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