1 /* 2 * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/callnode.hpp" 28 #include "opto/loopnode.hpp" 29 #include "opto/movenode.hpp" 30 31 32 //------------------------------split_thru_region------------------------------ 33 // Split Node 'n' through merge point. 34 Node *PhaseIdealLoop::split_thru_region( Node *n, Node *region ) { 35 uint wins = 0; 36 assert( n->is_CFG(), "" ); 37 assert( region->is_Region(), "" ); 38 Node *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 _igvn.register_new_node_with_optimizer(x); 50 set_loop(x, loop); 51 set_idom(x, x->in(0), dom_depth(x->in(0))+1); 52 r->init_req(i, x); 53 } 54 55 // Record region 56 r->set_req(0,region); // Not a TRUE RegionNode 57 _igvn.register_new_node_with_optimizer(r); 58 set_loop(r, loop); 59 if( !loop->_child ) 60 loop->_body.push(r); 61 return r; 62 } 63 64 //------------------------------split_up--------------------------------------- 65 // Split block-local op up through the phis to empty the current block 66 bool PhaseIdealLoop::split_up( Node *n, Node *blk1, Node *blk2 ) { 67 if( n->is_CFG() ) { 68 assert( n->in(0) != blk1, "Lousy candidate for split-if" ); 69 return false; 70 } 71 if( get_ctrl(n) != blk1 && get_ctrl(n) != blk2 ) 72 return false; // Not block local 73 if( n->is_Phi() ) return false; // Local PHIs are expected 74 75 // Recursively split-up inputs 76 for (uint i = 1; i < n->req(); i++) { 77 if( split_up( n->in(i), blk1, blk2 ) ) { 78 // Got split recursively and self went dead? 79 if (n->outcnt() == 0) 80 _igvn.remove_dead_node(n); 81 return true; 82 } 83 } 84 85 // Check for needing to clone-up a compare. Can't do that, it forces 86 // another (nested) split-if transform. Instead, clone it "down". 87 if( n->is_Cmp() ) { 88 assert(get_ctrl(n) == blk2 || get_ctrl(n) == blk1, "must be in block with IF"); 89 // Check for simple Cmp/Bool/CMove which we can clone-up. Cmp/Bool/CMove 90 // sequence can have no other users and it must all reside in the split-if 91 // block. Non-simple Cmp/Bool/CMove sequences are 'cloned-down' below - 92 // private, per-use versions of the Cmp and Bool are made. These sink to 93 // the CMove block. If the CMove is in the split-if block, then in the 94 // next iteration this will become a simple Cmp/Bool/CMove set to clone-up. 95 Node *bol, *cmov; 96 if( !(n->outcnt() == 1 && n->unique_out()->is_Bool() && 97 (bol = n->unique_out()->as_Bool()) && 98 (get_ctrl(bol) == blk1 || 99 get_ctrl(bol) == blk2) && 100 bol->outcnt() == 1 && 101 bol->unique_out()->is_CMove() && 102 (cmov = bol->unique_out()->as_CMove()) && 103 (get_ctrl(cmov) == blk1 || 104 get_ctrl(cmov) == blk2) ) ) { 105 106 // Must clone down 107 #ifndef PRODUCT 108 if( PrintOpto && VerifyLoopOptimizations ) { 109 tty->print("Cloning down: "); 110 n->dump(); 111 } 112 #endif 113 // Clone down any block-local BoolNode uses of this CmpNode 114 for (DUIterator i = n->outs(); n->has_out(i); i++) { 115 Node* bol = n->out(i); 116 assert( bol->is_Bool(), "" ); 117 if (bol->outcnt() == 1) { 118 Node* use = bol->unique_out(); 119 if (use->Opcode() == Op_Opaque4) { 120 if (use->outcnt() == 1) { 121 Node* iff = use->unique_out(); 122 assert(iff->is_If(), "unexpected node type"); 123 Node *use_c = iff->in(0); 124 if (use_c == blk1 || use_c == blk2) { 125 continue; 126 } 127 } 128 } else { 129 // We might see an Opaque1 from a loop limit check here 130 assert(use->is_If() || use->is_CMove() || use->Opcode() == Op_Opaque1, "unexpected node type"); 131 Node *use_c = use->is_If() ? use->in(0) : get_ctrl(use); 132 if (use_c == blk1 || use_c == blk2) { 133 assert(use->is_CMove(), "unexpected node type"); 134 continue; 135 } 136 } 137 } 138 if (get_ctrl(bol) == blk1 || get_ctrl(bol) == blk2) { 139 // Recursively sink any BoolNode 140 #ifndef PRODUCT 141 if( PrintOpto && VerifyLoopOptimizations ) { 142 tty->print("Cloning down: "); 143 bol->dump(); 144 } 145 #endif 146 for (DUIterator j = bol->outs(); bol->has_out(j); j++) { 147 Node* u = bol->out(j); 148 // Uses are either IfNodes, CMoves or Opaque4 149 if (u->Opcode() == Op_Opaque4) { 150 assert(u->in(1) == bol, "bad input"); 151 for (DUIterator_Last kmin, k = u->last_outs(kmin); k >= kmin; --k) { 152 Node* iff = u->last_out(k); 153 assert(iff->is_If() || iff->is_CMove(), "unexpected node type"); 154 assert( iff->in(1) == u, "" ); 155 // Get control block of either the CMove or the If input 156 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff); 157 Node *x1 = bol->clone(); 158 Node *x2 = u->clone(); 159 register_new_node(x1, iff_ctrl); 160 register_new_node(x2, iff_ctrl); 161 _igvn.replace_input_of(x2, 1, x1); 162 _igvn.replace_input_of(iff, 1, x2); 163 } 164 _igvn.remove_dead_node(u); 165 --j; 166 } else { 167 // We might see an Opaque1 from a loop limit check here 168 assert(u->is_If() || u->is_CMove() || u->Opcode() == Op_Opaque1, "unexpected node type"); 169 assert(u->in(1) == bol, ""); 170 // Get control block of either the CMove or the If input 171 Node *u_ctrl = u->is_If() ? u->in(0) : get_ctrl(u); 172 assert((u_ctrl != blk1 && u_ctrl != blk2) || u->is_CMove(), "won't converge"); 173 Node *x = bol->clone(); 174 register_new_node(x, u_ctrl); 175 _igvn.replace_input_of(u, 1, x); 176 --j; 177 } 178 } 179 _igvn.remove_dead_node(bol); 180 --i; 181 } 182 } 183 // Clone down this CmpNode 184 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) { 185 Node* bol = n->last_out(j); 186 assert( bol->in(1) == n, "" ); 187 Node *x = n->clone(); 188 register_new_node(x, get_ctrl(bol)); 189 _igvn.replace_input_of(bol, 1, x); 190 } 191 _igvn.remove_dead_node( n ); 192 193 return true; 194 } 195 } 196 197 // See if splitting-up a Store. Any anti-dep loads must go up as 198 // well. An anti-dep load might be in the wrong block, because in 199 // this particular layout/schedule we ignored anti-deps and allow 200 // memory to be alive twice. This only works if we do the same 201 // operations on anti-dep loads as we do their killing stores. 202 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) { 203 // Get store's memory slice 204 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr()); 205 206 // Get memory-phi anti-dep loads will be using 207 Node *memphi = n->in(MemNode::Memory); 208 assert( memphi->is_Phi(), "" ); 209 // Hoist any anti-dep load to the splitting block; 210 // it will then "split-up". 211 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) { 212 Node *load = memphi->fast_out(i); 213 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) ) 214 set_ctrl(load,blk1); 215 } 216 } 217 218 // Found some other Node; must clone it up 219 #ifndef PRODUCT 220 if( PrintOpto && VerifyLoopOptimizations ) { 221 tty->print("Cloning up: "); 222 n->dump(); 223 } 224 #endif 225 226 // ConvI2L may have type information on it which becomes invalid if 227 // it moves up in the graph so change any clones so widen the type 228 // to TypeLong::INT when pushing it up. 229 const Type* rtype = NULL; 230 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) { 231 rtype = TypeLong::INT; 232 } 233 234 // Now actually split-up this guy. One copy per control path merging. 235 Node *phi = PhiNode::make_blank(blk1, n); 236 for( uint j = 1; j < blk1->req(); j++ ) { 237 Node *x = n->clone(); 238 // Widen the type of the ConvI2L when pushing up. 239 if (rtype != NULL) x->as_Type()->set_type(rtype); 240 if( n->in(0) && n->in(0) == blk1 ) 241 x->set_req( 0, blk1->in(j) ); 242 for( uint i = 1; i < n->req(); i++ ) { 243 Node *m = n->in(i); 244 if( get_ctrl(m) == blk1 ) { 245 assert( m->in(0) == blk1, "" ); 246 x->set_req( i, m->in(j) ); 247 } 248 } 249 register_new_node( x, blk1->in(j) ); 250 phi->init_req( j, x ); 251 } 252 // Announce phi to optimizer 253 register_new_node(phi, blk1); 254 255 // Remove cloned-up value from optimizer; use phi instead 256 _igvn.replace_node( n, phi ); 257 258 // (There used to be a self-recursive call to split_up() here, 259 // but it is not needed. All necessary forward walking is done 260 // by do_split_if() below.) 261 262 return true; 263 } 264 265 //------------------------------register_new_node------------------------------ 266 void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) { 267 assert(!n->is_CFG(), "must be data node"); 268 _igvn.register_new_node_with_optimizer(n); 269 set_ctrl(n, blk); 270 IdealLoopTree *loop = get_loop(blk); 271 if( !loop->_child ) 272 loop->_body.push(n); 273 } 274 275 //------------------------------small_cache------------------------------------ 276 struct small_cache : public Dict { 277 278 small_cache() : Dict( cmpkey, hashptr ) {} 279 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); } 280 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); } 281 }; 282 283 //------------------------------spinup----------------------------------------- 284 // "Spin up" the dominator tree, starting at the use site and stopping when we 285 // find the post-dominating point. 286 287 // We must be at the merge point which post-dominates 'new_false' and 288 // 'new_true'. Figure out which edges into the RegionNode eventually lead up 289 // to false and which to true. Put in a PhiNode to merge values; plug in 290 // the appropriate false-arm or true-arm values. If some path leads to the 291 // original IF, then insert a Phi recursively. 292 Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) { 293 if (use_blk->is_top()) // Handle dead uses 294 return use_blk; 295 Node *prior_n = (Node*)((intptr_t)0xdeadbeef); 296 Node *n = use_blk; // Get path input 297 assert( use_blk != iff_dom, "" ); 298 // Here's the "spinup" the dominator tree loop. Do a cache-check 299 // along the way, in case we've come this way before. 300 while( n != iff_dom ) { // Found post-dominating point? 301 prior_n = n; 302 n = idom(n); // Search higher 303 Node *s = cache->probe( prior_n ); // Check cache 304 if( s ) return s; // Cache hit! 305 } 306 307 Node *phi_post; 308 if( prior_n == new_false || prior_n == new_true ) { 309 phi_post = def->clone(); 310 phi_post->set_req(0, prior_n ); 311 register_new_node(phi_post, prior_n); 312 } else { 313 // This method handles both control uses (looking for Regions) or data 314 // uses (looking for Phis). If looking for a control use, then we need 315 // to insert a Region instead of a Phi; however Regions always exist 316 // previously (the hash_find_insert below would always hit) so we can 317 // return the existing Region. 318 if( def->is_CFG() ) { 319 phi_post = prior_n; // If looking for CFG, return prior 320 } else { 321 assert( def->is_Phi(), "" ); 322 assert( prior_n->is_Region(), "must be a post-dominating merge point" ); 323 324 // Need a Phi here 325 phi_post = PhiNode::make_blank(prior_n, def); 326 // Search for both true and false on all paths till find one. 327 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths 328 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) ); 329 Node *t = _igvn.hash_find_insert(phi_post); 330 if( t ) { // See if we already have this one 331 // phi_post will not be used, so kill it 332 _igvn.remove_dead_node(phi_post); 333 phi_post->destruct(); 334 phi_post = t; 335 } else { 336 register_new_node( phi_post, prior_n ); 337 } 338 } 339 } 340 341 // Update cache everywhere 342 prior_n = (Node*)((intptr_t)0xdeadbeef); // Reset IDOM walk 343 n = use_blk; // Get path input 344 // Spin-up the idom tree again, basically doing path-compression. 345 // Insert cache entries along the way, so that if we ever hit this 346 // point in the IDOM tree again we'll stop immediately on a cache hit. 347 while( n != iff_dom ) { // Found post-dominating point? 348 prior_n = n; 349 n = idom(n); // Search higher 350 cache->lru_insert( prior_n, phi_post ); // Fill cache 351 } // End of while not gone high enough 352 353 return phi_post; 354 } 355 356 //------------------------------find_use_block--------------------------------- 357 // Find the block a USE is in. Normally USE's are in the same block as the 358 // using instruction. For Phi-USE's, the USE is in the predecessor block 359 // along the corresponding path. 360 Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) { 361 // CFG uses are their own block 362 if( use->is_CFG() ) 363 return use; 364 365 if( use->is_Phi() ) { // Phi uses in prior block 366 // Grab the first Phi use; there may be many. 367 // Each will be handled as a separate iteration of 368 // the "while( phi->outcnt() )" loop. 369 uint j; 370 for( j = 1; j < use->req(); j++ ) 371 if( use->in(j) == def ) 372 break; 373 assert( j < use->req(), "def should be among use's inputs" ); 374 return use->in(0)->in(j); 375 } 376 // Normal (non-phi) use 377 Node *use_blk = get_ctrl(use); 378 // Some uses are directly attached to the old (and going away) 379 // false and true branches. 380 if( use_blk == old_false ) { 381 use_blk = new_false; 382 set_ctrl(use, new_false); 383 } 384 if( use_blk == old_true ) { 385 use_blk = new_true; 386 set_ctrl(use, new_true); 387 } 388 389 if (use_blk == NULL) { // He's dead, Jim 390 _igvn.replace_node(use, C->top()); 391 } 392 393 return use_blk; 394 } 395 396 //------------------------------handle_use------------------------------------- 397 // Handle uses of the merge point. Basically, split-if makes the merge point 398 // go away so all uses of the merge point must go away as well. Most block 399 // local uses have already been split-up, through the merge point. Uses from 400 // far below the merge point can't always be split up (e.g., phi-uses are 401 // pinned) and it makes too much stuff live. Instead we use a path-based 402 // solution to move uses down. 403 // 404 // If the use is along the pre-split-CFG true branch, then the new use will 405 // be from the post-split-CFG true merge point. Vice-versa for the false 406 // path. Some uses will be along both paths; then we sink the use to the 407 // post-dominating location; we may need to insert a Phi there. 408 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 ) { 409 410 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true); 411 if( !use_blk ) return; // He's dead, Jim 412 413 // Walk up the dominator tree until I hit either the old IfFalse, the old 414 // IfTrue or the old If. Insert Phis where needed. 415 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache ); 416 417 // Found where this USE goes. Re-point him. 418 uint i; 419 for( i = 0; i < use->req(); i++ ) 420 if( use->in(i) == def ) 421 break; 422 assert( i < use->req(), "def should be among use's inputs" ); 423 _igvn.replace_input_of(use, i, new_def); 424 } 425 426 //------------------------------do_split_if------------------------------------ 427 // Found an If getting its condition-code input from a Phi in the same block. 428 // Split thru the Region. 429 void PhaseIdealLoop::do_split_if( Node *iff ) { 430 if (PrintOpto && VerifyLoopOptimizations) { 431 tty->print_cr("Split-if"); 432 } 433 if (TraceLoopOpts) { 434 tty->print_cr("SplitIf"); 435 } 436 437 C->set_major_progress(); 438 Node *region = iff->in(0); 439 Node *region_dom = idom(region); 440 441 // We are going to clone this test (and the control flow with it) up through 442 // the incoming merge point. We need to empty the current basic block. 443 // Clone any instructions which must be in this block up through the merge 444 // point. 445 DUIterator i, j; 446 bool progress = true; 447 while (progress) { 448 progress = false; 449 for (i = region->outs(); region->has_out(i); i++) { 450 Node* n = region->out(i); 451 if( n == region ) continue; 452 // The IF to be split is OK. 453 if( n == iff ) continue; 454 if( !n->is_Phi() ) { // Found pinned memory op or such 455 if (split_up(n, region, iff)) { 456 i = region->refresh_out_pos(i); 457 progress = true; 458 } 459 continue; 460 } 461 assert( n->in(0) == region, "" ); 462 463 // Recursively split up all users of a Phi 464 for (j = n->outs(); n->has_out(j); j++) { 465 Node* m = n->out(j); 466 // If m is dead, throw it away, and declare progress 467 if (_nodes[m->_idx] == NULL) { 468 _igvn.remove_dead_node(m); 469 // fall through 470 } 471 else if (m != iff && split_up(m, region, iff)) { 472 // fall through 473 } else { 474 continue; 475 } 476 // Something unpredictable changed. 477 // Tell the iterators to refresh themselves, and rerun the loop. 478 i = region->refresh_out_pos(i); 479 j = region->refresh_out_pos(j); 480 progress = true; 481 } 482 } 483 } 484 485 // Now we have no instructions in the block containing the IF. 486 // Split the IF. 487 Node *new_iff = split_thru_region( iff, region ); 488 489 // Replace both uses of 'new_iff' with Regions merging True/False 490 // paths. This makes 'new_iff' go dead. 491 Node *old_false = NULL, *old_true = NULL; 492 Node *new_false = NULL, *new_true = NULL; 493 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) { 494 Node *ifp = iff->last_out(j2); 495 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" ); 496 ifp->set_req(0, new_iff); 497 Node *ifpx = split_thru_region( ifp, region ); 498 499 // Replace 'If' projection of a Region with a Region of 500 // 'If' projections. 501 ifpx->set_req(0, ifpx); // A TRUE RegionNode 502 503 // Setup dominator info 504 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1); 505 506 // Check for splitting loop tails 507 if( get_loop(iff)->tail() == ifp ) 508 get_loop(iff)->_tail = ifpx; 509 510 // Replace in the graph with lazy-update mechanism 511 new_iff->set_req(0, new_iff); // hook self so it does not go dead 512 lazy_replace(ifp, ifpx); 513 new_iff->set_req(0, region); 514 515 // Record bits for later xforms 516 if( ifp->Opcode() == Op_IfFalse ) { 517 old_false = ifp; 518 new_false = ifpx; 519 } else { 520 old_true = ifp; 521 new_true = ifpx; 522 } 523 } 524 _igvn.remove_dead_node(new_iff); 525 // Lazy replace IDOM info with the region's dominator 526 lazy_replace( iff, region_dom ); 527 528 // Now make the original merge point go dead, by handling all its uses. 529 small_cache region_cache; 530 // Preload some control flow in region-cache 531 region_cache.lru_insert( new_false, new_false ); 532 region_cache.lru_insert( new_true , new_true ); 533 // Now handle all uses of the splitting block 534 for (DUIterator k = region->outs(); region->has_out(k); k++) { 535 Node* phi = region->out(k); 536 if (!phi->in(0)) { // Dead phi? Remove it 537 _igvn.remove_dead_node(phi); 538 } else if (phi == region) { // Found the self-reference 539 continue; // No roll-back of DUIterator 540 } else if (phi->is_Phi()) { // Expected common case: Phi hanging off of Region 541 assert(phi->in(0) == region, "Inconsistent graph"); 542 // Need a per-def cache. Phi represents a def, so make a cache 543 small_cache phi_cache; 544 545 // Inspect all Phi uses to make the Phi go dead 546 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) { 547 Node* use = phi->last_out(l); 548 // Compute the new DEF for this USE. New DEF depends on the path 549 // taken from the original DEF to the USE. The new DEF may be some 550 // collection of PHI's merging values from different paths. The Phis 551 // inserted depend only on the location of the USE. We use a 552 // 2-element cache to handle multiple uses from the same block. 553 handle_use(use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true); 554 } // End of while phi has uses 555 // Remove the dead Phi 556 _igvn.remove_dead_node( phi ); 557 } else { 558 assert(phi->in(0) == region, "Inconsistent graph"); 559 // Random memory op guarded by Region. Compute new DEF for USE. 560 handle_use(phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true); 561 } 562 // Every path above deletes a use of the region, except for the region 563 // self-cycle (which is needed by handle_use calling find_use_block 564 // calling get_ctrl calling get_ctrl_no_update looking for dead 565 // regions). So roll back the DUIterator innards. 566 --k; 567 } // End of while merge point has phis 568 569 assert(region->outcnt() == 1, "Only self reference should remain"); // Just Self on the Region 570 region->set_req(0, NULL); // Break the self-cycle 571 572 // Any leftover bits in the splitting block must not have depended on local 573 // Phi inputs (these have already been split-up). Hence it's safe to hoist 574 // these guys to the dominating point. 575 lazy_replace( region, region_dom ); 576 #ifndef PRODUCT 577 if( VerifyLoopOptimizations ) verify(); 578 #endif 579 }