1 #ifdef USE_PRAGMA_IDENT_SRC 2 #pragma ident "@(#)loopTransform.cpp 1.116 07/06/01 11:35:03 JVM" 3 #endif 4 /* 5 * Copyright 2000-2007 Sun Microsystems, Inc. All Rights Reserved. 6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 7 * 8 * This code is free software; you can redistribute it and/or modify it 9 * under the terms of the GNU General Public License version 2 only, as 10 * published by the Free Software Foundation. 11 * 12 * This code is distributed in the hope that it will be useful, but WITHOUT 13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 * version 2 for more details (a copy is included in the LICENSE file that 16 * accompanied this code). 17 * 18 * You should have received a copy of the GNU General Public License version 19 * 2 along with this work; if not, write to the Free Software Foundation, 20 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 21 * 22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 23 * CA 95054 USA or visit www.sun.com if you need additional information or 24 * have any questions. 25 * 26 */ 27 28 #include "incls/_precompiled.incl" 29 #include "incls/_loopTransform.cpp.incl" 30 31 //------------------------------is_loop_exit----------------------------------- 32 // Given an IfNode, return the loop-exiting projection or NULL if both 33 // arms remain in the loop. 34 Node *IdealLoopTree::is_loop_exit(Node *iff) const { 35 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests 36 PhaseIdealLoop *phase = _phase; 37 // Test is an IfNode, has 2 projections. If BOTH are in the loop 38 // we need loop unswitching instead of peeling. 39 if( !is_member(phase->get_loop( iff->raw_out(0) )) ) 40 return iff->raw_out(0); 41 if( !is_member(phase->get_loop( iff->raw_out(1) )) ) 42 return iff->raw_out(1); 43 return NULL; 44 } 45 46 47 //============================================================================= 48 49 50 //------------------------------record_for_igvn---------------------------- 51 // Put loop body on igvn work list 52 void IdealLoopTree::record_for_igvn() { 53 for( uint i = 0; i < _body.size(); i++ ) { 54 Node *n = _body.at(i); 55 _phase->_igvn._worklist.push(n); 56 } 57 } 58 59 //------------------------------compute_profile_trip_cnt---------------------------- 60 // Compute loop trip count from profile data as 61 // (backedge_count + loop_exit_count) / loop_exit_count 62 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { 63 if (!_head->is_CountedLoop()) { 64 return; 65 } 66 CountedLoopNode* head = _head->as_CountedLoop(); 67 if (head->profile_trip_cnt() != COUNT_UNKNOWN) { 68 return; // Already computed 69 } 70 float trip_cnt = (float)max_jint; // default is big 71 72 Node* back = head->in(LoopNode::LoopBackControl); 73 while (back != head) { 74 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 75 back->in(0) && 76 back->in(0)->is_If() && 77 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && 78 back->in(0)->as_If()->_prob != PROB_UNKNOWN) { 79 break; 80 } 81 back = phase->idom(back); 82 } 83 if (back != head) { 84 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && 85 back->in(0), "if-projection exists"); 86 IfNode* back_if = back->in(0)->as_If(); 87 float loop_back_cnt = back_if->_fcnt * back_if->_prob; 88 89 // Now compute a loop exit count 90 float loop_exit_cnt = 0.0f; 91 for( uint i = 0; i < _body.size(); i++ ) { 92 Node *n = _body[i]; 93 if( n->is_If() ) { 94 IfNode *iff = n->as_If(); 95 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { 96 Node *exit = is_loop_exit(iff); 97 if( exit ) { 98 float exit_prob = iff->_prob; 99 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; 100 if (exit_prob > PROB_MIN) { 101 float exit_cnt = iff->_fcnt * exit_prob; 102 loop_exit_cnt += exit_cnt; 103 } 104 } 105 } 106 } 107 } 108 if (loop_exit_cnt > 0.0f) { 109 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; 110 } else { 111 // No exit count so use 112 trip_cnt = loop_back_cnt; 113 } 114 } 115 #ifndef PRODUCT 116 if (TraceProfileTripCount) { 117 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); 118 } 119 #endif 120 head->set_profile_trip_cnt(trip_cnt); 121 } 122 123 //---------------------is_invariant_addition----------------------------- 124 // Return nonzero index of invariant operand for an Add or Sub 125 // of (nonconstant) invariant and variant values. Helper for reassoicate_invariants. 126 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { 127 int op = n->Opcode(); 128 if (op == Op_AddI || op == Op_SubI) { 129 bool in1_invar = this->is_invariant(n->in(1)); 130 bool in2_invar = this->is_invariant(n->in(2)); 131 if (in1_invar && !in2_invar) return 1; 132 if (!in1_invar && in2_invar) return 2; 133 } 134 return 0; 135 } 136 137 //---------------------reassociate_add_sub----------------------------- 138 // Reassociate invariant add and subtract expressions: 139 // 140 // inv1 + (x + inv2) => ( inv1 + inv2) + x 141 // (x + inv2) + inv1 => ( inv1 + inv2) + x 142 // inv1 + (x - inv2) => ( inv1 - inv2) + x 143 // inv1 - (inv2 - x) => ( inv1 - inv2) + x 144 // (x + inv2) - inv1 => (-inv1 + inv2) + x 145 // (x - inv2) + inv1 => ( inv1 - inv2) + x 146 // (x - inv2) - inv1 => (-inv1 - inv2) + x 147 // inv1 + (inv2 - x) => ( inv1 + inv2) - x 148 // inv1 - (x - inv2) => ( inv1 + inv2) - x 149 // (inv2 - x) + inv1 => ( inv1 + inv2) - x 150 // (inv2 - x) - inv1 => (-inv1 + inv2) - x 151 // inv1 - (x + inv2) => ( inv1 - inv2) - x 152 // 153 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { 154 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; 155 if (is_invariant(n1)) return NULL; 156 int inv1_idx = is_invariant_addition(n1, phase); 157 if (!inv1_idx) return NULL; 158 // Don't mess with add of constant (igvn moves them to expression tree root.) 159 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; 160 Node* inv1 = n1->in(inv1_idx); 161 Node* n2 = n1->in(3 - inv1_idx); 162 int inv2_idx = is_invariant_addition(n2, phase); 163 if (!inv2_idx) return NULL; 164 Node* x = n2->in(3 - inv2_idx); 165 Node* inv2 = n2->in(inv2_idx); 166 167 bool neg_x = n2->is_Sub() && inv2_idx == 1; 168 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; 169 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; 170 if (n1->is_Sub() && inv1_idx == 1) { 171 neg_x = !neg_x; 172 neg_inv2 = !neg_inv2; 173 } 174 Node* inv1_c = phase->get_ctrl(inv1); 175 Node* inv2_c = phase->get_ctrl(inv2); 176 Node* n_inv1; 177 if (neg_inv1) { 178 Node *zero = phase->_igvn.intcon(0); 179 phase->set_ctrl(zero, phase->C->root()); 180 n_inv1 = new (phase->C, 3) SubINode(zero, inv1); 181 phase->register_new_node(n_inv1, inv1_c); 182 } else { 183 n_inv1 = inv1; 184 } 185 Node* inv; 186 if (neg_inv2) { 187 inv = new (phase->C, 3) SubINode(n_inv1, inv2); 188 } else { 189 inv = new (phase->C, 3) AddINode(n_inv1, inv2); 190 } 191 phase->register_new_node(inv, phase->get_early_ctrl(inv)); 192 193 Node* addx; 194 if (neg_x) { 195 addx = new (phase->C, 3) SubINode(inv, x); 196 } else { 197 addx = new (phase->C, 3) AddINode(x, inv); 198 } 199 phase->register_new_node(addx, phase->get_ctrl(x)); 200 phase->_igvn.hash_delete(n1); 201 phase->_igvn.subsume_node(n1, addx); 202 return addx; 203 } 204 205 //---------------------reassociate_invariants----------------------------- 206 // Reassociate invariant expressions: 207 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { 208 for (int i = _body.size() - 1; i >= 0; i--) { 209 Node *n = _body.at(i); 210 for (int j = 0; j < 5; j++) { 211 Node* nn = reassociate_add_sub(n, phase); 212 if (nn == NULL) break; 213 n = nn; // again 214 }; 215 } 216 } 217 218 //------------------------------policy_peeling--------------------------------- 219 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can 220 // make some loop-invariant test (usually a null-check) happen before the loop. 221 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { 222 Node *test = ((IdealLoopTree*)this)->tail(); 223 int body_size = ((IdealLoopTree*)this)->_body.size(); 224 int uniq = phase->C->unique(); 225 // Peeling does loop cloning which can result in O(N^2) node construction 226 if( body_size > 255 /* Prevent overflow for large body_size */ 227 || (body_size * body_size + uniq > MaxNodeLimit) ) { 228 return false; // too large to safely clone 229 } 230 while( test != _head ) { // Scan till run off top of loop 231 if( test->is_If() ) { // Test? 232 Node *ctrl = phase->get_ctrl(test->in(1)); 233 if (ctrl->is_top()) 234 return false; // Found dead test on live IF? No peeling! 235 // Standard IF only has one input value to check for loop invariance 236 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); 237 // Condition is not a member of this loop? 238 if( !is_member(phase->get_loop(ctrl)) && 239 is_loop_exit(test) ) 240 return true; // Found reason to peel! 241 } 242 // Walk up dominators to loop _head looking for test which is 243 // executed on every path thru loop. 244 test = phase->idom(test); 245 } 246 return false; 247 } 248 249 //------------------------------peeled_dom_test_elim--------------------------- 250 // If we got the effect of peeling, either by actually peeling or by making 251 // a pre-loop which must execute at least once, we can remove all 252 // loop-invariant dominated tests in the main body. 253 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { 254 bool progress = true; 255 while( progress ) { 256 progress = false; // Reset for next iteration 257 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); 258 Node *test = prev->in(0); 259 while( test != loop->_head ) { // Scan till run off top of loop 260 261 int p_op = prev->Opcode(); 262 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && 263 test->is_If() && // Test? 264 !test->in(1)->is_Con() && // And not already obvious? 265 // Condition is not a member of this loop? 266 !loop->is_member(get_loop(get_ctrl(test->in(1))))){ 267 // Walk loop body looking for instances of this test 268 for( uint i = 0; i < loop->_body.size(); i++ ) { 269 Node *n = loop->_body.at(i); 270 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { 271 // IfNode was dominated by version in peeled loop body 272 progress = true; 273 dominated_by( old_new[prev->_idx], n ); 274 } 275 } 276 } 277 prev = test; 278 test = idom(test); 279 } // End of scan tests in loop 280 281 } // End of while( progress ) 282 } 283 284 //------------------------------do_peeling------------------------------------- 285 // Peel the first iteration of the given loop. 286 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 287 // The pre-loop illegally has 2 control users (old & new loops). 288 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 289 // Do this by making the old-loop fall-in edges act as if they came 290 // around the loopback from the prior iteration (follow the old-loop 291 // backedges) and then map to the new peeled iteration. This leaves 292 // the pre-loop with only 1 user (the new peeled iteration), but the 293 // peeled-loop backedge has 2 users. 294 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 295 // extra backedge user. 296 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { 297 298 C->set_major_progress(); 299 // Peeling a 'main' loop in a pre/main/post situation obfuscates the 300 // 'pre' loop from the main and the 'pre' can no longer have it's 301 // iterations adjusted. Therefore, we need to declare this loop as 302 // no longer a 'main' loop; it will need new pre and post loops before 303 // we can do further RCE. 304 Node *h = loop->_head; 305 if( h->is_CountedLoop() ) { 306 CountedLoopNode *cl = h->as_CountedLoop(); 307 assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); 308 cl->set_trip_count(cl->trip_count() - 1); 309 if( cl->is_main_loop() ) { 310 cl->set_normal_loop(); 311 #ifndef PRODUCT 312 if( PrintOpto && VerifyLoopOptimizations ) { 313 tty->print("Peeling a 'main' loop; resetting to 'normal' "); 314 loop->dump_head(); 315 } 316 #endif 317 } 318 } 319 320 // Step 1: Clone the loop body. The clone becomes the peeled iteration. 321 // The pre-loop illegally has 2 control users (old & new loops). 322 clone_loop( loop, old_new, dom_depth(loop->_head) ); 323 324 325 // Step 2: Make the old-loop fall-in edges point to the peeled iteration. 326 // Do this by making the old-loop fall-in edges act as if they came 327 // around the loopback from the prior iteration (follow the old-loop 328 // backedges) and then map to the new peeled iteration. This leaves 329 // the pre-loop with only 1 user (the new peeled iteration), but the 330 // peeled-loop backedge has 2 users. 331 for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) { 332 Node* old = loop->_head->fast_out(j); 333 if( old->in(0) == loop->_head && old->req() == 3 && 334 (old->is_Loop() || old->is_Phi()) ) { 335 Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; 336 if( !new_exit_value ) // Backedge value is ALSO loop invariant? 337 // Then loop body backedge value remains the same. 338 new_exit_value = old->in(LoopNode::LoopBackControl); 339 _igvn.hash_delete(old); 340 old->set_req(LoopNode::EntryControl, new_exit_value); 341 } 342 } 343 344 345 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the 346 // extra backedge user. 347 Node *nnn = old_new[loop->_head->_idx]; 348 _igvn.hash_delete(nnn); 349 nnn->set_req(LoopNode::LoopBackControl, C->top()); 350 for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) { 351 Node* use = nnn->fast_out(j2); 352 if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) { 353 _igvn.hash_delete(use); 354 use->set_req(LoopNode::LoopBackControl, C->top()); 355 } 356 } 357 358 359 // Step 4: Correct dom-depth info. Set to loop-head depth. 360 int dd = dom_depth(loop->_head); 361 set_idom(loop->_head, loop->_head->in(1), dd); 362 for (uint j3 = 0; j3 < loop->_body.size(); j3++) { 363 Node *old = loop->_body.at(j3); 364 Node *nnn = old_new[old->_idx]; 365 if (!has_ctrl(nnn)) 366 set_idom(nnn, idom(nnn), dd-1); 367 // While we're at it, remove any SafePoints from the peeled code 368 if( old->Opcode() == Op_SafePoint ) { 369 Node *nnn = old_new[old->_idx]; 370 lazy_replace(nnn,nnn->in(TypeFunc::Control)); 371 } 372 } 373 374 // Now force out all loop-invariant dominating tests. The optimizer 375 // finds some, but we _know_ they are all useless. 376 peeled_dom_test_elim(loop,old_new); 377 378 loop->record_for_igvn(); 379 } 380 381 //------------------------------policy_maximally_unroll------------------------ 382 // Return exact loop trip count, or 0 if not maximally unrolling 383 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { 384 CountedLoopNode *cl = _head->as_CountedLoop(); 385 assert( cl->is_normal_loop(), "" ); 386 387 Node *init_n = cl->init_trip(); 388 Node *limit_n = cl->limit(); 389 390 // Non-constant bounds 391 if( init_n == NULL || !init_n->is_Con() || 392 limit_n == NULL || !limit_n->is_Con() || 393 // protect against stride not being a constant 394 !cl->stride_is_con() ) { 395 return false; 396 } 397 int init = init_n->get_int(); 398 int limit = limit_n->get_int(); 399 int span = limit - init; 400 int stride = cl->stride_con(); 401 402 if (init >= limit || stride > span) { 403 // return a false (no maximally unroll) and the regular unroll/peel 404 // route will make a small mess which CCP will fold away. 405 return false; 406 } 407 uint trip_count = span/stride; // trip_count can be greater than 2 Gig. 408 assert( (int)trip_count*stride == span, "must divide evenly" ); 409 410 // Real policy: if we maximally unroll, does it get too big? 411 // Allow the unrolled mess to get larger than standard loop 412 // size. After all, it will no longer be a loop. 413 uint body_size = _body.size(); 414 uint unroll_limit = (uint)LoopUnrollLimit * 4; 415 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); 416 cl->set_trip_count(trip_count); 417 if( trip_count <= unroll_limit && body_size <= unroll_limit ) { 418 uint new_body_size = body_size * trip_count; 419 if (new_body_size <= unroll_limit && 420 body_size == new_body_size / trip_count && 421 // Unrolling can result in a large amount of node construction 422 new_body_size < MaxNodeLimit - phase->C->unique()) { 423 return true; // maximally unroll 424 } 425 } 426 427 return false; // Do not maximally unroll 428 } 429 430 431 //------------------------------policy_unroll---------------------------------- 432 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if 433 // the loop is a CountedLoop and the body is small enough. 434 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { 435 436 CountedLoopNode *cl = _head->as_CountedLoop(); 437 assert( cl->is_normal_loop() || cl->is_main_loop(), "" ); 438 439 // protect against stride not being a constant 440 if( !cl->stride_is_con() ) return false; 441 442 // protect against over-unrolling 443 if( cl->trip_count() <= 1 ) return false; 444 445 int future_unroll_ct = cl->unrolled_count() * 2; 446 447 // Don't unroll if the next round of unrolling would push us 448 // over the expected trip count of the loop. One is subtracted 449 // from the expected trip count because the pre-loop normally 450 // executes 1 iteration. 451 if (UnrollLimitForProfileCheck > 0 && 452 cl->profile_trip_cnt() != COUNT_UNKNOWN && 453 future_unroll_ct > UnrollLimitForProfileCheck && 454 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { 455 return false; 456 } 457 458 // When unroll count is greater than LoopUnrollMin, don't unroll if: 459 // the residual iterations are more than 10% of the trip count 460 // and rounds of "unroll,optimize" are not making significant progress 461 // Progress defined as current size less than 20% larger than previous size. 462 if (UseSuperWord && cl->node_count_before_unroll() > 0 && 463 future_unroll_ct > LoopUnrollMin && 464 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && 465 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { 466 return false; 467 } 468 469 Node *init_n = cl->init_trip(); 470 Node *limit_n = cl->limit(); 471 // Non-constant bounds. 472 // Protect against over-unrolling when init or/and limit are not constant 473 // (so that trip_count's init value is maxint) but iv range is known. 474 if( init_n == NULL || !init_n->is_Con() || 475 limit_n == NULL || !limit_n->is_Con() ) { 476 Node* phi = cl->phi(); 477 if( phi != NULL ) { 478 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); 479 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); 480 int next_stride = cl->stride_con() * 2; // stride after this unroll 481 if( next_stride > 0 ) { 482 if( iv_type->_lo + next_stride <= iv_type->_lo || // overflow 483 iv_type->_lo + next_stride > iv_type->_hi ) { 484 return false; // over-unrolling 485 } 486 } else if( next_stride < 0 ) { 487 if( iv_type->_hi + next_stride >= iv_type->_hi || // overflow 488 iv_type->_hi + next_stride < iv_type->_lo ) { 489 return false; // over-unrolling 490 } 491 } 492 } 493 } 494 495 // Adjust body_size to determine if we unroll or not 496 uint body_size = _body.size(); 497 // Key test to unroll CaffeineMark's Logic test 498 int xors_in_loop = 0; 499 // Also count ModL, DivL and MulL which expand mightly 500 for( uint k = 0; k < _body.size(); k++ ) { 501 switch( _body.at(k)->Opcode() ) { 502 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test 503 case Op_ModL: body_size += 30; break; 504 case Op_DivL: body_size += 30; break; 505 case Op_MulL: body_size += 10; break; 506 } 507 } 508 509 // Check for being too big 510 if( body_size > (uint)LoopUnrollLimit ) { 511 if( xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; 512 // Normal case: loop too big 513 return false; 514 } 515 516 // Check for stride being a small enough constant 517 if( abs(cl->stride_con()) > (1<<3) ) return false; 518 519 // Unroll once! (Each trip will soon do double iterations) 520 return true; 521 } 522 523 //------------------------------policy_align----------------------------------- 524 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the 525 // expression that does the alignment. Note that only one array base can be 526 // aligned in a loop (unless the VM guarentees mutual alignment). Note that 527 // if we vectorize short memory ops into longer memory ops, we may want to 528 // increase alignment. 529 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { 530 return false; 531 } 532 533 //------------------------------policy_range_check----------------------------- 534 // Return TRUE or FALSE if the loop should be range-check-eliminated. 535 // Actually we do iteration-splitting, a more powerful form of RCE. 536 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { 537 if( !RangeCheckElimination ) return false; 538 539 CountedLoopNode *cl = _head->as_CountedLoop(); 540 // If we unrolled with no intention of doing RCE and we later 541 // changed our minds, we got no pre-loop. Either we need to 542 // make a new pre-loop, or we gotta disallow RCE. 543 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now. 544 Node *trip_counter = cl->phi(); 545 546 // Check loop body for tests of trip-counter plus loop-invariant vs 547 // loop-invariant. 548 for( uint i = 0; i < _body.size(); i++ ) { 549 Node *iff = _body[i]; 550 if( iff->Opcode() == Op_If ) { // Test? 551 552 // Comparing trip+off vs limit 553 Node *bol = iff->in(1); 554 if( bol->req() != 2 ) continue; // dead constant test 555 Node *cmp = bol->in(1); 556 557 Node *rc_exp = cmp->in(1); 558 Node *limit = cmp->in(2); 559 560 Node *limit_c = phase->get_ctrl(limit); 561 if( limit_c == phase->C->top() ) 562 return false; // Found dead test on live IF? No RCE! 563 if( is_member(phase->get_loop(limit_c) ) ) { 564 // Compare might have operands swapped; commute them 565 rc_exp = cmp->in(2); 566 limit = cmp->in(1); 567 limit_c = phase->get_ctrl(limit); 568 if( is_member(phase->get_loop(limit_c) ) ) 569 continue; // Both inputs are loop varying; cannot RCE 570 } 571 572 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { 573 continue; 574 } 575 // Yeah! Found a test like 'trip+off vs limit' 576 // Test is an IfNode, has 2 projections. If BOTH are in the loop 577 // we need loop unswitching instead of iteration splitting. 578 if( is_loop_exit(iff) ) 579 return true; // Found reason to split iterations 580 } // End of is IF 581 } 582 583 return false; 584 } 585 586 //------------------------------policy_peel_only------------------------------- 587 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful 588 // for unrolling loops with NO array accesses. 589 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { 590 591 for( uint i = 0; i < _body.size(); i++ ) 592 if( _body[i]->is_Mem() ) 593 return false; 594 595 // No memory accesses at all! 596 return true; 597 } 598 599 //------------------------------clone_up_backedge_goo-------------------------- 600 // If Node n lives in the back_ctrl block and cannot float, we clone a private 601 // version of n in preheader_ctrl block and return that, otherwise return n. 602 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) { 603 if( get_ctrl(n) != back_ctrl ) return n; 604 605 Node *x = NULL; // If required, a clone of 'n' 606 // Check for 'n' being pinned in the backedge. 607 if( n->in(0) && n->in(0) == back_ctrl ) { 608 x = n->clone(); // Clone a copy of 'n' to preheader 609 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader 610 } 611 612 // Recursive fixup any other input edges into x. 613 // If there are no changes we can just return 'n', otherwise 614 // we need to clone a private copy and change it. 615 for( uint i = 1; i < n->req(); i++ ) { 616 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) ); 617 if( g != n->in(i) ) { 618 if( !x ) 619 x = n->clone(); 620 x->set_req(i, g); 621 } 622 } 623 if( x ) { // x can legally float to pre-header location 624 register_new_node( x, preheader_ctrl ); 625 return x; 626 } else { // raise n to cover LCA of uses 627 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); 628 } 629 return n; 630 } 631 632 //------------------------------insert_pre_post_loops-------------------------- 633 // Insert pre and post loops. If peel_only is set, the pre-loop can not have 634 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no 635 // alignment. Useful to unroll loops that do no array accesses. 636 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { 637 638 C->set_major_progress(); 639 640 // Find common pieces of the loop being guarded with pre & post loops 641 CountedLoopNode *main_head = loop->_head->as_CountedLoop(); 642 assert( main_head->is_normal_loop(), "" ); 643 CountedLoopEndNode *main_end = main_head->loopexit(); 644 assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); 645 uint dd_main_head = dom_depth(main_head); 646 uint max = main_head->outcnt(); 647 648 Node *pre_header= main_head->in(LoopNode::EntryControl); 649 Node *init = main_head->init_trip(); 650 Node *incr = main_end ->incr(); 651 Node *limit = main_end ->limit(); 652 Node *stride = main_end ->stride(); 653 Node *cmp = main_end ->cmp_node(); 654 BoolTest::mask b_test = main_end->test_trip(); 655 656 // Need only 1 user of 'bol' because I will be hacking the loop bounds. 657 Node *bol = main_end->in(CountedLoopEndNode::TestValue); 658 if( bol->outcnt() != 1 ) { 659 bol = bol->clone(); 660 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); 661 _igvn.hash_delete(main_end); 662 main_end->set_req(CountedLoopEndNode::TestValue, bol); 663 } 664 // Need only 1 user of 'cmp' because I will be hacking the loop bounds. 665 if( cmp->outcnt() != 1 ) { 666 cmp = cmp->clone(); 667 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); 668 _igvn.hash_delete(bol); 669 bol->set_req(1, cmp); 670 } 671 672 //------------------------------ 673 // Step A: Create Post-Loop. 674 Node* main_exit = main_end->proj_out(false); 675 assert( main_exit->Opcode() == Op_IfFalse, "" ); 676 int dd_main_exit = dom_depth(main_exit); 677 678 // Step A1: Clone the loop body. The clone becomes the post-loop. The main 679 // loop pre-header illegally has 2 control users (old & new loops). 680 clone_loop( loop, old_new, dd_main_exit ); 681 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); 682 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); 683 post_head->set_post_loop(main_head); 684 685 // Build the main-loop normal exit. 686 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end); 687 _igvn.register_new_node_with_optimizer( new_main_exit ); 688 set_idom(new_main_exit, main_end, dd_main_exit ); 689 set_loop(new_main_exit, loop->_parent); 690 691 // Step A2: Build a zero-trip guard for the post-loop. After leaving the 692 // main-loop, the post-loop may not execute at all. We 'opaque' the incr 693 // (the main-loop trip-counter exit value) because we will be changing 694 // the exit value (via unrolling) so we cannot constant-fold away the zero 695 // trip guard until all unrolling is done. 696 Node *zer_opaq = new (C, 2) Opaque1Node(incr); 697 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit ); 698 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test ); 699 register_new_node( zer_opaq, new_main_exit ); 700 register_new_node( zer_cmp , new_main_exit ); 701 register_new_node( zer_bol , new_main_exit ); 702 703 // Build the IfNode 704 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); 705 _igvn.register_new_node_with_optimizer( zer_iff ); 706 set_idom(zer_iff, new_main_exit, dd_main_exit); 707 set_loop(zer_iff, loop->_parent); 708 709 // Plug in the false-path, taken if we need to skip post-loop 710 _igvn.hash_delete( main_exit ); 711 main_exit->set_req(0, zer_iff); 712 _igvn._worklist.push(main_exit); 713 set_idom(main_exit, zer_iff, dd_main_exit); 714 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); 715 // Make the true-path, must enter the post loop 716 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff ); 717 _igvn.register_new_node_with_optimizer( zer_taken ); 718 set_idom(zer_taken, zer_iff, dd_main_exit); 719 set_loop(zer_taken, loop->_parent); 720 // Plug in the true path 721 _igvn.hash_delete( post_head ); 722 post_head->set_req(LoopNode::EntryControl, zer_taken); 723 set_idom(post_head, zer_taken, dd_main_exit); 724 725 // Step A3: Make the fall-in values to the post-loop come from the 726 // fall-out values of the main-loop. 727 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { 728 Node* main_phi = main_head->fast_out(i); 729 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { 730 Node *post_phi = old_new[main_phi->_idx]; 731 Node *fallmain = clone_up_backedge_goo(main_head->back_control(), 732 post_head->init_control(), 733 main_phi->in(LoopNode::LoopBackControl)); 734 _igvn.hash_delete(post_phi); 735 post_phi->set_req( LoopNode::EntryControl, fallmain ); 736 } 737 } 738 739 // Update local caches for next stanza 740 main_exit = new_main_exit; 741 742 743 //------------------------------ 744 // Step B: Create Pre-Loop. 745 746 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main 747 // loop pre-header illegally has 2 control users (old & new loops). 748 clone_loop( loop, old_new, dd_main_head ); 749 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); 750 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); 751 pre_head->set_pre_loop(main_head); 752 Node *pre_incr = old_new[incr->_idx]; 753 754 // Find the pre-loop normal exit. 755 Node* pre_exit = pre_end->proj_out(false); 756 assert( pre_exit->Opcode() == Op_IfFalse, "" ); 757 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end); 758 _igvn.register_new_node_with_optimizer( new_pre_exit ); 759 set_idom(new_pre_exit, pre_end, dd_main_head); 760 set_loop(new_pre_exit, loop->_parent); 761 762 // Step B2: Build a zero-trip guard for the main-loop. After leaving the 763 // pre-loop, the main-loop may not execute at all. Later in life this 764 // zero-trip guard will become the minimum-trip guard when we unroll 765 // the main-loop. 766 Node *min_opaq = new (C, 2) Opaque1Node(limit); 767 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq ); 768 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test ); 769 register_new_node( min_opaq, new_pre_exit ); 770 register_new_node( min_cmp , new_pre_exit ); 771 register_new_node( min_bol , new_pre_exit ); 772 773 // Build the IfNode 774 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_FAIR, COUNT_UNKNOWN ); 775 _igvn.register_new_node_with_optimizer( min_iff ); 776 set_idom(min_iff, new_pre_exit, dd_main_head); 777 set_loop(min_iff, loop->_parent); 778 779 // Plug in the false-path, taken if we need to skip main-loop 780 _igvn.hash_delete( pre_exit ); 781 pre_exit->set_req(0, min_iff); 782 set_idom(pre_exit, min_iff, dd_main_head); 783 set_idom(pre_exit->unique_out(), min_iff, dd_main_head); 784 // Make the true-path, must enter the main loop 785 Node *min_taken = new (C, 1) IfTrueNode( min_iff ); 786 _igvn.register_new_node_with_optimizer( min_taken ); 787 set_idom(min_taken, min_iff, dd_main_head); 788 set_loop(min_taken, loop->_parent); 789 // Plug in the true path 790 _igvn.hash_delete( main_head ); 791 main_head->set_req(LoopNode::EntryControl, min_taken); 792 set_idom(main_head, min_taken, dd_main_head); 793 794 // Step B3: Make the fall-in values to the main-loop come from the 795 // fall-out values of the pre-loop. 796 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { 797 Node* main_phi = main_head->fast_out(i2); 798 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { 799 Node *pre_phi = old_new[main_phi->_idx]; 800 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), 801 main_head->init_control(), 802 pre_phi->in(LoopNode::LoopBackControl)); 803 _igvn.hash_delete(main_phi); 804 main_phi->set_req( LoopNode::EntryControl, fallpre ); 805 } 806 } 807 808 // Step B4: Shorten the pre-loop to run only 1 iteration (for now). 809 // RCE and alignment may change this later. 810 Node *cmp_end = pre_end->cmp_node(); 811 assert( cmp_end->in(2) == limit, "" ); 812 Node *pre_limit = new (C, 3) AddINode( init, stride ); 813 814 // Save the original loop limit in this Opaque1 node for 815 // use by range check elimination. 816 Node *pre_opaq = new (C, 3) Opaque1Node(pre_limit, limit); 817 818 register_new_node( pre_limit, pre_head->in(0) ); 819 register_new_node( pre_opaq , pre_head->in(0) ); 820 821 // Since no other users of pre-loop compare, I can hack limit directly 822 assert( cmp_end->outcnt() == 1, "no other users" ); 823 _igvn.hash_delete(cmp_end); 824 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); 825 826 // Special case for not-equal loop bounds: 827 // Change pre loop test, main loop test, and the 828 // main loop guard test to use lt or gt depending on stride 829 // direction: 830 // positive stride use < 831 // negative stride use > 832 833 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { 834 835 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; 836 // Modify pre loop end condition 837 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 838 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test); 839 register_new_node( new_bol0, pre_head->in(0) ); 840 _igvn.hash_delete(pre_end); 841 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); 842 // Modify main loop guard condition 843 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); 844 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test); 845 register_new_node( new_bol1, new_pre_exit ); 846 _igvn.hash_delete(min_iff); 847 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); 848 // Modify main loop end condition 849 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); 850 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test); 851 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); 852 _igvn.hash_delete(main_end); 853 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); 854 } 855 856 // Flag main loop 857 main_head->set_main_loop(); 858 if( peel_only ) main_head->set_main_no_pre_loop(); 859 860 // It's difficult to be precise about the trip-counts 861 // for the pre/post loops. They are usually very short, 862 // so guess that 4 trips is a reasonable value. 863 post_head->set_profile_trip_cnt(4.0); 864 pre_head->set_profile_trip_cnt(4.0); 865 866 // Now force out all loop-invariant dominating tests. The optimizer 867 // finds some, but we _know_ they are all useless. 868 peeled_dom_test_elim(loop,old_new); 869 } 870 871 //------------------------------is_invariant----------------------------- 872 // Return true if n is invariant 873 bool IdealLoopTree::is_invariant(Node* n) const { 874 Node *n_c = _phase->get_ctrl(n); 875 if (n_c->is_top()) return false; 876 return !is_member(_phase->get_loop(n_c)); 877 } 878 879 880 //------------------------------do_unroll-------------------------------------- 881 // Unroll the loop body one step - make each trip do 2 iterations. 882 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { 883 assert( LoopUnrollLimit, "" ); 884 #ifndef PRODUCT 885 if( PrintOpto && VerifyLoopOptimizations ) { 886 tty->print("Unrolling "); 887 loop->dump_head(); 888 } 889 #endif 890 CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); 891 CountedLoopEndNode *loop_end = loop_head->loopexit(); 892 assert( loop_end, "" ); 893 894 // Remember loop node count before unrolling to detect 895 // if rounds of unroll,optimize are making progress 896 loop_head->set_node_count_before_unroll(loop->_body.size()); 897 898 Node *ctrl = loop_head->in(LoopNode::EntryControl); 899 Node *limit = loop_head->limit(); 900 Node *init = loop_head->init_trip(); 901 Node *strid = loop_head->stride(); 902 903 Node *opaq = NULL; 904 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment 905 assert( loop_head->is_main_loop(), "" ); 906 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 907 Node *iff = ctrl->in(0); 908 assert( iff->Opcode() == Op_If, "" ); 909 Node *bol = iff->in(1); 910 assert( bol->Opcode() == Op_Bool, "" ); 911 Node *cmp = bol->in(1); 912 assert( cmp->Opcode() == Op_CmpI, "" ); 913 opaq = cmp->in(2); 914 // Occasionally it's possible for a pre-loop Opaque1 node to be 915 // optimized away and then another round of loop opts attempted. 916 // We can not optimize this particular loop in that case. 917 if( opaq->Opcode() != Op_Opaque1 ) 918 return; // Cannot find pre-loop! Bail out! 919 } 920 921 C->set_major_progress(); 922 923 // Adjust max trip count. The trip count is intentionally rounded 924 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, 925 // the main, unrolled, part of the loop will never execute as it is protected 926 // by the min-trip test. See bug 4834191 for a case where we over-unrolled 927 // and later determined that part of the unrolled loop was dead. 928 loop_head->set_trip_count(loop_head->trip_count() / 2); 929 930 // Double the count of original iterations in the unrolled loop body. 931 loop_head->double_unrolled_count(); 932 933 // ----------- 934 // Step 2: Cut back the trip counter for an unroll amount of 2. 935 // Loop will normally trip (limit - init)/stride_con. Since it's a 936 // CountedLoop this is exact (stride divides limit-init exactly). 937 // We are going to double the loop body, so we want to knock off any 938 // odd iteration: (trip_cnt & ~1). Then back compute a new limit. 939 Node *span = new (C, 3) SubINode( limit, init ); 940 register_new_node( span, ctrl ); 941 Node *trip = new (C, 3) DivINode( 0, span, strid ); 942 register_new_node( trip, ctrl ); 943 Node *mtwo = _igvn.intcon(-2); 944 set_ctrl(mtwo, C->root()); 945 Node *rond = new (C, 3) AndINode( trip, mtwo ); 946 register_new_node( rond, ctrl ); 947 Node *spn2 = new (C, 3) MulINode( rond, strid ); 948 register_new_node( spn2, ctrl ); 949 Node *lim2 = new (C, 3) AddINode( spn2, init ); 950 register_new_node( lim2, ctrl ); 951 952 // Hammer in the new limit 953 Node *ctrl2 = loop_end->in(0); 954 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 ); 955 register_new_node( cmp2, ctrl2 ); 956 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() ); 957 register_new_node( bol2, ctrl2 ); 958 _igvn.hash_delete(loop_end); 959 loop_end->set_req(CountedLoopEndNode::TestValue, bol2); 960 961 // Step 3: Find the min-trip test guaranteed before a 'main' loop. 962 // Make it a 1-trip test (means at least 2 trips). 963 if( adjust_min_trip ) { 964 // Guard test uses an 'opaque' node which is not shared. Hence I 965 // can edit it's inputs directly. Hammer in the new limit for the 966 // minimum-trip guard. 967 assert( opaq->outcnt() == 1, "" ); 968 _igvn.hash_delete(opaq); 969 opaq->set_req(1, lim2); 970 } 971 972 // --------- 973 // Step 4: Clone the loop body. Move it inside the loop. This loop body 974 // represents the odd iterations; since the loop trips an even number of 975 // times its backedge is never taken. Kill the backedge. 976 uint dd = dom_depth(loop_head); 977 clone_loop( loop, old_new, dd ); 978 979 // Make backedges of the clone equal to backedges of the original. 980 // Make the fall-in from the original come from the fall-out of the clone. 981 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { 982 Node* phi = loop_head->fast_out(j); 983 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { 984 Node *newphi = old_new[phi->_idx]; 985 _igvn.hash_delete( phi ); 986 _igvn.hash_delete( newphi ); 987 988 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); 989 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); 990 phi ->set_req(LoopNode::LoopBackControl, C->top()); 991 } 992 } 993 Node *clone_head = old_new[loop_head->_idx]; 994 _igvn.hash_delete( clone_head ); 995 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); 996 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); 997 loop_head ->set_req(LoopNode::LoopBackControl, C->top()); 998 loop->_head = clone_head; // New loop header 999 1000 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); 1001 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); 1002 1003 // Kill the clone's backedge 1004 Node *newcle = old_new[loop_end->_idx]; 1005 _igvn.hash_delete( newcle ); 1006 Node *one = _igvn.intcon(1); 1007 set_ctrl(one, C->root()); 1008 newcle->set_req(1, one); 1009 // Force clone into same loop body 1010 uint max = loop->_body.size(); 1011 for( uint k = 0; k < max; k++ ) { 1012 Node *old = loop->_body.at(k); 1013 Node *nnn = old_new[old->_idx]; 1014 loop->_body.push(nnn); 1015 if (!has_ctrl(old)) 1016 set_loop(nnn, loop); 1017 } 1018 } 1019 1020 //------------------------------do_maximally_unroll---------------------------- 1021 1022 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { 1023 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1024 assert( cl->trip_count() > 0, ""); 1025 1026 // If loop is tripping an odd number of times, peel odd iteration 1027 if( (cl->trip_count() & 1) == 1 ) { 1028 do_peeling( loop, old_new ); 1029 } 1030 1031 // Now its tripping an even number of times remaining. Double loop body. 1032 // Do not adjust pre-guards; they are not needed and do not exist. 1033 if( cl->trip_count() > 0 ) { 1034 do_unroll( loop, old_new, false ); 1035 } 1036 } 1037 1038 //------------------------------dominates_backedge--------------------------------- 1039 // Returns true if ctrl is executed on every complete iteration 1040 bool IdealLoopTree::dominates_backedge(Node* ctrl) { 1041 assert(ctrl->is_CFG(), "must be control"); 1042 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); 1043 return _phase->dom_lca_internal(ctrl, backedge) == ctrl; 1044 } 1045 1046 //------------------------------add_constraint--------------------------------- 1047 // Constrain the main loop iterations so the condition: 1048 // scale_con * I + offset < limit 1049 // always holds true. That is, either increase the number of iterations in 1050 // the pre-loop or the post-loop until the condition holds true in the main 1051 // loop. Stride, scale, offset and limit are all loop invariant. Further, 1052 // stride and scale are constants (offset and limit often are). 1053 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { 1054 1055 // Compute "I :: (limit-offset)/scale_con" 1056 Node *con = new (C, 3) SubINode( limit, offset ); 1057 register_new_node( con, pre_ctrl ); 1058 Node *scale = _igvn.intcon(scale_con); 1059 set_ctrl(scale, C->root()); 1060 Node *X = new (C, 3) DivINode( 0, con, scale ); 1061 register_new_node( X, pre_ctrl ); 1062 1063 // For positive stride, the pre-loop limit always uses a MAX function 1064 // and the main loop a MIN function. For negative stride these are 1065 // reversed. 1066 1067 // Also for positive stride*scale the affine function is increasing, so the 1068 // pre-loop must check for underflow and the post-loop for overflow. 1069 // Negative stride*scale reverses this; pre-loop checks for overflow and 1070 // post-loop for underflow. 1071 if( stride_con*scale_con > 0 ) { 1072 // Compute I < (limit-offset)/scale_con 1073 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X) 1074 *main_limit = (stride_con > 0) 1075 ? (Node*)(new (C, 3) MinINode( *main_limit, X )) 1076 : (Node*)(new (C, 3) MaxINode( *main_limit, X )); 1077 register_new_node( *main_limit, pre_ctrl ); 1078 1079 } else { 1080 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I 1081 // Add the negation of the main-loop constraint to the pre-loop. 1082 // See footnote [++] below for a derivation of the limit expression. 1083 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1); 1084 set_ctrl(incr, C->root()); 1085 Node *adj = new (C, 3) AddINode( X, incr ); 1086 register_new_node( adj, pre_ctrl ); 1087 *pre_limit = (scale_con > 0) 1088 ? (Node*)new (C, 3) MinINode( *pre_limit, adj ) 1089 : (Node*)new (C, 3) MaxINode( *pre_limit, adj ); 1090 register_new_node( *pre_limit, pre_ctrl ); 1091 1092 // [++] Here's the algebra that justifies the pre-loop limit expression: 1093 // 1094 // NOT( scale_con * I + offset < limit ) 1095 // == 1096 // scale_con * I + offset >= limit 1097 // == 1098 // SGN(scale_con) * I >= (limit-offset)/|scale_con| 1099 // == 1100 // (limit-offset)/|scale_con| <= I * SGN(scale_con) 1101 // == 1102 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con) 1103 // == 1104 // ( if (scale_con > 0) /*common case*/ 1105 // (limit-offset)/scale_con - 1 < I 1106 // else 1107 // (limit-offset)/scale_con + 1 > I 1108 // ) 1109 // ( if (scale_con > 0) /*common case*/ 1110 // (limit-offset)/scale_con + SGN(-scale_con) < I 1111 // else 1112 // (limit-offset)/scale_con + SGN(-scale_con) > I 1113 } 1114 } 1115 1116 1117 //------------------------------is_scaled_iv--------------------------------- 1118 // Return true if exp is a constant times an induction var 1119 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { 1120 if (exp == iv) { 1121 if (p_scale != NULL) { 1122 *p_scale = 1; 1123 } 1124 return true; 1125 } 1126 int opc = exp->Opcode(); 1127 if (opc == Op_MulI) { 1128 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1129 if (p_scale != NULL) { 1130 *p_scale = exp->in(2)->get_int(); 1131 } 1132 return true; 1133 } 1134 if (exp->in(2) == iv && exp->in(1)->is_Con()) { 1135 if (p_scale != NULL) { 1136 *p_scale = exp->in(1)->get_int(); 1137 } 1138 return true; 1139 } 1140 } else if (opc == Op_LShiftI) { 1141 if (exp->in(1) == iv && exp->in(2)->is_Con()) { 1142 if (p_scale != NULL) { 1143 *p_scale = 1 << exp->in(2)->get_int(); 1144 } 1145 return true; 1146 } 1147 } 1148 return false; 1149 } 1150 1151 //-----------------------------is_scaled_iv_plus_offset------------------------------ 1152 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) 1153 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { 1154 if (is_scaled_iv(exp, iv, p_scale)) { 1155 if (p_offset != NULL) { 1156 Node *zero = _igvn.intcon(0); 1157 set_ctrl(zero, C->root()); 1158 *p_offset = zero; 1159 } 1160 return true; 1161 } 1162 int opc = exp->Opcode(); 1163 if (opc == Op_AddI) { 1164 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1165 if (p_offset != NULL) { 1166 *p_offset = exp->in(2); 1167 } 1168 return true; 1169 } 1170 if (exp->in(2)->is_Con()) { 1171 Node* offset2 = NULL; 1172 if (depth < 2 && 1173 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, 1174 p_offset != NULL ? &offset2 : NULL, depth+1)) { 1175 if (p_offset != NULL) { 1176 Node *ctrl_off2 = get_ctrl(offset2); 1177 Node* offset = new (C, 3) AddINode(offset2, exp->in(2)); 1178 register_new_node(offset, ctrl_off2); 1179 *p_offset = offset; 1180 } 1181 return true; 1182 } 1183 } 1184 } else if (opc == Op_SubI) { 1185 if (is_scaled_iv(exp->in(1), iv, p_scale)) { 1186 if (p_offset != NULL) { 1187 Node *zero = _igvn.intcon(0); 1188 set_ctrl(zero, C->root()); 1189 Node *ctrl_off = get_ctrl(exp->in(2)); 1190 Node* offset = new (C, 3) SubINode(zero, exp->in(2)); 1191 register_new_node(offset, ctrl_off); 1192 *p_offset = offset; 1193 } 1194 return true; 1195 } 1196 if (is_scaled_iv(exp->in(2), iv, p_scale)) { 1197 if (p_offset != NULL) { 1198 *p_scale *= -1; 1199 *p_offset = exp->in(1); 1200 } 1201 return true; 1202 } 1203 } 1204 return false; 1205 } 1206 1207 //------------------------------do_range_check--------------------------------- 1208 // Eliminate range-checks and other trip-counter vs loop-invariant tests. 1209 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { 1210 #ifndef PRODUCT 1211 if( PrintOpto && VerifyLoopOptimizations ) { 1212 tty->print("Range Check Elimination "); 1213 loop->dump_head(); 1214 } 1215 #endif 1216 assert( RangeCheckElimination, "" ); 1217 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 1218 assert( cl->is_main_loop(), "" ); 1219 1220 // Find the trip counter; we are iteration splitting based on it 1221 Node *trip_counter = cl->phi(); 1222 // Find the main loop limit; we will trim it's iterations 1223 // to not ever trip end tests 1224 Node *main_limit = cl->limit(); 1225 // Find the pre-loop limit; we will expand it's iterations to 1226 // not ever trip low tests. 1227 Node *ctrl = cl->in(LoopNode::EntryControl); 1228 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 1229 Node *iffm = ctrl->in(0); 1230 assert( iffm->Opcode() == Op_If, "" ); 1231 Node *p_f = iffm->in(0); 1232 assert( p_f->Opcode() == Op_IfFalse, "" ); 1233 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); 1234 assert( pre_end->loopnode()->is_pre_loop(), "" ); 1235 Node *pre_opaq1 = pre_end->limit(); 1236 // Occasionally it's possible for a pre-loop Opaque1 node to be 1237 // optimized away and then another round of loop opts attempted. 1238 // We can not optimize this particular loop in that case. 1239 if( pre_opaq1->Opcode() != Op_Opaque1 ) 1240 return; 1241 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; 1242 Node *pre_limit = pre_opaq->in(1); 1243 1244 // Where do we put new limit calculations 1245 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); 1246 1247 // Ensure the original loop limit is available from the 1248 // pre-loop Opaque1 node. 1249 Node *orig_limit = pre_opaq->original_loop_limit(); 1250 if( orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP ) 1251 return; 1252 1253 // Need to find the main-loop zero-trip guard 1254 Node *bolzm = iffm->in(1); 1255 assert( bolzm->Opcode() == Op_Bool, "" ); 1256 Node *cmpzm = bolzm->in(1); 1257 assert( cmpzm->is_Cmp(), "" ); 1258 Node *opqzm = cmpzm->in(2); 1259 if( opqzm->Opcode() != Op_Opaque1 ) 1260 return; 1261 assert( opqzm->in(1) == main_limit, "do not understand situation" ); 1262 1263 // Must know if its a count-up or count-down loop 1264 1265 // protect against stride not being a constant 1266 if ( !cl->stride_is_con() ) { 1267 return; 1268 } 1269 int stride_con = cl->stride_con(); 1270 Node *zero = _igvn.intcon(0); 1271 Node *one = _igvn.intcon(1); 1272 set_ctrl(zero, C->root()); 1273 set_ctrl(one, C->root()); 1274 1275 // Range checks that do not dominate the loop backedge (ie. 1276 // conditionally executed) can lengthen the pre loop limit beyond 1277 // the original loop limit. To prevent this, the pre limit is 1278 // (for stride > 0) MINed with the original loop limit (MAXed 1279 // stride < 0) when some range_check (rc) is conditionally 1280 // executed. 1281 bool conditional_rc = false; 1282 1283 // Check loop body for tests of trip-counter plus loop-invariant vs 1284 // loop-invariant. 1285 for( uint i = 0; i < loop->_body.size(); i++ ) { 1286 Node *iff = loop->_body[i]; 1287 if( iff->Opcode() == Op_If ) { // Test? 1288 1289 // Test is an IfNode, has 2 projections. If BOTH are in the loop 1290 // we need loop unswitching instead of iteration splitting. 1291 Node *exit = loop->is_loop_exit(iff); 1292 if( !exit ) continue; 1293 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; 1294 1295 // Get boolean condition to test 1296 Node *i1 = iff->in(1); 1297 if( !i1->is_Bool() ) continue; 1298 BoolNode *bol = i1->as_Bool(); 1299 BoolTest b_test = bol->_test; 1300 // Flip sense of test if exit condition is flipped 1301 if( flip ) 1302 b_test = b_test.negate(); 1303 1304 // Get compare 1305 Node *cmp = bol->in(1); 1306 1307 // Look for trip_counter + offset vs limit 1308 Node *rc_exp = cmp->in(1); 1309 Node *limit = cmp->in(2); 1310 jint scale_con= 1; // Assume trip counter not scaled 1311 1312 Node *limit_c = get_ctrl(limit); 1313 if( loop->is_member(get_loop(limit_c) ) ) { 1314 // Compare might have operands swapped; commute them 1315 b_test = b_test.commute(); 1316 rc_exp = cmp->in(2); 1317 limit = cmp->in(1); 1318 limit_c = get_ctrl(limit); 1319 if( loop->is_member(get_loop(limit_c) ) ) 1320 continue; // Both inputs are loop varying; cannot RCE 1321 } 1322 // Here we know 'limit' is loop invariant 1323 1324 // 'limit' maybe pinned below the zero trip test (probably from a 1325 // previous round of rce), in which case, it can't be used in the 1326 // zero trip test expression which must occur before the zero test's if. 1327 if( limit_c == ctrl ) { 1328 continue; // Don't rce this check but continue looking for other candidates. 1329 } 1330 1331 // Check for scaled induction variable plus an offset 1332 Node *offset = NULL; 1333 1334 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { 1335 continue; 1336 } 1337 1338 Node *offset_c = get_ctrl(offset); 1339 if( loop->is_member( get_loop(offset_c) ) ) 1340 continue; // Offset is not really loop invariant 1341 // Here we know 'offset' is loop invariant. 1342 1343 // As above for the 'limit', the 'offset' maybe pinned below the 1344 // zero trip test. 1345 if( offset_c == ctrl ) { 1346 continue; // Don't rce this check but continue looking for other candidates. 1347 } 1348 1349 // At this point we have the expression as: 1350 // scale_con * trip_counter + offset :: limit 1351 // where scale_con, offset and limit are loop invariant. Trip_counter 1352 // monotonically increases by stride_con, a constant. Both (or either) 1353 // stride_con and scale_con can be negative which will flip about the 1354 // sense of the test. 1355 1356 // Adjust pre and main loop limits to guard the correct iteration set 1357 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests 1358 if( b_test._test == BoolTest::lt ) { // Range checks always use lt 1359 // The overflow limit: scale*I+offset < limit 1360 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1361 // The underflow limit: 0 <= scale*I+offset. 1362 // Some math yields: -scale*I-(offset+1) < 0 1363 Node *plus_one = new (C, 3) AddINode( offset, one ); 1364 register_new_node( plus_one, pre_ctrl ); 1365 Node *neg_offset = new (C, 3) SubINode( zero, plus_one ); 1366 register_new_node( neg_offset, pre_ctrl ); 1367 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit ); 1368 if (!conditional_rc) { 1369 conditional_rc = !loop->dominates_backedge(iff); 1370 } 1371 } else { 1372 #ifndef PRODUCT 1373 if( PrintOpto ) 1374 tty->print_cr("missed RCE opportunity"); 1375 #endif 1376 continue; // In release mode, ignore it 1377 } 1378 } else { // Otherwise work on normal compares 1379 switch( b_test._test ) { 1380 case BoolTest::ge: // Convert X >= Y to -X <= -Y 1381 scale_con = -scale_con; 1382 offset = new (C, 3) SubINode( zero, offset ); 1383 register_new_node( offset, pre_ctrl ); 1384 limit = new (C, 3) SubINode( zero, limit ); 1385 register_new_node( limit, pre_ctrl ); 1386 // Fall into LE case 1387 case BoolTest::le: // Convert X <= Y to X < Y+1 1388 limit = new (C, 3) AddINode( limit, one ); 1389 register_new_node( limit, pre_ctrl ); 1390 // Fall into LT case 1391 case BoolTest::lt: 1392 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); 1393 if (!conditional_rc) { 1394 conditional_rc = !loop->dominates_backedge(iff); 1395 } 1396 break; 1397 default: 1398 #ifndef PRODUCT 1399 if( PrintOpto ) 1400 tty->print_cr("missed RCE opportunity"); 1401 #endif 1402 continue; // Unhandled case 1403 } 1404 } 1405 1406 // Kill the eliminated test 1407 C->set_major_progress(); 1408 Node *kill_con = _igvn.intcon( 1-flip ); 1409 set_ctrl(kill_con, C->root()); 1410 _igvn.hash_delete(iff); 1411 iff->set_req(1, kill_con); 1412 _igvn._worklist.push(iff); 1413 // Find surviving projection 1414 assert(iff->is_If(), ""); 1415 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); 1416 // Find loads off the surviving projection; remove their control edge 1417 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { 1418 Node* cd = dp->fast_out(i); // Control-dependent node 1419 if( cd->is_Load() ) { // Loads can now float around in the loop 1420 _igvn.hash_delete(cd); 1421 // Allow the load to float around in the loop, or before it 1422 // but NOT before the pre-loop. 1423 cd->set_req(0, ctrl); // ctrl, not NULL 1424 _igvn._worklist.push(cd); 1425 --i; 1426 --imax; 1427 } 1428 } 1429 1430 } // End of is IF 1431 1432 } 1433 1434 // Update loop limits 1435 if (conditional_rc) { 1436 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit) 1437 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit); 1438 register_new_node(pre_limit, pre_ctrl); 1439 } 1440 _igvn.hash_delete(pre_opaq); 1441 pre_opaq->set_req(1, pre_limit); 1442 1443 // Note:: we are making the main loop limit no longer precise; 1444 // need to round up based on stride. 1445 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case 1446 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init 1447 // Hopefully, compiler will optimize for powers of 2. 1448 Node *ctrl = get_ctrl(main_limit); 1449 Node *stride = cl->stride(); 1450 Node *init = cl->init_trip(); 1451 Node *span = new (C, 3) SubINode(main_limit,init); 1452 register_new_node(span,ctrl); 1453 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); 1454 Node *add = new (C, 3) AddINode(span,rndup); 1455 register_new_node(add,ctrl); 1456 Node *div = new (C, 3) DivINode(0,add,stride); 1457 register_new_node(div,ctrl); 1458 Node *mul = new (C, 3) MulINode(div,stride); 1459 register_new_node(mul,ctrl); 1460 Node *newlim = new (C, 3) AddINode(mul,init); 1461 register_new_node(newlim,ctrl); 1462 main_limit = newlim; 1463 } 1464 1465 Node *main_cle = cl->loopexit(); 1466 Node *main_bol = main_cle->in(1); 1467 // Hacking loop bounds; need private copies of exit test 1468 if( main_bol->outcnt() > 1 ) {// BoolNode shared? 1469 _igvn.hash_delete(main_cle); 1470 main_bol = main_bol->clone();// Clone a private BoolNode 1471 register_new_node( main_bol, main_cle->in(0) ); 1472 main_cle->set_req(1,main_bol); 1473 } 1474 Node *main_cmp = main_bol->in(1); 1475 if( main_cmp->outcnt() > 1 ) { // CmpNode shared? 1476 _igvn.hash_delete(main_bol); 1477 main_cmp = main_cmp->clone();// Clone a private CmpNode 1478 register_new_node( main_cmp, main_cle->in(0) ); 1479 main_bol->set_req(1,main_cmp); 1480 } 1481 // Hack the now-private loop bounds 1482 _igvn.hash_delete(main_cmp); 1483 main_cmp->set_req(2, main_limit); 1484 _igvn._worklist.push(main_cmp); 1485 // The OpaqueNode is unshared by design 1486 _igvn.hash_delete(opqzm); 1487 assert( opqzm->outcnt() == 1, "cannot hack shared node" ); 1488 opqzm->set_req(1,main_limit); 1489 _igvn._worklist.push(opqzm); 1490 } 1491 1492 //------------------------------DCE_loop_body---------------------------------- 1493 // Remove simplistic dead code from loop body 1494 void IdealLoopTree::DCE_loop_body() { 1495 for( uint i = 0; i < _body.size(); i++ ) 1496 if( _body.at(i)->outcnt() == 0 ) 1497 _body.map( i--, _body.pop() ); 1498 } 1499 1500 1501 //------------------------------adjust_loop_exit_prob-------------------------- 1502 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. 1503 // Replace with a 1-in-10 exit guess. 1504 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { 1505 Node *test = tail(); 1506 while( test != _head ) { 1507 uint top = test->Opcode(); 1508 if( top == Op_IfTrue || top == Op_IfFalse ) { 1509 int test_con = ((ProjNode*)test)->_con; 1510 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); 1511 IfNode *iff = test->in(0)->as_If(); 1512 if( iff->outcnt() == 2 ) { // Ignore dead tests 1513 Node *bol = iff->in(1); 1514 if( bol && bol->req() > 1 && bol->in(1) && 1515 ((bol->in(1)->Opcode() == Op_StorePConditional ) || 1516 (bol->in(1)->Opcode() == Op_StoreLConditional ) || 1517 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || 1518 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || 1519 (bol->in(1)->Opcode() == Op_CompareAndSwapP ))) 1520 return; // Allocation loops RARELY take backedge 1521 // Find the OTHER exit path from the IF 1522 Node* ex = iff->proj_out(1-test_con); 1523 float p = iff->_prob; 1524 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { 1525 if( top == Op_IfTrue ) { 1526 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { 1527 iff->_prob = PROB_STATIC_FREQUENT; 1528 } 1529 } else { 1530 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { 1531 iff->_prob = PROB_STATIC_INFREQUENT; 1532 } 1533 } 1534 } 1535 } 1536 } 1537 test = phase->idom(test); 1538 } 1539 } 1540 1541 1542 //------------------------------policy_do_remove_empty_loop-------------------- 1543 // Micro-benchmark spamming. Policy is to always remove empty loops. 1544 // The 'DO' part is to replace the trip counter with the value it will 1545 // have on the last iteration. This will break the loop. 1546 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { 1547 // Minimum size must be empty loop 1548 if( _body.size() > 7/*number of nodes in an empty loop*/ ) return false; 1549 1550 if( !_head->is_CountedLoop() ) return false; // Dead loop 1551 CountedLoopNode *cl = _head->as_CountedLoop(); 1552 if( !cl->loopexit() ) return false; // Malformed loop 1553 if( !phase->is_member(this,phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)) ) ) 1554 return false; // Infinite loop 1555 #ifndef PRODUCT 1556 if( PrintOpto ) 1557 tty->print_cr("Removing empty loop"); 1558 #endif 1559 #ifdef ASSERT 1560 // Ensure only one phi which is the iv. 1561 Node* iv = NULL; 1562 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { 1563 Node* n = cl->fast_out(i); 1564 if (n->Opcode() == Op_Phi) { 1565 assert(iv == NULL, "Too many phis" ); 1566 iv = n; 1567 } 1568 } 1569 assert(iv == cl->phi(), "Wrong phi" ); 1570 #endif 1571 // Replace the phi at loop head with the final value of the last 1572 // iteration. Then the CountedLoopEnd will collapse (backedge never 1573 // taken) and all loop-invariant uses of the exit values will be correct. 1574 Node *phi = cl->phi(); 1575 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() ); 1576 phase->register_new_node(final,cl->in(LoopNode::EntryControl)); 1577 phase->_igvn.hash_delete(phi); 1578 phase->_igvn.subsume_node(phi,final); 1579 phase->C->set_major_progress(); 1580 return true; 1581 } 1582 1583 1584 //============================================================================= 1585 //------------------------------iteration_split_impl--------------------------- 1586 void IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { 1587 // Check and remove empty loops (spam micro-benchmarks) 1588 if( policy_do_remove_empty_loop(phase) ) 1589 return; // Here we removed an empty loop 1590 1591 bool should_peel = policy_peeling(phase); // Should we peel? 1592 1593 bool should_unswitch = policy_unswitching(phase); 1594 1595 // Non-counted loops may be peeled; exactly 1 iteration is peeled. 1596 // This removes loop-invariant tests (usually null checks). 1597 if( !_head->is_CountedLoop() ) { // Non-counted loop 1598 if (PartialPeelLoop && phase->partial_peel(this, old_new)) { 1599 return; 1600 } 1601 if( should_peel ) { // Should we peel? 1602 #ifndef PRODUCT 1603 if (PrintOpto) tty->print_cr("should_peel"); 1604 #endif 1605 phase->do_peeling(this,old_new); 1606 } else if( should_unswitch ) { 1607 phase->do_unswitching(this, old_new); 1608 } 1609 return; 1610 } 1611 CountedLoopNode *cl = _head->as_CountedLoop(); 1612 1613 if( !cl->loopexit() ) return; // Ignore various kinds of broken loops 1614 1615 // Do nothing special to pre- and post- loops 1616 if( cl->is_pre_loop() || cl->is_post_loop() ) return; 1617 1618 // Compute loop trip count from profile data 1619 compute_profile_trip_cnt(phase); 1620 1621 // Before attempting fancy unrolling, RCE or alignment, see if we want 1622 // to completely unroll this loop or do loop unswitching. 1623 if( cl->is_normal_loop() ) { 1624 bool should_maximally_unroll = policy_maximally_unroll(phase); 1625 if( should_maximally_unroll ) { 1626 // Here we did some unrolling and peeling. Eventually we will 1627 // completely unroll this loop and it will no longer be a loop. 1628 phase->do_maximally_unroll(this,old_new); 1629 return; 1630 } 1631 if (should_unswitch) { 1632 phase->do_unswitching(this, old_new); 1633 return; 1634 } 1635 } 1636 1637 1638 // Counted loops may be peeled, may need some iterations run up 1639 // front for RCE, and may want to align loop refs to a cache 1640 // line. Thus we clone a full loop up front whose trip count is 1641 // at least 1 (if peeling), but may be several more. 1642 1643 // The main loop will start cache-line aligned with at least 1 1644 // iteration of the unrolled body (zero-trip test required) and 1645 // will have some range checks removed. 1646 1647 // A post-loop will finish any odd iterations (leftover after 1648 // unrolling), plus any needed for RCE purposes. 1649 1650 bool should_unroll = policy_unroll(phase); 1651 1652 bool should_rce = policy_range_check(phase); 1653 1654 bool should_align = policy_align(phase); 1655 1656 // If not RCE'ing (iteration splitting) or Aligning, then we do not 1657 // need a pre-loop. We may still need to peel an initial iteration but 1658 // we will not be needing an unknown number of pre-iterations. 1659 // 1660 // Basically, if may_rce_align reports FALSE first time through, 1661 // we will not be able to later do RCE or Aligning on this loop. 1662 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; 1663 1664 // If we have any of these conditions (RCE, alignment, unrolling) met, then 1665 // we switch to the pre-/main-/post-loop model. This model also covers 1666 // peeling. 1667 if( should_rce || should_align || should_unroll ) { 1668 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops 1669 phase->insert_pre_post_loops(this,old_new, !may_rce_align); 1670 1671 // Adjust the pre- and main-loop limits to let the pre and post loops run 1672 // with full checks, but the main-loop with no checks. Remove said 1673 // checks from the main body. 1674 if( should_rce ) 1675 phase->do_range_check(this,old_new); 1676 1677 // Double loop body for unrolling. Adjust the minimum-trip test (will do 1678 // twice as many iterations as before) and the main body limit (only do 1679 // an even number of trips). If we are peeling, we might enable some RCE 1680 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if 1681 // peeling. 1682 if( should_unroll && !should_peel ) 1683 phase->do_unroll(this,old_new, true); 1684 1685 // Adjust the pre-loop limits to align the main body 1686 // iterations. 1687 if( should_align ) 1688 Unimplemented(); 1689 1690 } else { // Else we have an unchanged counted loop 1691 if( should_peel ) // Might want to peel but do nothing else 1692 phase->do_peeling(this,old_new); 1693 } 1694 } 1695 1696 1697 //============================================================================= 1698 //------------------------------iteration_split-------------------------------- 1699 void IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { 1700 // Recursively iteration split nested loops 1701 if( _child ) _child->iteration_split( phase, old_new ); 1702 1703 // Clean out prior deadwood 1704 DCE_loop_body(); 1705 1706 1707 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. 1708 // Replace with a 1-in-10 exit guess. 1709 if( _parent /*not the root loop*/ && 1710 !_irreducible && 1711 // Also ignore the occasional dead backedge 1712 !tail()->is_top() ) { 1713 adjust_loop_exit_prob(phase); 1714 } 1715 1716 1717 // Gate unrolling, RCE and peeling efforts. 1718 if( !_child && // If not an inner loop, do not split 1719 !_irreducible && 1720 !tail()->is_top() ) { // Also ignore the occasional dead backedge 1721 if (!_has_call) { 1722 iteration_split_impl( phase, old_new ); 1723 } else if (policy_unswitching(phase)) { 1724 phase->do_unswitching(this, old_new); 1725 } 1726 } 1727 1728 // Minor offset re-organization to remove loop-fallout uses of 1729 // trip counter. 1730 if( _head->is_CountedLoop() ) phase->reorg_offsets( this ); 1731 if( _next ) _next->iteration_split( phase, old_new ); 1732 }