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