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 "opto/addnode.hpp" 26 #include "opto/node.hpp" 27 #include "precompiled.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/inlinetypenode.hpp" 31 #include "opto/loopnode.hpp" 32 #include "opto/movenode.hpp" 33 #include "opto/opaquenode.hpp" 34 35 36 //------------------------------split_thru_region------------------------------ 37 // Split Node 'n' through merge point. 38 RegionNode* PhaseIdealLoop::split_thru_region(Node* n, RegionNode* region) { 39 assert(n->is_CFG(), ""); 40 RegionNode* r = new RegionNode(region->req()); 41 IdealLoopTree* loop = get_loop(n); 42 for (uint i = 1; i < region->req(); i++) { 43 Node* x = n->clone(); 44 Node* in0 = n->in(0); 45 if (in0->in(0) == region) x->set_req(0, in0->in(i)); 46 for (uint j = 1; j < n->req(); j++) { 47 Node* in = n->in(j); 48 if (get_ctrl(in) == region) { 49 x->set_req(j, in->in(i)); 50 } 51 } 52 _igvn.register_new_node_with_optimizer(x); 53 set_loop(x, loop); 54 set_idom(x, x->in(0), dom_depth(x->in(0))+1); 55 r->init_req(i, x); 56 } 57 58 // Record region 59 r->set_req(0,region); // Not a TRUE RegionNode 60 _igvn.register_new_node_with_optimizer(r); 61 set_loop(r, loop); 62 if (!loop->_child) { 63 loop->_body.push(r); 64 } 65 return r; 66 } 67 68 //------------------------------split_up--------------------------------------- 69 // Split block-local op up through the phis to empty the current block 70 bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) { 71 if( n->is_CFG() ) { 72 assert( n->in(0) != blk1, "Lousy candidate for split-if" ); 73 return false; 74 } 75 if (!at_relevant_ctrl(n, blk1, blk2)) 76 return false; // Not block local 77 if( n->is_Phi() ) return false; // Local PHIs are expected 78 79 // Recursively split-up inputs 80 for (uint i = 1; i < n->req(); i++) { 81 if( split_up( n->in(i), blk1, blk2 ) ) { 82 // Got split recursively and self went dead? 83 if (n->outcnt() == 0) 84 _igvn.remove_dead_node(n); 85 return true; 86 } 87 } 88 89 if (clone_cmp_loadklass_down(n, blk1, blk2)) { 90 return true; 91 } 92 93 // Check for needing to clone-up a compare. Can't do that, it forces 94 // another (nested) split-if transform. Instead, clone it "down". 95 if (clone_cmp_down(n, blk1, blk2)) { 96 return true; 97 } 98 99 if (subgraph_has_opaque(n)) { 100 Unique_Node_List wq; 101 wq.push(n); 102 for (uint i = 0; i < wq.size(); i++) { 103 Node* m = wq.at(i); 104 if (m->is_If()) { 105 assert(skeleton_predicate_has_opaque(m->as_If()), "opaque node not reachable from if?"); 106 Node* bol = clone_skeleton_predicate_bool(m, NULL, NULL, m->in(0)); 107 _igvn.replace_input_of(m, 1, bol); 108 } else { 109 assert(!m->is_CFG(), "not CFG expected"); 110 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 111 Node* u = m->fast_out(j); 112 wq.push(u); 113 } 114 } 115 } 116 } 117 118 if (n->Opcode() == Op_OpaqueZeroTripGuard) { 119 // If this Opaque1 is part of the zero trip guard for a loop: 120 // 1- it can't be shared 121 // 2- the zero trip guard can't be the if that's being split 122 // As a consequence, this node could be assigned control anywhere between its current control and the zero trip guard. 123 // Move it down to get it out of the way of split if and avoid breaking the zero trip guard shape. 124 Node* cmp = n->unique_out(); 125 assert(cmp->Opcode() == Op_CmpI, "bad zero trip guard shape"); 126 Node* bol = cmp->unique_out(); 127 assert(bol->Opcode() == Op_Bool, "bad zero trip guard shape"); 128 Node* iff = bol->unique_out(); 129 assert(iff->Opcode() == Op_If, "bad zero trip guard shape"); 130 set_ctrl(n, iff->in(0)); 131 set_ctrl(cmp, iff->in(0)); 132 set_ctrl(bol, iff->in(0)); 133 return true; 134 } 135 136 // See if splitting-up a Store. Any anti-dep loads must go up as 137 // well. An anti-dep load might be in the wrong block, because in 138 // this particular layout/schedule we ignored anti-deps and allow 139 // memory to be alive twice. This only works if we do the same 140 // operations on anti-dep loads as we do their killing stores. 141 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) { 142 // Get store's memory slice 143 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr()); 144 145 // Get memory-phi anti-dep loads will be using 146 Node *memphi = n->in(MemNode::Memory); 147 assert( memphi->is_Phi(), "" ); 148 // Hoist any anti-dep load to the splitting block; 149 // it will then "split-up". 150 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) { 151 Node *load = memphi->fast_out(i); 152 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) ) 153 set_ctrl(load,blk1); 154 } 155 } 156 157 // Found some other Node; must clone it up 158 #ifndef PRODUCT 159 if( PrintOpto && VerifyLoopOptimizations ) { 160 tty->print("Cloning up: "); 161 n->dump(); 162 } 163 #endif 164 165 // ConvI2L may have type information on it which becomes invalid if 166 // it moves up in the graph so change any clones so widen the type 167 // to TypeLong::INT when pushing it up. 168 const Type* rtype = NULL; 169 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) { 170 rtype = TypeLong::INT; 171 } 172 173 // Now actually split-up this guy. One copy per control path merging. 174 Node *phi = PhiNode::make_blank(blk1, n); 175 for( uint j = 1; j < blk1->req(); j++ ) { 176 Node *x = n->clone(); 177 // Widen the type of the ConvI2L when pushing up. 178 if (rtype != NULL) x->as_Type()->set_type(rtype); 179 if( n->in(0) && n->in(0) == blk1 ) 180 x->set_req( 0, blk1->in(j) ); 181 for( uint i = 1; i < n->req(); i++ ) { 182 Node *m = n->in(i); 183 if( get_ctrl(m) == blk1 ) { 184 assert( m->in(0) == blk1, "" ); 185 x->set_req( i, m->in(j) ); 186 } 187 } 188 register_new_node( x, blk1->in(j) ); 189 phi->init_req( j, x ); 190 } 191 // Announce phi to optimizer 192 register_new_node(phi, blk1); 193 194 // Remove cloned-up value from optimizer; use phi instead 195 _igvn.replace_node( n, phi ); 196 197 // (There used to be a self-recursive call to split_up() here, 198 // but it is not needed. All necessary forward walking is done 199 // by do_split_if() below.) 200 201 return true; 202 } 203 204 // Look for a (If .. (Bool(CmpP (LoadKlass .. (AddP obj ..)) ..))) and clone all of it down. 205 // There's likely a CheckCastPP on one of the branches of the If, with obj as input. 206 // If the (LoadKlass .. (AddP obj ..)) is not cloned down, then split if transforms this to: (If .. (Bool(CmpP phi1 ..))) 207 // and the CheckCastPP to (CheckCastPP phi2). It's possible then that phi2 is transformed to a CheckCastPP 208 // (through PhiNode::Ideal) and that that CheckCastPP is replaced by another narrower CheckCastPP at the same control 209 // (through ConstraintCastNode::Identity). That could cause the CheckCastPP at the If to become top while (CmpP phi1) 210 // wouldn't constant fold because it's using a different data path. Cloning the whole subgraph down guarantees both the 211 // AddP and CheckCastPP have the same obj input after split if. 212 bool PhaseIdealLoop::clone_cmp_loadklass_down(Node* n, const Node* blk1, const Node* blk2) { 213 if (n->Opcode() == Op_AddP && at_relevant_ctrl(n, blk1, blk2)) { 214 Node_List cmp_nodes; 215 uint old = C->unique(); 216 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 217 Node* u1 = n->fast_out(i); 218 if (u1->Opcode() == Op_LoadNKlass && at_relevant_ctrl(u1, blk1, blk2)) { 219 for (DUIterator_Fast jmax, j = u1->fast_outs(jmax); j < jmax; j++) { 220 Node* u2 = u1->fast_out(j); 221 if (u2->Opcode() == Op_DecodeNKlass && at_relevant_ctrl(u2, blk1, blk2)) { 222 for (DUIterator k = u2->outs(); u2->has_out(k); k++) { 223 Node* u3 = u2->out(k); 224 if (at_relevant_ctrl(u3, blk1, blk2) && clone_cmp_down(u3, blk1, blk2)) { 225 --k; 226 } 227 } 228 for (DUIterator_Fast kmax, k = u2->fast_outs(kmax); k < kmax; k++) { 229 Node* u3 = u2->fast_out(k); 230 if (u3->_idx >= old) { 231 cmp_nodes.push(u3); 232 } 233 } 234 } 235 } 236 } else if (u1->Opcode() == Op_LoadKlass && at_relevant_ctrl(u1, blk1, blk2)) { 237 for (DUIterator j = u1->outs(); u1->has_out(j); j++) { 238 Node* u2 = u1->out(j); 239 if (at_relevant_ctrl(u2, blk1, blk2) && clone_cmp_down(u2, blk1, blk2)) { 240 --j; 241 } 242 } 243 for (DUIterator_Fast kmax, k = u1->fast_outs(kmax); k < kmax; k++) { 244 Node* u2 = u1->fast_out(k); 245 if (u2->_idx >= old) { 246 cmp_nodes.push(u2); 247 } 248 } 249 } 250 } 251 252 for (uint i = 0; i < cmp_nodes.size(); ++i) { 253 Node* cmp = cmp_nodes.at(i); 254 clone_loadklass_nodes_at_cmp_index(n, cmp, 1); 255 clone_loadklass_nodes_at_cmp_index(n, cmp, 2); 256 } 257 if (n->outcnt() == 0) { 258 assert(n->is_dead(), ""); 259 return true; 260 } 261 } 262 return false; 263 } 264 265 bool PhaseIdealLoop::at_relevant_ctrl(Node* n, const Node* blk1, const Node* blk2) { 266 return ctrl_or_self(n) == blk1 || ctrl_or_self(n) == blk2; 267 } 268 269 void PhaseIdealLoop::clone_loadklass_nodes_at_cmp_index(const Node* n, Node* cmp, int i) { 270 Node* decode = cmp->in(i); 271 if (decode->Opcode() == Op_DecodeNKlass) { 272 Node* loadklass = decode->in(1); 273 if (loadklass->Opcode() == Op_LoadNKlass) { 274 Node* addp = loadklass->in(MemNode::Address); 275 if (addp == n) { 276 Node* ctrl = get_ctrl(cmp); 277 Node* decode_clone = decode->clone(); 278 Node* loadklass_clone = loadklass->clone(); 279 Node* addp_clone = addp->clone(); 280 register_new_node(decode_clone, ctrl); 281 register_new_node(loadklass_clone, ctrl); 282 register_new_node(addp_clone, ctrl); 283 _igvn.replace_input_of(cmp, i, decode_clone); 284 _igvn.replace_input_of(decode_clone, 1, loadklass_clone); 285 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone); 286 if (decode->outcnt() == 0) { 287 _igvn.remove_dead_node(decode); 288 } 289 } 290 } 291 } else { 292 Node* loadklass = cmp->in(i); 293 if (loadklass->Opcode() == Op_LoadKlass) { 294 Node* addp = loadklass->in(MemNode::Address); 295 if (addp == n) { 296 Node* ctrl = get_ctrl(cmp); 297 Node* loadklass_clone = loadklass->clone(); 298 Node* addp_clone = addp->clone(); 299 register_new_node(loadklass_clone, ctrl); 300 register_new_node(addp_clone, ctrl); 301 _igvn.replace_input_of(cmp, i, loadklass_clone); 302 _igvn.replace_input_of(loadklass_clone, MemNode::Address, addp_clone); 303 if (loadklass->outcnt() == 0) { 304 _igvn.remove_dead_node(loadklass); 305 } 306 } 307 } 308 } 309 } 310 311 bool PhaseIdealLoop::clone_cmp_down(Node* n, const Node* blk1, const Node* blk2) { 312 if( n->is_Cmp() ) { 313 assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF"); 314 // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove 315 // sequence can have no other users and it must all reside in the split-if 316 // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below - 317 // private, per-use versions of the Cmp and Bool are made. These sink to 318 // the CMove block. If the CMove is in the split-if block, then in the 319 // next iteration this will become a simple Cmp/Bool/CMove set to clone-up. 320 Node *bol, *cmov; 321 if (!(n->outcnt() == 1 && n->unique_out()->is_Bool() && 322 (bol = n->unique_out()->as_Bool()) && 323 (at_relevant_ctrl(bol, blk1, blk2) && 324 bol->outcnt() == 1 && 325 bol->unique_out()->is_CMove() && 326 (cmov = bol->unique_out()->as_CMove()) && 327 at_relevant_ctrl(cmov, blk1, blk2)))) { 328 329 // Must clone down 330 #ifndef PRODUCT 331 if( PrintOpto && VerifyLoopOptimizations ) { 332 tty->print("Cloning down: "); 333 n->dump(); 334 } 335 #endif 336 if (!n->is_FastLock()) { 337 // Clone down any block-local BoolNode uses of this CmpNode 338 for (DUIterator i = n->outs(); n->has_out(i); i++) { 339 Node* bol = n->out(i); 340 assert( bol->is_Bool(), "" ); 341 if (bol->outcnt() == 1) { 342 Node* use = bol->unique_out(); 343 if (use->Opcode() == Op_Opaque4) { 344 if (use->outcnt() == 1) { 345 Node* iff = use->unique_out(); 346 assert(iff->is_If(), "unexpected node type"); 347 Node *use_c = iff->in(0); 348 if (use_c == blk1 || use_c == blk2) { 349 continue; 350 } 351 } 352 } else { 353 // We might see an Opaque1 from a loop limit check here 354 assert(use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque1 || use->is_AllocateArray(), "unexpected node type"); 355 Node *use_c = (use->is_If() || use->is_AllocateArray()) ? use->in(0) : get_ctrl(use); 356 if (use_c == blk1 || use_c == blk2) { 357 assert(use->is_CMove(), "unexpected node type"); 358 continue; 359 } 360 } 361 } 362 if (at_relevant_ctrl(bol, blk1, blk2)) { 363 // Recursively sink any BoolNode 364 #ifndef PRODUCT 365 if( PrintOpto && VerifyLoopOptimizations ) { 366 tty->print("Cloning down: "); 367 bol->dump(); 368 } 369 #endif 370 for (DUIterator j = bol->outs(); bol->has_out(j); j++) { 371 Node* u = bol->out(j); 372 // Uses are either IfNodes, CMoves or Opaque4 373 if (u->Opcode() == Op_Opaque4) { 374 assert(u->in(1) == bol, "bad input"); 375 for (DUIterator_Last kmin, k = u->last_outs(kmin); k >= kmin; --k) { 376 Node* iff = u->last_out(k); 377 assert(iff->is_If() || iff->is_CMove(), "unexpected node type"); 378 assert( iff->in(1) == u, "" ); 379 // Get control block of either the CMove or the If input 380 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff); 381 Node *x1 = bol->clone(); 382 Node *x2 = u->clone(); 383 register_new_node(x1, iff_ctrl); 384 register_new_node(x2, iff_ctrl); 385 _igvn.replace_input_of(x2, 1, x1); 386 _igvn.replace_input_of(iff, 1, x2); 387 } 388 _igvn.remove_dead_node(u); 389 --j; 390 } else { 391 // We might see an Opaque1 from a loop limit check here 392 assert(u->is_If() || u->is_CMove() || u->Opcode() == Op_Opaque1 || u->is_AllocateArray(), "unexpected node type"); 393 assert(u->is_AllocateArray() || u->in(1) == bol, ""); 394 assert(!u->is_AllocateArray() || u->in(AllocateNode::ValidLengthTest) == bol, "wrong input to AllocateArray"); 395 // Get control block of either the CMove or the If input 396 Node *u_ctrl = (u->is_If() || u->is_AllocateArray()) ? u->in(0) : get_ctrl(u); 397 assert((u_ctrl != blk1 && u_ctrl != blk2) || u->is_CMove(), "won't converge"); 398 Node *x = bol->clone(); 399 register_new_node(x, u_ctrl); 400 _igvn.replace_input_of(u, u->is_AllocateArray() ? AllocateNode::ValidLengthTest : 1, x); 401 --j; 402 } 403 } 404 _igvn.remove_dead_node(bol); 405 --i; 406 } 407 } 408 } 409 // Clone down this CmpNode 410 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) { 411 Node* use = n->last_out(j); 412 uint pos = 1; 413 if (n->is_FastLock()) { 414 pos = TypeFunc::Parms + 2; 415 assert(use->is_Lock(), "FastLock only used by LockNode"); 416 } 417 assert(use->in(pos) == n, "" ); 418 Node *x = n->clone(); 419 register_new_node(x, ctrl_or_self(use)); 420 _igvn.replace_input_of(use, pos, x); 421 } 422 _igvn.remove_dead_node(n); 423 424 return true; 425 } 426 } 427 return false; 428 } 429 430 //------------------------------register_new_node------------------------------ 431 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) { 432 assert(!n->is_CFG(), "must be data node"); 433 _igvn.register_new_node_with_optimizer(n); 434 set_ctrl(n, blk); 435 IdealLoopTree *loop = get_loop(blk); 436 if( !loop->_child ) 437 loop->_body.push(n); 438 } 439 440 //------------------------------small_cache------------------------------------ 441 struct small_cache : public Dict { 442 443 small_cache() : Dict( cmpkey, hashptr ) {} 444 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); } 445 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); } 446 }; 447 448 //------------------------------spinup----------------------------------------- 449 // "Spin up" the dominator tree, starting at the use site and stopping when we 450 // find the post-dominating point. 451 452 // We must be at the merge point which post-dominates 'new_false' and 453 // 'new_true'. Figure out which edges into the RegionNode eventually lead up 454 // to false and which to true. Put in a PhiNode to merge values; plug in 455 // the appropriate false-arm or true-arm values. If some path leads to the 456 // original IF, then insert a Phi recursively. 457 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) { 458 if (use_blk->is_top()) // Handle dead uses 459 return use_blk; 460 Node *prior_n = (Node*)((intptr_t)0xdeadbeef); 461 Node *n = use_blk; // Get path input 462 assert( use_blk != iff_dom, "" ); 463 // Here's the "spinup" the dominator tree loop. Do a cache-check 464 // along the way, in case we've come this way before. 465 while( n != iff_dom ) { // Found post-dominating point? 466 prior_n = n; 467 n = idom(n); // Search higher 468 Node *s = cache->probe( prior_n ); // Check cache 469 if( s ) return s; // Cache hit! 470 } 471 472 Node *phi_post; 473 if( prior_n == new_false || prior_n == new_true ) { 474 phi_post = def->clone(); 475 phi_post->set_req(0, prior_n ); 476 register_new_node(phi_post, prior_n); 477 } else { 478 // This method handles both control uses (looking for Regions) or data 479 // uses (looking for Phis). If looking for a control use, then we need 480 // to insert a Region instead of a Phi; however Regions always exist 481 // previously (the hash_find_insert below would always hit) so we can 482 // return the existing Region. 483 if( def->is_CFG() ) { 484 phi_post = prior_n; // If looking for CFG, return prior 485 } else { 486 assert( def->is_Phi(), "" ); 487 assert( prior_n->is_Region(), "must be a post-dominating merge point" ); 488 489 // Need a Phi here 490 phi_post = PhiNode::make_blank(prior_n, def); 491 // Search for both true and false on all paths till find one. 492 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths 493 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) ); 494 Node *t = _igvn.hash_find_insert(phi_post); 495 if( t ) { // See if we already have this one 496 // phi_post will not be used, so kill it 497 _igvn.remove_dead_node(phi_post); 498 phi_post->destruct(&_igvn); 499 phi_post = t; 500 } else { 501 register_new_node( phi_post, prior_n ); 502 } 503 } 504 } 505 506 // Update cache everywhere 507 prior_n = (Node*)((intptr_t)0xdeadbeef); // Reset IDOM walk 508 n = use_blk; // Get path input 509 // Spin-up the idom tree again, basically doing path-compression. 510 // Insert cache entries along the way, so that if we ever hit this 511 // point in the IDOM tree again we'll stop immediately on a cache hit. 512 while( n != iff_dom ) { // Found post-dominating point? 513 prior_n = n; 514 n = idom(n); // Search higher 515 cache->lru_insert( prior_n, phi_post ); // Fill cache 516 } // End of while not gone high enough 517 518 return phi_post; 519 } 520 521 //------------------------------find_use_block--------------------------------- 522 // Find the block a USE is in. Normally USE's are in the same block as the 523 // using instruction. For Phi-USE's, the USE is in the predecessor block 524 // along the corresponding path. 525 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) { 526 // CFG uses are their own block 527 if( use->is_CFG() ) 528 return use; 529 530 if( use->is_Phi() ) { // Phi uses in prior block 531 // Grab the first Phi use; there may be many. 532 // Each will be handled as a separate iteration of 533 // the "while( phi->outcnt() )" loop. 534 uint j; 535 for( j = 1; j < use->req(); j++ ) 536 if( use->in(j) == def ) 537 break; 538 assert( j < use->req(), "def should be among use's inputs" ); 539 return use->in(0)->in(j); 540 } 541 // Normal (non-phi) use 542 Node *use_blk = get_ctrl(use); 543 // Some uses are directly attached to the old (and going away) 544 // false and true branches. 545 if( use_blk == old_false ) { 546 use_blk = new_false; 547 set_ctrl(use, new_false); 548 } 549 if( use_blk == old_true ) { 550 use_blk = new_true; 551 set_ctrl(use, new_true); 552 } 553 554 if (use_blk == NULL) { // He's dead, Jim 555 _igvn.replace_node(use, C->top()); 556 } 557 558 return use_blk; 559 } 560 561 //------------------------------handle_use------------------------------------- 562 // Handle uses of the merge point. Basically, split-if makes the merge point 563 // go away so all uses of the merge point must go away as well. Most block 564 // local uses have already been split-up, through the merge point. Uses from 565 // far below the merge point can't always be split up (e.g., phi-uses are 566 // pinned) and it makes too much stuff live. Instead we use a path-based 567 // solution to move uses down. 568 // 569 // If the use is along the pre-split-CFG true branch, then the new use will 570 // be from the post-split-CFG true merge point. Vice-versa for the false 571 // path. Some uses will be along both paths; then we sink the use to the 572 // post-dominating location; we may need to insert a Phi there. 573 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 ) { 574 575 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true); 576 if( !use_blk ) return; // He's dead, Jim 577 578 // Walk up the dominator tree until I hit either the old IfFalse, the old 579 // IfTrue or the old If. Insert Phis where needed. 580 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache ); 581 582 // Found where this USE goes. Re-point him. 583 uint i; 584 for( i = 0; i < use->req(); i++ ) 585 if( use->in(i) == def ) 586 break; 587 assert( i < use->req(), "def should be among use's inputs" ); 588 _igvn.replace_input_of(use, i, new_def); 589 } 590 591 //------------------------------do_split_if------------------------------------ 592 // Found an If getting its condition-code input from a Phi in the same block. 593 // Split thru the Region. 594 void PhaseIdealLoop::do_split_if(Node* iff, RegionNode** new_false_region, RegionNode** new_true_region) { 595 if (PrintOpto && VerifyLoopOptimizations) { 596 tty->print_cr("Split-if"); 597 } 598 if (TraceLoopOpts) { 599 tty->print_cr("SplitIf"); 600 } 601 602 C->set_major_progress(); 603 RegionNode *region = iff->in(0)->as_Region(); 604 Node *region_dom = idom(region); 605 606 // We are going to clone this test (and the control flow with it) up through 607 // the incoming merge point. We need to empty the current basic block. 608 // Clone any instructions which must be in this block up through the merge 609 // point. 610 DUIterator i, j; 611 bool progress = true; 612 while (progress) { 613 progress = false; 614 for (i = region->outs(); region->has_out(i); i++) { 615 Node* n = region->out(i); 616 if( n == region ) continue; 617 // The IF to be split is OK. 618 if( n == iff ) continue; 619 if( !n->is_Phi() ) { // Found pinned memory op or such 620 if (split_up(n, region, iff)) { 621 i = region->refresh_out_pos(i); 622 progress = true; 623 } 624 continue; 625 } 626 assert( n->in(0) == region, "" ); 627 628 // Recursively split up all users of a Phi 629 for (j = n->outs(); n->has_out(j); j++) { 630 Node* m = n->out(j); 631 // If m is dead, throw it away, and declare progress 632 if (_nodes[m->_idx] == NULL) { 633 _igvn.remove_dead_node(m); 634 // fall through 635 } else if (m != iff && split_up(m, region, iff)) { 636 // fall through 637 } else { 638 continue; 639 } 640 // Something unpredictable changed. 641 // Tell the iterators to refresh themselves, and rerun the loop. 642 i = region->refresh_out_pos(i); 643 j = region->refresh_out_pos(j); 644 progress = true; 645 } 646 } 647 } 648 649 // Now we have no instructions in the block containing the IF. 650 // Split the IF. 651 RegionNode *new_iff = split_thru_region(iff, region); 652 653 // Replace both uses of 'new_iff' with Regions merging True/False 654 // paths. This makes 'new_iff' go dead. 655 Node *old_false = NULL, *old_true = NULL; 656 RegionNode* new_false = NULL; 657 RegionNode* new_true = NULL; 658 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) { 659 Node *ifp = iff->last_out(j2); 660 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" ); 661 ifp->set_req(0, new_iff); 662 RegionNode* ifpx = split_thru_region(ifp, region); 663 664 // Replace 'If' projection of a Region with a Region of 665 // 'If' projections. 666 ifpx->set_req(0, ifpx); // A TRUE RegionNode 667 668 // Setup dominator info 669 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1); 670 671 // Check for splitting loop tails 672 if( get_loop(iff)->tail() == ifp ) 673 get_loop(iff)->_tail = ifpx; 674 675 // Replace in the graph with lazy-update mechanism 676 new_iff->set_req(0, new_iff); // hook self so it does not go dead 677 lazy_replace(ifp, ifpx); 678 new_iff->set_req(0, region); 679 680 // Record bits for later xforms 681 if( ifp->Opcode() == Op_IfFalse ) { 682 old_false = ifp; 683 new_false = ifpx; 684 } else { 685 old_true = ifp; 686 new_true = ifpx; 687 } 688 } 689 _igvn.remove_dead_node(new_iff); 690 // Lazy replace IDOM info with the region's dominator 691 lazy_replace(iff, region_dom); 692 lazy_update(region, region_dom); // idom must be update before handle_uses 693 region->set_req(0, NULL); // Break the self-cycle. Required for lazy_update to work on region 694 695 // Now make the original merge point go dead, by handling all its uses. 696 small_cache region_cache; 697 // Preload some control flow in region-cache 698 region_cache.lru_insert( new_false, new_false ); 699 region_cache.lru_insert( new_true , new_true ); 700 // Now handle all uses of the splitting block 701 for (DUIterator k = region->outs(); region->has_out(k); k++) { 702 Node* phi = region->out(k); 703 if (!phi->in(0)) { // Dead phi? Remove it 704 _igvn.remove_dead_node(phi); 705 } else if (phi == region) { // Found the self-reference 706 continue; // No roll-back of DUIterator 707 } else if (phi->is_Phi()) { // Expected common case: Phi hanging off of Region 708 assert(phi->in(0) == region, "Inconsistent graph"); 709 // Need a per-def cache. Phi represents a def, so make a cache 710 small_cache phi_cache; 711 712 // Inspect all Phi uses to make the Phi go dead 713 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) { 714 Node* use = phi->last_out(l); 715 // Compute the new DEF for this USE. New DEF depends on the path 716 // taken from the original DEF to the USE. The new DEF may be some 717 // collection of PHI's merging values from different paths. The Phis 718 // inserted depend only on the location of the USE. We use a 719 // 2-element cache to handle multiple uses from the same block. 720 handle_use(use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true); 721 } // End of while phi has uses 722 // Remove the dead Phi 723 _igvn.remove_dead_node( phi ); 724 } else { 725 assert(phi->in(0) == region, "Inconsistent graph"); 726 // Random memory op guarded by Region. Compute new DEF for USE. 727 handle_use(phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true); 728 } 729 // Every path above deletes a use of the region, except for the region 730 // self-cycle (which is needed by handle_use calling find_use_block 731 // calling get_ctrl calling get_ctrl_no_update looking for dead 732 // regions). So roll back the DUIterator innards. 733 --k; 734 } // End of while merge point has phis 735 736 _igvn.remove_dead_node(region); 737 738 if (new_false_region != NULL) { 739 *new_false_region = new_false; 740 } 741 if (new_true_region != NULL) { 742 *new_true_region = new_true; 743 } 744 745 #ifndef PRODUCT 746 if( VerifyLoopOptimizations ) verify(); 747 #endif 748 }