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