1/* 2 * Copyright (c) 1999, 2012, 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. 34Node *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 66bool 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 Node *use_c = use->is_If() ? use->in(0) : get_ctrl(use); 120 if (use_c == blk1 || use_c == blk2) { 121 continue; 122 } 123 } 124 if (get_ctrl(bol) == blk1 || get_ctrl(bol) == blk2) { 125 // Recursively sink any BoolNode 126#ifndef PRODUCT 127 if( PrintOpto && VerifyLoopOptimizations ) { 128 tty->print("Cloning down: "); 129 bol->dump(); 130 } 131#endif 132 for (DUIterator_Last jmin, j = bol->last_outs(jmin); j >= jmin; --j) { 133 // Uses are either IfNodes or CMoves 134 Node* iff = bol->last_out(j); 135 assert( iff->in(1) == bol, "" ); 136 // Get control block of either the CMove or the If input 137 Node *iff_ctrl = iff->is_If() ? iff->in(0) : get_ctrl(iff); 138 Node *x = bol->clone(); 139 register_new_node(x, iff_ctrl); 140 _igvn.replace_input_of(iff, 1, x); 141 } 142 _igvn.remove_dead_node( bol ); 143 --i; 144 } 145 } 146 // Clone down this CmpNode 147 for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; --j) { 148 Node* bol = n->last_out(j); 149 assert( bol->in(1) == n, "" ); 150 Node *x = n->clone(); 151 register_new_node(x, get_ctrl(bol)); 152 _igvn.replace_input_of(bol, 1, x); 153 } 154 _igvn.remove_dead_node( n ); 155 156 return true; 157 } 158 } 159 160 // See if splitting-up a Store. Any anti-dep loads must go up as 161 // well. An anti-dep load might be in the wrong block, because in 162 // this particular layout/schedule we ignored anti-deps and allow 163 // memory to be alive twice. This only works if we do the same 164 // operations on anti-dep loads as we do their killing stores. 165 if( n->is_Store() && n->in(MemNode::Memory)->in(0) == n->in(0) ) { 166 // Get store's memory slice 167 int alias_idx = C->get_alias_index(_igvn.type(n->in(MemNode::Address))->is_ptr()); 168 169 // Get memory-phi anti-dep loads will be using 170 Node *memphi = n->in(MemNode::Memory); 171 assert( memphi->is_Phi(), "" ); 172 // Hoist any anti-dep load to the splitting block; 173 // it will then "split-up". 174 for (DUIterator_Fast imax,i = memphi->fast_outs(imax); i < imax; i++) { 175 Node *load = memphi->fast_out(i); 176 if( load->is_Load() && alias_idx == C->get_alias_index(_igvn.type(load->in(MemNode::Address))->is_ptr()) ) 177 set_ctrl(load,blk1); 178 } 179 } 180 181 // Found some other Node; must clone it up 182#ifndef PRODUCT 183 if( PrintOpto && VerifyLoopOptimizations ) { 184 tty->print("Cloning up: "); 185 n->dump(); 186 } 187#endif 188 189 // ConvI2L may have type information on it which becomes invalid if 190 // it moves up in the graph so change any clones so widen the type 191 // to TypeLong::INT when pushing it up. 192 const Type* rtype = NULL; 193 if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::INT) { 194 rtype = TypeLong::INT; 195 } 196 197 // Now actually split-up this guy. One copy per control path merging. 198 Node *phi = PhiNode::make_blank(blk1, n); 199 for( uint j = 1; j < blk1->req(); j++ ) { 200 Node *x = n->clone(); 201 // Widen the type of the ConvI2L when pushing up. 202 if (rtype != NULL) x->as_Type()->set_type(rtype); 203 if( n->in(0) && n->in(0) == blk1 ) 204 x->set_req( 0, blk1->in(j) ); 205 for( uint i = 1; i < n->req(); i++ ) { 206 Node *m = n->in(i); 207 if( get_ctrl(m) == blk1 ) { 208 assert( m->in(0) == blk1, "" ); 209 x->set_req( i, m->in(j) ); 210 } 211 } 212 register_new_node( x, blk1->in(j) ); 213 phi->init_req( j, x ); 214 } 215 // Announce phi to optimizer 216 register_new_node(phi, blk1); 217 218 // Remove cloned-up value from optimizer; use phi instead 219 _igvn.replace_node( n, phi ); 220 221 // (There used to be a self-recursive call to split_up() here, 222 // but it is not needed. All necessary forward walking is done 223 // by do_split_if() below.) 224 225 return true; 226} 227 228//------------------------------register_new_node------------------------------ 229void PhaseIdealLoop::register_new_node( Node *n, Node *blk ) { 230 assert(!n->is_CFG(), "must be data node"); 231 _igvn.register_new_node_with_optimizer(n); 232 set_ctrl(n, blk); 233 IdealLoopTree *loop = get_loop(blk); 234 if( !loop->_child ) 235 loop->_body.push(n); 236} 237 238//------------------------------small_cache------------------------------------ 239struct small_cache : public Dict { 240 241 small_cache() : Dict( cmpkey, hashptr ) {} 242 Node *probe( Node *use_blk ) { return (Node*)((*this)[use_blk]); } 243 void lru_insert( Node *use_blk, Node *new_def ) { Insert(use_blk,new_def); } 244}; 245 246//------------------------------spinup----------------------------------------- 247// "Spin up" the dominator tree, starting at the use site and stopping when we 248// find the post-dominating point. 249 250// We must be at the merge point which post-dominates 'new_false' and 251// 'new_true'. Figure out which edges into the RegionNode eventually lead up 252// to false and which to true. Put in a PhiNode to merge values; plug in 253// the appropriate false-arm or true-arm values. If some path leads to the 254// original IF, then insert a Phi recursively. 255Node *PhaseIdealLoop::spinup( Node *iff_dom, Node *new_false, Node *new_true, Node *use_blk, Node *def, small_cache *cache ) { 256 if (use_blk->is_top()) // Handle dead uses 257 return use_blk; 258 Node *prior_n = (Node*)0xdeadbeef; 259 Node *n = use_blk; // Get path input 260 assert( use_blk != iff_dom, "" ); 261 // Here's the "spinup" the dominator tree loop. Do a cache-check 262 // along the way, in case we've come this way before. 263 while( n != iff_dom ) { // Found post-dominating point? 264 prior_n = n; 265 n = idom(n); // Search higher 266 Node *s = cache->probe( prior_n ); // Check cache 267 if( s ) return s; // Cache hit! 268 } 269 270 Node *phi_post; 271 if( prior_n == new_false || prior_n == new_true ) { 272 phi_post = def->clone(); 273 phi_post->set_req(0, prior_n ); 274 register_new_node(phi_post, prior_n); 275 } else { 276 // This method handles both control uses (looking for Regions) or data 277 // uses (looking for Phis). If looking for a control use, then we need 278 // to insert a Region instead of a Phi; however Regions always exist 279 // previously (the hash_find_insert below would always hit) so we can 280 // return the existing Region. 281 if( def->is_CFG() ) { 282 phi_post = prior_n; // If looking for CFG, return prior 283 } else { 284 assert( def->is_Phi(), "" ); 285 assert( prior_n->is_Region(), "must be a post-dominating merge point" ); 286 287 // Need a Phi here 288 phi_post = PhiNode::make_blank(prior_n, def); 289 // Search for both true and false on all paths till find one. 290 for( uint i = 1; i < phi_post->req(); i++ ) // For all paths 291 phi_post->init_req( i, spinup( iff_dom, new_false, new_true, prior_n->in(i), def, cache ) ); 292 Node *t = _igvn.hash_find_insert(phi_post); 293 if( t ) { // See if we already have this one 294 // phi_post will not be used, so kill it 295 _igvn.remove_dead_node(phi_post); 296 phi_post->destruct(); 297 phi_post = t; 298 } else { 299 register_new_node( phi_post, prior_n ); 300 } 301 } 302 } 303 304 // Update cache everywhere 305 prior_n = (Node*)0xdeadbeef; // Reset IDOM walk 306 n = use_blk; // Get path input 307 // Spin-up the idom tree again, basically doing path-compression. 308 // Insert cache entries along the way, so that if we ever hit this 309 // point in the IDOM tree again we'll stop immediately on a cache hit. 310 while( n != iff_dom ) { // Found post-dominating point? 311 prior_n = n; 312 n = idom(n); // Search higher 313 cache->lru_insert( prior_n, phi_post ); // Fill cache 314 } // End of while not gone high enough 315 316 return phi_post; 317} 318 319//------------------------------find_use_block--------------------------------- 320// Find the block a USE is in. Normally USE's are in the same block as the 321// using instruction. For Phi-USE's, the USE is in the predecessor block 322// along the corresponding path. 323Node *PhaseIdealLoop::find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ) { 324 // CFG uses are their own block 325 if( use->is_CFG() ) 326 return use; 327 328 if( use->is_Phi() ) { // Phi uses in prior block 329 // Grab the first Phi use; there may be many. 330 // Each will be handled as a separate iteration of 331 // the "while( phi->outcnt() )" loop. 332 uint j; 333 for( j = 1; j < use->req(); j++ ) 334 if( use->in(j) == def ) 335 break; 336 assert( j < use->req(), "def should be among use's inputs" ); 337 return use->in(0)->in(j); 338 } 339 // Normal (non-phi) use 340 Node *use_blk = get_ctrl(use); 341 // Some uses are directly attached to the old (and going away) 342 // false and true branches. 343 if( use_blk == old_false ) { 344 use_blk = new_false; 345 set_ctrl(use, new_false); 346 } 347 if( use_blk == old_true ) { 348 use_blk = new_true; 349 set_ctrl(use, new_true); 350 } 351 352 if (use_blk == NULL) { // He's dead, Jim 353 _igvn.replace_node(use, C->top()); 354 } 355 356 return use_blk; 357} 358 359//------------------------------handle_use------------------------------------- 360// Handle uses of the merge point. Basically, split-if makes the merge point 361// go away so all uses of the merge point must go away as well. Most block 362// local uses have already been split-up, through the merge point. Uses from 363// far below the merge point can't always be split up (e.g., phi-uses are 364// pinned) and it makes too much stuff live. Instead we use a path-based 365// solution to move uses down. 366// 367// If the use is along the pre-split-CFG true branch, then the new use will 368// be from the post-split-CFG true merge point. Vice-versa for the false 369// path. Some uses will be along both paths; then we sink the use to the 370// post-dominating location; we may need to insert a Phi there. 371void 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 ) { 372 373 Node *use_blk = find_use_block(use,def,old_false,new_false,old_true,new_true); 374 if( !use_blk ) return; // He's dead, Jim 375 376 // Walk up the dominator tree until I hit either the old IfFalse, the old 377 // IfTrue or the old If. Insert Phis where needed. 378 Node *new_def = spinup( region_dom, new_false, new_true, use_blk, def, cache ); 379 380 // Found where this USE goes. Re-point him. 381 uint i; 382 for( i = 0; i < use->req(); i++ ) 383 if( use->in(i) == def ) 384 break; 385 assert( i < use->req(), "def should be among use's inputs" ); 386 _igvn.replace_input_of(use, i, new_def); 387} 388 389//------------------------------do_split_if------------------------------------ 390// Found an If getting its condition-code input from a Phi in the same block. 391// Split thru the Region. 392void PhaseIdealLoop::do_split_if( Node *iff ) { 393 if (PrintOpto && VerifyLoopOptimizations) { 394 tty->print_cr("Split-if"); 395 } 396 if (TraceLoopOpts) { 397 tty->print_cr("SplitIf"); 398 } 399 400 C->set_major_progress(); 401 Node *region = iff->in(0); 402 Node *region_dom = idom(region); 403 404 // We are going to clone this test (and the control flow with it) up through 405 // the incoming merge point. We need to empty the current basic block. 406 // Clone any instructions which must be in this block up through the merge 407 // point. 408 DUIterator i, j; 409 bool progress = true; 410 while (progress) { 411 progress = false; 412 for (i = region->outs(); region->has_out(i); i++) { 413 Node* n = region->out(i); 414 if( n == region ) continue; 415 // The IF to be split is OK. 416 if( n == iff ) continue; 417 if( !n->is_Phi() ) { // Found pinned memory op or such 418 if (split_up(n, region, iff)) { 419 i = region->refresh_out_pos(i); 420 progress = true; 421 } 422 continue; 423 } 424 assert( n->in(0) == region, "" ); 425 426 // Recursively split up all users of a Phi 427 for (j = n->outs(); n->has_out(j); j++) { 428 Node* m = n->out(j); 429 // If m is dead, throw it away, and declare progress 430 if (_nodes[m->_idx] == NULL) { 431 _igvn.remove_dead_node(m); 432 // fall through 433 } 434 else if (m != iff && split_up(m, region, iff)) { 435 // fall through 436 } else { 437 continue; 438 } 439 // Something unpredictable changed. 440 // Tell the iterators to refresh themselves, and rerun the loop. 441 i = region->refresh_out_pos(i); 442 j = region->refresh_out_pos(j); 443 progress = true; 444 } 445 } 446 } 447 448 // Now we have no instructions in the block containing the IF. 449 // Split the IF. 450 Node *new_iff = split_thru_region( iff, region ); 451 452 // Replace both uses of 'new_iff' with Regions merging True/False 453 // paths. This makes 'new_iff' go dead. 454 Node *old_false = NULL, *old_true = NULL; 455 Node *new_false = NULL, *new_true = NULL; 456 for (DUIterator_Last j2min, j2 = iff->last_outs(j2min); j2 >= j2min; --j2) { 457 Node *ifp = iff->last_out(j2); 458 assert( ifp->Opcode() == Op_IfFalse || ifp->Opcode() == Op_IfTrue, "" ); 459 ifp->set_req(0, new_iff); 460 Node *ifpx = split_thru_region( ifp, region ); 461 462 // Replace 'If' projection of a Region with a Region of 463 // 'If' projections. 464 ifpx->set_req(0, ifpx); // A TRUE RegionNode 465 466 // Setup dominator info 467 set_idom(ifpx, region_dom, dom_depth(region_dom) + 1); 468 469 // Check for splitting loop tails 470 if( get_loop(iff)->tail() == ifp ) 471 get_loop(iff)->_tail = ifpx; 472 473 // Replace in the graph with lazy-update mechanism 474 new_iff->set_req(0, new_iff); // hook self so it does not go dead 475 lazy_replace(ifp, ifpx); 476 new_iff->set_req(0, region); 477 478 // Record bits for later xforms 479 if( ifp->Opcode() == Op_IfFalse ) { 480 old_false = ifp; 481 new_false = ifpx; 482 } else { 483 old_true = ifp; 484 new_true = ifpx; 485 } 486 } 487 _igvn.remove_dead_node(new_iff); 488 // Lazy replace IDOM info with the region's dominator 489 lazy_replace( iff, region_dom ); 490 491 // Now make the original merge point go dead, by handling all its uses. 492 small_cache region_cache; 493 // Preload some control flow in region-cache 494 region_cache.lru_insert( new_false, new_false ); 495 region_cache.lru_insert( new_true , new_true ); 496 // Now handle all uses of the splitting block 497 for (DUIterator k = region->outs(); region->has_out(k); k++) { 498 Node* phi = region->out(k); 499 if (!phi->in(0)) { // Dead phi? Remove it 500 _igvn.remove_dead_node(phi); 501 } else if (phi == region) { // Found the self-reference 502 continue; // No roll-back of DUIterator 503 } else if (phi->is_Phi()) { // Expected common case: Phi hanging off of Region 504 assert(phi->in(0) == region, "Inconsistent graph"); 505 // Need a per-def cache. Phi represents a def, so make a cache 506 small_cache phi_cache; 507 508 // Inspect all Phi uses to make the Phi go dead 509 for (DUIterator_Last lmin, l = phi->last_outs(lmin); l >= lmin; --l) { 510 Node* use = phi->last_out(l); 511 // Compute the new DEF for this USE. New DEF depends on the path 512 // taken from the original DEF to the USE. The new DEF may be some 513 // collection of PHI's merging values from different paths. The Phis 514 // inserted depend only on the location of the USE. We use a 515 // 2-element cache to handle multiple uses from the same block. 516 handle_use(use, phi, &phi_cache, region_dom, new_false, new_true, old_false, old_true); 517 } // End of while phi has uses 518 // Remove the dead Phi 519 _igvn.remove_dead_node( phi ); 520 } else { 521 assert(phi->in(0) == region, "Inconsistent graph"); 522 // Random memory op guarded by Region. Compute new DEF for USE. 523 handle_use(phi, region, ®ion_cache, region_dom, new_false, new_true, old_false, old_true); 524 } 525 // Every path above deletes a use of the region, except for the region 526 // self-cycle (which is needed by handle_use calling find_use_block 527 // calling get_ctrl calling get_ctrl_no_update looking for dead 528 // regions). So roll back the DUIterator innards. 529 --k; 530 } // End of while merge point has phis 531 532 assert(region->outcnt() == 1, "Only self reference should remain"); // Just Self on the Region 533 region->set_req(0, NULL); // Break the self-cycle 534 535 // Any leftover bits in the splitting block must not have depended on local 536 // Phi inputs (these have already been split-up). Hence it's safe to hoist 537 // these guys to the dominating point. 538 lazy_replace( region, region_dom ); 539#ifndef PRODUCT 540 if( VerifyLoopOptimizations ) verify(); 541#endif 542} 543