1 /* 2 * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/connode.hpp" 31 #include "opto/divnode.hpp" 32 #include "opto/loopnode.hpp" 33 #include "opto/mulnode.hpp" 34 #include "opto/rootnode.hpp" 35 #include "opto/runtime.hpp" 36 #include "opto/subnode.hpp" 37 38 //------------------------------is_loop_exit----------------------------------- 39 // Given an IfNode, return the loop-exiting projection or NULL if both 40 // arms remain in the loop. 41 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 43 PhaseIdealLoop *phase = _phase; 44 // Test is an IfNode, has 2 projections. If BOTH are in the loop 45 // we need loop unswitching instead of peeling. 46 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 47 return iff->raw_out(0); 48 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 49 return iff->raw_out(1); 50 return NULL; 51 } 52 53 54 //============================================================================= 55 56 57 //------------------------------record_for_igvn---------------------------- 58 // Put loop body on igvn work list 59 void IdealLoopTree::record_for_igvn() { 60 for( uint i = 0; i < _body.size(); i++ ) { 61 Node *n = _body.at(i); 62 _phase->_igvn._worklist.push(n); 63 } 64 } 65 66 //------------------------------compute_profile_trip_cnt---------------------------- 67 // Compute loop trip count from profile data as 68 // (backedge_count + loop_exit_count) / loop_exit_count 69 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 70 if (!_head->is_CountedLoop()) { 71 return; 72 } 73 CountedLoopNode* head = _head->as_CountedLoop(); 74 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 75 return; // Already computed 76 } 77 float trip_cnt = (float)max_jint; // default is big 78 79 Node* back = head->in(LoopNode::LoopBackControl); 80 while (back != head) { 81 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 82 back->in(0) && 83 back->in(0)->is_If() && 84 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 85 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 86 break; 87 } 88 back = phase->idom(back); 89 } 90 if (back != head) { 91 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 92 back->in(0), "if-projection exists"); 93 IfNode* back_if = back->in(0)->as_If(); 94 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 95 96 // Now compute a loop exit count 97 float loop_exit_cnt = 0.0f; 98 for( uint i = 0; i < _body.size(); i++ ) { 99 Node *n = _body[i]; 100 if( n->is_If() ) { 101 IfNode *iff = n->as_If(); 102 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 103 Node *exit = is_loop_exit(iff); 104 if( exit ) { 105 float exit_prob = iff->_prob; 106 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 107 if (exit_prob > PROB_MIN) { 108 float exit_cnt = iff->_fcnt * exit_prob; 109 loop_exit_cnt += exit_cnt; 110 } 111 } 112 } 113 } 114 } 115 if (loop_exit_cnt > 0.0f) { 116 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 117 } else { 118 // No exit count so use 119 trip_cnt = loop_back_cnt; 120 } 121 } 122 #ifndef PRODUCT 123 if (TraceProfileTripCount) { 124 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 125 } 126 #endif 127 head->set_profile_trip_cnt(trip_cnt); 128 } 129 130 //---------------------is_invariant_addition----------------------------- 131 // Return nonzero index of invariant operand for an Add or Sub 132 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. 133 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 134 int op = n->Opcode(); 135 if (op == Op_AddI || op == Op_SubI) { 136 bool in1_invar = this->is_invariant(n->in(1)); 137 bool in2_invar = this->is_invariant(n->in(2)); 138 if (in1_invar && !in2_invar) return 1; 139 if (!in1_invar && in2_invar) return 2; 140 } 141 return 0; 142 } 143 144 //---------------------reassociate_add_sub----------------------------- 145 // Reassociate invariant add and subtract expressions: 146 // 147 // inv1 + (x + inv2) => ( inv1 + inv2) + x 148 // (x + inv2) + inv1 => ( inv1 + inv2) + x 149 // inv1 + (x - inv2) => ( inv1 - inv2) + x 150 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 151 // (x + inv2) - inv1 => (-inv1 + inv2) + x 152 // (x - inv2) + inv1 => ( inv1 - inv2) + x 153 // (x - inv2) - inv1 => (-inv1 - inv2) + x 154 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 155 // inv1 - (x - inv2) => ( inv1 + inv2) - x 156 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 157 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 158 // inv1 - (x + inv2) => ( inv1 - inv2) - x 159 // 160 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 161 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 162 if (is_invariant(n1)) return NULL; 163 int inv1_idx = is_invariant_addition(n1, phase); 164 if (!inv1_idx) return NULL; 165 // Don't mess with add of constant (igvn moves them to expression tree root.) 166 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 167 Node* inv1 = n1->in(inv1_idx); 168 Node* n2 = n1->in(3 - inv1_idx); 169 int inv2_idx = is_invariant_addition(n2, phase); 170 if (!inv2_idx) return NULL; 171 Node* x = n2->in(3 - inv2_idx); 172 Node* inv2 = n2->in(inv2_idx); 173 174 bool neg_x = n2->is_Sub() && inv2_idx == 1; 175 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 176 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 177 if (n1->is_Sub() && inv1_idx == 1) { 178 neg_x = !neg_x; 179 neg_inv2 = !neg_inv2; 180 } 181 Node* inv1_c = phase->get_ctrl(inv1); 182 Node* inv2_c = phase->get_ctrl(inv2); 183 Node* n_inv1; 184 if (neg_inv1) { 185 Node *zero = phase->_igvn.intcon(0); 186 phase->set_ctrl(zero, phase->C->root()); 187 n_inv1 = new (phase->C, 3) SubINode(zero, inv1); 188 phase->register_new_node(n_inv1, inv1_c); 189 } else { 190 n_inv1 = inv1; 191 } 192 Node* inv; 193 if (neg_inv2) { 194 inv = new (phase->C, 3) SubINode(n_inv1, inv2); 195 } else { 196 inv = new (phase->C, 3) AddINode(n_inv1, inv2); 197 } 198 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 199 200 Node* addx; 201 if (neg_x) { 202 addx = new (phase->C, 3) SubINode(inv, x); 203 } else { 204 addx = new (phase->C, 3) AddINode(x, inv); 205 } 206 phase->register_new_node(addx, phase->get_ctrl(x)); 207 phase->_igvn.replace_node(n1, addx); 208 assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); 209 _body.yank(n1); 210 return addx; 211 } 212 213 //---------------------reassociate_invariants----------------------------- 214 // Reassociate invariant expressions: 215 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 216 for (int i = _body.size() - 1; i >= 0; i--) { 217 Node *n = _body.at(i); 218 for (int j = 0; j < 5; j++) { 219 Node* nn = reassociate_add_sub(n, phase); 220 if (nn == NULL) break; 221 n = nn; // again 222 }; 223 } 224 } 225 226 //------------------------------policy_peeling--------------------------------- 227 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 228 // make some loop-invariant test (usually a null-check) happen before the loop. 229 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 230 Node *test = ((IdealLoopTree*)this)->tail(); 231 int body_size = ((IdealLoopTree*)this)->_body.size(); 232 int uniq = phase->C->unique(); 233 // Peeling does loop cloning which can result in O(N^2) node construction 234 if( body_size > 255 /* Prevent overflow for large body_size */ 235 || (body_size * body_size + uniq > MaxNodeLimit) ) { 236 return false; // too large to safely clone 237 } 238 while( test != _head ) { // Scan till run off top of loop 239 if( test->is_If() ) { // Test? 240 Node *ctrl = phase->get_ctrl(test->in(1)); 241 if (ctrl->is_top()) 242 return false; // Found dead test on live IF? No peeling! 243 // Standard IF only has one input value to check for loop invariance 244 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 245 // Condition is not a member of this loop? 246 if( !is_member(phase->get_loop(ctrl)) && 247 is_loop_exit(test) ) 248 return true; // Found reason to peel! 249 } 250 // Walk up dominators to loop _head looking for test which is 251 // executed on every path thru loop. 252 test = phase->idom(test); 253 } 254 return false; 255 } 256 257 //------------------------------peeled_dom_test_elim--------------------------- 258 // If we got the effect of peeling, either by actually peeling or by making 259 // a pre-loop which must execute at least once, we can remove all 260 // loop-invariant dominated tests in the main body. 261 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 262 bool progress = true; 263 while( progress ) { 264 progress = false; // Reset for next iteration 265 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 266 Node *test = prev->in(0); 267 while( test != loop->_head ) { // Scan till run off top of loop 268 269 int p_op = prev->Opcode(); 270 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 271 test->is_If() && // Test? 272 !test->in(1)->is_Con() && // And not already obvious? 273 // Condition is not a member of this loop? 274 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 275 // Walk loop body looking for instances of this test 276 for( uint i = 0; i < loop->_body.size(); i++ ) { 277 Node *n = loop->_body.at(i); 278 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 279 // IfNode was dominated by version in peeled loop body 280 progress = true; 281 dominated_by( old_new[prev->_idx], n ); 282 } 283 } 284 } 285 prev = test; 286 test = idom(test); 287 } // End of scan tests in loop 288 289 } // End of while( progress ) 290 } 291 292 //------------------------------do_peeling------------------------------------- 293 // Peel the first iteration of the given loop. 294 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 295 // The pre-loop illegally has 2 control users (old & new loops). 296 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 297 // Do this by making the old-loop fall-in edges act as if they came 298 // around the loopback from the prior iteration (follow the old-loop 299 // backedges) and then map to the new peeled iteration. This leaves 300 // the pre-loop with only 1 user (the new peeled iteration), but the 301 // peeled-loop backedge has 2 users. 302 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 303 // extra backedge user. 304 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 305 306 C->set_major_progress(); 307 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 308 // 'pre' loop from the main and the 'pre' can no longer have it's 309 // iterations adjusted. Therefore, we need to declare this loop as 310 // no longer a 'main' loop; it will need new pre and post loops before 311 // we can do further RCE. 312 #ifndef PRODUCT 313 if (TraceLoopOpts) { 314 tty->print("Peel "); 315 loop->dump_head(); 316 } 317 #endif 318 Node *h = loop->_head; 319 if (h->is_CountedLoop()) { 320 CountedLoopNode *cl = h->as_CountedLoop(); 321 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 322 cl->set_trip_count(cl->trip_count() - 1); 323 if (cl->is_main_loop()) { 324 cl->set_normal_loop(); 325 #ifndef PRODUCT 326 if (PrintOpto && VerifyLoopOptimizations) { 327 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 328 loop->dump_head(); 329 } 330 #endif 331 } 332 } 333 334 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 335 // The pre-loop illegally has 2 control users (old & new loops). 336 clone_loop( loop, old_new, dom_depth(loop->_head) ); 337 338 339 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 340 // Do this by making the old-loop fall-in edges act as if they came 341 // around the loopback from the prior iteration (follow the old-loop 342 // backedges) and then map to the new peeled iteration. This leaves 343 // the pre-loop with only 1 user (the new peeled iteration), but the 344 // peeled-loop backedge has 2 users. 345 for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) { 346 Node* old = loop->_head->fast_out(j); 347 if( old->in(0) == loop->_head && old->req() == 3 && 348 (old->is_Loop() || old->is_Phi()) ) { 349 Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 350 if( !new_exit_value ) // Backedge value is ALSO loop invariant? 351 // Then loop body backedge value remains the same. 352 new_exit_value = old->in(LoopNode::LoopBackControl); 353 _igvn.hash_delete(old); 354 old->set_req(LoopNode::EntryControl, new_exit_value); 355 } 356 } 357 358 359 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 360 // extra backedge user. 361 Node *nnn = old_new[loop->_head->_idx]; 362 _igvn.hash_delete(nnn); 363 nnn->set_req(LoopNode::LoopBackControl, C->top()); 364 for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) { 365 Node* use = nnn->fast_out(j2); 366 if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) { 367 _igvn.hash_delete(use); 368 use->set_req(LoopNode::LoopBackControl, C->top()); 369 } 370 } 371 372 373 // Step 4: Correct dom-depth info. Set to loop-head depth. 374 int dd = dom_depth(loop->_head); 375 set_idom(loop->_head, loop->_head->in(1), dd); 376 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 377 Node *old = loop->_body.at(j3); 378 Node *nnn = old_new[old->_idx]; 379 if (!has_ctrl(nnn)) 380 set_idom(nnn, idom(nnn), dd-1); 381 // While we're at it, remove any SafePoints from the peeled code 382 if( old->Opcode() == Op_SafePoint ) { 383 Node *nnn = old_new[old->_idx]; 384 lazy_replace(nnn,nnn->in(TypeFunc::Control)); 385 } 386 } 387 388 // Now force out all loop-invariant dominating tests. The optimizer 389 // finds some, but we _know_ they are all useless. 390 peeled_dom_test_elim(loop,old_new); 391 392 loop->record_for_igvn(); 393 } 394 395 //------------------------------policy_maximally_unroll------------------------ 396 // Return exact loop trip count, or 0 if not maximally unrolling 397 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 398 CountedLoopNode *cl = _head->as_CountedLoop(); 399 assert(cl->is_normal_loop(), ""); 400 401 Node *init_n = cl->init_trip(); 402 Node *limit_n = cl->limit(); 403 404 // Non-constant bounds 405 if (init_n == NULL || !init_n->is_Con() || 406 limit_n == NULL || !limit_n->is_Con() || 407 // protect against stride not being a constant 408 !cl->stride_is_con()) { 409 return false; 410 } 411 int init = init_n->get_int(); 412 int limit = limit_n->get_int(); 413 int span = limit - init; 414 int stride = cl->stride_con(); 415 416 if (init >= limit || stride > span) { 417 // return a false (no maximally unroll) and the regular unroll/peel 418 // route will make a small mess which CCP will fold away. 419 return false; 420 } 421 uint trip_count = span/stride; // trip_count can be greater than 2 Gig. 422 assert( (int)trip_count*stride == span, "must divide evenly" ); 423 424 // Real policy: if we maximally unroll, does it get too big? 425 // Allow the unrolled mess to get larger than standard loop 426 // size. After all, it will no longer be a loop. 427 uint body_size = _body.size(); 428 uint unroll_limit = (uint)LoopUnrollLimit * 4; 429 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 430 cl->set_trip_count(trip_count); 431 if (trip_count > unroll_limit || body_size > unroll_limit) { 432 return false; 433 } 434 435 // Do not unroll a loop with String intrinsics code. 436 // String intrinsics are large and have loops. 437 for (uint k = 0; k < _body.size(); k++) { 438 Node* n = _body.at(k); 439 switch (n->Opcode()) { 440 case Op_StrComp: 441 case Op_StrEquals: 442 case Op_StrIndexOf: 443 case Op_AryEq: { 444 return false; 445 } 446 } // switch 447 } 448 449 if (body_size <= unroll_limit) { 450 uint new_body_size = body_size * trip_count; 451 if (new_body_size <= unroll_limit && 452 body_size == new_body_size / trip_count && 453 // Unrolling can result in a large amount of node construction 454 new_body_size < MaxNodeLimit - phase->C->unique()) { 455 return true; // maximally unroll 456 } 457 } 458 459 return false; // Do not maximally unroll 460 } 461 462 463 //------------------------------policy_unroll---------------------------------- 464 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 465 // the loop is a CountedLoop and the body is small enough. 466 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { 467 468 CountedLoopNode *cl = _head->as_CountedLoop(); 469 assert(cl->is_normal_loop() || cl->is_main_loop(), ""); 470 471 // protect against stride not being a constant 472 if (!cl->stride_is_con()) return false; 473 474 // protect against over-unrolling 475 if (cl->trip_count() <= 1) return false; 476 477 int future_unroll_ct = cl->unrolled_count() * 2; 478 479 // Don't unroll if the next round of unrolling would push us 480 // over the expected trip count of the loop. One is subtracted 481 // from the expected trip count because the pre-loop normally 482 // executes 1 iteration. 483 if (UnrollLimitForProfileCheck > 0 && 484 cl->profile_trip_cnt() != COUNT_UNKNOWN && 485 future_unroll_ct > UnrollLimitForProfileCheck && 486 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 487 return false; 488 } 489 490 // When unroll count is greater than LoopUnrollMin, don't unroll if: 491 // the residual iterations are more than 10% of the trip count 492 // and rounds of "unroll,optimize" are not making significant progress 493 // Progress defined as current size less than 20% larger than previous size. 494 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 495 future_unroll_ct > LoopUnrollMin && 496 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 497 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 498 return false; 499 } 500 501 Node *init_n = cl->init_trip(); 502 Node *limit_n = cl->limit(); 503 // Non-constant bounds. 504 // Protect against over-unrolling when init or/and limit are not constant 505 // (so that trip_count's init value is maxint) but iv range is known. 506 if (init_n == NULL || !init_n->is_Con() || 507 limit_n == NULL || !limit_n->is_Con()) { 508 Node* phi = cl->phi(); 509 if (phi != NULL) { 510 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 511 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 512 int next_stride = cl->stride_con() * 2; // stride after this unroll 513 if (next_stride > 0) { 514 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow 515 iv_type->_lo + next_stride > iv_type->_hi) { 516 return false; // over-unrolling 517 } 518 } else if (next_stride < 0) { 519 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow 520 iv_type->_hi + next_stride < iv_type->_lo) { 521 return false; // over-unrolling 522 } 523 } 524 } 525 } 526 527 // Adjust body_size to determine if we unroll or not 528 uint body_size = _body.size(); 529 // Key test to unroll CaffeineMark's Logic test 530 int xors_in_loop = 0; 531 // Also count ModL, DivL and MulL which expand mightly 532 for (uint k = 0; k < _body.size(); k++) { 533 Node* n = _body.at(k); 534 switch (n->Opcode()) { 535 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test 536 case Op_ModL: body_size += 30; break; 537 case Op_DivL: body_size += 30; break; 538 case Op_MulL: body_size += 10; break; 539 case Op_StrComp: 540 case Op_StrEquals: 541 case Op_StrIndexOf: 542 case Op_AryEq: { 543 // Do not unroll a loop with String intrinsics code. 544 // String intrinsics are large and have loops. 545 return false; 546 } 547 } // switch 548 } 549 550 // Check for being too big 551 if (body_size > (uint)LoopUnrollLimit) { 552 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 553 // Normal case: loop too big 554 return false; 555 } 556 557 // Check for stride being a small enough constant 558 if (abs(cl->stride_con()) > (1<<3)) return false; 559 560 // Unroll once! (Each trip will soon do double iterations) 561 return true; 562 } 563 564 //------------------------------policy_align----------------------------------- 565 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 566 // expression that does the alignment. Note that only one array base can be 567 // aligned in a loop (unless the VM guarantees mutual alignment). Note that 568 // if we vectorize short memory ops into longer memory ops, we may want to 569 // increase alignment. 570 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 571 return false; 572 } 573 574 //------------------------------policy_range_check----------------------------- 575 // Return TRUE or FALSE if the loop should be range-check-eliminated. 576 // Actually we do iteration-splitting, a more powerful form of RCE. 577 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 578 if( !RangeCheckElimination ) return false; 579 580 CountedLoopNode *cl = _head->as_CountedLoop(); 581 // If we unrolled with no intention of doing RCE and we later 582 // changed our minds, we got no pre-loop. Either we need to 583 // make a new pre-loop, or we gotta disallow RCE. 584 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now. 585 Node *trip_counter = cl->phi(); 586 587 // Check loop body for tests of trip-counter plus loop-invariant vs 588 // loop-invariant. 589 for( uint i = 0; i < _body.size(); i++ ) { 590 Node *iff = _body[i]; 591 if( iff->Opcode() == Op_If ) { // Test? 592 593 // Comparing trip+off vs limit 594 Node *bol = iff->in(1); 595 if( bol->req() != 2 ) continue; // dead constant test 596 if (!bol->is_Bool()) { 597 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); 598 continue; 599 } 600 Node *cmp = bol->in(1); 601 602 Node *rc_exp = cmp->in(1); 603 Node *limit = cmp->in(2); 604 605 Node *limit_c = phase->get_ctrl(limit); 606 if( limit_c == phase->C->top() ) 607 return false; // Found dead test on live IF? No RCE! 608 if( is_member(phase->get_loop(limit_c) ) ) { 609 // Compare might have operands swapped; commute them 610 rc_exp = cmp->in(2); 611 limit = cmp->in(1); 612 limit_c = phase->get_ctrl(limit); 613 if( is_member(phase->get_loop(limit_c) ) ) 614 continue; // Both inputs are loop varying; cannot RCE 615 } 616 617 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 618 continue; 619 } 620 // Yeah! Found a test like 'trip+off vs limit' 621 // Test is an IfNode, has 2 projections. If BOTH are in the loop 622 // we need loop unswitching instead of iteration splitting. 623 if( is_loop_exit(iff) ) 624 return true; // Found reason to split iterations 625 } // End of is IF 626 } 627 628 return false; 629 } 630 631 //------------------------------policy_peel_only------------------------------- 632 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 633 // for unrolling loops with NO array accesses. 634 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 635 636 for( uint i = 0; i < _body.size(); i++ ) 637 if( _body[i]->is_Mem() ) 638 return false; 639 640 // No memory accesses at all! 641 return true; 642 } 643 644 //------------------------------clone_up_backedge_goo-------------------------- 645 // If Node n lives in the back_ctrl block and cannot float, we clone a private 646 // version of n in preheader_ctrl block and return that, otherwise return n. 647 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) { 648 if( get_ctrl(n) != back_ctrl ) return n; 649 650 Node *x = NULL; // If required, a clone of 'n' 651 // Check for 'n' being pinned in the backedge. 652 if( n->in(0) && n->in(0) == back_ctrl ) { 653 x = n->clone(); // Clone a copy of 'n' to preheader 654 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 655 } 656 657 // Recursive fixup any other input edges into x. 658 // If there are no changes we can just return 'n', otherwise 659 // we need to clone a private copy and change it. 660 for( uint i = 1; i < n->req(); i++ ) { 661 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) ); 662 if( g != n->in(i) ) { 663 if( !x ) 664 x = n->clone(); 665 x->set_req(i, g); 666 } 667 } 668 if( x ) { // x can legally float to pre-header location 669 register_new_node( x, preheader_ctrl ); 670 return x; 671 } else { // raise n to cover LCA of uses 672 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 673 } 674 return n; 675 } 676 677 //------------------------------insert_pre_post_loops-------------------------- 678 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 679 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 680 // alignment. Useful to unroll loops that do no array accesses. 681 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 682 683 #ifndef PRODUCT 684 if (TraceLoopOpts) { 685 if (peel_only) 686 tty->print("PeelMainPost "); 687 else 688 tty->print("PreMainPost "); 689 loop->dump_head(); 690 } 691 #endif 692 C->set_major_progress(); 693 694 // Find common pieces of the loop being guarded with pre & post loops 695 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 696 assert( main_head->is_normal_loop(), "" ); 697 CountedLoopEndNode *main_end = main_head->loopexit(); 698 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 699 uint dd_main_head = dom_depth(main_head); 700 uint max = main_head->outcnt(); 701 702 Node *pre_header= main_head->in(LoopNode::EntryControl); 703 Node *init = main_head->init_trip(); 704 Node *incr = main_end ->incr(); 705 Node *limit = main_end ->limit(); 706 Node *stride = main_end ->stride(); 707 Node *cmp = main_end ->cmp_node(); 708 BoolTest::mask b_test = main_end->test_trip(); 709 710 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 711 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 712 if( bol->outcnt() != 1 ) { 713 bol = bol->clone(); 714 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 715 _igvn.hash_delete(main_end); 716 main_end->set_req(CountedLoopEndNode::TestValue, bol); 717 } 718 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 719 if( cmp->outcnt() != 1 ) { 720 cmp = cmp->clone(); 721 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 722 _igvn.hash_delete(bol); 723 bol->set_req(1, cmp); 724 } 725 726 //------------------------------ 727 // Step A: Create Post-Loop. 728 Node* main_exit = main_end->proj_out(false); 729 assert( main_exit->Opcode() == Op_IfFalse, "" ); 730 int dd_main_exit = dom_depth(main_exit); 731 732 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 733 // loop pre-header illegally has 2 control users (old & new loops). 734 clone_loop( loop, old_new, dd_main_exit ); 735 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 736 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 737 post_head->set_post_loop(main_head); 738 739 // Reduce the post-loop trip count. 740 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 741 post_end->_prob = PROB_FAIR; 742 743 // Build the main-loop normal exit. 744 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end); 745 _igvn.register_new_node_with_optimizer( new_main_exit ); 746 set_idom(new_main_exit, main_end, dd_main_exit ); 747 set_loop(new_main_exit, loop->_parent); 748 749 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 750 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 751 // (the main-loop trip-counter exit value) because we will be changing 752 // the exit value (via unrolling) so we cannot constant-fold away the zero 753 // trip guard until all unrolling is done. 754 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr); 755 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit ); 756 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test ); 757 register_new_node( zer_opaq, new_main_exit ); 758 register_new_node( zer_cmp , new_main_exit ); 759 register_new_node( zer_bol , new_main_exit ); 760 761 // Build the IfNode 762 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 763 _igvn.register_new_node_with_optimizer( zer_iff ); 764 set_idom(zer_iff, new_main_exit, dd_main_exit); 765 set_loop(zer_iff, loop->_parent); 766 767 // Plug in the false-path, taken if we need to skip post-loop 768 _igvn.hash_delete( main_exit ); 769 main_exit->set_req(0, zer_iff); 770 _igvn._worklist.push(main_exit); 771 set_idom(main_exit, zer_iff, dd_main_exit); 772 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 773 // Make the true-path, must enter the post loop 774 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff ); 775 _igvn.register_new_node_with_optimizer( zer_taken ); 776 set_idom(zer_taken, zer_iff, dd_main_exit); 777 set_loop(zer_taken, loop->_parent); 778 // Plug in the true path 779 _igvn.hash_delete( post_head ); 780 post_head->set_req(LoopNode::EntryControl, zer_taken); 781 set_idom(post_head, zer_taken, dd_main_exit); 782 783 // Step A3: Make the fall-in values to the post-loop come from the 784 // fall-out values of the main-loop. 785 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 786 Node* main_phi = main_head->fast_out(i); 787 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 788 Node *post_phi = old_new[main_phi->_idx]; 789 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 790 post_head->init_control(), 791 main_phi->in(LoopNode::LoopBackControl)); 792 _igvn.hash_delete(post_phi); 793 post_phi->set_req( LoopNode::EntryControl, fallmain ); 794 } 795 } 796 797 // Update local caches for next stanza 798 main_exit = new_main_exit; 799 800 801 //------------------------------ 802 // Step B: Create Pre-Loop. 803 804 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 805 // loop pre-header illegally has 2 control users (old & new loops). 806 clone_loop( loop, old_new, dd_main_head ); 807 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 808 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 809 pre_head->set_pre_loop(main_head); 810 Node *pre_incr = old_new[incr->_idx]; 811 812 // Reduce the pre-loop trip count. 813 pre_end->_prob = PROB_FAIR; 814 815 // Find the pre-loop normal exit. 816 Node* pre_exit = pre_end->proj_out(false); 817 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 818 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end); 819 _igvn.register_new_node_with_optimizer( new_pre_exit ); 820 set_idom(new_pre_exit, pre_end, dd_main_head); 821 set_loop(new_pre_exit, loop->_parent); 822 823 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 824 // pre-loop, the main-loop may not execute at all. Later in life this 825 // zero-trip guard will become the minimum-trip guard when we unroll 826 // the main-loop. 827 Node *min_opaq = new (C, 2) Opaque1Node(C, limit); 828 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq ); 829 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test ); 830 register_new_node( min_opaq, new_pre_exit ); 831 register_new_node( min_cmp , new_pre_exit ); 832 register_new_node( min_bol , new_pre_exit ); 833 834 // Build the IfNode (assume the main-loop is executed always). 835 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); 836 _igvn.register_new_node_with_optimizer( min_iff ); 837 set_idom(min_iff, new_pre_exit, dd_main_head); 838 set_loop(min_iff, loop->_parent); 839 840 // Plug in the false-path, taken if we need to skip main-loop 841 _igvn.hash_delete( pre_exit ); 842 pre_exit->set_req(0, min_iff); 843 set_idom(pre_exit, min_iff, dd_main_head); 844 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 845 // Make the true-path, must enter the main loop 846 Node *min_taken = new (C, 1) IfTrueNode( min_iff ); 847 _igvn.register_new_node_with_optimizer( min_taken ); 848 set_idom(min_taken, min_iff, dd_main_head); 849 set_loop(min_taken, loop->_parent); 850 // Plug in the true path 851 _igvn.hash_delete( main_head ); 852 main_head->set_req(LoopNode::EntryControl, min_taken); 853 set_idom(main_head, min_taken, dd_main_head); 854 855 // Step B3: Make the fall-in values to the main-loop come from the 856 // fall-out values of the pre-loop. 857 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 858 Node* main_phi = main_head->fast_out(i2); 859 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 860 Node *pre_phi = old_new[main_phi->_idx]; 861 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 862 main_head->init_control(), 863 pre_phi->in(LoopNode::LoopBackControl)); 864 _igvn.hash_delete(main_phi); 865 main_phi->set_req( LoopNode::EntryControl, fallpre ); 866 } 867 } 868 869 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 870 // RCE and alignment may change this later. 871 Node *cmp_end = pre_end->cmp_node(); 872 assert( cmp_end->in(2) == limit, "" ); 873 Node *pre_limit = new (C, 3) AddINode( init, stride ); 874 875 // Save the original loop limit in this Opaque1 node for 876 // use by range check elimination. 877 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit); 878 879 register_new_node( pre_limit, pre_head->in(0) ); 880 register_new_node( pre_opaq , pre_head->in(0) ); 881 882 // Since no other users of pre-loop compare, I can hack limit directly 883 assert( cmp_end->outcnt() == 1, "no other users" ); 884 _igvn.hash_delete(cmp_end); 885 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 886 887 // Special case for not-equal loop bounds: 888 // Change pre loop test, main loop test, and the 889 // main loop guard test to use lt or gt depending on stride 890 // direction: 891 // positive stride use < 892 // negative stride use > 893 894 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 895 896 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 897 // Modify pre loop end condition 898 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 899 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test); 900 register_new_node( new_bol0, pre_head->in(0) ); 901 _igvn.hash_delete(pre_end); 902 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); 903 // Modify main loop guard condition 904 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 905 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test); 906 register_new_node( new_bol1, new_pre_exit ); 907 _igvn.hash_delete(min_iff); 908 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 909 // Modify main loop end condition 910 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 911 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test); 912 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 913 _igvn.hash_delete(main_end); 914 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); 915 } 916 917 // Flag main loop 918 main_head->set_main_loop(); 919 if( peel_only ) main_head->set_main_no_pre_loop(); 920 921 // It's difficult to be precise about the trip-counts 922 // for the pre/post loops. They are usually very short, 923 // so guess that 4 trips is a reasonable value. 924 post_head->set_profile_trip_cnt(4.0); 925 pre_head->set_profile_trip_cnt(4.0); 926 927 // Now force out all loop-invariant dominating tests. The optimizer 928 // finds some, but we _know_ they are all useless. 929 peeled_dom_test_elim(loop,old_new); 930 } 931 932 //------------------------------is_invariant----------------------------- 933 // Return true if n is invariant 934 bool IdealLoopTree::is_invariant(Node* n) const { 935 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; 936 if (n_c->is_top()) return false; 937 return !is_member(_phase->get_loop(n_c)); 938 } 939 940 941 //------------------------------do_unroll-------------------------------------- 942 // Unroll the loop body one step - make each trip do 2 iterations. 943 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 944 assert(LoopUnrollLimit, ""); 945 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 946 CountedLoopEndNode *loop_end = loop_head->loopexit(); 947 assert(loop_end, ""); 948 #ifndef PRODUCT 949 if (PrintOpto && VerifyLoopOptimizations) { 950 tty->print("Unrolling "); 951 loop->dump_head(); 952 } else if (TraceLoopOpts) { 953 tty->print("Unroll %d ", loop_head->unrolled_count()*2); 954 loop->dump_head(); 955 } 956 #endif 957 958 // Remember loop node count before unrolling to detect 959 // if rounds of unroll,optimize are making progress 960 loop_head->set_node_count_before_unroll(loop->_body.size()); 961 962 Node *ctrl = loop_head->in(LoopNode::EntryControl); 963 Node *limit = loop_head->limit(); 964 Node *init = loop_head->init_trip(); 965 Node *stride = loop_head->stride(); 966 967 Node *opaq = NULL; 968 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment 969 assert( loop_head->is_main_loop(), "" ); 970 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 971 Node *iff = ctrl->in(0); 972 assert( iff->Opcode() == Op_If, "" ); 973 Node *bol = iff->in(1); 974 assert( bol->Opcode() == Op_Bool, "" ); 975 Node *cmp = bol->in(1); 976 assert( cmp->Opcode() == Op_CmpI, "" ); 977 opaq = cmp->in(2); 978 // Occasionally it's possible for a pre-loop Opaque1 node to be 979 // optimized away and then another round of loop opts attempted. 980 // We can not optimize this particular loop in that case. 981 if( opaq->Opcode() != Op_Opaque1 ) 982 return; // Cannot find pre-loop! Bail out! 983 } 984 985 C->set_major_progress(); 986 987 // Adjust max trip count. The trip count is intentionally rounded 988 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 989 // the main, unrolled, part of the loop will never execute as it is protected 990 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 991 // and later determined that part of the unrolled loop was dead. 992 loop_head->set_trip_count(loop_head->trip_count() / 2); 993 994 // Double the count of original iterations in the unrolled loop body. 995 loop_head->double_unrolled_count(); 996 997 // ----------- 998 // Step 2: Cut back the trip counter for an unroll amount of 2. 999 // Loop will normally trip (limit - init)/stride_con. Since it's a 1000 // CountedLoop this is exact (stride divides limit-init exactly). 1001 // We are going to double the loop body, so we want to knock off any 1002 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 1003 Node *span = new (C, 3) SubINode( limit, init ); 1004 register_new_node( span, ctrl ); 1005 Node *trip = new (C, 3) DivINode( 0, span, stride ); 1006 register_new_node( trip, ctrl ); 1007 Node *mtwo = _igvn.intcon(-2); 1008 set_ctrl(mtwo, C->root()); 1009 Node *rond = new (C, 3) AndINode( trip, mtwo ); 1010 register_new_node( rond, ctrl ); 1011 Node *spn2 = new (C, 3) MulINode( rond, stride ); 1012 register_new_node( spn2, ctrl ); 1013 Node *lim2 = new (C, 3) AddINode( spn2, init ); 1014 register_new_node( lim2, ctrl ); 1015 1016 // Hammer in the new limit 1017 Node *ctrl2 = loop_end->in(0); 1018 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 ); 1019 register_new_node( cmp2, ctrl2 ); 1020 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() ); 1021 register_new_node( bol2, ctrl2 ); 1022 _igvn.hash_delete(loop_end); 1023 loop_end->set_req(CountedLoopEndNode::TestValue, bol2); 1024 1025 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 1026 // Make it a 1-trip test (means at least 2 trips). 1027 if( adjust_min_trip ) { 1028 // Guard test uses an 'opaque' node which is not shared. Hence I 1029 // can edit it's inputs directly. Hammer in the new limit for the 1030 // minimum-trip guard. 1031 assert( opaq->outcnt() == 1, "" ); 1032 _igvn.hash_delete(opaq); 1033 opaq->set_req(1, lim2); 1034 } 1035 1036 // --------- 1037 // Step 4: Clone the loop body. Move it inside the loop. This loop body 1038 // represents the odd iterations; since the loop trips an even number of 1039 // times its backedge is never taken. Kill the backedge. 1040 uint dd = dom_depth(loop_head); 1041 clone_loop( loop, old_new, dd ); 1042 1043 // Make backedges of the clone equal to backedges of the original. 1044 // Make the fall-in from the original come from the fall-out of the clone. 1045 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 1046 Node* phi = loop_head->fast_out(j); 1047 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 1048 Node *newphi = old_new[phi->_idx]; 1049 _igvn.hash_delete( phi ); 1050 _igvn.hash_delete( newphi ); 1051 1052 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 1053 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 1054 phi ->set_req(LoopNode::LoopBackControl, C->top()); 1055 } 1056 } 1057 Node *clone_head = old_new[loop_head->_idx]; 1058 _igvn.hash_delete( clone_head ); 1059 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 1060 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 1061 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 1062 loop->_head = clone_head; // New loop header 1063 1064 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1065 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1066 1067 // Kill the clone's backedge 1068 Node *newcle = old_new[loop_end->_idx]; 1069 _igvn.hash_delete( newcle ); 1070 Node *one = _igvn.intcon(1); 1071 set_ctrl(one, C->root()); 1072 newcle->set_req(1, one); 1073 // Force clone into same loop body 1074 uint max = loop->_body.size(); 1075 for( uint k = 0; k < max; k++ ) { 1076 Node *old = loop->_body.at(k); 1077 Node *nnn = old_new[old->_idx]; 1078 loop->_body.push(nnn); 1079 if (!has_ctrl(old)) 1080 set_loop(nnn, loop); 1081 } 1082 1083 loop->record_for_igvn(); 1084 } 1085 1086 //------------------------------do_maximally_unroll---------------------------- 1087 1088 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1089 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1090 assert(cl->trip_count() > 0, ""); 1091 #ifndef PRODUCT 1092 if (TraceLoopOpts) { 1093 tty->print("MaxUnroll %d ", cl->trip_count()); 1094 loop->dump_head(); 1095 } 1096 #endif 1097 1098 // If loop is tripping an odd number of times, peel odd iteration 1099 if ((cl->trip_count() & 1) == 1) { 1100 do_peeling(loop, old_new); 1101 } 1102 1103 // Now its tripping an even number of times remaining. Double loop body. 1104 // Do not adjust pre-guards; they are not needed and do not exist. 1105 if (cl->trip_count() > 0) { 1106 do_unroll(loop, old_new, false); 1107 } 1108 } 1109 1110 //------------------------------dominates_backedge--------------------------------- 1111 // Returns true if ctrl is executed on every complete iteration 1112 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1113 assert(ctrl->is_CFG(), "must be control"); 1114 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1115 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1116 } 1117 1118 //------------------------------add_constraint--------------------------------- 1119 // Constrain the main loop iterations so the condition: 1120 // scale_con * I + offset < limit 1121 // always holds true. That is, either increase the number of iterations in 1122 // the pre-loop or the post-loop until the condition holds true in the main 1123 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1124 // stride and scale are constants (offset and limit often are). 1125 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { 1126 1127 // Compute "I :: (limit-offset)/scale_con" 1128 Node *con = new (C, 3) SubINode( limit, offset ); 1129 register_new_node( con, pre_ctrl ); 1130 Node *scale = _igvn.intcon(scale_con); 1131 set_ctrl(scale, C->root()); 1132 Node *X = new (C, 3) DivINode( 0, con, scale ); 1133 register_new_node( X, pre_ctrl ); 1134 1135 // For positive stride, the pre-loop limit always uses a MAX function 1136 // and the main loop a MIN function. For negative stride these are 1137 // reversed. 1138 1139 // Also for positive stride*scale the affine function is increasing, so the 1140 // pre-loop must check for underflow and the post-loop for overflow. 1141 // Negative stride*scale reverses this; pre-loop checks for overflow and 1142 // post-loop for underflow. 1143 if( stride_con*scale_con > 0 ) { 1144 // Compute I < (limit-offset)/scale_con 1145 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X) 1146 *main_limit = (stride_con > 0) 1147 ? (Node*)(new (C, 3) MinINode( *main_limit, X )) 1148 : (Node*)(new (C, 3) MaxINode( *main_limit, X )); 1149 register_new_node( *main_limit, pre_ctrl ); 1150 1151 } else { 1152 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I 1153 // Add the negation of the main-loop constraint to the pre-loop. 1154 // See footnote [++] below for a derivation of the limit expression. 1155 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1); 1156 set_ctrl(incr, C->root()); 1157 Node *adj = new (C, 3) AddINode( X, incr ); 1158 register_new_node( adj, pre_ctrl ); 1159 *pre_limit = (scale_con > 0) 1160 ? (Node*)new (C, 3) MinINode( *pre_limit, adj ) 1161 : (Node*)new (C, 3) MaxINode( *pre_limit, adj ); 1162 register_new_node( *pre_limit, pre_ctrl ); 1163 1164 // [++] Here's the algebra that justifies the pre-loop limit expression: 1165 // 1166 // NOT( scale_con * I + offset < limit ) 1167 // == 1168 // scale_con * I + offset >= limit 1169 // == 1170 // SGN(scale_con) * I >= (limit-offset)/|scale_con| 1171 // == 1172 // (limit-offset)/|scale_con| <= I * SGN(scale_con) 1173 // == 1174 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con) 1175 // == 1176 // ( if (scale_con > 0) /*common case*/ 1177 // (limit-offset)/scale_con - 1 < I 1178 // else 1179 // (limit-offset)/scale_con + 1 > I 1180 // ) 1181 // ( if (scale_con > 0) /*common case*/ 1182 // (limit-offset)/scale_con + SGN(-scale_con) < I 1183 // else 1184 // (limit-offset)/scale_con + SGN(-scale_con) > I 1185 } 1186 } 1187 1188 1189 //------------------------------is_scaled_iv--------------------------------- 1190 // Return true if exp is a constant times an induction var 1191 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1192 if (exp == iv) { 1193 if (p_scale != NULL) { 1194 *p_scale = 1; 1195 } 1196 return true; 1197 } 1198 int opc = exp->Opcode(); 1199 if (opc == Op_MulI) { 1200 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1201 if (p_scale != NULL) { 1202 *p_scale = exp->in(2)->get_int(); 1203 } 1204 return true; 1205 } 1206 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1207 if (p_scale != NULL) { 1208 *p_scale = exp->in(1)->get_int(); 1209 } 1210 return true; 1211 } 1212 } else if (opc == Op_LShiftI) { 1213 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1214 if (p_scale != NULL) { 1215 *p_scale = 1 << exp->in(2)->get_int(); 1216 } 1217 return true; 1218 } 1219 } 1220 return false; 1221 } 1222 1223 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1224 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1225 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1226 if (is_scaled_iv(exp, iv, p_scale)) { 1227 if (p_offset != NULL) { 1228 Node *zero = _igvn.intcon(0); 1229 set_ctrl(zero, C->root()); 1230 *p_offset = zero; 1231 } 1232 return true; 1233 } 1234 int opc = exp->Opcode(); 1235 if (opc == Op_AddI) { 1236 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1237 if (p_offset != NULL) { 1238 *p_offset = exp->in(2); 1239 } 1240 return true; 1241 } 1242 if (exp->in(2)->is_Con()) { 1243 Node* offset2 = NULL; 1244 if (depth < 2 && 1245 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1246 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1247 if (p_offset != NULL) { 1248 Node *ctrl_off2 = get_ctrl(offset2); 1249 Node* offset = new (C, 3) AddINode(offset2, exp->in(2)); 1250 register_new_node(offset, ctrl_off2); 1251 *p_offset = offset; 1252 } 1253 return true; 1254 } 1255 } 1256 } else if (opc == Op_SubI) { 1257 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1258 if (p_offset != NULL) { 1259 Node *zero = _igvn.intcon(0); 1260 set_ctrl(zero, C->root()); 1261 Node *ctrl_off = get_ctrl(exp->in(2)); 1262 Node* offset = new (C, 3) SubINode(zero, exp->in(2)); 1263 register_new_node(offset, ctrl_off); 1264 *p_offset = offset; 1265 } 1266 return true; 1267 } 1268 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1269 if (p_offset != NULL) { 1270 *p_scale *= -1; 1271 *p_offset = exp->in(1); 1272 } 1273 return true; 1274 } 1275 } 1276 return false; 1277 } 1278 1279 //------------------------------do_range_check--------------------------------- 1280 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1281 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1282 #ifndef PRODUCT 1283 if (PrintOpto && VerifyLoopOptimizations) { 1284 tty->print("Range Check Elimination "); 1285 loop->dump_head(); 1286 } else if (TraceLoopOpts) { 1287 tty->print("RangeCheck "); 1288 loop->dump_head(); 1289 } 1290 #endif 1291 assert(RangeCheckElimination, ""); 1292 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1293 assert(cl->is_main_loop(), ""); 1294 1295 // protect against stride not being a constant 1296 if (!cl->stride_is_con()) 1297 return; 1298 1299 // Find the trip counter; we are iteration splitting based on it 1300 Node *trip_counter = cl->phi(); 1301 // Find the main loop limit; we will trim it's iterations 1302 // to not ever trip end tests 1303 Node *main_limit = cl->limit(); 1304 1305 // Need to find the main-loop zero-trip guard 1306 Node *ctrl = cl->in(LoopNode::EntryControl); 1307 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, ""); 1308 Node *iffm = ctrl->in(0); 1309 assert(iffm->Opcode() == Op_If, ""); 1310 Node *bolzm = iffm->in(1); 1311 assert(bolzm->Opcode() == Op_Bool, ""); 1312 Node *cmpzm = bolzm->in(1); 1313 assert(cmpzm->is_Cmp(), ""); 1314 Node *opqzm = cmpzm->in(2); 1315 // Can not optimize a loop if pre-loop Opaque1 node is optimized 1316 // away and then another round of loop opts attempted. 1317 if (opqzm->Opcode() != Op_Opaque1) 1318 return; 1319 assert(opqzm->in(1) == main_limit, "do not understand situation"); 1320 1321 // Find the pre-loop limit; we will expand it's iterations to 1322 // not ever trip low tests. 1323 Node *p_f = iffm->in(0); 1324 assert(p_f->Opcode() == Op_IfFalse, ""); 1325 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 1326 assert(pre_end->loopnode()->is_pre_loop(), ""); 1327 Node *pre_opaq1 = pre_end->limit(); 1328 // Occasionally it's possible for a pre-loop Opaque1 node to be 1329 // optimized away and then another round of loop opts attempted. 1330 // We can not optimize this particular loop in that case. 1331 if (pre_opaq1->Opcode() != Op_Opaque1) 1332 return; 1333 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 1334 Node *pre_limit = pre_opaq->in(1); 1335 1336 // Where do we put new limit calculations 1337 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 1338 1339 // Ensure the original loop limit is available from the 1340 // pre-loop Opaque1 node. 1341 Node *orig_limit = pre_opaq->original_loop_limit(); 1342 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) 1343 return; 1344 1345 // Must know if its a count-up or count-down loop 1346 1347 int stride_con = cl->stride_con(); 1348 Node *zero = _igvn.intcon(0); 1349 Node *one = _igvn.intcon(1); 1350 set_ctrl(zero, C->root()); 1351 set_ctrl(one, C->root()); 1352 1353 // Range checks that do not dominate the loop backedge (ie. 1354 // conditionally executed) can lengthen the pre loop limit beyond 1355 // the original loop limit. To prevent this, the pre limit is 1356 // (for stride > 0) MINed with the original loop limit (MAXed 1357 // stride < 0) when some range_check (rc) is conditionally 1358 // executed. 1359 bool conditional_rc = false; 1360 1361 // Check loop body for tests of trip-counter plus loop-invariant vs 1362 // loop-invariant. 1363 for( uint i = 0; i < loop->_body.size(); i++ ) { 1364 Node *iff = loop->_body[i]; 1365 if( iff->Opcode() == Op_If ) { // Test? 1366 1367 // Test is an IfNode, has 2 projections. If BOTH are in the loop 1368 // we need loop unswitching instead of iteration splitting. 1369 Node *exit = loop->is_loop_exit(iff); 1370 if( !exit ) continue; 1371 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 1372 1373 // Get boolean condition to test 1374 Node *i1 = iff->in(1); 1375 if( !i1->is_Bool() ) continue; 1376 BoolNode *bol = i1->as_Bool(); 1377 BoolTest b_test = bol->_test; 1378 // Flip sense of test if exit condition is flipped 1379 if( flip ) 1380 b_test = b_test.negate(); 1381 1382 // Get compare 1383 Node *cmp = bol->in(1); 1384 1385 // Look for trip_counter + offset vs limit 1386 Node *rc_exp = cmp->in(1); 1387 Node *limit = cmp->in(2); 1388 jint scale_con= 1; // Assume trip counter not scaled 1389 1390 Node *limit_c = get_ctrl(limit); 1391 if( loop->is_member(get_loop(limit_c) ) ) { 1392 // Compare might have operands swapped; commute them 1393 b_test = b_test.commute(); 1394 rc_exp = cmp->in(2); 1395 limit = cmp->in(1); 1396 limit_c = get_ctrl(limit); 1397 if( loop->is_member(get_loop(limit_c) ) ) 1398 continue; // Both inputs are loop varying; cannot RCE 1399 } 1400 // Here we know 'limit' is loop invariant 1401 1402 // 'limit' maybe pinned below the zero trip test (probably from a 1403 // previous round of rce), in which case, it can't be used in the 1404 // zero trip test expression which must occur before the zero test's if. 1405 if( limit_c == ctrl ) { 1406 continue; // Don't rce this check but continue looking for other candidates. 1407 } 1408 1409 // Check for scaled induction variable plus an offset 1410 Node *offset = NULL; 1411 1412 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 1413 continue; 1414 } 1415 1416 Node *offset_c = get_ctrl(offset); 1417 if( loop->is_member( get_loop(offset_c) ) ) 1418 continue; // Offset is not really loop invariant 1419 // Here we know 'offset' is loop invariant. 1420 1421 // As above for the 'limit', the 'offset' maybe pinned below the 1422 // zero trip test. 1423 if( offset_c == ctrl ) { 1424 continue; // Don't rce this check but continue looking for other candidates. 1425 } 1426 1427 // At this point we have the expression as: 1428 // scale_con * trip_counter + offset :: limit 1429 // where scale_con, offset and limit are loop invariant. Trip_counter 1430 // monotonically increases by stride_con, a constant. Both (or either) 1431 // stride_con and scale_con can be negative which will flip about the 1432 // sense of the test. 1433 1434 // Adjust pre and main loop limits to guard the correct iteration set 1435 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 1436 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 1437 // The overflow limit: scale*I+offset < limit 1438 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1439 // The underflow limit: 0 <= scale*I+offset. 1440 // Some math yields: -scale*I-(offset+1) < 0 1441 Node *plus_one = new (C, 3) AddINode( offset, one ); 1442 register_new_node( plus_one, pre_ctrl ); 1443 Node *neg_offset = new (C, 3) SubINode( zero, plus_one ); 1444 register_new_node( neg_offset, pre_ctrl ); 1445 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit ); 1446 if (!conditional_rc) { 1447 conditional_rc = !loop->dominates_backedge(iff); 1448 } 1449 } else { 1450 #ifndef PRODUCT 1451 if( PrintOpto ) 1452 tty->print_cr("missed RCE opportunity"); 1453 #endif 1454 continue; // In release mode, ignore it 1455 } 1456 } else { // Otherwise work on normal compares 1457 switch( b_test._test ) { 1458 case BoolTest::ge: // Convert X >= Y to -X <= -Y 1459 scale_con = -scale_con; 1460 offset = new (C, 3) SubINode( zero, offset ); 1461 register_new_node( offset, pre_ctrl ); 1462 limit = new (C, 3) SubINode( zero, limit ); 1463 register_new_node( limit, pre_ctrl ); 1464 // Fall into LE case 1465 case BoolTest::le: // Convert X <= Y to X < Y+1 1466 limit = new (C, 3) AddINode( limit, one ); 1467 register_new_node( limit, pre_ctrl ); 1468 // Fall into LT case 1469 case BoolTest::lt: 1470 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1471 if (!conditional_rc) { 1472 conditional_rc = !loop->dominates_backedge(iff); 1473 } 1474 break; 1475 default: 1476 #ifndef PRODUCT 1477 if( PrintOpto ) 1478 tty->print_cr("missed RCE opportunity"); 1479 #endif 1480 continue; // Unhandled case 1481 } 1482 } 1483 1484 // Kill the eliminated test 1485 C->set_major_progress(); 1486 Node *kill_con = _igvn.intcon( 1-flip ); 1487 set_ctrl(kill_con, C->root()); 1488 _igvn.hash_delete(iff); 1489 iff->set_req(1, kill_con); 1490 _igvn._worklist.push(iff); 1491 // Find surviving projection 1492 assert(iff->is_If(), ""); 1493 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 1494 // Find loads off the surviving projection; remove their control edge 1495 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 1496 Node* cd = dp->fast_out(i); // Control-dependent node 1497 if( cd->is_Load() ) { // Loads can now float around in the loop 1498 _igvn.hash_delete(cd); 1499 // Allow the load to float around in the loop, or before it 1500 // but NOT before the pre-loop. 1501 cd->set_req(0, ctrl); // ctrl, not NULL 1502 _igvn._worklist.push(cd); 1503 --i; 1504 --imax; 1505 } 1506 } 1507 1508 } // End of is IF 1509 1510 } 1511 1512 // Update loop limits 1513 if (conditional_rc) { 1514 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit) 1515 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit); 1516 register_new_node(pre_limit, pre_ctrl); 1517 } 1518 _igvn.hash_delete(pre_opaq); 1519 pre_opaq->set_req(1, pre_limit); 1520 1521 // Note:: we are making the main loop limit no longer precise; 1522 // need to round up based on stride. 1523 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case 1524 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 1525 // Hopefully, compiler will optimize for powers of 2. 1526 Node *ctrl = get_ctrl(main_limit); 1527 Node *stride = cl->stride(); 1528 Node *init = cl->init_trip(); 1529 Node *span = new (C, 3) SubINode(main_limit,init); 1530 register_new_node(span,ctrl); 1531 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 1532 Node *add = new (C, 3) AddINode(span,rndup); 1533 register_new_node(add,ctrl); 1534 Node *div = new (C, 3) DivINode(0,add,stride); 1535 register_new_node(div,ctrl); 1536 Node *mul = new (C, 3) MulINode(div,stride); 1537 register_new_node(mul,ctrl); 1538 Node *newlim = new (C, 3) AddINode(mul,init); 1539 register_new_node(newlim,ctrl); 1540 main_limit = newlim; 1541 } 1542 1543 Node *main_cle = cl->loopexit(); 1544 Node *main_bol = main_cle->in(1); 1545 // Hacking loop bounds; need private copies of exit test 1546 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 1547 _igvn.hash_delete(main_cle); 1548 main_bol = main_bol->clone();// Clone a private BoolNode 1549 register_new_node( main_bol, main_cle->in(0) ); 1550 main_cle->set_req(1,main_bol); 1551 } 1552 Node *main_cmp = main_bol->in(1); 1553 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 1554 _igvn.hash_delete(main_bol); 1555 main_cmp = main_cmp->clone();// Clone a private CmpNode 1556 register_new_node( main_cmp, main_cle->in(0) ); 1557 main_bol->set_req(1,main_cmp); 1558 } 1559 // Hack the now-private loop bounds 1560 _igvn.hash_delete(main_cmp); 1561 main_cmp->set_req(2, main_limit); 1562 _igvn._worklist.push(main_cmp); 1563 // The OpaqueNode is unshared by design 1564 _igvn.hash_delete(opqzm); 1565 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 1566 opqzm->set_req(1,main_limit); 1567 _igvn._worklist.push(opqzm); 1568 } 1569 1570 //------------------------------DCE_loop_body---------------------------------- 1571 // Remove simplistic dead code from loop body 1572 void IdealLoopTree::DCE_loop_body() { 1573 for( uint i = 0; i < _body.size(); i++ ) 1574 if( _body.at(i)->outcnt() == 0 ) 1575 _body.map( i--, _body.pop() ); 1576 } 1577 1578 1579 //------------------------------adjust_loop_exit_prob-------------------------- 1580 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 1581 // Replace with a 1-in-10 exit guess. 1582 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 1583 Node *test = tail(); 1584 while( test != _head ) { 1585 uint top = test->Opcode(); 1586 if( top == Op_IfTrue || top == Op_IfFalse ) { 1587 int test_con = ((ProjNode*)test)->_con; 1588 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 1589 IfNode *iff = test->in(0)->as_If(); 1590 if( iff->outcnt() == 2 ) { // Ignore dead tests 1591 Node *bol = iff->in(1); 1592 if( bol && bol->req() > 1 && bol->in(1) && 1593 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 1594 (bol->in(1)->Opcode() == Op_StoreIConditional ) || 1595 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 1596 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 1597 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 1598 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || 1599 (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) 1600 return; // Allocation loops RARELY take backedge 1601 // Find the OTHER exit path from the IF 1602 Node* ex = iff->proj_out(1-test_con); 1603 float p = iff->_prob; 1604 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 1605 if( top == Op_IfTrue ) { 1606 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 1607 iff->_prob = PROB_STATIC_FREQUENT; 1608 } 1609 } else { 1610 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 1611 iff->_prob = PROB_STATIC_INFREQUENT; 1612 } 1613 } 1614 } 1615 } 1616 } 1617 test = phase->idom(test); 1618 } 1619 } 1620 1621 1622 //------------------------------policy_do_remove_empty_loop-------------------- 1623 // Micro-benchmark spamming. Policy is to always remove empty loops. 1624 // The 'DO' part is to replace the trip counter with the value it will 1625 // have on the last iteration. This will break the loop. 1626 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 1627 // Minimum size must be empty loop 1628 if (_body.size() > 7/*number of nodes in an empty loop*/) 1629 return false; 1630 1631 if (!_head->is_CountedLoop()) 1632 return false; // Dead loop 1633 CountedLoopNode *cl = _head->as_CountedLoop(); 1634 if (!cl->loopexit()) 1635 return false; // Malformed loop 1636 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) 1637 return false; // Infinite loop 1638 1639 #ifdef ASSERT 1640 // Ensure only one phi which is the iv. 1641 Node* iv = NULL; 1642 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 1643 Node* n = cl->fast_out(i); 1644 if (n->Opcode() == Op_Phi) { 1645 assert(iv == NULL, "Too many phis" ); 1646 iv = n; 1647 } 1648 } 1649 assert(iv == cl->phi(), "Wrong phi" ); 1650 #endif 1651 1652 // main and post loops have explicitly created zero trip guard 1653 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); 1654 if (needs_guard) { 1655 // Check for an obvious zero trip guard. 1656 Node* inctrl = cl->in(LoopNode::EntryControl); 1657 if (inctrl->Opcode() == Op_IfTrue) { 1658 // The test should look like just the backedge of a CountedLoop 1659 Node* iff = inctrl->in(0); 1660 if (iff->is_If()) { 1661 Node* bol = iff->in(1); 1662 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { 1663 Node* cmp = bol->in(1); 1664 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { 1665 needs_guard = false; 1666 } 1667 } 1668 } 1669 } 1670 } 1671 1672 #ifndef PRODUCT 1673 if (PrintOpto) { 1674 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); 1675 this->dump_head(); 1676 } else if (TraceLoopOpts) { 1677 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); 1678 this->dump_head(); 1679 } 1680 #endif 1681 1682 if (needs_guard) { 1683 // Peel the loop to ensure there's a zero trip guard 1684 Node_List old_new; 1685 phase->do_peeling(this, old_new); 1686 } 1687 1688 // Replace the phi at loop head with the final value of the last 1689 // iteration. Then the CountedLoopEnd will collapse (backedge never 1690 // taken) and all loop-invariant uses of the exit values will be correct. 1691 Node *phi = cl->phi(); 1692 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() ); 1693 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 1694 phase->_igvn.replace_node(phi,final); 1695 phase->C->set_major_progress(); 1696 return true; 1697 } 1698 1699 1700 //============================================================================= 1701 //------------------------------iteration_split_impl--------------------------- 1702 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 1703 // Check and remove empty loops (spam micro-benchmarks) 1704 if( policy_do_remove_empty_loop(phase) ) 1705 return true; // Here we removed an empty loop 1706 1707 bool should_peel = policy_peeling(phase); // Should we peel? 1708 1709 bool should_unswitch = policy_unswitching(phase); 1710 1711 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 1712 // This removes loop-invariant tests (usually null checks). 1713 if( !_head->is_CountedLoop() ) { // Non-counted loop 1714 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 1715 // Partial peel succeeded so terminate this round of loop opts 1716 return false; 1717 } 1718 if( should_peel ) { // Should we peel? 1719 #ifndef PRODUCT 1720 if (PrintOpto) tty->print_cr("should_peel"); 1721 #endif 1722 phase->do_peeling(this,old_new); 1723 } else if( should_unswitch ) { 1724 phase->do_unswitching(this, old_new); 1725 } 1726 return true; 1727 } 1728 CountedLoopNode *cl = _head->as_CountedLoop(); 1729 1730 if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops 1731 1732 // Do nothing special to pre- and post- loops 1733 if( cl->is_pre_loop() || cl->is_post_loop() ) return true; 1734 1735 // Compute loop trip count from profile data 1736 compute_profile_trip_cnt(phase); 1737 1738 // Before attempting fancy unrolling, RCE or alignment, see if we want 1739 // to completely unroll this loop or do loop unswitching. 1740 if( cl->is_normal_loop() ) { 1741 if (should_unswitch) { 1742 phase->do_unswitching(this, old_new); 1743 return true; 1744 } 1745 bool should_maximally_unroll = policy_maximally_unroll(phase); 1746 if( should_maximally_unroll ) { 1747 // Here we did some unrolling and peeling. Eventually we will 1748 // completely unroll this loop and it will no longer be a loop. 1749 phase->do_maximally_unroll(this,old_new); 1750 return true; 1751 } 1752 } 1753 1754 1755 // Counted loops may be peeled, may need some iterations run up 1756 // front for RCE, and may want to align loop refs to a cache 1757 // line. Thus we clone a full loop up front whose trip count is 1758 // at least 1 (if peeling), but may be several more. 1759 1760 // The main loop will start cache-line aligned with at least 1 1761 // iteration of the unrolled body (zero-trip test required) and 1762 // will have some range checks removed. 1763 1764 // A post-loop will finish any odd iterations (leftover after 1765 // unrolling), plus any needed for RCE purposes. 1766 1767 bool should_unroll = policy_unroll(phase); 1768 1769 bool should_rce = policy_range_check(phase); 1770 1771 bool should_align = policy_align(phase); 1772 1773 // If not RCE'ing (iteration splitting) or Aligning, then we do not 1774 // need a pre-loop. We may still need to peel an initial iteration but 1775 // we will not be needing an unknown number of pre-iterations. 1776 // 1777 // Basically, if may_rce_align reports FALSE first time through, 1778 // we will not be able to later do RCE or Aligning on this loop. 1779 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 1780 1781 // If we have any of these conditions (RCE, alignment, unrolling) met, then 1782 // we switch to the pre-/main-/post-loop model. This model also covers 1783 // peeling. 1784 if( should_rce || should_align || should_unroll ) { 1785 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops 1786 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 1787 1788 // Adjust the pre- and main-loop limits to let the pre and post loops run 1789 // with full checks, but the main-loop with no checks. Remove said 1790 // checks from the main body. 1791 if( should_rce ) 1792 phase->do_range_check(this,old_new); 1793 1794 // Double loop body for unrolling. Adjust the minimum-trip test (will do 1795 // twice as many iterations as before) and the main body limit (only do 1796 // an even number of trips). If we are peeling, we might enable some RCE 1797 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 1798 // peeling. 1799 if( should_unroll && !should_peel ) 1800 phase->do_unroll(this,old_new, true); 1801 1802 // Adjust the pre-loop limits to align the main body 1803 // iterations. 1804 if( should_align ) 1805 Unimplemented(); 1806 1807 } else { // Else we have an unchanged counted loop 1808 if( should_peel ) // Might want to peel but do nothing else 1809 phase->do_peeling(this,old_new); 1810 } 1811 return true; 1812 } 1813 1814 1815 //============================================================================= 1816 //------------------------------iteration_split-------------------------------- 1817 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 1818 // Recursively iteration split nested loops 1819 if (_child && !_child->iteration_split(phase, old_new)) 1820 return false; 1821 1822 // Clean out prior deadwood 1823 DCE_loop_body(); 1824 1825 1826 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 1827 // Replace with a 1-in-10 exit guess. 1828 if (_parent /*not the root loop*/ && 1829 !_irreducible && 1830 // Also ignore the occasional dead backedge 1831 !tail()->is_top()) { 1832 adjust_loop_exit_prob(phase); 1833 } 1834 1835 // Gate unrolling, RCE and peeling efforts. 1836 if (!_child && // If not an inner loop, do not split 1837 !_irreducible && 1838 _allow_optimizations && 1839 !tail()->is_top()) { // Also ignore the occasional dead backedge 1840 if (!_has_call) { 1841 if (!iteration_split_impl(phase, old_new)) { 1842 return false; 1843 } 1844 } else if (policy_unswitching(phase)) { 1845 phase->do_unswitching(this, old_new); 1846 } 1847 } 1848 1849 // Minor offset re-organization to remove loop-fallout uses of 1850 // trip counter when there was no major reshaping. 1851 phase->reorg_offsets(this); 1852 1853 if (_next && !_next->iteration_split(phase, old_new)) 1854 return false; 1855 return true; 1856 } 1857 1858 //-------------------------------is_uncommon_trap_proj---------------------------- 1859 // Return true if proj is the form of "proj->[region->..]call_uct" 1860 bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) { 1861 int path_limit = 10; 1862 assert(proj, "invalid argument"); 1863 Node* out = proj; 1864 for (int ct = 0; ct < path_limit; ct++) { 1865 out = out->unique_ctrl_out(); 1866 if (out == NULL || out->is_Root() || out->is_Start()) 1867 return false; 1868 if (out->is_CallStaticJava()) { 1869 int req = out->as_CallStaticJava()->uncommon_trap_request(); 1870 if (req != 0) { 1871 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req); 1872 if (trap_reason == reason || reason == Deoptimization::Reason_none) { 1873 return true; 1874 } 1875 } 1876 return false; // don't do further after call 1877 } 1878 } 1879 return false; 1880 } 1881 1882 //-------------------------------is_uncommon_trap_if_pattern------------------------- 1883 // Return true for "if(test)-> proj -> ... 1884 // | 1885 // V 1886 // other_proj->[region->..]call_uct" 1887 // 1888 // "must_reason_predicate" means the uct reason must be Reason_predicate 1889 bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) { 1890 Node *in0 = proj->in(0); 1891 if (!in0->is_If()) return false; 1892 // Variation of a dead If node. 1893 if (in0->outcnt() < 2) return false; 1894 IfNode* iff = in0->as_If(); 1895 1896 // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate 1897 if (reason != Deoptimization::Reason_none) { 1898 if (iff->in(1)->Opcode() != Op_Conv2B || 1899 iff->in(1)->in(1)->Opcode() != Op_Opaque1) { 1900 return false; 1901 } 1902 } 1903 1904 ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj(); 1905 return is_uncommon_trap_proj(other_proj, reason); 1906 } 1907 1908 //-------------------------------register_control------------------------- 1909 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) { 1910 assert(n->is_CFG(), "must be control node"); 1911 _igvn.register_new_node_with_optimizer(n); 1912 loop->_body.push(n); 1913 set_loop(n, loop); 1914 // When called from beautify_loops() idom is not constructed yet. 1915 if (_idom != NULL) { 1916 set_idom(n, pred, dom_depth(pred)); 1917 } 1918 } 1919 1920 //------------------------------create_new_if_for_predicate------------------------ 1921 // create a new if above the uct_if_pattern for the predicate to be promoted. 1922 // 1923 // before after 1924 // ---------- ---------- 1925 // ctrl ctrl 1926 // | | 1927 // | | 1928 // v v 1929 // iff new_iff 1930 // / \ / \ 1931 // / \ / \ 1932 // v v v v 1933 // uncommon_proj cont_proj if_uct if_cont 1934 // \ | | | | 1935 // \ | | | | 1936 // v v v | v 1937 // rgn loop | iff 1938 // | | / \ 1939 // | | / \ 1940 // v | v v 1941 // uncommon_trap | uncommon_proj cont_proj 1942 // \ \ | | 1943 // \ \ | | 1944 // v v v v 1945 // rgn loop 1946 // | 1947 // | 1948 // v 1949 // uncommon_trap 1950 // 1951 // 1952 // We will create a region to guard the uct call if there is no one there. 1953 // The true projecttion (if_cont) of the new_iff is returned. 1954 // This code is also used to clone predicates to clonned loops. 1955 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, 1956 Deoptimization::DeoptReason reason) { 1957 assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!"); 1958 IfNode* iff = cont_proj->in(0)->as_If(); 1959 1960 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); 1961 Node *rgn = uncommon_proj->unique_ctrl_out(); 1962 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); 1963 1964 if (!rgn->is_Region()) { // create a region to guard the call 1965 assert(rgn->is_Call(), "must be call uct"); 1966 CallNode* call = rgn->as_Call(); 1967 IdealLoopTree* loop = get_loop(call); 1968 rgn = new (C, 1) RegionNode(1); 1969 rgn->add_req(uncommon_proj); 1970 register_control(rgn, loop, uncommon_proj); 1971 _igvn.hash_delete(call); 1972 call->set_req(0, rgn); 1973 // When called from beautify_loops() idom is not constructed yet. 1974 if (_idom != NULL) { 1975 set_idom(call, rgn, dom_depth(rgn)); 1976 } 1977 } 1978 1979 Node* entry = iff->in(0); 1980 if (new_entry != NULL) { 1981 // Clonning the predicate to new location. 1982 entry = new_entry; 1983 } 1984 // Create new_iff 1985 IdealLoopTree* lp = get_loop(entry); 1986 IfNode *new_iff = new (C, 2) IfNode(entry, NULL, iff->_prob, iff->_fcnt); 1987 register_control(new_iff, lp, entry); 1988 Node *if_cont = new (C, 1) IfTrueNode(new_iff); 1989 Node *if_uct = new (C, 1) IfFalseNode(new_iff); 1990 if (cont_proj->is_IfFalse()) { 1991 // Swap 1992 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; 1993 } 1994 register_control(if_cont, lp, new_iff); 1995 register_control(if_uct, get_loop(rgn), new_iff); 1996 1997 // if_uct to rgn 1998 _igvn.hash_delete(rgn); 1999 rgn->add_req(if_uct); 2000 // When called from beautify_loops() idom is not constructed yet. 2001 if (_idom != NULL) { 2002 Node* ridom = idom(rgn); 2003 Node* nrdom = dom_lca(ridom, new_iff); 2004 set_idom(rgn, nrdom, dom_depth(rgn)); 2005 } 2006 // rgn must have no phis 2007 assert(!rgn->as_Region()->has_phi(), "region must have no phis"); 2008 2009 if (new_entry == NULL) { 2010 // Attach if_cont to iff 2011 _igvn.hash_delete(iff); 2012 iff->set_req(0, if_cont); 2013 if (_idom != NULL) { 2014 set_idom(iff, if_cont, dom_depth(iff)); 2015 } 2016 } 2017 return if_cont->as_Proj(); 2018 } 2019 2020 //--------------------------find_predicate_insertion_point------------------- 2021 // Find a good location to insert a predicate 2022 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) { 2023 if (start_c == NULL || !start_c->is_Proj()) 2024 return NULL; 2025 if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) { 2026 return start_c->as_Proj(); 2027 } 2028 return NULL; 2029 } 2030 2031 //--------------------------find_predicate------------------------------------ 2032 // Find a predicate 2033 Node* PhaseIdealLoop::find_predicate(Node* entry) { 2034 Node* predicate = NULL; 2035 if (UseLoopPredicate) { 2036 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 2037 if (predicate != NULL) { // right pattern that can be used by loop predication 2038 assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); 2039 return entry; 2040 } 2041 } 2042 return NULL; 2043 } 2044 2045 //------------------------------Invariance----------------------------------- 2046 // Helper class for loop_predication_impl to compute invariance on the fly and 2047 // clone invariants. 2048 class Invariance : public StackObj { 2049 VectorSet _visited, _invariant; 2050 Node_Stack _stack; 2051 VectorSet _clone_visited; 2052 Node_List _old_new; // map of old to new (clone) 2053 IdealLoopTree* _lpt; 2054 PhaseIdealLoop* _phase; 2055 2056 // Helper function to set up the invariance for invariance computation 2057 // If n is a known invariant, set up directly. Otherwise, look up the 2058 // the possibility to push n onto the stack for further processing. 2059 void visit(Node* use, Node* n) { 2060 if (_lpt->is_invariant(n)) { // known invariant 2061 _invariant.set(n->_idx); 2062 } else if (!n->is_CFG()) { 2063 Node *n_ctrl = _phase->ctrl_or_self(n); 2064 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG 2065 if (_phase->is_dominator(n_ctrl, u_ctrl)) { 2066 _stack.push(n, n->in(0) == NULL ? 1 : 0); 2067 } 2068 } 2069 } 2070 2071 // Compute invariance for "the_node" and (possibly) all its inputs recursively 2072 // on the fly 2073 void compute_invariance(Node* n) { 2074 assert(_visited.test(n->_idx), "must be"); 2075 visit(n, n); 2076 while (_stack.is_nonempty()) { 2077 Node* n = _stack.node(); 2078 uint idx = _stack.index(); 2079 if (idx == n->req()) { // all inputs are processed 2080 _stack.pop(); 2081 // n is invariant if it's inputs are all invariant 2082 bool all_inputs_invariant = true; 2083 for (uint i = 0; i < n->req(); i++) { 2084 Node* in = n->in(i); 2085 if (in == NULL) continue; 2086 assert(_visited.test(in->_idx), "must have visited input"); 2087 if (!_invariant.test(in->_idx)) { // bad guy 2088 all_inputs_invariant = false; 2089 break; 2090 } 2091 } 2092 if (all_inputs_invariant) { 2093 _invariant.set(n->_idx); // I am a invariant too 2094 } 2095 } else { // process next input 2096 _stack.set_index(idx + 1); 2097 Node* m = n->in(idx); 2098 if (m != NULL && !_visited.test_set(m->_idx)) { 2099 visit(n, m); 2100 } 2101 } 2102 } 2103 } 2104 2105 // Helper function to set up _old_new map for clone_nodes. 2106 // If n is a known invariant, set up directly ("clone" of n == n). 2107 // Otherwise, push n onto the stack for real cloning. 2108 void clone_visit(Node* n) { 2109 assert(_invariant.test(n->_idx), "must be invariant"); 2110 if (_lpt->is_invariant(n)) { // known invariant 2111 _old_new.map(n->_idx, n); 2112 } else{ // to be cloned 2113 assert (!n->is_CFG(), "should not see CFG here"); 2114 _stack.push(n, n->in(0) == NULL ? 1 : 0); 2115 } 2116 } 2117 2118 // Clone "n" and (possibly) all its inputs recursively 2119 void clone_nodes(Node* n, Node* ctrl) { 2120 clone_visit(n); 2121 while (_stack.is_nonempty()) { 2122 Node* n = _stack.node(); 2123 uint idx = _stack.index(); 2124 if (idx == n->req()) { // all inputs processed, clone n! 2125 _stack.pop(); 2126 // clone invariant node 2127 Node* n_cl = n->clone(); 2128 _old_new.map(n->_idx, n_cl); 2129 _phase->register_new_node(n_cl, ctrl); 2130 for (uint i = 0; i < n->req(); i++) { 2131 Node* in = n_cl->in(i); 2132 if (in == NULL) continue; 2133 n_cl->set_req(i, _old_new[in->_idx]); 2134 } 2135 } else { // process next input 2136 _stack.set_index(idx + 1); 2137 Node* m = n->in(idx); 2138 if (m != NULL && !_clone_visited.test_set(m->_idx)) { 2139 clone_visit(m); // visit the input 2140 } 2141 } 2142 } 2143 } 2144 2145 public: 2146 Invariance(Arena* area, IdealLoopTree* lpt) : 2147 _lpt(lpt), _phase(lpt->_phase), 2148 _visited(area), _invariant(area), _stack(area, 10 /* guess */), 2149 _clone_visited(area), _old_new(area) 2150 {} 2151 2152 // Map old to n for invariance computation and clone 2153 void map_ctrl(Node* old, Node* n) { 2154 assert(old->is_CFG() && n->is_CFG(), "must be"); 2155 _old_new.map(old->_idx, n); // "clone" of old is n 2156 _invariant.set(old->_idx); // old is invariant 2157 _clone_visited.set(old->_idx); 2158 } 2159 2160 // Driver function to compute invariance 2161 bool is_invariant(Node* n) { 2162 if (!_visited.test_set(n->_idx)) 2163 compute_invariance(n); 2164 return (_invariant.test(n->_idx) != 0); 2165 } 2166 2167 // Driver function to clone invariant 2168 Node* clone(Node* n, Node* ctrl) { 2169 assert(ctrl->is_CFG(), "must be"); 2170 assert(_invariant.test(n->_idx), "must be an invariant"); 2171 if (!_clone_visited.test(n->_idx)) 2172 clone_nodes(n, ctrl); 2173 return _old_new[n->_idx]; 2174 } 2175 }; 2176 2177 //------------------------------is_range_check_if ----------------------------------- 2178 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format 2179 // Note: this function is particularly designed for loop predication. We require load_range 2180 // and offset to be loop invariant computed on the fly by "invar" 2181 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { 2182 if (!is_loop_exit(iff)) { 2183 return false; 2184 } 2185 if (!iff->in(1)->is_Bool()) { 2186 return false; 2187 } 2188 const BoolNode *bol = iff->in(1)->as_Bool(); 2189 if (bol->_test._test != BoolTest::lt) { 2190 return false; 2191 } 2192 if (!bol->in(1)->is_Cmp()) { 2193 return false; 2194 } 2195 const CmpNode *cmp = bol->in(1)->as_Cmp(); 2196 if (cmp->Opcode() != Op_CmpU ) { 2197 return false; 2198 } 2199 Node* range = cmp->in(2); 2200 if (range->Opcode() != Op_LoadRange) { 2201 const TypeInt* tint = phase->_igvn.type(range)->isa_int(); 2202 if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) { 2203 // Allow predication on positive values that aren't LoadRanges. 2204 // This allows optimization of loops where the length of the 2205 // array is a known value and doesn't need to be loaded back 2206 // from the array. 2207 return false; 2208 } 2209 } 2210 if (!invar.is_invariant(range)) { 2211 return false; 2212 } 2213 Node *iv = _head->as_CountedLoop()->phi(); 2214 int scale = 0; 2215 Node *offset = NULL; 2216 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { 2217 return false; 2218 } 2219 if(offset && !invar.is_invariant(offset)) { // offset must be invariant 2220 return false; 2221 } 2222 return true; 2223 } 2224 2225 //------------------------------rc_predicate----------------------------------- 2226 // Create a range check predicate 2227 // 2228 // for (i = init; i < limit; i += stride) { 2229 // a[scale*i+offset] 2230 // } 2231 // 2232 // Compute max(scale*i + offset) for init <= i < limit and build the predicate 2233 // as "max(scale*i + offset) u< a.length". 2234 // 2235 // There are two cases for max(scale*i + offset): 2236 // (1) stride*scale > 0 2237 // max(scale*i + offset) = scale*(limit-stride) + offset 2238 // (2) stride*scale < 0 2239 // max(scale*i + offset) = scale*init + offset 2240 BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl, 2241 int scale, Node* offset, 2242 Node* init, Node* limit, Node* stride, 2243 Node* range, bool upper) { 2244 DEBUG_ONLY(ttyLocker ttyl); 2245 if (TraceLoopPredicate) tty->print("rc_predicate "); 2246 2247 Node* max_idx_expr = init; 2248 int stride_con = stride->get_int(); 2249 if ((stride_con > 0) == (scale > 0) == upper) { 2250 max_idx_expr = new (C, 3) SubINode(limit, stride); 2251 register_new_node(max_idx_expr, ctrl); 2252 if (TraceLoopPredicate) tty->print("(limit - stride) "); 2253 } else { 2254 if (TraceLoopPredicate) tty->print("init "); 2255 } 2256 2257 if (scale != 1) { 2258 ConNode* con_scale = _igvn.intcon(scale); 2259 max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale); 2260 register_new_node(max_idx_expr, ctrl); 2261 if (TraceLoopPredicate) tty->print("* %d ", scale); 2262 } 2263 2264 if (offset && (!offset->is_Con() || offset->get_int() != 0)){ 2265 max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset); 2266 register_new_node(max_idx_expr, ctrl); 2267 if (TraceLoopPredicate) 2268 if (offset->is_Con()) tty->print("+ %d ", offset->get_int()); 2269 else tty->print("+ offset "); 2270 } 2271 2272 CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range); 2273 register_new_node(cmp, ctrl); 2274 BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt); 2275 register_new_node(bol, ctrl); 2276 2277 if (TraceLoopPredicate) tty->print_cr("<u range"); 2278 return bol; 2279 } 2280 2281 //------------------------------ loop_predication_impl-------------------------- 2282 // Insert loop predicates for null checks and range checks 2283 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) { 2284 if (!UseLoopPredicate) return false; 2285 2286 if (!loop->_head->is_Loop()) { 2287 // Could be a simple region when irreducible loops are present. 2288 return false; 2289 } 2290 2291 if (loop->_head->unique_ctrl_out()->Opcode() == Op_NeverBranch) { 2292 // do nothing for infinite loops 2293 return false; 2294 } 2295 2296 CountedLoopNode *cl = NULL; 2297 if (loop->_head->is_CountedLoop()) { 2298 cl = loop->_head->as_CountedLoop(); 2299 // do nothing for iteration-splitted loops 2300 if (!cl->is_normal_loop()) return false; 2301 } 2302 2303 LoopNode *lpn = loop->_head->as_Loop(); 2304 Node* entry = lpn->in(LoopNode::EntryControl); 2305 2306 ProjNode *predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 2307 if (!predicate_proj) { 2308 #ifndef PRODUCT 2309 if (TraceLoopPredicate) { 2310 tty->print("missing predicate:"); 2311 loop->dump_head(); 2312 lpn->dump(1); 2313 } 2314 #endif 2315 return false; 2316 } 2317 ConNode* zero = _igvn.intcon(0); 2318 set_ctrl(zero, C->root()); 2319 2320 ResourceArea *area = Thread::current()->resource_area(); 2321 Invariance invar(area, loop); 2322 2323 // Create list of if-projs such that a newer proj dominates all older 2324 // projs in the list, and they all dominate loop->tail() 2325 Node_List if_proj_list(area); 2326 LoopNode *head = loop->_head->as_Loop(); 2327 Node *current_proj = loop->tail(); //start from tail 2328 while ( current_proj != head ) { 2329 if (loop == get_loop(current_proj) && // still in the loop ? 2330 current_proj->is_Proj() && // is a projection ? 2331 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ? 2332 if_proj_list.push(current_proj); 2333 } 2334 current_proj = idom(current_proj); 2335 } 2336 2337 bool hoisted = false; // true if at least one proj is promoted 2338 while (if_proj_list.size() > 0) { 2339 // Following are changed to nonnull when a predicate can be hoisted 2340 ProjNode* new_predicate_proj = NULL; 2341 2342 ProjNode* proj = if_proj_list.pop()->as_Proj(); 2343 IfNode* iff = proj->in(0)->as_If(); 2344 2345 if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) { 2346 if (loop->is_loop_exit(iff)) { 2347 // stop processing the remaining projs in the list because the execution of them 2348 // depends on the condition of "iff" (iff->in(1)). 2349 break; 2350 } else { 2351 // Both arms are inside the loop. There are two cases: 2352 // (1) there is one backward branch. In this case, any remaining proj 2353 // in the if_proj list post-dominates "iff". So, the condition of "iff" 2354 // does not determine the execution the remining projs directly, and we 2355 // can safely continue. 2356 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" 2357 // does not dominate loop->tail(), so it can not be in the if_proj list. 2358 continue; 2359 } 2360 } 2361 2362 Node* test = iff->in(1); 2363 if (!test->is_Bool()){ //Conv2B, ... 2364 continue; 2365 } 2366 BoolNode* bol = test->as_Bool(); 2367 if (invar.is_invariant(bol)) { 2368 // Invariant test 2369 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL, 2370 Deoptimization::Reason_predicate); 2371 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); 2372 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); 2373 2374 // Negate test if necessary 2375 bool negated = false; 2376 if (proj->_con != predicate_proj->_con) { 2377 new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); 2378 register_new_node(new_predicate_bol, ctrl); 2379 negated = true; 2380 } 2381 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); 2382 _igvn.hash_delete(new_predicate_iff); 2383 new_predicate_iff->set_req(1, new_predicate_bol); 2384 #ifndef PRODUCT 2385 if (TraceLoopPredicate) { 2386 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx); 2387 loop->dump_head(); 2388 } else if (TraceLoopOpts) { 2389 tty->print("Predicate IC "); 2390 loop->dump_head(); 2391 } 2392 #endif 2393 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) { 2394 assert(proj->_con == predicate_proj->_con, "must match"); 2395 2396 // Range check for counted loops 2397 const Node* cmp = bol->in(1)->as_Cmp(); 2398 Node* idx = cmp->in(1); 2399 assert(!invar.is_invariant(idx), "index is variant"); 2400 assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be"); 2401 Node* rng = cmp->in(2); 2402 assert(invar.is_invariant(rng), "range must be invariant"); 2403 int scale = 1; 2404 Node* offset = zero; 2405 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); 2406 assert(ok, "must be index expression"); 2407 2408 Node* init = cl->init_trip(); 2409 Node* limit = cl->limit(); 2410 Node* stride = cl->stride(); 2411 2412 // Build if's for the upper and lower bound tests. The 2413 // lower_bound test will dominate the upper bound test and all 2414 // cloned or created nodes will use the lower bound test as 2415 // their declared control. 2416 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); 2417 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate); 2418 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate"); 2419 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0); 2420 2421 // Perform cloning to keep Invariance state correct since the 2422 // late schedule will place invariant things in the loop. 2423 rng = invar.clone(rng, ctrl); 2424 if (offset && offset != zero) { 2425 assert(invar.is_invariant(offset), "offset must be loop invariant"); 2426 offset = invar.clone(offset, ctrl); 2427 } 2428 2429 // Test the lower bound 2430 Node* lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false); 2431 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); 2432 _igvn.hash_delete(lower_bound_iff); 2433 lower_bound_iff->set_req(1, lower_bound_bol); 2434 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx); 2435 2436 // Test the upper bound 2437 Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true); 2438 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); 2439 _igvn.hash_delete(upper_bound_iff); 2440 upper_bound_iff->set_req(1, upper_bound_bol); 2441 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx); 2442 2443 // Fall through into rest of the clean up code which will move 2444 // any dependent nodes onto the upper bound test. 2445 new_predicate_proj = upper_bound_proj; 2446 2447 #ifndef PRODUCT 2448 if (TraceLoopOpts && !TraceLoopPredicate) { 2449 tty->print("Predicate RC "); 2450 loop->dump_head(); 2451 } 2452 #endif 2453 } else { 2454 // Loop variant check (for example, range check in non-counted loop) 2455 // with uncommon trap. 2456 continue; 2457 } 2458 assert(new_predicate_proj != NULL, "sanity"); 2459 // Success - attach condition (new_predicate_bol) to predicate if 2460 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate 2461 2462 // Eliminate the old If in the loop body 2463 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con ); 2464 2465 hoisted = true; 2466 C->set_major_progress(); 2467 } // end while 2468 2469 #ifndef PRODUCT 2470 // report that the loop predication has been actually performed 2471 // for this loop 2472 if (TraceLoopPredicate && hoisted) { 2473 tty->print("Loop Predication Performed:"); 2474 loop->dump_head(); 2475 } 2476 #endif 2477 2478 return hoisted; 2479 } 2480 2481 //------------------------------loop_predication-------------------------------- 2482 // driver routine for loop predication optimization 2483 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { 2484 bool hoisted = false; 2485 // Recursively promote predicates 2486 if ( _child ) { 2487 hoisted = _child->loop_predication( phase); 2488 } 2489 2490 // self 2491 if (!_irreducible && !tail()->is_top()) { 2492 hoisted |= phase->loop_predication_impl(this); 2493 } 2494 2495 if ( _next ) { //sibling 2496 hoisted |= _next->loop_predication( phase); 2497 } 2498 2499 return hoisted; 2500 } 2501 2502 2503 // Process all the loops in the loop tree and replace any fill 2504 // patterns with an intrisc version. 2505 bool PhaseIdealLoop::do_intrinsify_fill() { 2506 bool changed = false; 2507 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2508 IdealLoopTree* lpt = iter.current(); 2509 changed |= intrinsify_fill(lpt); 2510 } 2511 return changed; 2512 } 2513 2514 2515 // Examine an inner loop looking for a a single store of an invariant 2516 // value in a unit stride loop, 2517 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, 2518 Node*& shift, Node*& con) { 2519 const char* msg = NULL; 2520 Node* msg_node = NULL; 2521 2522 store_value = NULL; 2523 con = NULL; 2524 shift = NULL; 2525 2526 // Process the loop looking for stores. If there are multiple 2527 // stores or extra control flow give at this point. 2528 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2529 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2530 Node* n = lpt->_body.at(i); 2531 if (n->outcnt() == 0) continue; // Ignore dead 2532 if (n->is_Store()) { 2533 if (store != NULL) { 2534 msg = "multiple stores"; 2535 break; 2536 } 2537 int opc = n->Opcode(); 2538 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) { 2539 msg = "oop fills not handled"; 2540 break; 2541 } 2542 Node* value = n->in(MemNode::ValueIn); 2543 if (!lpt->is_invariant(value)) { 2544 msg = "variant store value"; 2545 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { 2546 msg = "not array address"; 2547 } 2548 store = n; 2549 store_value = value; 2550 } else if (n->is_If() && n != head->loopexit()) { 2551 msg = "extra control flow"; 2552 msg_node = n; 2553 } 2554 } 2555 2556 if (store == NULL) { 2557 // No store in loop 2558 return false; 2559 } 2560 2561 if (msg == NULL && head->stride_con() != 1) { 2562 // could handle negative strides too 2563 if (head->stride_con() < 0) { 2564 msg = "negative stride"; 2565 } else { 2566 msg = "non-unit stride"; 2567 } 2568 } 2569 2570 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { 2571 msg = "can't handle store address"; 2572 msg_node = store->in(MemNode::Address); 2573 } 2574 2575 if (msg == NULL && 2576 (!store->in(MemNode::Memory)->is_Phi() || 2577 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { 2578 msg = "store memory isn't proper phi"; 2579 msg_node = store->in(MemNode::Memory); 2580 } 2581 2582 // Make sure there is an appropriate fill routine 2583 BasicType t = store->as_Mem()->memory_type(); 2584 const char* fill_name; 2585 if (msg == NULL && 2586 StubRoutines::select_fill_function(t, false, fill_name) == NULL) { 2587 msg = "unsupported store"; 2588 msg_node = store; 2589 } 2590 2591 if (msg != NULL) { 2592 #ifndef PRODUCT 2593 if (TraceOptimizeFill) { 2594 tty->print_cr("not fill intrinsic candidate: %s", msg); 2595 if (msg_node != NULL) msg_node->dump(); 2596 } 2597 #endif 2598 return false; 2599 } 2600 2601 // Make sure the address expression can be handled. It should be 2602 // head->phi * elsize + con. head->phi might have a ConvI2L. 2603 Node* elements[4]; 2604 Node* conv = NULL; 2605 bool found_index = false; 2606 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); 2607 for (int e = 0; e < count; e++) { 2608 Node* n = elements[e]; 2609 if (n->is_Con() && con == NULL) { 2610 con = n; 2611 } else if (n->Opcode() == Op_LShiftX && shift == NULL) { 2612 Node* value = n->in(1); 2613 #ifdef _LP64 2614 if (value->Opcode() == Op_ConvI2L) { 2615 conv = value; 2616 value = value->in(1); 2617 } 2618 #endif 2619 if (value != head->phi()) { 2620 msg = "unhandled shift in address"; 2621 } else { 2622 found_index = true; 2623 shift = n; 2624 assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int(), "scale should match"); 2625 } 2626 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { 2627 if (n->in(1) == head->phi()) { 2628 found_index = true; 2629 conv = n; 2630 } else { 2631 msg = "unhandled input to ConvI2L"; 2632 } 2633 } else if (n == head->phi()) { 2634 // no shift, check below for allowed cases 2635 found_index = true; 2636 } else { 2637 msg = "unhandled node in address"; 2638 msg_node = n; 2639 } 2640 } 2641 2642 if (count == -1) { 2643 msg = "malformed address expression"; 2644 msg_node = store; 2645 } 2646 2647 if (!found_index) { 2648 msg = "missing use of index"; 2649 } 2650 2651 // byte sized items won't have a shift 2652 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { 2653 msg = "can't find shift"; 2654 msg_node = store; 2655 } 2656 2657 if (msg != NULL) { 2658 #ifndef PRODUCT 2659 if (TraceOptimizeFill) { 2660 tty->print_cr("not fill intrinsic: %s", msg); 2661 if (msg_node != NULL) msg_node->dump(); 2662 } 2663 #endif 2664 return false; 2665 } 2666 2667 // No make sure all the other nodes in the loop can be handled 2668 VectorSet ok(Thread::current()->resource_area()); 2669 2670 // store related values are ok 2671 ok.set(store->_idx); 2672 ok.set(store->in(MemNode::Memory)->_idx); 2673 2674 // Loop structure is ok 2675 ok.set(head->_idx); 2676 ok.set(head->loopexit()->_idx); 2677 ok.set(head->phi()->_idx); 2678 ok.set(head->incr()->_idx); 2679 ok.set(head->loopexit()->cmp_node()->_idx); 2680 ok.set(head->loopexit()->in(1)->_idx); 2681 2682 // Address elements are ok 2683 if (con) ok.set(con->_idx); 2684 if (shift) ok.set(shift->_idx); 2685 if (conv) ok.set(conv->_idx); 2686 2687 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2688 Node* n = lpt->_body.at(i); 2689 if (n->outcnt() == 0) continue; // Ignore dead 2690 if (ok.test(n->_idx)) continue; 2691 // Backedge projection is ok 2692 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue; 2693 if (!n->is_AddP()) { 2694 msg = "unhandled node"; 2695 msg_node = n; 2696 break; 2697 } 2698 } 2699 2700 // Make sure no unexpected values are used outside the loop 2701 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { 2702 Node* n = lpt->_body.at(i); 2703 // These values can be replaced with other nodes if they are used 2704 // outside the loop. 2705 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue; 2706 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { 2707 Node* use = iter.get(); 2708 if (!lpt->_body.contains(use)) { 2709 msg = "node is used outside loop"; 2710 // lpt->_body.dump(); 2711 msg_node = n; 2712 break; 2713 } 2714 } 2715 } 2716 2717 #ifdef ASSERT 2718 if (TraceOptimizeFill) { 2719 if (msg != NULL) { 2720 tty->print_cr("no fill intrinsic: %s", msg); 2721 if (msg_node != NULL) msg_node->dump(); 2722 } else { 2723 tty->print_cr("fill intrinsic for:"); 2724 } 2725 store->dump(); 2726 if (Verbose) { 2727 lpt->_body.dump(); 2728 } 2729 } 2730 #endif 2731 2732 return msg == NULL; 2733 } 2734 2735 2736 2737 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { 2738 // Only for counted inner loops 2739 if (!lpt->is_counted() || !lpt->is_inner()) { 2740 return false; 2741 } 2742 2743 // Must have constant stride 2744 CountedLoopNode* head = lpt->_head->as_CountedLoop(); 2745 if (!head->stride_is_con() || !head->is_normal_loop()) { 2746 return false; 2747 } 2748 2749 // Check that the body only contains a store of a loop invariant 2750 // value that is indexed by the loop phi. 2751 Node* store = NULL; 2752 Node* store_value = NULL; 2753 Node* shift = NULL; 2754 Node* offset = NULL; 2755 if (!match_fill_loop(lpt, store, store_value, shift, offset)) { 2756 return false; 2757 } 2758 2759 // Now replace the whole loop body by a call to a fill routine that 2760 // covers the same region as the loop. 2761 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); 2762 2763 // Build an expression for the beginning of the copy region 2764 Node* index = head->init_trip(); 2765 #ifdef _LP64 2766 index = new (C, 2) ConvI2LNode(index); 2767 _igvn.register_new_node_with_optimizer(index); 2768 #endif 2769 if (shift != NULL) { 2770 // byte arrays don't require a shift but others do. 2771 index = new (C, 3) LShiftXNode(index, shift->in(2)); 2772 _igvn.register_new_node_with_optimizer(index); 2773 } 2774 index = new (C, 4) AddPNode(base, base, index); 2775 _igvn.register_new_node_with_optimizer(index); 2776 Node* from = new (C, 4) AddPNode(base, index, offset); 2777 _igvn.register_new_node_with_optimizer(from); 2778 // Compute the number of elements to copy 2779 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip()); 2780 _igvn.register_new_node_with_optimizer(len); 2781 2782 BasicType t = store->as_Mem()->memory_type(); 2783 bool aligned = false; 2784 if (offset != NULL && head->init_trip()->is_Con()) { 2785 int element_size = type2aelembytes(t); 2786 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; 2787 } 2788 2789 // Build a call to the fill routine 2790 const char* fill_name; 2791 address fill = StubRoutines::select_fill_function(t, aligned, fill_name); 2792 assert(fill != NULL, "what?"); 2793 2794 // Convert float/double to int/long for fill routines 2795 if (t == T_FLOAT) { 2796 store_value = new (C, 2) MoveF2INode(store_value); 2797 _igvn.register_new_node_with_optimizer(store_value); 2798 } else if (t == T_DOUBLE) { 2799 store_value = new (C, 2) MoveD2LNode(store_value); 2800 _igvn.register_new_node_with_optimizer(store_value); 2801 } 2802 2803 Node* mem_phi = store->in(MemNode::Memory); 2804 Node* result_ctrl; 2805 Node* result_mem; 2806 const TypeFunc* call_type = OptoRuntime::array_fill_Type(); 2807 int size = call_type->domain()->cnt(); 2808 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill, 2809 fill_name, TypeAryPtr::get_array_body_type(t)); 2810 call->init_req(TypeFunc::Parms+0, from); 2811 call->init_req(TypeFunc::Parms+1, store_value); 2812 #ifdef _LP64 2813 len = new (C, 2) ConvI2LNode(len); 2814 _igvn.register_new_node_with_optimizer(len); 2815 #endif 2816 call->init_req(TypeFunc::Parms+2, len); 2817 #ifdef _LP64 2818 call->init_req(TypeFunc::Parms+3, C->top()); 2819 #endif 2820 call->init_req( TypeFunc::Control, head->init_control()); 2821 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o 2822 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) ); 2823 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) ); 2824 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) ); 2825 _igvn.register_new_node_with_optimizer(call); 2826 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); 2827 _igvn.register_new_node_with_optimizer(result_ctrl); 2828 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory); 2829 _igvn.register_new_node_with_optimizer(result_mem); 2830 2831 // If this fill is tightly coupled to an allocation and overwrites 2832 // the whole body, allow it to take over the zeroing. 2833 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); 2834 if (alloc != NULL && alloc->is_AllocateArray()) { 2835 Node* length = alloc->as_AllocateArray()->Ideal_length(); 2836 if (head->limit() == length && 2837 head->init_trip() == _igvn.intcon(0)) { 2838 if (TraceOptimizeFill) { 2839 tty->print_cr("Eliminated zeroing in allocation"); 2840 } 2841 alloc->maybe_set_complete(&_igvn); 2842 } else { 2843 #ifdef ASSERT 2844 if (TraceOptimizeFill) { 2845 tty->print_cr("filling array but bounds don't match"); 2846 alloc->dump(); 2847 head->init_trip()->dump(); 2848 head->limit()->dump(); 2849 length->dump(); 2850 } 2851 #endif 2852 } 2853 } 2854 2855 // Redirect the old control and memory edges that are outside the loop. 2856 Node* exit = head->loopexit()->proj_out(0); 2857 // Sometimes the memory phi of the head is used as the outgoing 2858 // state of the loop. It's safe in this case to replace it with the 2859 // result_mem. 2860 _igvn.replace_node(store->in(MemNode::Memory), result_mem); 2861 _igvn.replace_node(exit, result_ctrl); 2862 _igvn.replace_node(store, result_mem); 2863 // Any uses the increment outside of the loop become the loop limit. 2864 _igvn.replace_node(head->incr(), head->limit()); 2865 2866 // Disconnect the head from the loop. 2867 for (uint i = 0; i < lpt->_body.size(); i++) { 2868 Node* n = lpt->_body.at(i); 2869 _igvn.replace_node(n, C->top()); 2870 } 2871 2872 return true; 2873 }