1 /* 2 * Copyright (c) 1999, 2023, 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 "memory/allocation.inline.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/callnode.hpp" 29 #include "opto/loopnode.hpp" 30 #include "opto/movenode.hpp" 31 #include "opto/node.hpp" 32 #include "opto/opaquenode.hpp" 33 34 //------------------------------split_thru_region------------------------------ 35 // Split Node 'n' through merge point. 36 RegionNode* PhaseIdealLoop::split_thru_region(Node* n, RegionNode* region) { 37 assert(n->is_CFG(), ""); 38 RegionNode* r = new RegionNode(region->req()); 39 IdealLoopTree* loop = get_loop(n); 40 for (uint i = 1; i < region->req(); i++) { 41 Node* x = n->clone(); 42 Node* in0 = n->in(0); 43 if (in0->in(0) == region) x->set_req(0, in0->in(i)); 44 for (uint j = 1; j < n->req(); j++) { 45 Node* in = n->in(j); 46 if (get_ctrl(in) == region) { 47 x->set_req(j, in->in(i)); 48 } 49 } 50 _igvn.register_new_node_with_optimizer(x); 51 set_loop(x, loop); 52 set_idom(x, x->in(0), dom_depth(x->in(0))+1); 53 r->init_req(i, x); 54 } 55 56 // Record region 57 r->set_req(0,region); // Not a TRUE RegionNode 58 _igvn.register_new_node_with_optimizer(r); 59 set_loop(r, loop); 60 if (!loop->_child) { 61 loop->_body.push(r); 62 } 63 return r; 64 } 65 66 //------------------------------split_up--------------------------------------- 67 // Split block-local op up through the phis to empty the current block 68 bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) { 69 if( n->is_CFG() ) { 70 assert( n->in(0) != blk1, "Lousy candidate for split-if" ); 71 return false; 72 } 73 if (!at_relevant_ctrl(n, blk1, blk2)) 74 return false; // Not block local 75 if( n->is_Phi() ) return false; // Local PHIs are expected 76 77 // Recursively split-up inputs 78 for (uint i = 1; i < n->req(); i++) { 79 if( split_up( n->in(i), blk1, blk2 ) ) { 80 // Got split recursively and self went dead? 81 if (n->outcnt() == 0) 82 _igvn.remove_dead_node(n); 83 return true; 84 } 85 } 86 87 if (clone_cmp_loadklass_down(n, blk1, blk2)) { 88 return true; 89 } 90 91 // Check for needing to clone-up a compare. Can't do that, it forces 92 // another (nested) split-if transform. Instead, clone it "down". 93 if (clone_cmp_down(n, blk1, blk2)) { 94 return true; 95 } 96 97 if (subgraph_has_opaque(n)) { 98 Unique_Node_List wq; 99 wq.push(n); 100 for (uint i = 0; i < wq.size(); i++) { 101 Node* m = wq.at(i); 102 if (m->is_If()) { 103 assert(assertion_predicate_has_loop_opaque_node(m->as_If()), "opaque node not reachable from if?"); 104 Node* bol = create_bool_from_template_assertion_predicate(m, nullptr, nullptr, m->in(0)); 105 _igvn.replace_input_of(m, 1, bol); 106 } else { 107 assert(!m->is_CFG(), "not CFG expected"); 108 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 109 Node* u = m->fast_out(j); 110 wq.push(u); 111 } 112 } 113 } 114 } 115 116 if (n->Opcode() == Op_OpaqueZeroTripGuard) { 117 // If this Opaque1 is part of the zero trip guard for a loop: 118 // 1- it can't be shared 119 // 2- the zero trip guard can't be the if that's being split 120 // As a consequence, this node could be assigned control anywhere between its current control and the zero trip guard. 121 // Move it down to get it out of the way of split if and avoid breaking the zero trip guard shape. 122 Node* cmp = n->unique_out(); 123 assert(cmp->Opcode() == Op_CmpI, "bad zero trip guard shape"); 124 Node* bol = cmp->unique_out(); 125 assert(bol->Opcode() == Op_Bool, "bad zero trip guard shape"); 126 Node* iff = bol->unique_out(); 127 assert(iff->Opcode() == Op_If, "bad zero trip guard shape"); 128 set_ctrl(n, iff->in(0)); 129 set_ctrl(cmp, iff->in(0)); 130 set_ctrl(bol, iff->in(0)); 131 return true; 132 } 133 134 // See if splitting-up a Store. Any anti-dep loads must go up as 135 // well. An anti-dep load might be in the wrong block, because in 136 // this particular layout/schedule we ignored anti-deps and allow 137 // memory to be alive twice. This only works if we do the same 138 // operations on anti-dep loads as we do their killing stores. 139 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) { 140 // Get store's memory slice 141 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr()); 142 143 // Get memory-phi anti-dep loads will be using 144 Node *memphi = n->in(MemNode::Memory); 145 assert( memphi->is_Phi(), "" ); 146 // Hoist any anti-dep load to the splitting block; 147 // it will then "split-up". 148 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) { 149 Node *load = memphi->fast_out(i); 150 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) ) 151 set_ctrl(load,blk1); 152 } 153 } 154 155 // Found some other Node; must clone it up 156 #ifndef PRODUCT 157 if( PrintOpto && VerifyLoopOptimizations ) { 158 tty->print("Cloning up: "); 159 n->dump(); 160 } 161 #endif 162 163 // ConvI2L may have type information on it which becomes invalid if 164 // it moves up in the graph so change any clones so widen the type 165 // to TypeLong::INT when pushing it up. 166 const Type* rtype = nullptr; 167 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) { 168 rtype = TypeLong::INT; 169 } 170 171 // Now actually split-up this guy. One copy per control path merging. 172 Node *phi = PhiNode::make_blank(blk1, n); 173 for( uint j = 1; j < blk1->req(); j++ ) { 174 Node *x = n->clone(); 175 // Widen the type of the ConvI2L when pushing up. 176 if (rtype != nullptr) x->as_Type()->set_type(rtype); 177 if( n->in(0) && n->in(0) == blk1 ) 178 x->set_req( 0, blk1->in(j) ); 179 for( uint i = 1; i < n->req(); i++ ) { 180 Node *m = n->in(i); 181 if( get_ctrl(m) == blk1 ) { 182 assert( m->in(0) == blk1, "" ); 183 x->set_req( i, m->in(j) ); 184 } 185 } 186 register_new_node( x, blk1->in(j) ); 187 phi->init_req( j, x ); 188 } 189 // Announce phi to optimizer 190 register_new_node(phi, blk1); 191 192 // Remove cloned-up value from optimizer; use phi instead 193 _igvn.replace_node( n, phi ); 194 195 // (There used to be a self-recursive call to split_up() here, 196 // but it is not needed. All necessary forward walking is done 197 // by do_split_if() below.) 198 199 return true; 200 } 201 202 // Look for a (If .. (Bool(CmpP (LoadKlass .. (AddP obj ..)) ..))) and clone all of it down. 203 // There's likely a CheckCastPP on one of the branches of the If, with obj as input. 204 // If the (LoadKlass .. (AddP obj ..)) is not cloned down, then split if transforms this to: (If .. (Bool(CmpP phi1 ..))) 205 // and the CheckCastPP to (CheckCastPP phi2). It's possible then that phi2 is transformed to a CheckCastPP 206 // (through PhiNode::Ideal) and that that CheckCastPP is replaced by another narrower CheckCastPP at the same control 207 // (through ConstraintCastNode::Identity). That could cause the CheckCastPP at the If to become top while (CmpP phi1) 208 // wouldn't constant fold because it's using a different data path. Cloning the whole subgraph down guarantees both the 209 // AddP and CheckCastPP have the same obj input after split if. 210 bool PhaseIdealLoop::clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2) { 211 if (n->Opcode() == Op_AddP && at_relevant_ctrl(n, blk1, blk2)) { 212 Node_List cmp_nodes; 213 uint old = C->unique(); 214 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 215 Node* u1 = n->fast_out(i); 216 if (u1->Opcode() == Op_LoadNKlass && at_relevant_ctrl(u1, blk1, blk2)) { 217 for (DUIterator_Fast jmax, j = u1->fast_outs(jmax); j < jmax; j++) { 218 Node* u2 = u1->fast_out(j); 219 if (u2->Opcode() == Op_DecodeNKlass && at_relevant_ctrl(u2, blk1, blk2)) { 220 for (DUIterator k = u2->outs(); u2->has_out(k); k++) { 221 Node* u3 = u2->out(k); 222 if (at_relevant_ctrl(u3, blk1, blk2) && clone_cmp_down(u3, blk1, blk2)) { 223 --k; 224 } 225 } 226 for (DUIterator_Fast kmax, k = u2->fast_outs(kmax); k < kmax; k++) { 227 Node* u3 = u2->fast_out(k); 228 if (u3->_idx >= old) { 229 cmp_nodes.push(u3); 230 } 231 } 232 } 233 } 234 } else if (u1->Opcode() == Op_LoadKlass && at_relevant_ctrl(u1, blk1, blk2)) { 235 for (DUIterator j = u1->outs(); u1->has_out(j); j++) { 236 Node* u2 = u1->out(j); 237 if (at_relevant_ctrl(u2, blk1, blk2) && clone_cmp_down(u2, blk1, blk2)) { 238 --j; 239 } 240 } 241 for (DUIterator_Fast kmax, k = u1->fast_outs(kmax); k < kmax; k++) { 242 Node* u2 = u1->fast_out(k); 243 if (u2->_idx >= old) { 244 cmp_nodes.push(u2); 245 } 246 } 247 } 248 } 249 250 for (uint i = 0; i < cmp_nodes.size(); ++i) { 251 Node* cmp = cmp_nodes.at(i); 252 clone_loadklass_nodes_at_cmp_index(n, cmp, 1); 253 clone_loadklass_nodes_at_cmp_index(n, cmp, 2); 254 } 255 if (n->outcnt() == 0) { 256 assert(n->is_dead(), ""); 257 return true; 258 } 259 } 260 return false; 261 } 262 263 bool PhaseIdealLoop::at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2) { 264 return ctrl_or_self(n) == blk1 || ctrl_or_self(n) == blk2; 265 } 266 267 void PhaseIdealLoop::clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i) { 268 Node* decode = cmp->in(i); 269 if (decode->Opcode() == Op_DecodeNKlass) { 270 Node* loadklass = decode->in(1); 271 if (loadklass->Opcode() == Op_LoadNKlass) { 272 Node* addp = loadklass->in(MemNode::Address); 273 if (addp == n) { 274 Node* ctrl = get_ctrl(cmp); 275 Node* decode_clone = decode->clone(); 276 Node* loadklass_clone = loadklass->clone(); 277 Node* addp_clone = addp->clone(); 278 register_new_node(decode_clone, ctrl); 279 register_new_node(loadklass_clone, ctrl); 280 register_new_node(addp_clone, ctrl); 281 _igvn.replace_input_of(cmp, i, decode_clone); 282 _igvn.replace_input_of(decode_clone, 1, loadklass_clone); 283 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone); 284 if (decode->outcnt() == 0) { 285 _igvn.remove_dead_node(decode); 286 } 287 } 288 } 289 } else { 290 Node* loadklass = cmp->in(i); 291 if (loadklass->Opcode() == Op_LoadKlass) { 292 Node* addp = loadklass->in(MemNode::Address); 293 if (addp == n) { 294 Node* ctrl = get_ctrl(cmp); 295 Node* loadklass_clone = loadklass->clone(); 296 Node* addp_clone = addp->clone(); 297 register_new_node(loadklass_clone, ctrl); 298 register_new_node(addp_clone, ctrl); 299 _igvn.replace_input_of(cmp, i, loadklass_clone); 300 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone); 301 if (loadklass->outcnt() == 0) { 302 _igvn.remove_dead_node(loadklass); 303 } 304 } 305 } 306 } 307 } 308 309 bool PhaseIdealLoop::clone_cmp_down(Node* n, const Node* blk1, const Node* blk2) { 310 if( n->is_Cmp() ) { 311 assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF"); 312 // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove 313 // sequence can have no other users and it must all reside in the split-if 314 // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below - 315 // private, per-use versions of the Cmp and Bool are made. These sink to 316 // the CMove block. If the CMove is in the split-if block, then in the 317 // next iteration this will become a simple Cmp/Bool/CMove set to clone-up. 318 Node *bol, *cmov; 319 if (!(n->outcnt() == 1 && n->unique_out()->is_Bool() && 320 (bol = n->unique_out()->as_Bool()) && 321 (at_relevant_ctrl(bol, blk1, blk2) && 322 bol->outcnt() == 1 && 323 bol->unique_out()->is_CMove() && 324 (cmov = bol->unique_out()->as_CMove()) && 325 at_relevant_ctrl(cmov, blk1, blk2)))) { 326 327 // Must clone down 328 #ifndef PRODUCT 329 if( PrintOpto && VerifyLoopOptimizations ) { 330 tty->print("Cloning down: "); 331 n->dump(); 332 } 333 #endif 334 if (!n->is_FastLock()) { 335 // Clone down any block-local BoolNode uses of this CmpNode 336 for (DUIterator i = n->outs(); n->has_out(i); i++) { 337 Node* bol = n->out(i); 338 assert( bol->is_Bool(), "" ); 339 if (bol->outcnt() == 1) { 340 Node* use = bol->unique_out(); 341 if (use->Opcode() == Op_Opaque4) { 342 if (use->outcnt() == 1) { 343 Node* iff = use->unique_out(); 344 assert(iff->is_If(), "unexpected node type"); 345 Node *use_c = iff->in(0); 346 if (use_c == blk1 || use_c == blk2) { 347 continue; 348 } 349 } 350 } else { 351 // We might see an Opaque1 from a loop limit check here 352 assert(use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque1 || use->is_AllocateArray(), "unexpected node type"); 353 Node *use_c = (use->is_If() || use->is_AllocateArray()) ? use->in(0) : get_ctrl(use); 354 if (use_c == blk1 || use_c == blk2) { 355 assert(use->is_CMove(), "unexpected node type"); 356 continue; 357 } 358 } 359 } 360 if (at_relevant_ctrl(bol, blk1, blk2)) { 361 // Recursively sink any BoolNode 362 #ifndef PRODUCT 363 if( PrintOpto && VerifyLoopOptimizations ) { 364 tty->print("Cloning down: "); 365 bol->dump(); 366 } 367 #endif 368 for (DUIterator j = bol->outs(); bol->has_out(j); j++) { 369 Node* u = bol->out(j); 370 // Uses are either IfNodes, CMoves or Opaque4 371 if (u->Opcode() == Op_Opaque4) { 372 assert(u->in(1) == bol, "bad input"); 373 for (DUIterator_Last kmin, k = u->last_outs(kmin); k >= kmin; --k) { 374 Node* iff = u->last_out(k); 375 assert(iff->is_If() || iff->is_CMove(), "unexpected node type"); 376 assert( iff->in(1) == u, "" ); 377 // Get control block of either the CMove or the If input 378 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff); 379 Node *x1 = bol->clone(); 380 Node *x2 = u->clone(); 381 register_new_node(x1, iff_ctrl); 382 register_new_node(x2, iff_ctrl); 383 _igvn.replace_input_of(x2, 1, x1); 384 _igvn.replace_input_of(iff, 1, x2); 385 } 386 _igvn.remove_dead_node(u); 387 --j; 388 } else { 389 // We might see an Opaque1 from a loop limit check here 390 assert(u->is_If() || u->is_CMove() || u->Opcode() == Op_Opaque1 || u->is_AllocateArray(), "unexpected node type"); 391 assert(u->is_AllocateArray() || u->in(1) == bol, ""); 392 assert(!u->is_AllocateArray() || u->in(AllocateNode::ValidLengthTest) == bol, "wrong input to AllocateArray"); 393 // Get control block of either the CMove or the If input 394 Node *u_ctrl = (u->is_If() || u->is_AllocateArray()) ? u->in(0) : get_ctrl(u); 395 assert((u_ctrl != blk1 && u_ctrl != blk2) || u->is_CMove(), "won't converge"); 396 Node *x = bol->clone(); 397 register_new_node(x, u_ctrl); 398 _igvn.replace_input_of(u, u->is_AllocateArray() ? AllocateNode::ValidLengthTest : 1, x); 399 --j; 400 } 401 } 402 _igvn.remove_dead_node(bol); 403 --i; 404 } 405 } 406 } 407 // Clone down this CmpNode 408 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) { 409 Node* use = n->last_out(j); 410 uint pos = 1; 411 if (n->is_FastLock()) { 412 pos = TypeFunc::Parms + 2; 413 assert(use->is_Lock(), "FastLock only used by LockNode"); 414 } 415 assert(use->in(pos) == n, "" ); 416 Node *x = n->clone(); 417 register_new_node(x, ctrl_or_self(use)); 418 _igvn.replace_input_of(use, pos, x); 419 } 420 _igvn.remove_dead_node(n); 421 422 return true; 423 } 424 } 425 return false; 426 } 427 428 //------------------------------register_new_node------------------------------ 429 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) { 430 assert(!n->is_CFG(), "must be data node"); 431 _igvn.register_new_node_with_optimizer(n); 432 set_ctrl(n, blk); 433 IdealLoopTree *loop = get_loop(blk); 434 if( !loop->_child ) 435 loop->_body.push(n); 436 } 437 438 //------------------------------small_cache------------------------------------ 439 struct small_cache : public Dict { 440 441 small_cache() : Dict( cmpkey, hashptr ) {} 442 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); } 443 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); } 444 }; 445 446 //------------------------------spinup----------------------------------------- 447 // "Spin up" the dominator tree, starting at the use site and stopping when we 448 // find the post-dominating point. 449 450 // We must be at the merge point which post-dominates 'new_false' and 451 // 'new_true'. Figure out which edges into the RegionNode eventually lead up 452 // to false and which to true. Put in a PhiNode to merge values; plug in 453 // the appropriate false-arm or true-arm values. If some path leads to the 454 // original IF, then insert a Phi recursively. 455 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) { 456 if (use_blk->is_top()) // Handle dead uses 457 return use_blk; 458 Node *prior_n = (Node*)((intptr_t)0xdeadbeef); 459 Node *n = use_blk; // Get path input 460 assert( use_blk != iff_dom, "" ); 461 // Here's the "spinup" the dominator tree loop. Do a cache-check 462 // along the way, in case we've come this way before. 463 while( n != iff_dom ) { // Found post-dominating point? 464 prior_n = n; 465 n = idom(n); // Search higher 466 Node *s = cache->probe( prior_n ); // Check cache 467 if( s ) return s; // Cache hit! 468 } 469 470 Node *phi_post; 471 if( prior_n == new_false || prior_n == new_true ) { 472 phi_post = def->clone(); 473 phi_post->set_req(0, prior_n ); 474 register_new_node(phi_post, prior_n); 475 } else { 476 // This method handles both control uses (looking for Regions) or data 477 // uses (looking for Phis). If looking for a control use, then we need 478 // to insert a Region instead of a Phi; however Regions always exist 479 // previously (the hash_find_insert below would always hit) so we can 480 // return the existing Region. 481 if( def->is_CFG() ) { 482 phi_post = prior_n; // If looking for CFG, return prior 483 } else { 484 assert( def->is_Phi(), "" ); 485 assert( prior_n->is_Region(), "must be a post-dominating merge point" ); 486 487 // Need a Phi here 488 phi_post = PhiNode::make_blank(prior_n, def); 489 // Search for both true and false on all paths till find one. 490 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths 491 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) ); 492 Node *t = _igvn.hash_find_insert(phi_post); 493 if( t ) { // See if we already have this one 494 // phi_post will not be used, so kill it 495 _igvn.remove_dead_node(phi_post); 496 phi_post->destruct(&_igvn); 497 phi_post = t; 498 } else { 499 register_new_node( phi_post, prior_n ); 500 } 501 } 502 } 503 504 // Update cache everywhere 505 prior_n = (Node*)((intptr_t)0xdeadbeef); // Reset IDOM walk 506 n = use_blk; // Get path input 507 // Spin-up the idom tree again, basically doing path-compression. 508 // Insert cache entries along the way, so that if we ever hit this 509 // point in the IDOM tree again we'll stop immediately on a cache hit. 510 while( n != iff_dom ) { // Found post-dominating point? 511 prior_n = n; 512 n = idom(n); // Search higher 513 cache->lru_insert( prior_n, phi_post ); // Fill cache 514 } // End of while not gone high enough 515 516 return phi_post; 517 } 518 519 //------------------------------find_use_block--------------------------------- 520 // Find the block a USE is in. Normally USE's are in the same block as the 521 // using instruction. For Phi-USE's, the USE is in the predecessor block 522 // along the corresponding path. 523 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) { 524 // CFG uses are their own block 525 if( use->is_CFG() ) 526 return use; 527 528 if( use->is_Phi() ) { // Phi uses in prior block 529 // Grab the first Phi use; there may be many. 530 // Each will be handled as a separate iteration of 531 // the "while( phi->outcnt() )" loop. 532 uint j; 533 for( j = 1; j < use->req(); j++ ) 534 if( use->in(j) == def ) 535 break; 536 assert( j < use->req(), "def should be among use's inputs" ); 537 return use->in(0)->in(j); 538 } 539 // Normal (non-phi) use 540 Node *use_blk = get_ctrl(use); 541 // Some uses are directly attached to the old (and going away) 542 // false and true branches. 543 if( use_blk == old_false ) { 544 use_blk = new_false; 545 set_ctrl(use, new_false); 546 } 547 if( use_blk == old_true ) { 548 use_blk = new_true; 549 set_ctrl(use, new_true); 550 } 551 552 if (use_blk == nullptr) { // He's dead, Jim 553 _igvn.replace_node(use, C->top()); 554 } 555 556 return use_blk; 557 } 558 559 //------------------------------handle_use------------------------------------- 560 // Handle uses of the merge point. Basically, split-if makes the merge point 561 // go away so all uses of the merge point must go away as well. Most block 562 // local uses have already been split-up, through the merge point. Uses from 563 // far below the merge point can't always be split up (e.g., phi-uses are 564 // pinned) and it makes too much stuff live. Instead we use a path-based 565 // solution to move uses down. 566 // 567 // If the use is along the pre-split-CFG true branch, then the new use will 568 // be from the post-split-CFG true merge point. Vice-versa for the false 569 // path. Some uses will be along both paths; then we sink the use to the 570 // post-dominating location; we may need to insert a Phi there. 571 void PhaseIdealLoop::handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true ) { 572 573 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true); 574 if( !use_blk ) return; // He's dead, Jim 575 576 // Walk up the dominator tree until I hit either the old IfFalse, the old 577 // IfTrue or the old If. Insert Phis where needed. 578 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache ); 579 580 // Found where this USE goes. Re-point him. 581 uint i; 582 for( i = 0; i < use->req(); i++ ) 583 if( use->in(i) == def ) 584 break; 585 assert( i < use->req(), "def should be among use's inputs" ); 586 _igvn.replace_input_of(use, i, new_def); 587 } 588 589 //------------------------------do_split_if------------------------------------ 590 // Found an If getting its condition-code input from a Phi in the same block. 591 // Split thru the Region. 592 void PhaseIdealLoop::do_split_if(Node* iff, RegionNode** new_false_region, RegionNode** new_true_region) { 593 594 C->set_major_progress(); 595 RegionNode *region = iff->in(0)->as_Region(); 596 Node *region_dom = idom(region); 597 598 // We are going to clone this test (and the control flow with it) up through 599 // the incoming merge point. We need to empty the current basic block. 600 // Clone any instructions which must be in this block up through the merge 601 // point. 602 DUIterator i, j; 603 bool progress = true; 604 while (progress) { 605 progress = false; 606 for (i = region->outs(); region->has_out(i); i++) { 607 Node* n = region->out(i); 608 if( n == region ) continue; 609 // The IF to be split is OK. 610 if( n == iff ) continue; 611 if( !n->is_Phi() ) { // Found pinned memory op or such 612 if (split_up(n, region, iff)) { 613 i = region->refresh_out_pos(i); 614 progress = true; 615 } 616 continue; 617 } 618 assert( n->in(0) == region, "" ); 619 620 // Recursively split up all users of a Phi 621 for (j = n->outs(); n->has_out(j); j++) { 622 Node* m = n->out(j); 623 // If m is dead, throw it away, and declare progress 624 if (_loop_or_ctrl[m->_idx] == nullptr) { 625 _igvn.remove_dead_node(m); 626 // fall through 627 } 628 else if (m != iff && split_up(m, region, iff)) { 629 // fall through 630 } else { 631 continue; 632 } 633 // Something unpredictable changed. 634 // Tell the iterators to refresh themselves, and rerun the loop. 635 i = region->refresh_out_pos(i); 636 j = region->refresh_out_pos(j); 637 progress = true; 638 } 639 } 640 } 641 642 // Now we have no instructions in the block containing the IF. 643 // Split the IF. 644 RegionNode *new_iff = split_thru_region(iff, region); 645 646 // Replace both uses of 'new_iff' with Regions merging True/False 647 // paths. This makes 'new_iff' go dead. 648 Node *old_false = nullptr, *old_true = nullptr; 649 RegionNode* new_false = nullptr; 650 RegionNode* new_true = nullptr; 651 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) { 652 Node *ifp = iff->last_out(j2); 653 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" ); 654 ifp->set_req(0, new_iff); 655 RegionNode* ifpx = split_thru_region(ifp, region); 656 657 // Replace 'If' projection of a Region with a Region of 658 // 'If' projections. 659 ifpx->set_req(0, ifpx); // A TRUE RegionNode 660 661 // Setup dominator info 662 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1); 663 664 // Check for splitting loop tails 665 if( get_loop(iff)->tail() == ifp ) 666 get_loop(iff)->_tail = ifpx; 667 668 // Replace in the graph with lazy-update mechanism 669 new_iff->set_req(0, new_iff); // hook self so it does not go dead 670 lazy_replace(ifp, ifpx); 671 new_iff->set_req(0, region); 672 673 // Record bits for later xforms 674 if( ifp->Opcode() == Op_IfFalse ) { 675 old_false = ifp; 676 new_false = ifpx; 677 } else { 678 old_true = ifp; 679 new_true = ifpx; 680 } 681 } 682 _igvn.remove_dead_node(new_iff); 683 // Lazy replace IDOM info with the region's dominator 684 lazy_replace(iff, region_dom); 685 lazy_update(region, region_dom); // idom must be update before handle_uses 686 region->set_req(0, nullptr); // Break the self-cycle. Required for lazy_update to work on region 687 688 // Now make the original merge point go dead, by handling all its uses. 689 small_cache region_cache; 690 // Preload some control flow in region-cache 691 region_cache.lru_insert( new_false, new_false ); 692 region_cache.lru_insert( new_true , new_true ); 693 // Now handle all uses of the splitting block 694 for (DUIterator k = region->outs(); region->has_out(k); k++) { 695 Node* phi = region->out(k); 696 if (!phi->in(0)) { // Dead phi? Remove it 697 _igvn.remove_dead_node(phi); 698 } else if (phi == region) { // Found the self-reference 699 continue; // No roll-back of DUIterator 700 } else if (phi->is_Phi()) { // Expected common case: Phi hanging off of Region 701 assert(phi->in(0) == region, "Inconsistent graph"); 702 // Need a per-def cache. Phi represents a def, so make a cache 703 small_cache phi_cache; 704 705 // Inspect all Phi uses to make the Phi go dead 706 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) { 707 Node* use = phi->last_out(l); 708 // Compute the new DEF for this USE. New DEF depends on the path 709 // taken from the original DEF to the USE. The new DEF may be some 710 // collection of PHI's merging values from different paths. The Phis 711 // inserted depend only on the location of the USE. We use a 712 // 2-element cache to handle multiple uses from the same block. 713 handle_use(use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true); 714 } // End of while phi has uses 715 // Remove the dead Phi 716 _igvn.remove_dead_node( phi ); 717 } else { 718 assert(phi->in(0) == region, "Inconsistent graph"); 719 // Random memory op guarded by Region. Compute new DEF for USE. 720 handle_use(phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true); 721 } 722 // Every path above deletes a use of the region, except for the region 723 // self-cycle (which is needed by handle_use calling find_use_block 724 // calling get_ctrl calling get_ctrl_no_update looking for dead 725 // regions). So roll back the DUIterator innards. 726 --k; 727 } // End of while merge point has phis 728 729 _igvn.remove_dead_node(region); 730 731 if (new_false_region != nullptr) { 732 *new_false_region = new_false; 733 } 734 if (new_true_region != nullptr) { 735 *new_true_region = new_true; 736 } 737 738 DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } ); 739 }