()) return multi_version_succeeded; 2785 Node* covering_region = post_loop_region->in(RegionNode::Control+1); 2786 if (covering_region == NULL) return multi_version_succeeded; 2787 if (!covering_region->is_Region()) return multi_version_succeeded; 2788 Node* p_f = covering_region->in(RegionNode::Control); 2789 if (p_f == NULL) return multi_version_succeeded; 2790 if (!p_f->is_IfFalse()) return multi_version_succeeded; 2791 if (!p_f->in(0)->is_CountedLoopEnd()) return multi_version_succeeded; 2792 CountedLoopEndNode* rce_loop_end = p_f->in(0)->as_CountedLoopEnd(); 2793 if (rce_loop_end == NULL) return multi_version_succeeded; 2794 CountedLoopNode* rce_cl = rce_loop_end->loopnode(); 2795 if (rce_cl == NULL || !rce_cl->is_post_loop()) return multi_version_succeeded; 2796 CountedLoopNode *known_rce_cl = rce_loop->_head->as_CountedLoop(); 2797 if (rce_cl != known_rce_cl) return multi_version_succeeded; 2798 2799 // Then we fetch the cover entry test 2800 ctrl = rce_cl->in(LoopNode::EntryControl); 2801 if (!ctrl->is_IfTrue() && !ctrl->is_IfFalse()) return multi_version_succeeded; 2802 2803 #ifndef PRODUCT 2804 if (TraceLoopOpts) { 2805 tty->print("PostMultiVersion\n"); 2806 rce_loop->dump_head(); 2807 legacy_loop->dump_head(); 2808 } 2809 #endif 2810 2811 // Now fetch the limit we want to compare against 2812 Node *limit = rce_cl->limit(); 2813 bool first_time = true; 2814 2815 // If we got this far, we identified the post loop which has been RCE'd and 2816 // we have a work list. Now we will try to transform the if guard to cause 2817 // the loop pair to be multi version executed with the determination left to runtime 2818 // or the optimizer if full information is known about the given arrays at compile time. 2819 Node *last_min = NULL; 2820 multi_version_succeeded = true; 2821 while (worklist.size()) { 2822 Node* rc_iffm = worklist.pop(); 2823 if (rc_iffm->is_If()) { 2824 Node *rc_bolzm = rc_iffm->in(1); 2825 if (rc_bolzm->is_Bool()) { 2826 Node *rc_cmpzm = rc_bolzm->in(1); 2827 if (rc_cmpzm->is_Cmp()) { 2828 Node *rc_left = rc_cmpzm->in(2); 2829 if (rc_left->Opcode() != Op_LoadRange) { 2830 multi_version_succeeded = false; 2831 break; 2832 } 2833 if (first_time) { 2834 last_min = rc_left; 2835 first_time = false; 2836 } else { 2837 Node *cur_min = new MinINode(last_min, rc_left); 2838 last_min = cur_min; 2839 _igvn.register_new_node_with_optimizer(last_min); 2840 } 2841 } 2842 } 2843 } 2844 } 2845 2846 // All we have to do is update the limit of the rce loop 2847 // with the min of our expression and the current limit. 2848 // We will use this expression to replace the current limit. 2849 if (last_min && multi_version_succeeded) { 2850 Node *cur_min = new MinINode(last_min, limit); 2851 _igvn.register_new_node_with_optimizer(cur_min); 2852 Node *cmp_node = rce_loop_end->cmp_node(); 2853 _igvn.replace_input_of(cmp_node, 2, cur_min); 2854 set_ctrl(cur_min, ctrl); 2855 set_loop(cur_min, rce_loop->_parent); 2856 2857 legacy_cl->mark_is_multiversioned(); 2858 rce_cl->mark_is_multiversioned(); 2859 multi_version_succeeded = true; 2860 2861 C->set_major_progress(); 2862 } 2863 2864 return multi_version_succeeded; 2865 } 2866 2867 //-------------------------poison_rce_post_loop-------------------------------- 2868 // Causes the rce'd post loop to be optimized away if multiversioning fails 2869 void PhaseIdealLoop::poison_rce_post_loop(IdealLoopTree *rce_loop) { 2870 CountedLoopNode *rce_cl = rce_loop->_head->as_CountedLoop(); 2871 Node* ctrl = rce_cl->in(LoopNode::EntryControl); 2872 if (ctrl->is_IfTrue() || ctrl->is_IfFalse()) { 2873 Node* iffm = ctrl->in(0); 2874 if (iffm->is_If()) { 2875 Node* cur_bool = iffm->in(1); 2876 if (cur_bool->is_Bool()) { 2877 Node* cur_cmp = cur_bool->in(1); 2878 if (cur_cmp->is_Cmp()) { 2879 BoolTest::mask new_test = BoolTest::gt; 2880 BoolNode *new_bool = new BoolNode(cur_cmp, new_test); 2881 _igvn.replace_node(cur_bool, new_bool); 2882 _igvn._worklist.push(new_bool); 2883 Node* left_op = cur_cmp->in(1); 2884 _igvn.replace_input_of(cur_cmp, 2, left_op); 2885 C->set_major_progress(); 2886 } 2887 } 2888 } 2889 } 2890 } 2891 2892 //------------------------------DCE_loop_body---------------------------------- 2893 // Remove simplistic dead code from loop body 2894 void IdealLoopTree::DCE_loop_body() { 2895 for (uint i = 0; i < _body.size(); i++) { 2896 if (_body.at(i)->outcnt() == 0) { 2897 _body.map(i, _body.pop()); 2898 i--; // Ensure we revisit the updated index. 2899 } 2900 } 2901 } 2902 2903 2904 //------------------------------adjust_loop_exit_prob-------------------------- 2905 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 2906 // Replace with a 1-in-10 exit guess. 2907 void IdealLoopTree::adjust_loop_exit_prob(PhaseIdealLoop *phase) { 2908 Node *test = tail(); 2909 while (test != _head) { 2910 uint top = test->Opcode(); 2911 if (top == Op_IfTrue || top == Op_IfFalse) { 2912 int test_con = ((ProjNode*)test)->_con; 2913 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 2914 IfNode *iff = test->in(0)->as_If(); 2915 if (iff->outcnt() == 2) { // Ignore dead tests 2916 Node *bol = iff->in(1); 2917 if (bol && bol->req() > 1 && bol->in(1) && 2918 ((bol->in(1)->Opcode() == Op_StorePConditional) || 2919 (bol->in(1)->Opcode() == Op_StoreIConditional) || 2920 (bol->in(1)->Opcode() == Op_StoreLConditional) || 2921 (bol->in(1)->Opcode() == Op_CompareAndExchangeB) || 2922 (bol->in(1)->Opcode() == Op_CompareAndExchangeS) || 2923 (bol->in(1)->Opcode() == Op_CompareAndExchangeI) || 2924 (bol->in(1)->Opcode() == Op_CompareAndExchangeL) || 2925 (bol->in(1)->Opcode() == Op_CompareAndExchangeP) || 2926 (bol->in(1)->Opcode() == Op_CompareAndExchangeN) || 2927 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapB) || 2928 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapS) || 2929 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapI) || 2930 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapL) || 2931 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapP) || 2932 (bol->in(1)->Opcode() == Op_WeakCompareAndSwapN) || 2933 (bol->in(1)->Opcode() == Op_CompareAndSwapB) || 2934 (bol->in(1)->Opcode() == Op_CompareAndSwapS) || 2935 (bol->in(1)->Opcode() == Op_CompareAndSwapI) || 2936 (bol->in(1)->Opcode() == Op_CompareAndSwapL) || 2937 (bol->in(1)->Opcode() == Op_CompareAndSwapP) || 2938 (bol->in(1)->Opcode() == Op_CompareAndSwapN) || 2939 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndExchangeP) || 2940 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndExchangeN) || 2941 (bol->in(1)->Opcode() == Op_ShenandoahWeakCompareAndSwapP) || 2942 (bol->in(1)->Opcode() == Op_ShenandoahWeakCompareAndSwapN) || 2943 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndSwapP) || 2944 (bol->in(1)->Opcode() == Op_ShenandoahCompareAndSwapN))) 2945 return; // Allocation loops RARELY take backedge 2946 // Find the OTHER exit path from the IF 2947 Node* ex = iff->proj_out(1-test_con); 2948 float p = iff->_prob; 2949 if (!phase->is_member(this, ex) && iff->_fcnt == COUNT_UNKNOWN) { 2950 if (top == Op_IfTrue) { 2951 if (p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 2952 iff->_prob = PROB_STATIC_FREQUENT; 2953 } 2954 } else { 2955 if (p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 2956 iff->_prob = PROB_STATIC_INFREQUENT; 2957 } 2958 } 2959 } 2960 } 2961 } 2962 test = phase->idom(test); 2963 } 2964 } 2965 2966 #ifdef ASSERT 2967 static CountedLoopNode* locate_pre_from_main(CountedLoopNode *cl) { 2968 Node *ctrl = cl->skip_predicates(); 2969 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 2970 Node *iffm = ctrl->in(0); 2971 assert(iffm->Opcode() == Op_If, ""); 2972 Node *p_f = iffm->in(0); 2973 assert(p_f->Opcode() == Op_IfFalse, ""); 2974 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 2975 assert(pre_end->loopnode()->is_pre_loop(), ""); 2976 return pre_end->loopnode(); 2977 } 2978 #endif 2979 2980 // Remove the main and post loops and make the pre loop execute all 2981 // iterations. Useful when the pre loop is found empty. 2982 void IdealLoopTree::remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase) { 2983 CountedLoopEndNode* pre_end = cl->loopexit(); 2984 Node* pre_cmp = pre_end->cmp_node(); 2985 if (pre_cmp->in(2)->Opcode() != Op_Opaque1) { 2986 // Only safe to remove the main loop if the compiler optimized it 2987 // out based on an unknown number of iterations 2988 return; 2989 } 2990 2991 // Can we find the main loop? 2992 if (_next == NULL) { 2993 return; 2994 } 2995 2996 Node* next_head = _next->_head; 2997 if (!next_head->is_CountedLoop()) { 2998 return; 2999 } 3000 3001 CountedLoopNode* main_head = next_head->as_CountedLoop(); 3002 if (!main_head->is_main_loop()) { 3003 return; 3004 } 3005 3006 assert(locate_pre_from_main(main_head) == cl, "bad main loop"); 3007 Node* main_iff = main_head->skip_predicates()->in(0); 3008 3009 // Remove the Opaque1Node of the pre loop and make it execute all iterations 3010 phase->_igvn.replace_input_of(pre_cmp, 2, pre_cmp->in(2)->in(2)); 3011 // Remove the Opaque1Node of the main loop so it can be optimized out 3012 Node* main_cmp = main_iff->in(1)->in(1); 3013 assert(main_cmp->in(2)->Opcode() == Op_Opaque1, "main loop has no opaque node?"); 3014 phase->_igvn.replace_input_of(main_cmp, 2, main_cmp->in(2)->in(1)); 3015 } 3016 3017 //------------------------------do_remove_empty_loop--------------------------- 3018 // We always attempt remove empty loops. The approach is to replace the trip 3019 // counter with the value it will have on the last iteration. This will break 3020 // the loop. 3021 bool IdealLoopTree::do_remove_empty_loop(PhaseIdealLoop *phase) { 3022 // Minimum size must be empty loop 3023 if (_body.size() > EMPTY_LOOP_SIZE) { 3024 return false; 3025 } 3026 if (!_head->is_CountedLoop()) { 3027 return false; // Dead loop 3028 } 3029 CountedLoopNode *cl = _head->as_CountedLoop(); 3030 if (!cl->is_valid_counted_loop()) { 3031 return false; // Malformed loop 3032 } 3033 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) { 3034 return false; // Infinite loop 3035 } 3036 if (cl->is_pre_loop()) { 3037 // If the loop we are removing is a pre-loop then the main and post loop 3038 // can be removed as well. 3039 remove_main_post_loops(cl, phase); 3040 } 3041 3042 #ifdef ASSERT 3043 // Ensure only one phi which is the iv. 3044 Node* iv = NULL; 3045 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 3046 Node* n = cl->fast_out(i); 3047 if (n->Opcode() == Op_Phi) { 3048 assert(iv == NULL, "Too many phis"); 3049 iv = n; 3050 } 3051 } 3052 assert(iv == cl->phi(), "Wrong phi"); 3053 #endif 3054 3055 // main and post loops have explicitly created zero trip guard 3056 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 3057 if (needs_guard) { 3058 // Skip guard if values not overlap. 3059 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); 3060 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); 3061 int stride_con = cl->stride_con(); 3062 if (stride_con > 0) { 3063 needs_guard = (init_t->_hi >= limit_t->_lo); 3064 } else { 3065 needs_guard = (init_t->_lo <= limit_t->_hi); 3066 } 3067 } 3068 if (needs_guard) { 3069 // Check for an obvious zero trip guard. 3070 Node* inctrl = PhaseIdealLoop::skip_all_loop_predicates(cl->skip_predicates()); 3071 if (inctrl->Opcode() == Op_IfTrue || inctrl->Opcode() == Op_IfFalse) { 3072 bool maybe_swapped = (inctrl->Opcode() == Op_IfFalse); 3073 // The test should look like just the backedge of a CountedLoop 3074 Node* iff = inctrl->in(0); 3075 if (iff->is_If()) { 3076 Node* bol = iff->in(1); 3077 if (bol->is_Bool()) { 3078 BoolTest test = bol->as_Bool()->_test; 3079 if (maybe_swapped) { 3080 test._test = test.commute(); 3081 test._test = test.negate(); 3082 } 3083 if (test._test == cl->loopexit()->test_trip()) { 3084 Node* cmp = bol->in(1); 3085 int init_idx = maybe_swapped ? 2 : 1; 3086 int limit_idx = maybe_swapped ? 1 : 2; 3087 if (cmp->is_Cmp() && cmp->in(init_idx) == cl->init_trip() && cmp->in(limit_idx) == cl->limit()) { 3088 needs_guard = false; 3089 } 3090 } 3091 } 3092 } 3093 } 3094 } 3095 3096 #ifndef PRODUCT 3097 if (PrintOpto) { 3098 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 3099 this->dump_head(); 3100 } else if (TraceLoopOpts) { 3101 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 3102 this->dump_head(); 3103 } 3104 #endif 3105 3106 if (needs_guard) { 3107 // Peel the loop to ensure there's a zero trip guard 3108 Node_List old_new; 3109 phase->do_peeling(this, old_new); 3110 } 3111 3112 // Replace the phi at loop head with the final value of the last 3113 // iteration. Then the CountedLoopEnd will collapse (backedge never 3114 // taken) and all loop-invariant uses of the exit values will be correct. 3115 Node *phi = cl->phi(); 3116 Node *exact_limit = phase->exact_limit(this); 3117 if (exact_limit != cl->limit()) { 3118 // We also need to replace the original limit to collapse loop exit. 3119 Node* cmp = cl->loopexit()->cmp_node(); 3120 assert(cl->limit() == cmp->in(2), "sanity"); 3121 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist 3122 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist 3123 } 3124 // Note: the final value after increment should not overflow since 3125 // counted loop has limit check predicate. 3126 Node *final = new SubINode(exact_limit, cl->stride()); 3127 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 3128 phase->_igvn.replace_node(phi,final); 3129 phase->C->set_major_progress(); 3130 return true; 3131 } 3132 3133 //------------------------------do_one_iteration_loop-------------------------- 3134 // Convert one iteration loop into normal code. 3135 bool IdealLoopTree::do_one_iteration_loop(PhaseIdealLoop *phase) { 3136 if (!_head->as_Loop()->is_valid_counted_loop()) { 3137 return false; // Only for counted loop 3138 } 3139 CountedLoopNode *cl = _head->as_CountedLoop(); 3140 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { 3141 return false; 3142 } 3143 3144 #ifndef PRODUCT 3145 if (TraceLoopOpts) { 3146 tty->print("OneIteration "); 3147 this->dump_head(); 3148 } 3149 #endif 3150 3151 Node *init_n = cl->init_trip(); 3152 #ifdef ASSERT 3153 // Loop boundaries should be constant since trip count is exact. 3154 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); 3155 #endif 3156 // Replace the phi at loop head with the value of the init_trip. 3157 // Then the CountedLoopEnd will collapse (backedge will not be taken) 3158 // and all loop-invariant uses of the exit values will be correct. 3159 phase->_igvn.replace_node(cl->phi(), cl->init_trip()); 3160 phase->C->set_major_progress(); 3161 return true; 3162 } 3163 3164 //============================================================================= 3165 //------------------------------iteration_split_impl--------------------------- 3166 bool IdealLoopTree::iteration_split_impl(PhaseIdealLoop *phase, Node_List &old_new) { 3167 // Compute loop trip count if possible. 3168 compute_trip_count(phase); 3169 3170 // Convert one iteration loop into normal code. 3171 if (do_one_iteration_loop(phase)) { 3172 return true; 3173 } 3174 // Check and remove empty loops (spam micro-benchmarks) 3175 if (do_remove_empty_loop(phase)) { 3176 return true; // Here we removed an empty loop 3177 } 3178 3179 AutoNodeBudget node_budget(phase); 3180 3181 bool should_peel = policy_peeling(phase); 3182 bool should_unswitch = policy_unswitching(phase); 3183 3184 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 3185 // This removes loop-invariant tests (usually null checks). 3186 if (!_head->is_CountedLoop()) { // Non-counted loop 3187 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 3188 // Partial peel succeeded so terminate this round of loop opts 3189 return false; 3190 } 3191 if (should_peel) { // Should we peel? 3192 if (PrintOpto) { tty->print_cr("should_peel"); } 3193 phase->do_peeling(this,old_new); 3194 } else if (should_unswitch) { 3195 phase->do_unswitching(this, old_new); 3196 } 3197 return true; 3198 } 3199 CountedLoopNode *cl = _head->as_CountedLoop(); 3200 3201 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops 3202 3203 // Do nothing special to pre- and post- loops 3204 if (cl->is_pre_loop() || cl->is_post_loop()) return true; 3205 3206 // Compute loop trip count from profile data 3207 compute_profile_trip_cnt(phase); 3208 3209 // Before attempting fancy unrolling, RCE or alignment, see if we want 3210 // to completely unroll this loop or do loop unswitching. 3211 if (cl->is_normal_loop()) { 3212 if (should_unswitch) { 3213 phase->do_unswitching(this, old_new); 3214 return true; 3215 } 3216 bool should_maximally_unroll = policy_maximally_unroll(phase); 3217 if (should_maximally_unroll) { 3218 // Here we did some unrolling and peeling. Eventually we will 3219 // completely unroll this loop and it will no longer be a loop. 3220 phase->do_maximally_unroll(this, old_new); 3221 return true; 3222 } 3223 } 3224 3225 // Counted loops may be peeled, may need some iterations run up 3226 // front for RCE, and may want to align loop refs to a cache 3227 // line. Thus we clone a full loop up front whose trip count is 3228 // at least 1 (if peeling), but may be several more. 3229 3230 // The main loop will start cache-line aligned with at least 1 3231 // iteration of the unrolled body (zero-trip test required) and 3232 // will have some range checks removed. 3233 3234 // A post-loop will finish any odd iterations (leftover after 3235 // unrolling), plus any needed for RCE purposes. 3236 3237 bool should_unroll = policy_unroll(phase); 3238 bool should_rce = policy_range_check(phase); 3239 // TODO: Remove align -- not used. 3240 bool should_align = policy_align(phase); 3241 3242 // If not RCE'ing (iteration splitting) or Aligning, then we do not need a 3243 // pre-loop. We may still need to peel an initial iteration but we will not 3244 // be needing an unknown number of pre-iterations. 3245 // 3246 // Basically, if may_rce_align reports FALSE first time through, we will not 3247 // be able to later do RCE or Aligning on this loop. 3248 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 3249 3250 // If we have any of these conditions (RCE, alignment, unrolling) met, then 3251 // we switch to the pre-/main-/post-loop model. This model also covers 3252 // peeling. 3253 if (should_rce || should_align || should_unroll) { 3254 if (cl->is_normal_loop()) { // Convert to 'pre/main/post' loops 3255 if (!phase->may_require_nodes(est_loop_clone_sz(3, _body.size()))) { 3256 return false; 3257 } 3258 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 3259 } 3260 // Adjust the pre- and main-loop limits to let the pre and post loops run 3261 // with full checks, but the main-loop with no checks. Remove said 3262 // checks from the main body. 3263 if (should_rce) { 3264 if (phase->do_range_check(this, old_new) != 0) { 3265 cl->mark_has_range_checks(); 3266 } 3267 } else if (PostLoopMultiversioning) { 3268 phase->has_range_checks(this); 3269 } 3270 3271 if (should_unroll && !should_peel && PostLoopMultiversioning) { 3272 // Try to setup multiversioning on main loops before they are unrolled 3273 if (cl->is_main_loop() && (cl->unrolled_count() == 1)) { 3274 phase->insert_scalar_rced_post_loop(this, old_new); 3275 } 3276 } 3277 3278 // Double loop body for unrolling. Adjust the minimum-trip test (will do 3279 // twice as many iterations as before) and the main body limit (only do 3280 // an even number of trips). If we are peeling, we might enable some RCE 3281 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 3282 // peeling. 3283 if (should_unroll && !should_peel) { 3284 if (SuperWordLoopUnrollAnalysis) { 3285 phase->insert_vector_post_loop(this, old_new); 3286 } 3287 phase->do_unroll(this, old_new, true); 3288 } 3289 3290 // Adjust the pre-loop limits to align the main body iterations. 3291 if (should_align) { 3292 Unimplemented(); 3293 } 3294 } else { // Else we have an unchanged counted loop 3295 if (should_peel) { // Might want to peel but do nothing else 3296 phase->do_peeling(this,old_new); 3297 } 3298 } 3299 return true; 3300 } 3301 3302 3303 //============================================================================= 3304 //------------------------------iteration_split-------------------------------- 3305 bool IdealLoopTree::iteration_split(PhaseIdealLoop* phase, Node_List &old_new) { 3306 // Recursively iteration split nested loops 3307 if (_child && !_child->iteration_split(phase, old_new)) { 3308 return false; 3309 } 3310 3311 // Clean out prior deadwood 3312 DCE_loop_body(); 3313 3314 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 3315 // Replace with a 1-in-10 exit guess. 3316 if (!is_root() && is_loop()) { 3317 adjust_loop_exit_prob(phase); 3318 } 3319 3320 // Unrolling, RCE and peeling efforts, iff innermost loop. 3321 if (_allow_optimizations && is_innermost()) { 3322 if (!_has_call) { 3323 if (!iteration_split_impl(phase, old_new)) { 3324 return false; 3325 } 3326 } else { 3327 AutoNodeBudget node_budget(phase); 3328 if (policy_unswitching(phase)) { 3329 phase->do_unswitching(this, old_new); 3330 } 3331 } 3332 } 3333 3334 // Minor offset re-organization to remove loop-fallout uses of 3335 // trip counter when there was no major reshaping. 3336 phase->reorg_offsets(this); 3337 3338 if (_next && !_next->iteration_split(phase, old_new)) { 3339 return false; 3340 } 3341 return true; 3342 } 3343 3344 3345 //============================================================================= 3346 // Process all the loops in the loop tree and replace any fill 3347 // patterns with an intrinsic version. 3348 bool PhaseIdealLoop::do_intrinsify_fill() { 3349 bool changed = false; 3350 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 3351 IdealLoopTree* lpt = iter.current(); 3352 changed |= intrinsify_fill(lpt); 3353 } 3354 return changed; 3355 } 3356 3357 3358 // Examine an inner loop looking for a a single store of an invariant 3359 // value in a unit stride loop, 3360 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 3361 Node*& shift, Node*& con) { 3362 const char* msg = NULL; 3363 Node* msg_node = NULL; 3364 3365 store_value = NULL; 3366 con = NULL; 3367 shift = NULL; 3368 3369 // Process the loop looking for stores. If there are multiple 3370 // stores or extra control flow give at this point. 3371 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 3372 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 3373 Node* n = lpt->_body.at(i); 3374 if (n->outcnt() == 0) continue; // Ignore dead 3375 if (n->is_Store()) { 3376 if (store != NULL) { 3377 msg = "multiple stores"; 3378 break; 3379 } 3380 int opc = n->Opcode(); 3381 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) { 3382 msg = "oop fills not handled"; 3383 break; 3384 } 3385 Node* value = n->in(MemNode::ValueIn); 3386 if (!lpt->is_invariant(value)) { 3387 msg = "variant store value"; 3388 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 3389 msg = "not array address"; 3390 } 3391 store = n; 3392 store_value = value; 3393 } else if (n->is_If() && n != head->loopexit_or_null()) { 3394 msg = "extra control flow"; 3395 msg_node = n; 3396 } 3397 } 3398 3399 if (store == NULL) { 3400 // No store in loop 3401 return false; 3402 } 3403 3404 if (msg == NULL && head->stride_con() != 1) { 3405 // could handle negative strides too 3406 if (head->stride_con() < 0) { 3407 msg = "negative stride"; 3408 } else { 3409 msg = "non-unit stride"; 3410 } 3411 } 3412 3413 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 3414 msg = "can't handle store address"; 3415 msg_node = store->in(MemNode::Address); 3416 } 3417 3418 if (msg == NULL && 3419 (!store->in(MemNode::Memory)->is_Phi() || 3420 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 3421 msg = "store memory isn't proper phi"; 3422 msg_node = store->in(MemNode::Memory); 3423 } 3424 3425 // Make sure there is an appropriate fill routine 3426 BasicType t = store->as_Mem()->memory_type(); 3427 const char* fill_name; 3428 if (msg == NULL && 3429 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 3430 msg = "unsupported store"; 3431 msg_node = store; 3432 } 3433 3434 if (msg != NULL) { 3435 #ifndef PRODUCT 3436 if (TraceOptimizeFill) { 3437 tty->print_cr("not fill intrinsic candidate: %s", msg); 3438 if (msg_node != NULL) msg_node->dump(); 3439 } 3440 #endif 3441 return false; 3442 } 3443 3444 // Make sure the address expression can be handled. It should be 3445 // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()). 3446 Node* elements[4]; 3447 Node* cast = NULL; 3448 Node* conv = NULL; 3449 bool found_index = false; 3450 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 3451 for (int e = 0; e < count; e++) { 3452 Node* n = elements[e]; 3453 if (n->is_Con() && con == NULL) { 3454 con = n; 3455 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 3456 Node* value = n->in(1); 3457 #ifdef _LP64 3458 if (value->Opcode() == Op_ConvI2L) { 3459 conv = value; 3460 value = value->in(1); 3461 } 3462 if (value->Opcode() == Op_CastII && 3463 value->as_CastII()->has_range_check()) { 3464 // Skip range check dependent CastII nodes 3465 cast = value; 3466 value = value->in(1); 3467 } 3468 #endif 3469 if (value != head->phi()) { 3470 msg = "unhandled shift in address"; 3471 } else { 3472 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { 3473 msg = "scale doesn't match"; 3474 } else { 3475 found_index = true; 3476 shift = n; 3477 } 3478 } 3479 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 3480 conv = n; 3481 n = n->in(1); 3482 if (n->Opcode() == Op_CastII && 3483 n->as_CastII()->has_range_check()) { 3484 // Skip range check dependent CastII nodes 3485 cast = n; 3486 n = n->in(1); 3487 } 3488 if (n == head->phi()) { 3489 found_index = true; 3490 } else { 3491 msg = "unhandled input to ConvI2L"; 3492 } 3493 } else if (n == head->phi()) { 3494 // no shift, check below for allowed cases 3495 found_index = true; 3496 } else { 3497 msg = "unhandled node in address"; 3498 msg_node = n; 3499 } 3500 } 3501 3502 if (count == -1) { 3503 msg = "malformed address expression"; 3504 msg_node = store; 3505 } 3506 3507 if (!found_index) { 3508 msg = "missing use of index"; 3509 } 3510 3511 // byte sized items won't have a shift 3512 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 3513 msg = "can't find shift"; 3514 msg_node = store; 3515 } 3516 3517 if (msg != NULL) { 3518 #ifndef PRODUCT 3519 if (TraceOptimizeFill) { 3520 tty->print_cr("not fill intrinsic: %s", msg); 3521 if (msg_node != NULL) msg_node->dump(); 3522 } 3523 #endif 3524 return false; 3525 } 3526 3527 // No make sure all the other nodes in the loop can be handled 3528 VectorSet ok(Thread::current()->resource_area()); 3529 3530 // store related values are ok 3531 ok.set(store->_idx); 3532 ok.set(store->in(MemNode::Memory)->_idx); 3533 3534 CountedLoopEndNode* loop_exit = head->loopexit(); 3535 3536 // Loop structure is ok 3537 ok.set(head->_idx); 3538 ok.set(loop_exit->_idx); 3539 ok.set(head->phi()->_idx); 3540 ok.set(head->incr()->_idx); 3541 ok.set(loop_exit->cmp_node()->_idx); 3542 ok.set(loop_exit->in(1)->_idx); 3543 3544 // Address elements are ok 3545 if (con) ok.set(con->_idx); 3546 if (shift) ok.set(shift->_idx); 3547 if (cast) ok.set(cast->_idx); 3548 if (conv) ok.set(conv->_idx); 3549 3550 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 3551 Node* n = lpt->_body.at(i); 3552 if (n->outcnt() == 0) continue; // Ignore dead 3553 if (ok.test(n->_idx)) continue; 3554 // Backedge projection is ok 3555 if (n->is_IfTrue() && n->in(0) == loop_exit) continue; 3556 if (!n->is_AddP()) { 3557 msg = "unhandled node"; 3558 msg_node = n; 3559 break; 3560 } 3561 } 3562 3563 // Make sure no unexpected values are used outside the loop 3564 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 3565 Node* n = lpt->_body.at(i); 3566 // These values can be replaced with other nodes if they are used 3567 // outside the loop. 3568 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue; 3569 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 3570 Node* use = iter.get(); 3571 if (!lpt->_body.contains(use)) { 3572 msg = "node is used outside loop"; 3573 // lpt->_body.dump(); 3574 msg_node = n; 3575 break; 3576 } 3577 } 3578 } 3579 3580 #ifdef ASSERT 3581 if (TraceOptimizeFill) { 3582 if (msg != NULL) { 3583 tty->print_cr("no fill intrinsic: %s", msg); 3584 if (msg_node != NULL) msg_node->dump(); 3585 } else { 3586 tty->print_cr("fill intrinsic for:"); 3587 } 3588 store->dump(); 3589 if (Verbose) { 3590 lpt->_body.dump(); 3591 } 3592 } 3593 #endif 3594 3595 return msg == NULL; 3596 } 3597 3598 3599 3600 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 3601 // Only for counted inner loops 3602 if (!lpt->is_counted() || !lpt->is_innermost()) { 3603 return false; 3604 } 3605 3606 // Must have constant stride 3607 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 3608 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) { 3609 return false; 3610 } 3611 3612 head->verify_strip_mined(1); 3613 3614 // Check that the body only contains a store of a loop invariant 3615 // value that is indexed by the loop phi. 3616 Node* store = NULL; 3617 Node* store_value = NULL; 3618 Node* shift = NULL; 3619 Node* offset = NULL; 3620 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 3621 return false; 3622 } 3623 3624 Node* exit = head->loopexit()->proj_out_or_null(0); 3625 if (exit == NULL) { 3626 return false; 3627 } 3628 3629 #ifndef PRODUCT 3630 if (TraceLoopOpts) { 3631 tty->print("ArrayFill "); 3632 lpt->dump_head(); 3633 } 3634 #endif 3635 3636 // Now replace the whole loop body by a call to a fill routine that 3637 // covers the same region as the loop. 3638 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 3639 3640 // Build an expression for the beginning of the copy region 3641 Node* index = head->init_trip(); 3642 #ifdef _LP64 3643 index = new ConvI2LNode(index); 3644 _igvn.register_new_node_with_optimizer(index); 3645 #endif 3646 if (shift != NULL) { 3647 // byte arrays don't require a shift but others do. 3648 index = new LShiftXNode(index, shift->in(2)); 3649 _igvn.register_new_node_with_optimizer(index); 3650 } 3651 index = new AddPNode(base, base, index); 3652 _igvn.register_new_node_with_optimizer(index); 3653 Node* from = new AddPNode(base, index, offset); 3654 _igvn.register_new_node_with_optimizer(from); 3655 // Compute the number of elements to copy 3656 Node* len = new SubINode(head->limit(), head->init_trip()); 3657 _igvn.register_new_node_with_optimizer(len); 3658 3659 BasicType t = store->as_Mem()->memory_type(); 3660 bool aligned = false; 3661 if (offset != NULL && head->init_trip()->is_Con()) { 3662 int element_size = type2aelembytes(t); 3663 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 3664 } 3665 3666 // Build a call to the fill routine 3667 const char* fill_name; 3668 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 3669 assert(fill != NULL, "what?"); 3670 3671 // Convert float/double to int/long for fill routines 3672 if (t == T_FLOAT) { 3673 store_value = new MoveF2INode(store_value); 3674 _igvn.register_new_node_with_optimizer(store_value); 3675 } else if (t == T_DOUBLE) { 3676 store_value = new MoveD2LNode(store_value); 3677 _igvn.register_new_node_with_optimizer(store_value); 3678 } 3679 3680 Node* mem_phi = store->in(MemNode::Memory); 3681 Node* result_ctrl; 3682 Node* result_mem; 3683 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 3684 CallLeafNode *call = new CallLeafNoFPNode(call_type, fill, 3685 fill_name, TypeAryPtr::get_array_body_type(t)); 3686 uint cnt = 0; 3687 call->init_req(TypeFunc::Parms + cnt++, from); 3688 call->init_req(TypeFunc::Parms + cnt++, store_value); 3689 #ifdef _LP64 3690 len = new ConvI2LNode(len); 3691 _igvn.register_new_node_with_optimizer(len); 3692 #endif 3693 call->init_req(TypeFunc::Parms + cnt++, len); 3694 #ifdef _LP64 3695 call->init_req(TypeFunc::Parms + cnt++, C->top()); 3696 #endif 3697 call->init_req(TypeFunc::Control, head->init_control()); 3698 call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. 3699 call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); 3700 call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out_or_null(TypeFunc::ReturnAdr)); 3701 call->init_req(TypeFunc::FramePtr, C->start()->proj_out_or_null(TypeFunc::FramePtr)); 3702 _igvn.register_new_node_with_optimizer(call); 3703 result_ctrl = new ProjNode(call,TypeFunc::Control); 3704 _igvn.register_new_node_with_optimizer(result_ctrl); 3705 result_mem = new ProjNode(call,TypeFunc::Memory); 3706 _igvn.register_new_node_with_optimizer(result_mem); 3707 3708 /* Disable following optimization until proper fix (add missing checks). 3709 3710 // If this fill is tightly coupled to an allocation and overwrites 3711 // the whole body, allow it to take over the zeroing. 3712 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 3713 if (alloc != NULL && alloc->is_AllocateArray()) { 3714 Node* length = alloc->as_AllocateArray()->Ideal_length(); 3715 if (head->limit() == length && 3716 head->init_trip() == _igvn.intcon(0)) { 3717 if (TraceOptimizeFill) { 3718 tty->print_cr("Eliminated zeroing in allocation"); 3719 } 3720 alloc->maybe_set_complete(&_igvn); 3721 } else { 3722 #ifdef ASSERT 3723 if (TraceOptimizeFill) { 3724 tty->print_cr("filling array but bounds don't match"); 3725 alloc->dump(); 3726 head->init_trip()->dump(); 3727 head->limit()->dump(); 3728 length->dump(); 3729 } 3730 #endif 3731 } 3732 } 3733 */ 3734 3735 if (head->is_strip_mined()) { 3736 // Inner strip mined loop goes away so get rid of outer strip 3737 // mined loop 3738 Node* outer_sfpt = head->outer_safepoint(); 3739 Node* in = outer_sfpt->in(0); 3740 Node* outer_out = head->outer_loop_exit(); 3741 lazy_replace(outer_out, in); 3742 _igvn.replace_input_of(outer_sfpt, 0, C->top()); 3743 } 3744 3745 // Redirect the old control and memory edges that are outside the loop. 3746 // Sometimes the memory phi of the head is used as the outgoing 3747 // state of the loop. It's safe in this case to replace it with the 3748 // result_mem. 3749 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 3750 lazy_replace(exit, result_ctrl); 3751 _igvn.replace_node(store, result_mem); 3752 // Any uses the increment outside of the loop become the loop limit. 3753 _igvn.replace_node(head->incr(), head->limit()); 3754 3755 // Disconnect the head from the loop. 3756 for (uint i = 0; i < lpt->_body.size(); i++) { 3757 Node* n = lpt->_body.at(i); 3758 _igvn.replace_node(n, C->top()); 3759 } 3760 3761 return true; 3762 } --- EOF ---