1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 // Portions of code courtesy of Clifford Click 26 27 // Optimization - Graph Style 28 29 #include "incls/_precompiled.incl" 30 #include "incls/_gcm.cpp.incl" 31 32 // To avoid float value underflow 33 #define MIN_BLOCK_FREQUENCY 1.e-35f 34 35 //----------------------------schedule_node_into_block------------------------- 36 // Insert node n into block b. Look for projections of n and make sure they 37 // are in b also. 38 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) { 39 // Set basic block of n, Add n to b, 40 _bbs.map(n->_idx, b); 41 b->add_inst(n); 42 43 // After Matching, nearly any old Node may have projections trailing it. 44 // These are usually machine-dependent flags. In any case, they might 45 // float to another block below this one. Move them up. 46 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 47 Node* use = n->fast_out(i); 48 if (use->is_Proj()) { 49 Block* buse = _bbs[use->_idx]; 50 if (buse != b) { // In wrong block? 51 if (buse != NULL) 52 buse->find_remove(use); // Remove from wrong block 53 _bbs.map(use->_idx, b); // Re-insert in this block 54 b->add_inst(use); 55 } 56 } 57 } 58 } 59 60 //----------------------------replace_block_proj_ctrl------------------------- 61 // Nodes that have is_block_proj() nodes as their control need to use 62 // the appropriate Region for their actual block as their control since 63 // the projection will be in a predecessor block. 64 void PhaseCFG::replace_block_proj_ctrl( Node *n ) { 65 const Node *in0 = n->in(0); 66 assert(in0 != NULL, "Only control-dependent"); 67 const Node *p = in0->is_block_proj(); 68 if (p != NULL && p != n) { // Control from a block projection? 69 assert(!n->pinned() || n->is_SafePointScalarObject(), "only SafePointScalarObject pinned node is expected here"); 70 // Find trailing Region 71 Block *pb = _bbs[in0->_idx]; // Block-projection already has basic block 72 uint j = 0; 73 if (pb->_num_succs != 1) { // More then 1 successor? 74 // Search for successor 75 uint max = pb->_nodes.size(); 76 assert( max > 1, "" ); 77 uint start = max - pb->_num_succs; 78 // Find which output path belongs to projection 79 for (j = start; j < max; j++) { 80 if( pb->_nodes[j] == in0 ) 81 break; 82 } 83 assert( j < max, "must find" ); 84 // Change control to match head of successor basic block 85 j -= start; 86 } 87 n->set_req(0, pb->_succs[j]->head()); 88 } 89 } 90 91 92 //------------------------------schedule_pinned_nodes-------------------------- 93 // Set the basic block for Nodes pinned into blocks 94 void PhaseCFG::schedule_pinned_nodes( VectorSet &visited ) { 95 // Allocate node stack of size C->unique()+8 to avoid frequent realloc 96 GrowableArray <Node *> spstack(C->unique()+8); 97 spstack.push(_root); 98 while ( spstack.is_nonempty() ) { 99 Node *n = spstack.pop(); 100 if( !visited.test_set(n->_idx) ) { // Test node and flag it as visited 101 if( n->pinned() && !_bbs.lookup(n->_idx) ) { // Pinned? Nail it down! 102 assert( n->in(0), "pinned Node must have Control" ); 103 // Before setting block replace block_proj control edge 104 replace_block_proj_ctrl(n); 105 Node *input = n->in(0); 106 while( !input->is_block_start() ) 107 input = input->in(0); 108 Block *b = _bbs[input->_idx]; // Basic block of controlling input 109 schedule_node_into_block(n, b); 110 } 111 for( int i = n->req() - 1; i >= 0; --i ) { // For all inputs 112 if( n->in(i) != NULL ) 113 spstack.push(n->in(i)); 114 } 115 } 116 } 117 } 118 119 #ifdef ASSERT 120 // Assert that new input b2 is dominated by all previous inputs. 121 // Check this by by seeing that it is dominated by b1, the deepest 122 // input observed until b2. 123 static void assert_dom(Block* b1, Block* b2, Node* n, Block_Array &bbs) { 124 if (b1 == NULL) return; 125 assert(b1->_dom_depth < b2->_dom_depth, "sanity"); 126 Block* tmp = b2; 127 while (tmp != b1 && tmp != NULL) { 128 tmp = tmp->_idom; 129 } 130 if (tmp != b1) { 131 // Detected an unschedulable graph. Print some nice stuff and die. 132 tty->print_cr("!!! Unschedulable graph !!!"); 133 for (uint j=0; j<n->len(); j++) { // For all inputs 134 Node* inn = n->in(j); // Get input 135 if (inn == NULL) continue; // Ignore NULL, missing inputs 136 Block* inb = bbs[inn->_idx]; 137 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order, 138 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth); 139 inn->dump(); 140 } 141 tty->print("Failing node: "); 142 n->dump(); 143 assert(false, "unscheduable graph"); 144 } 145 } 146 #endif 147 148 static Block* find_deepest_input(Node* n, Block_Array &bbs) { 149 // Find the last input dominated by all other inputs. 150 Block* deepb = NULL; // Deepest block so far 151 int deepb_dom_depth = 0; 152 for (uint k = 0; k < n->len(); k++) { // For all inputs 153 Node* inn = n->in(k); // Get input 154 if (inn == NULL) continue; // Ignore NULL, missing inputs 155 Block* inb = bbs[inn->_idx]; 156 assert(inb != NULL, "must already have scheduled this input"); 157 if (deepb_dom_depth < (int) inb->_dom_depth) { 158 // The new inb must be dominated by the previous deepb. 159 // The various inputs must be linearly ordered in the dom 160 // tree, or else there will not be a unique deepest block. 161 DEBUG_ONLY(assert_dom(deepb, inb, n, bbs)); 162 deepb = inb; // Save deepest block 163 deepb_dom_depth = deepb->_dom_depth; 164 } 165 } 166 assert(deepb != NULL, "must be at least one input to n"); 167 return deepb; 168 } 169 170 171 //------------------------------schedule_early--------------------------------- 172 // Find the earliest Block any instruction can be placed in. Some instructions 173 // are pinned into Blocks. Unpinned instructions can appear in last block in 174 // which all their inputs occur. 175 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) { 176 // Allocate stack with enough space to avoid frequent realloc 177 Node_Stack nstack(roots.Size() + 8); // (unique >> 1) + 24 from Java2D stats 178 // roots.push(_root); _root will be processed among C->top() inputs 179 roots.push(C->top()); 180 visited.set(C->top()->_idx); 181 182 while (roots.size() != 0) { 183 // Use local variables nstack_top_n & nstack_top_i to cache values 184 // on stack's top. 185 Node *nstack_top_n = roots.pop(); 186 uint nstack_top_i = 0; 187 //while_nstack_nonempty: 188 while (true) { 189 // Get parent node and next input's index from stack's top. 190 Node *n = nstack_top_n; 191 uint i = nstack_top_i; 192 193 if (i == 0) { 194 // Fixup some control. Constants without control get attached 195 // to root and nodes that use is_block_proj() nodes should be attached 196 // to the region that starts their block. 197 const Node *in0 = n->in(0); 198 if (in0 != NULL) { // Control-dependent? 199 replace_block_proj_ctrl(n); 200 } else { // n->in(0) == NULL 201 if (n->req() == 1) { // This guy is a constant with NO inputs? 202 n->set_req(0, _root); 203 } 204 } 205 } 206 207 // First, visit all inputs and force them to get a block. If an 208 // input is already in a block we quit following inputs (to avoid 209 // cycles). Instead we put that Node on a worklist to be handled 210 // later (since IT'S inputs may not have a block yet). 211 bool done = true; // Assume all n's inputs will be processed 212 while (i < n->len()) { // For all inputs 213 Node *in = n->in(i); // Get input 214 ++i; 215 if (in == NULL) continue; // Ignore NULL, missing inputs 216 int is_visited = visited.test_set(in->_idx); 217 if (!_bbs.lookup(in->_idx)) { // Missing block selection? 218 if (is_visited) { 219 // assert( !visited.test(in->_idx), "did not schedule early" ); 220 return false; 221 } 222 nstack.push(n, i); // Save parent node and next input's index. 223 nstack_top_n = in; // Process current input now. 224 nstack_top_i = 0; 225 done = false; // Not all n's inputs processed. 226 break; // continue while_nstack_nonempty; 227 } else if (!is_visited) { // Input not yet visited? 228 roots.push(in); // Visit this guy later, using worklist 229 } 230 } 231 if (done) { 232 // All of n's inputs have been processed, complete post-processing. 233 234 // Some instructions are pinned into a block. These include Region, 235 // Phi, Start, Return, and other control-dependent instructions and 236 // any projections which depend on them. 237 if (!n->pinned()) { 238 // Set earliest legal block. 239 _bbs.map(n->_idx, find_deepest_input(n, _bbs)); 240 } else { 241 assert(_bbs[n->_idx] == _bbs[n->in(0)->_idx], "Pinned Node should be at the same block as its control edge"); 242 } 243 244 if (nstack.is_empty()) { 245 // Finished all nodes on stack. 246 // Process next node on the worklist 'roots'. 247 break; 248 } 249 // Get saved parent node and next input's index. 250 nstack_top_n = nstack.node(); 251 nstack_top_i = nstack.index(); 252 nstack.pop(); 253 } // if (done) 254 } // while (true) 255 } // while (roots.size() != 0) 256 return true; 257 } 258 259 //------------------------------dom_lca---------------------------------------- 260 // Find least common ancestor in dominator tree 261 // LCA is a current notion of LCA, to be raised above 'this'. 262 // As a convenient boundary condition, return 'this' if LCA is NULL. 263 // Find the LCA of those two nodes. 264 Block* Block::dom_lca(Block* LCA) { 265 if (LCA == NULL || LCA == this) return this; 266 267 Block* anc = this; 268 while (anc->_dom_depth > LCA->_dom_depth) 269 anc = anc->_idom; // Walk up till anc is as high as LCA 270 271 while (LCA->_dom_depth > anc->_dom_depth) 272 LCA = LCA->_idom; // Walk up till LCA is as high as anc 273 274 while (LCA != anc) { // Walk both up till they are the same 275 LCA = LCA->_idom; 276 anc = anc->_idom; 277 } 278 279 return LCA; 280 } 281 282 //--------------------------raise_LCA_above_use-------------------------------- 283 // We are placing a definition, and have been given a def->use edge. 284 // The definition must dominate the use, so move the LCA upward in the 285 // dominator tree to dominate the use. If the use is a phi, adjust 286 // the LCA only with the phi input paths which actually use this def. 287 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, Block_Array &bbs) { 288 Block* buse = bbs[use->_idx]; 289 if (buse == NULL) return LCA; // Unused killing Projs have no use block 290 if (!use->is_Phi()) return buse->dom_lca(LCA); 291 uint pmax = use->req(); // Number of Phi inputs 292 // Why does not this loop just break after finding the matching input to 293 // the Phi? Well...it's like this. I do not have true def-use/use-def 294 // chains. Means I cannot distinguish, from the def-use direction, which 295 // of many use-defs lead from the same use to the same def. That is, this 296 // Phi might have several uses of the same def. Each use appears in a 297 // different predecessor block. But when I enter here, I cannot distinguish 298 // which use-def edge I should find the predecessor block for. So I find 299 // them all. Means I do a little extra work if a Phi uses the same value 300 // more than once. 301 for (uint j=1; j<pmax; j++) { // For all inputs 302 if (use->in(j) == def) { // Found matching input? 303 Block* pred = bbs[buse->pred(j)->_idx]; 304 LCA = pred->dom_lca(LCA); 305 } 306 } 307 return LCA; 308 } 309 310 //----------------------------raise_LCA_above_marks---------------------------- 311 // Return a new LCA that dominates LCA and any of its marked predecessors. 312 // Search all my parents up to 'early' (exclusive), looking for predecessors 313 // which are marked with the given index. Return the LCA (in the dom tree) 314 // of all marked blocks. If there are none marked, return the original 315 // LCA. 316 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, 317 Block* early, Block_Array &bbs) { 318 Block_List worklist; 319 worklist.push(LCA); 320 while (worklist.size() > 0) { 321 Block* mid = worklist.pop(); 322 if (mid == early) continue; // stop searching here 323 324 // Test and set the visited bit. 325 if (mid->raise_LCA_visited() == mark) continue; // already visited 326 327 // Don't process the current LCA, otherwise the search may terminate early 328 if (mid != LCA && mid->raise_LCA_mark() == mark) { 329 // Raise the LCA. 330 LCA = mid->dom_lca(LCA); 331 if (LCA == early) break; // stop searching everywhere 332 assert(early->dominates(LCA), "early is high enough"); 333 // Resume searching at that point, skipping intermediate levels. 334 worklist.push(LCA); 335 if (LCA == mid) 336 continue; // Don't mark as visited to avoid early termination. 337 } else { 338 // Keep searching through this block's predecessors. 339 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) { 340 Block* mid_parent = bbs[ mid->pred(j)->_idx ]; 341 worklist.push(mid_parent); 342 } 343 } 344 mid->set_raise_LCA_visited(mark); 345 } 346 return LCA; 347 } 348 349 //--------------------------memory_early_block-------------------------------- 350 // This is a variation of find_deepest_input, the heart of schedule_early. 351 // Find the "early" block for a load, if we considered only memory and 352 // address inputs, that is, if other data inputs were ignored. 353 // 354 // Because a subset of edges are considered, the resulting block will 355 // be earlier (at a shallower dom_depth) than the true schedule_early 356 // point of the node. We compute this earlier block as a more permissive 357 // site for anti-dependency insertion, but only if subsume_loads is enabled. 358 static Block* memory_early_block(Node* load, Block* early, Block_Array &bbs) { 359 Node* base; 360 Node* index; 361 Node* store = load->in(MemNode::Memory); 362 load->as_Mach()->memory_inputs(base, index); 363 364 assert(base != NodeSentinel && index != NodeSentinel, 365 "unexpected base/index inputs"); 366 367 Node* mem_inputs[4]; 368 int mem_inputs_length = 0; 369 if (base != NULL) mem_inputs[mem_inputs_length++] = base; 370 if (index != NULL) mem_inputs[mem_inputs_length++] = index; 371 if (store != NULL) mem_inputs[mem_inputs_length++] = store; 372 373 // In the comparision below, add one to account for the control input, 374 // which may be null, but always takes up a spot in the in array. 375 if (mem_inputs_length + 1 < (int) load->req()) { 376 // This "load" has more inputs than just the memory, base and index inputs. 377 // For purposes of checking anti-dependences, we need to start 378 // from the early block of only the address portion of the instruction, 379 // and ignore other blocks that may have factored into the wider 380 // schedule_early calculation. 381 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0); 382 383 Block* deepb = NULL; // Deepest block so far 384 int deepb_dom_depth = 0; 385 for (int i = 0; i < mem_inputs_length; i++) { 386 Block* inb = bbs[mem_inputs[i]->_idx]; 387 if (deepb_dom_depth < (int) inb->_dom_depth) { 388 // The new inb must be dominated by the previous deepb. 389 // The various inputs must be linearly ordered in the dom 390 // tree, or else there will not be a unique deepest block. 391 DEBUG_ONLY(assert_dom(deepb, inb, load, bbs)); 392 deepb = inb; // Save deepest block 393 deepb_dom_depth = deepb->_dom_depth; 394 } 395 } 396 early = deepb; 397 } 398 399 return early; 400 } 401 402 //--------------------------insert_anti_dependences--------------------------- 403 // A load may need to witness memory that nearby stores can overwrite. 404 // For each nearby store, either insert an "anti-dependence" edge 405 // from the load to the store, or else move LCA upward to force the 406 // load to (eventually) be scheduled in a block above the store. 407 // 408 // Do not add edges to stores on distinct control-flow paths; 409 // only add edges to stores which might interfere. 410 // 411 // Return the (updated) LCA. There will not be any possibly interfering 412 // store between the load's "early block" and the updated LCA. 413 // Any stores in the updated LCA will have new precedence edges 414 // back to the load. The caller is expected to schedule the load 415 // in the LCA, in which case the precedence edges will make LCM 416 // preserve anti-dependences. The caller may also hoist the load 417 // above the LCA, if it is not the early block. 418 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) { 419 assert(load->needs_anti_dependence_check(), "must be a load of some sort"); 420 assert(LCA != NULL, ""); 421 DEBUG_ONLY(Block* LCA_orig = LCA); 422 423 // Compute the alias index. Loads and stores with different alias indices 424 // do not need anti-dependence edges. 425 uint load_alias_idx = C->get_alias_index(load->adr_type()); 426 #ifdef ASSERT 427 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 && 428 (PrintOpto || VerifyAliases || 429 PrintMiscellaneous && (WizardMode || Verbose))) { 430 // Load nodes should not consume all of memory. 431 // Reporting a bottom type indicates a bug in adlc. 432 // If some particular type of node validly consumes all of memory, 433 // sharpen the preceding "if" to exclude it, so we can catch bugs here. 434 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory."); 435 load->dump(2); 436 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, ""); 437 } 438 #endif 439 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp), 440 "String compare is only known 'load' that does not conflict with any stores"); 441 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals), 442 "String equals is a 'load' that does not conflict with any stores"); 443 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf), 444 "String indexOf is a 'load' that does not conflict with any stores"); 445 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq), 446 "Arrays equals is a 'load' that do not conflict with any stores"); 447 448 if (!C->alias_type(load_alias_idx)->is_rewritable()) { 449 // It is impossible to spoil this load by putting stores before it, 450 // because we know that the stores will never update the value 451 // which 'load' must witness. 452 return LCA; 453 } 454 455 node_idx_t load_index = load->_idx; 456 457 // Note the earliest legal placement of 'load', as determined by 458 // by the unique point in the dom tree where all memory effects 459 // and other inputs are first available. (Computed by schedule_early.) 460 // For normal loads, 'early' is the shallowest place (dom graph wise) 461 // to look for anti-deps between this load and any store. 462 Block* early = _bbs[load_index]; 463 464 // If we are subsuming loads, compute an "early" block that only considers 465 // memory or address inputs. This block may be different than the 466 // schedule_early block in that it could be at an even shallower depth in the 467 // dominator tree, and allow for a broader discovery of anti-dependences. 468 if (C->subsume_loads()) { 469 early = memory_early_block(load, early, _bbs); 470 } 471 472 ResourceArea *area = Thread::current()->resource_area(); 473 Node_List worklist_mem(area); // prior memory state to store 474 Node_List worklist_store(area); // possible-def to explore 475 Node_List worklist_visited(area); // visited mergemem nodes 476 Node_List non_early_stores(area); // all relevant stores outside of early 477 bool must_raise_LCA = false; 478 479 #ifdef TRACK_PHI_INPUTS 480 // %%% This extra checking fails because MergeMem nodes are not GVNed. 481 // Provide "phi_inputs" to check if every input to a PhiNode is from the 482 // original memory state. This indicates a PhiNode for which should not 483 // prevent the load from sinking. For such a block, set_raise_LCA_mark 484 // may be overly conservative. 485 // Mechanism: count inputs seen for each Phi encountered in worklist_store. 486 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0)); 487 #endif 488 489 // 'load' uses some memory state; look for users of the same state. 490 // Recurse through MergeMem nodes to the stores that use them. 491 492 // Each of these stores is a possible definition of memory 493 // that 'load' needs to use. We need to force 'load' 494 // to occur before each such store. When the store is in 495 // the same block as 'load', we insert an anti-dependence 496 // edge load->store. 497 498 // The relevant stores "nearby" the load consist of a tree rooted 499 // at initial_mem, with internal nodes of type MergeMem. 500 // Therefore, the branches visited by the worklist are of this form: 501 // initial_mem -> (MergeMem ->)* store 502 // The anti-dependence constraints apply only to the fringe of this tree. 503 504 Node* initial_mem = load->in(MemNode::Memory); 505 worklist_store.push(initial_mem); 506 worklist_visited.push(initial_mem); 507 worklist_mem.push(NULL); 508 while (worklist_store.size() > 0) { 509 // Examine a nearby store to see if it might interfere with our load. 510 Node* mem = worklist_mem.pop(); 511 Node* store = worklist_store.pop(); 512 uint op = store->Opcode(); 513 514 // MergeMems do not directly have anti-deps. 515 // Treat them as internal nodes in a forward tree of memory states, 516 // the leaves of which are each a 'possible-def'. 517 if (store == initial_mem // root (exclusive) of tree we are searching 518 || op == Op_MergeMem // internal node of tree we are searching 519 ) { 520 mem = store; // It's not a possibly interfering store. 521 if (store == initial_mem) 522 initial_mem = NULL; // only process initial memory once 523 524 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 525 store = mem->fast_out(i); 526 if (store->is_MergeMem()) { 527 // Be sure we don't get into combinatorial problems. 528 // (Allow phis to be repeated; they can merge two relevant states.) 529 uint j = worklist_visited.size(); 530 for (; j > 0; j--) { 531 if (worklist_visited.at(j-1) == store) break; 532 } 533 if (j > 0) continue; // already on work list; do not repeat 534 worklist_visited.push(store); 535 } 536 worklist_mem.push(mem); 537 worklist_store.push(store); 538 } 539 continue; 540 } 541 542 if (op == Op_MachProj || op == Op_Catch) continue; 543 if (store->needs_anti_dependence_check()) continue; // not really a store 544 545 // Compute the alias index. Loads and stores with different alias 546 // indices do not need anti-dependence edges. Wide MemBar's are 547 // anti-dependent on everything (except immutable memories). 548 const TypePtr* adr_type = store->adr_type(); 549 if (!C->can_alias(adr_type, load_alias_idx)) continue; 550 551 // Most slow-path runtime calls do NOT modify Java memory, but 552 // they can block and so write Raw memory. 553 if (store->is_Mach()) { 554 MachNode* mstore = store->as_Mach(); 555 if (load_alias_idx != Compile::AliasIdxRaw) { 556 // Check for call into the runtime using the Java calling 557 // convention (and from there into a wrapper); it has no 558 // _method. Can't do this optimization for Native calls because 559 // they CAN write to Java memory. 560 if (mstore->ideal_Opcode() == Op_CallStaticJava) { 561 assert(mstore->is_MachSafePoint(), ""); 562 MachSafePointNode* ms = (MachSafePointNode*) mstore; 563 assert(ms->is_MachCallJava(), ""); 564 MachCallJavaNode* mcj = (MachCallJavaNode*) ms; 565 if (mcj->_method == NULL) { 566 // These runtime calls do not write to Java visible memory 567 // (other than Raw) and so do not require anti-dependence edges. 568 continue; 569 } 570 } 571 // Same for SafePoints: they read/write Raw but only read otherwise. 572 // This is basically a workaround for SafePoints only defining control 573 // instead of control + memory. 574 if (mstore->ideal_Opcode() == Op_SafePoint) 575 continue; 576 } else { 577 // Some raw memory, such as the load of "top" at an allocation, 578 // can be control dependent on the previous safepoint. See 579 // comments in GraphKit::allocate_heap() about control input. 580 // Inserting an anti-dep between such a safepoint and a use 581 // creates a cycle, and will cause a subsequent failure in 582 // local scheduling. (BugId 4919904) 583 // (%%% How can a control input be a safepoint and not a projection??) 584 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore) 585 continue; 586 } 587 } 588 589 // Identify a block that the current load must be above, 590 // or else observe that 'store' is all the way up in the 591 // earliest legal block for 'load'. In the latter case, 592 // immediately insert an anti-dependence edge. 593 Block* store_block = _bbs[store->_idx]; 594 assert(store_block != NULL, "unused killing projections skipped above"); 595 596 if (store->is_Phi()) { 597 // 'load' uses memory which is one (or more) of the Phi's inputs. 598 // It must be scheduled not before the Phi, but rather before 599 // each of the relevant Phi inputs. 600 // 601 // Instead of finding the LCA of all inputs to a Phi that match 'mem', 602 // we mark each corresponding predecessor block and do a combined 603 // hoisting operation later (raise_LCA_above_marks). 604 // 605 // Do not assert(store_block != early, "Phi merging memory after access") 606 // PhiNode may be at start of block 'early' with backedge to 'early' 607 DEBUG_ONLY(bool found_match = false); 608 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) { 609 if (store->in(j) == mem) { // Found matching input? 610 DEBUG_ONLY(found_match = true); 611 Block* pred_block = _bbs[store_block->pred(j)->_idx]; 612 if (pred_block != early) { 613 // If any predecessor of the Phi matches the load's "early block", 614 // we do not need a precedence edge between the Phi and 'load' 615 // since the load will be forced into a block preceding the Phi. 616 pred_block->set_raise_LCA_mark(load_index); 617 assert(!LCA_orig->dominates(pred_block) || 618 early->dominates(pred_block), "early is high enough"); 619 must_raise_LCA = true; 620 } 621 } 622 } 623 assert(found_match, "no worklist bug"); 624 #ifdef TRACK_PHI_INPUTS 625 #ifdef ASSERT 626 // This assert asks about correct handling of PhiNodes, which may not 627 // have all input edges directly from 'mem'. See BugId 4621264 628 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1; 629 // Increment by exactly one even if there are multiple copies of 'mem' 630 // coming into the phi, because we will run this block several times 631 // if there are several copies of 'mem'. (That's how DU iterators work.) 632 phi_inputs.at_put(store->_idx, num_mem_inputs); 633 assert(PhiNode::Input + num_mem_inputs < store->req(), 634 "Expect at least one phi input will not be from original memory state"); 635 #endif //ASSERT 636 #endif //TRACK_PHI_INPUTS 637 } else if (store_block != early) { 638 // 'store' is between the current LCA and earliest possible block. 639 // Label its block, and decide later on how to raise the LCA 640 // to include the effect on LCA of this store. 641 // If this store's block gets chosen as the raised LCA, we 642 // will find him on the non_early_stores list and stick him 643 // with a precedence edge. 644 // (But, don't bother if LCA is already raised all the way.) 645 if (LCA != early) { 646 store_block->set_raise_LCA_mark(load_index); 647 must_raise_LCA = true; 648 non_early_stores.push(store); 649 } 650 } else { 651 // Found a possibly-interfering store in the load's 'early' block. 652 // This means 'load' cannot sink at all in the dominator tree. 653 // Add an anti-dep edge, and squeeze 'load' into the highest block. 654 assert(store != load->in(0), "dependence cycle found"); 655 if (verify) { 656 assert(store->find_edge(load) != -1, "missing precedence edge"); 657 } else { 658 store->add_prec(load); 659 } 660 LCA = early; 661 // This turns off the process of gathering non_early_stores. 662 } 663 } 664 // (Worklist is now empty; all nearby stores have been visited.) 665 666 // Finished if 'load' must be scheduled in its 'early' block. 667 // If we found any stores there, they have already been given 668 // precedence edges. 669 if (LCA == early) return LCA; 670 671 // We get here only if there are no possibly-interfering stores 672 // in the load's 'early' block. Move LCA up above all predecessors 673 // which contain stores we have noted. 674 // 675 // The raised LCA block can be a home to such interfering stores, 676 // but its predecessors must not contain any such stores. 677 // 678 // The raised LCA will be a lower bound for placing the load, 679 // preventing the load from sinking past any block containing 680 // a store that may invalidate the memory state required by 'load'. 681 if (must_raise_LCA) 682 LCA = raise_LCA_above_marks(LCA, load->_idx, early, _bbs); 683 if (LCA == early) return LCA; 684 685 // Insert anti-dependence edges from 'load' to each store 686 // in the non-early LCA block. 687 // Mine the non_early_stores list for such stores. 688 if (LCA->raise_LCA_mark() == load_index) { 689 while (non_early_stores.size() > 0) { 690 Node* store = non_early_stores.pop(); 691 Block* store_block = _bbs[store->_idx]; 692 if (store_block == LCA) { 693 // add anti_dependence from store to load in its own block 694 assert(store != load->in(0), "dependence cycle found"); 695 if (verify) { 696 assert(store->find_edge(load) != -1, "missing precedence edge"); 697 } else { 698 store->add_prec(load); 699 } 700 } else { 701 assert(store_block->raise_LCA_mark() == load_index, "block was marked"); 702 // Any other stores we found must be either inside the new LCA 703 // or else outside the original LCA. In the latter case, they 704 // did not interfere with any use of 'load'. 705 assert(LCA->dominates(store_block) 706 || !LCA_orig->dominates(store_block), "no stray stores"); 707 } 708 } 709 } 710 711 // Return the highest block containing stores; any stores 712 // within that block have been given anti-dependence edges. 713 return LCA; 714 } 715 716 // This class is used to iterate backwards over the nodes in the graph. 717 718 class Node_Backward_Iterator { 719 720 private: 721 Node_Backward_Iterator(); 722 723 public: 724 // Constructor for the iterator 725 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs); 726 727 // Postincrement operator to iterate over the nodes 728 Node *next(); 729 730 private: 731 VectorSet &_visited; 732 Node_List &_stack; 733 Block_Array &_bbs; 734 }; 735 736 // Constructor for the Node_Backward_Iterator 737 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, Block_Array &bbs ) 738 : _visited(visited), _stack(stack), _bbs(bbs) { 739 // The stack should contain exactly the root 740 stack.clear(); 741 stack.push(root); 742 743 // Clear the visited bits 744 visited.Clear(); 745 } 746 747 // Iterator for the Node_Backward_Iterator 748 Node *Node_Backward_Iterator::next() { 749 750 // If the _stack is empty, then just return NULL: finished. 751 if ( !_stack.size() ) 752 return NULL; 753 754 // '_stack' is emulating a real _stack. The 'visit-all-users' loop has been 755 // made stateless, so I do not need to record the index 'i' on my _stack. 756 // Instead I visit all users each time, scanning for unvisited users. 757 // I visit unvisited not-anti-dependence users first, then anti-dependent 758 // children next. 759 Node *self = _stack.pop(); 760 761 // I cycle here when I am entering a deeper level of recursion. 762 // The key variable 'self' was set prior to jumping here. 763 while( 1 ) { 764 765 _visited.set(self->_idx); 766 767 // Now schedule all uses as late as possible. 768 uint src = self->is_Proj() ? self->in(0)->_idx : self->_idx; 769 uint src_rpo = _bbs[src]->_rpo; 770 771 // Schedule all nodes in a post-order visit 772 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any 773 774 // Scan for unvisited nodes 775 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) { 776 // For all uses, schedule late 777 Node* n = self->fast_out(i); // Use 778 779 // Skip already visited children 780 if ( _visited.test(n->_idx) ) 781 continue; 782 783 // do not traverse backward control edges 784 Node *use = n->is_Proj() ? n->in(0) : n; 785 uint use_rpo = _bbs[use->_idx]->_rpo; 786 787 if ( use_rpo < src_rpo ) 788 continue; 789 790 // Phi nodes always precede uses in a basic block 791 if ( use_rpo == src_rpo && use->is_Phi() ) 792 continue; 793 794 unvisited = n; // Found unvisited 795 796 // Check for possible-anti-dependent 797 if( !n->needs_anti_dependence_check() ) 798 break; // Not visited, not anti-dep; schedule it NOW 799 } 800 801 // Did I find an unvisited not-anti-dependent Node? 802 if ( !unvisited ) 803 break; // All done with children; post-visit 'self' 804 805 // Visit the unvisited Node. Contains the obvious push to 806 // indicate I'm entering a deeper level of recursion. I push the 807 // old state onto the _stack and set a new state and loop (recurse). 808 _stack.push(self); 809 self = unvisited; 810 } // End recursion loop 811 812 return self; 813 } 814 815 //------------------------------ComputeLatenciesBackwards---------------------- 816 // Compute the latency of all the instructions. 817 void PhaseCFG::ComputeLatenciesBackwards(VectorSet &visited, Node_List &stack) { 818 #ifndef PRODUCT 819 if (trace_opto_pipelining()) 820 tty->print("\n#---- ComputeLatenciesBackwards ----\n"); 821 #endif 822 823 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs); 824 Node *n; 825 826 // Walk over all the nodes from last to first 827 while (n = iter.next()) { 828 // Set the latency for the definitions of this instruction 829 partial_latency_of_defs(n); 830 } 831 } // end ComputeLatenciesBackwards 832 833 //------------------------------partial_latency_of_defs------------------------ 834 // Compute the latency impact of this node on all defs. This computes 835 // a number that increases as we approach the beginning of the routine. 836 void PhaseCFG::partial_latency_of_defs(Node *n) { 837 // Set the latency for this instruction 838 #ifndef PRODUCT 839 if (trace_opto_pipelining()) { 840 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", 841 n->_idx, _node_latency.at_grow(n->_idx)); 842 dump(); 843 } 844 #endif 845 846 if (n->is_Proj()) 847 n = n->in(0); 848 849 if (n->is_Root()) 850 return; 851 852 uint nlen = n->len(); 853 uint use_latency = _node_latency.at_grow(n->_idx); 854 uint use_pre_order = _bbs[n->_idx]->_pre_order; 855 856 for ( uint j=0; j<nlen; j++ ) { 857 Node *def = n->in(j); 858 859 if (!def || def == n) 860 continue; 861 862 // Walk backwards thru projections 863 if (def->is_Proj()) 864 def = def->in(0); 865 866 #ifndef PRODUCT 867 if (trace_opto_pipelining()) { 868 tty->print("# in(%2d): ", j); 869 def->dump(); 870 } 871 #endif 872 873 // If the defining block is not known, assume it is ok 874 Block *def_block = _bbs[def->_idx]; 875 uint def_pre_order = def_block ? def_block->_pre_order : 0; 876 877 if ( (use_pre_order < def_pre_order) || 878 (use_pre_order == def_pre_order && n->is_Phi()) ) 879 continue; 880 881 uint delta_latency = n->latency(j); 882 uint current_latency = delta_latency + use_latency; 883 884 if (_node_latency.at_grow(def->_idx) < current_latency) { 885 _node_latency.at_put_grow(def->_idx, current_latency); 886 } 887 888 #ifndef PRODUCT 889 if (trace_opto_pipelining()) { 890 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d", 891 use_latency, j, delta_latency, current_latency, def->_idx, 892 _node_latency.at_grow(def->_idx)); 893 } 894 #endif 895 } 896 } 897 898 //------------------------------latency_from_use------------------------------- 899 // Compute the latency of a specific use 900 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) { 901 // If self-reference, return no latency 902 if (use == n || use->is_Root()) 903 return 0; 904 905 uint def_pre_order = _bbs[def->_idx]->_pre_order; 906 uint latency = 0; 907 908 // If the use is not a projection, then it is simple... 909 if (!use->is_Proj()) { 910 #ifndef PRODUCT 911 if (trace_opto_pipelining()) { 912 tty->print("# out(): "); 913 use->dump(); 914 } 915 #endif 916 917 uint use_pre_order = _bbs[use->_idx]->_pre_order; 918 919 if (use_pre_order < def_pre_order) 920 return 0; 921 922 if (use_pre_order == def_pre_order && use->is_Phi()) 923 return 0; 924 925 uint nlen = use->len(); 926 uint nl = _node_latency.at_grow(use->_idx); 927 928 for ( uint j=0; j<nlen; j++ ) { 929 if (use->in(j) == n) { 930 // Change this if we want local latencies 931 uint ul = use->latency(j); 932 uint l = ul + nl; 933 if (latency < l) latency = l; 934 #ifndef PRODUCT 935 if (trace_opto_pipelining()) { 936 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d", 937 nl, j, ul, l, latency); 938 } 939 #endif 940 } 941 } 942 } else { 943 // This is a projection, just grab the latency of the use(s) 944 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) { 945 uint l = latency_from_use(use, def, use->fast_out(j)); 946 if (latency < l) latency = l; 947 } 948 } 949 950 return latency; 951 } 952 953 //------------------------------latency_from_uses------------------------------ 954 // Compute the latency of this instruction relative to all of it's uses. 955 // This computes a number that increases as we approach the beginning of the 956 // routine. 957 void PhaseCFG::latency_from_uses(Node *n) { 958 // Set the latency for this instruction 959 #ifndef PRODUCT 960 if (trace_opto_pipelining()) { 961 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", 962 n->_idx, _node_latency.at_grow(n->_idx)); 963 dump(); 964 } 965 #endif 966 uint latency=0; 967 const Node *def = n->is_Proj() ? n->in(0): n; 968 969 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 970 uint l = latency_from_use(n, def, n->fast_out(i)); 971 972 if (latency < l) latency = l; 973 } 974 975 _node_latency.at_put_grow(n->_idx, latency); 976 } 977 978 //------------------------------hoist_to_cheaper_block------------------------- 979 // Pick a block for node self, between early and LCA, that is a cheaper 980 // alternative to LCA. 981 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) { 982 const double delta = 1+PROB_UNLIKELY_MAG(4); 983 Block* least = LCA; 984 double least_freq = least->_freq; 985 uint target = _node_latency.at_grow(self->_idx); 986 uint start_latency = _node_latency.at_grow(LCA->_nodes[0]->_idx); 987 uint end_latency = _node_latency.at_grow(LCA->_nodes[LCA->end_idx()]->_idx); 988 bool in_latency = (target <= start_latency); 989 const Block* root_block = _bbs[_root->_idx]; 990 991 // Turn off latency scheduling if scheduling is just plain off 992 if (!C->do_scheduling()) 993 in_latency = true; 994 995 // Do not hoist (to cover latency) instructions which target a 996 // single register. Hoisting stretches the live range of the 997 // single register and may force spilling. 998 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL; 999 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty()) 1000 in_latency = true; 1001 1002 #ifndef PRODUCT 1003 if (trace_opto_pipelining()) { 1004 tty->print("# Find cheaper block for latency %d: ", 1005 _node_latency.at_grow(self->_idx)); 1006 self->dump(); 1007 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g", 1008 LCA->_pre_order, 1009 LCA->_nodes[0]->_idx, 1010 start_latency, 1011 LCA->_nodes[LCA->end_idx()]->_idx, 1012 end_latency, 1013 least_freq); 1014 } 1015 #endif 1016 1017 // Walk up the dominator tree from LCA (Lowest common ancestor) to 1018 // the earliest legal location. Capture the least execution frequency. 1019 while (LCA != early) { 1020 LCA = LCA->_idom; // Follow up the dominator tree 1021 1022 if (LCA == NULL) { 1023 // Bailout without retry 1024 C->record_method_not_compilable("late schedule failed: LCA == NULL"); 1025 return least; 1026 } 1027 1028 // Don't hoist machine instructions to the root basic block 1029 if (mach && LCA == root_block) 1030 break; 1031 1032 uint start_lat = _node_latency.at_grow(LCA->_nodes[0]->_idx); 1033 uint end_idx = LCA->end_idx(); 1034 uint end_lat = _node_latency.at_grow(LCA->_nodes[end_idx]->_idx); 1035 double LCA_freq = LCA->_freq; 1036 #ifndef PRODUCT 1037 if (trace_opto_pipelining()) { 1038 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g", 1039 LCA->_pre_order, LCA->_nodes[0]->_idx, start_lat, end_idx, end_lat, LCA_freq); 1040 } 1041 #endif 1042 if (LCA_freq < least_freq || // Better Frequency 1043 ( !in_latency && // No block containing latency 1044 LCA_freq < least_freq * delta && // No worse frequency 1045 target >= end_lat && // within latency range 1046 !self->is_iteratively_computed() ) // But don't hoist IV increments 1047 // because they may end up above other uses of their phi forcing 1048 // their result register to be different from their input. 1049 ) { 1050 least = LCA; // Found cheaper block 1051 least_freq = LCA_freq; 1052 start_latency = start_lat; 1053 end_latency = end_lat; 1054 if (target <= start_lat) 1055 in_latency = true; 1056 } 1057 } 1058 1059 #ifndef PRODUCT 1060 if (trace_opto_pipelining()) { 1061 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g", 1062 least->_pre_order, start_latency, least_freq); 1063 } 1064 #endif 1065 1066 // See if the latency needs to be updated 1067 if (target < end_latency) { 1068 #ifndef PRODUCT 1069 if (trace_opto_pipelining()) { 1070 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency); 1071 } 1072 #endif 1073 _node_latency.at_put_grow(self->_idx, end_latency); 1074 partial_latency_of_defs(self); 1075 } 1076 1077 return least; 1078 } 1079 1080 1081 //------------------------------schedule_late----------------------------------- 1082 // Now schedule all codes as LATE as possible. This is the LCA in the 1083 // dominator tree of all USES of a value. Pick the block with the least 1084 // loop nesting depth that is lowest in the dominator tree. 1085 extern const char must_clone[]; 1086 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) { 1087 #ifndef PRODUCT 1088 if (trace_opto_pipelining()) 1089 tty->print("\n#---- schedule_late ----\n"); 1090 #endif 1091 1092 Node_Backward_Iterator iter((Node *)_root, visited, stack, _bbs); 1093 Node *self; 1094 1095 // Walk over all the nodes from last to first 1096 while (self = iter.next()) { 1097 Block* early = _bbs[self->_idx]; // Earliest legal placement 1098 1099 if (self->is_top()) { 1100 // Top node goes in bb #2 with other constants. 1101 // It must be special-cased, because it has no out edges. 1102 early->add_inst(self); 1103 continue; 1104 } 1105 1106 // No uses, just terminate 1107 if (self->outcnt() == 0) { 1108 assert(self->Opcode() == Op_MachProj, "sanity"); 1109 continue; // Must be a dead machine projection 1110 } 1111 1112 // If node is pinned in the block, then no scheduling can be done. 1113 if( self->pinned() ) // Pinned in block? 1114 continue; 1115 1116 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL; 1117 if (mach) { 1118 switch (mach->ideal_Opcode()) { 1119 case Op_CreateEx: 1120 // Don't move exception creation 1121 early->add_inst(self); 1122 continue; 1123 break; 1124 case Op_CheckCastPP: 1125 // Don't move CheckCastPP nodes away from their input, if the input 1126 // is a rawptr (5071820). 1127 Node *def = self->in(1); 1128 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) { 1129 early->add_inst(self); 1130 continue; 1131 } 1132 break; 1133 } 1134 } 1135 1136 // Gather LCA of all uses 1137 Block *LCA = NULL; 1138 { 1139 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) { 1140 // For all uses, find LCA 1141 Node* use = self->fast_out(i); 1142 LCA = raise_LCA_above_use(LCA, use, self, _bbs); 1143 } 1144 } // (Hide defs of imax, i from rest of block.) 1145 1146 // Place temps in the block of their use. This isn't a 1147 // requirement for correctness but it reduces useless 1148 // interference between temps and other nodes. 1149 if (mach != NULL && mach->is_MachTemp()) { 1150 _bbs.map(self->_idx, LCA); 1151 LCA->add_inst(self); 1152 continue; 1153 } 1154 1155 // Check if 'self' could be anti-dependent on memory 1156 if (self->needs_anti_dependence_check()) { 1157 // Hoist LCA above possible-defs and insert anti-dependences to 1158 // defs in new LCA block. 1159 LCA = insert_anti_dependences(LCA, self); 1160 } 1161 1162 if (early->_dom_depth > LCA->_dom_depth) { 1163 // Somehow the LCA has moved above the earliest legal point. 1164 // (One way this can happen is via memory_early_block.) 1165 if (C->subsume_loads() == true && !C->failing()) { 1166 // Retry with subsume_loads == false 1167 // If this is the first failure, the sentinel string will "stick" 1168 // to the Compile object, and the C2Compiler will see it and retry. 1169 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 1170 } else { 1171 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth) 1172 C->record_method_not_compilable("late schedule failed: incorrect graph"); 1173 } 1174 return; 1175 } 1176 1177 // If there is no opportunity to hoist, then we're done. 1178 bool try_to_hoist = (LCA != early); 1179 1180 // Must clone guys stay next to use; no hoisting allowed. 1181 // Also cannot hoist guys that alter memory or are otherwise not 1182 // allocatable (hoisting can make a value live longer, leading to 1183 // anti and output dependency problems which are normally resolved 1184 // by the register allocator giving everyone a different register). 1185 if (mach != NULL && must_clone[mach->ideal_Opcode()]) 1186 try_to_hoist = false; 1187 1188 Block* late = NULL; 1189 if (try_to_hoist) { 1190 // Now find the block with the least execution frequency. 1191 // Start at the latest schedule and work up to the earliest schedule 1192 // in the dominator tree. Thus the Node will dominate all its uses. 1193 late = hoist_to_cheaper_block(LCA, early, self); 1194 } else { 1195 // Just use the LCA of the uses. 1196 late = LCA; 1197 } 1198 1199 // Put the node into target block 1200 schedule_node_into_block(self, late); 1201 1202 #ifdef ASSERT 1203 if (self->needs_anti_dependence_check()) { 1204 // since precedence edges are only inserted when we're sure they 1205 // are needed make sure that after placement in a block we don't 1206 // need any new precedence edges. 1207 verify_anti_dependences(late, self); 1208 } 1209 #endif 1210 } // Loop until all nodes have been visited 1211 1212 } // end ScheduleLate 1213 1214 //------------------------------GlobalCodeMotion------------------------------- 1215 void PhaseCFG::GlobalCodeMotion( Matcher &matcher, uint unique, Node_List &proj_list ) { 1216 ResourceMark rm; 1217 1218 #ifndef PRODUCT 1219 if (trace_opto_pipelining()) { 1220 tty->print("\n---- Start GlobalCodeMotion ----\n"); 1221 } 1222 #endif 1223 1224 // Initialize the bbs.map for things on the proj_list 1225 uint i; 1226 for( i=0; i < proj_list.size(); i++ ) 1227 _bbs.map(proj_list[i]->_idx, NULL); 1228 1229 // Set the basic block for Nodes pinned into blocks 1230 Arena *a = Thread::current()->resource_area(); 1231 VectorSet visited(a); 1232 schedule_pinned_nodes( visited ); 1233 1234 // Find the earliest Block any instruction can be placed in. Some 1235 // instructions are pinned into Blocks. Unpinned instructions can 1236 // appear in last block in which all their inputs occur. 1237 visited.Clear(); 1238 Node_List stack(a); 1239 stack.map( (unique >> 1) + 16, NULL); // Pre-grow the list 1240 if (!schedule_early(visited, stack)) { 1241 // Bailout without retry 1242 C->record_method_not_compilable("early schedule failed"); 1243 return; 1244 } 1245 1246 // Build Def-Use edges. 1247 proj_list.push(_root); // Add real root as another root 1248 proj_list.pop(); 1249 1250 // Compute the latency information (via backwards walk) for all the 1251 // instructions in the graph 1252 GrowableArray<uint> node_latency; 1253 _node_latency = node_latency; 1254 1255 if( C->do_scheduling() ) 1256 ComputeLatenciesBackwards(visited, stack); 1257 1258 // Now schedule all codes as LATE as possible. This is the LCA in the 1259 // dominator tree of all USES of a value. Pick the block with the least 1260 // loop nesting depth that is lowest in the dominator tree. 1261 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() ) 1262 schedule_late(visited, stack); 1263 if( C->failing() ) { 1264 // schedule_late fails only when graph is incorrect. 1265 assert(!VerifyGraphEdges, "verification should have failed"); 1266 return; 1267 } 1268 1269 unique = C->unique(); 1270 1271 #ifndef PRODUCT 1272 if (trace_opto_pipelining()) { 1273 tty->print("\n---- Detect implicit null checks ----\n"); 1274 } 1275 #endif 1276 1277 // Detect implicit-null-check opportunities. Basically, find NULL checks 1278 // with suitable memory ops nearby. Use the memory op to do the NULL check. 1279 // I can generate a memory op if there is not one nearby. 1280 if (C->is_method_compilation()) { 1281 // Don't do it for natives, adapters, or runtime stubs 1282 int allowed_reasons = 0; 1283 // ...and don't do it when there have been too many traps, globally. 1284 for (int reason = (int)Deoptimization::Reason_none+1; 1285 reason < Compile::trapHistLength; reason++) { 1286 assert(reason < BitsPerInt, "recode bit map"); 1287 if (!C->too_many_traps((Deoptimization::DeoptReason) reason)) 1288 allowed_reasons |= nth_bit(reason); 1289 } 1290 // By reversing the loop direction we get a very minor gain on mpegaudio. 1291 // Feel free to revert to a forward loop for clarity. 1292 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) { 1293 for( int i= matcher._null_check_tests.size()-2; i>=0; i-=2 ) { 1294 Node *proj = matcher._null_check_tests[i ]; 1295 Node *val = matcher._null_check_tests[i+1]; 1296 _bbs[proj->_idx]->implicit_null_check(this, proj, val, allowed_reasons); 1297 // The implicit_null_check will only perform the transformation 1298 // if the null branch is truly uncommon, *and* it leads to an 1299 // uncommon trap. Combined with the too_many_traps guards 1300 // above, this prevents SEGV storms reported in 6366351, 1301 // by recompiling offending methods without this optimization. 1302 } 1303 } 1304 1305 #ifndef PRODUCT 1306 if (trace_opto_pipelining()) { 1307 tty->print("\n---- Start Local Scheduling ----\n"); 1308 } 1309 #endif 1310 1311 // Schedule locally. Right now a simple topological sort. 1312 // Later, do a real latency aware scheduler. 1313 int *ready_cnt = NEW_RESOURCE_ARRAY(int,C->unique()); 1314 memset( ready_cnt, -1, C->unique() * sizeof(int) ); 1315 visited.Clear(); 1316 for (i = 0; i < _num_blocks; i++) { 1317 if (!_blocks[i]->schedule_local(this, matcher, ready_cnt, visited)) { 1318 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 1319 C->record_method_not_compilable("local schedule failed"); 1320 } 1321 return; 1322 } 1323 } 1324 1325 // If we inserted any instructions between a Call and his CatchNode, 1326 // clone the instructions on all paths below the Catch. 1327 for( i=0; i < _num_blocks; i++ ) 1328 _blocks[i]->call_catch_cleanup(_bbs); 1329 1330 #ifndef PRODUCT 1331 if (trace_opto_pipelining()) { 1332 tty->print("\n---- After GlobalCodeMotion ----\n"); 1333 for (uint i = 0; i < _num_blocks; i++) { 1334 _blocks[i]->dump(); 1335 } 1336 } 1337 #endif 1338 } 1339 1340 1341 //------------------------------Estimate_Block_Frequency----------------------- 1342 // Estimate block frequencies based on IfNode probabilities. 1343 void PhaseCFG::Estimate_Block_Frequency() { 1344 1345 // Force conditional branches leading to uncommon traps to be unlikely, 1346 // not because we get to the uncommon_trap with less relative frequency, 1347 // but because an uncommon_trap typically causes a deopt, so we only get 1348 // there once. 1349 if (C->do_freq_based_layout()) { 1350 Block_List worklist; 1351 Block* root_blk = _blocks[0]; 1352 for (uint i = 1; i < root_blk->num_preds(); i++) { 1353 Block *pb = _bbs[root_blk->pred(i)->_idx]; 1354 if (pb->has_uncommon_code()) { 1355 worklist.push(pb); 1356 } 1357 } 1358 while (worklist.size() > 0) { 1359 Block* uct = worklist.pop(); 1360 if (uct == _broot) continue; 1361 for (uint i = 1; i < uct->num_preds(); i++) { 1362 Block *pb = _bbs[uct->pred(i)->_idx]; 1363 if (pb->_num_succs == 1) { 1364 worklist.push(pb); 1365 } else if (pb->num_fall_throughs() == 2) { 1366 pb->update_uncommon_branch(uct); 1367 } 1368 } 1369 } 1370 } 1371 1372 // Create the loop tree and calculate loop depth. 1373 _root_loop = create_loop_tree(); 1374 _root_loop->compute_loop_depth(0); 1375 1376 // Compute block frequency of each block, relative to a single loop entry. 1377 _root_loop->compute_freq(); 1378 1379 // Adjust all frequencies to be relative to a single method entry 1380 _root_loop->_freq = 1.0; 1381 _root_loop->scale_freq(); 1382 1383 // Save outmost loop frequency for LRG frequency threshold 1384 _outer_loop_freq = _root_loop->outer_loop_freq(); 1385 1386 // force paths ending at uncommon traps to be infrequent 1387 if (!C->do_freq_based_layout()) { 1388 Block_List worklist; 1389 Block* root_blk = _blocks[0]; 1390 for (uint i = 1; i < root_blk->num_preds(); i++) { 1391 Block *pb = _bbs[root_blk->pred(i)->_idx]; 1392 if (pb->has_uncommon_code()) { 1393 worklist.push(pb); 1394 } 1395 } 1396 while (worklist.size() > 0) { 1397 Block* uct = worklist.pop(); 1398 uct->_freq = PROB_MIN; 1399 for (uint i = 1; i < uct->num_preds(); i++) { 1400 Block *pb = _bbs[uct->pred(i)->_idx]; 1401 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) { 1402 worklist.push(pb); 1403 } 1404 } 1405 } 1406 } 1407 1408 #ifdef ASSERT 1409 for (uint i = 0; i < _num_blocks; i++ ) { 1410 Block *b = _blocks[i]; 1411 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency"); 1412 } 1413 #endif 1414 1415 #ifndef PRODUCT 1416 if (PrintCFGBlockFreq) { 1417 tty->print_cr("CFG Block Frequencies"); 1418 _root_loop->dump_tree(); 1419 if (Verbose) { 1420 tty->print_cr("PhaseCFG dump"); 1421 dump(); 1422 tty->print_cr("Node dump"); 1423 _root->dump(99999); 1424 } 1425 } 1426 #endif 1427 } 1428 1429 //----------------------------create_loop_tree-------------------------------- 1430 // Create a loop tree from the CFG 1431 CFGLoop* PhaseCFG::create_loop_tree() { 1432 1433 #ifdef ASSERT 1434 assert( _blocks[0] == _broot, "" ); 1435 for (uint i = 0; i < _num_blocks; i++ ) { 1436 Block *b = _blocks[i]; 1437 // Check that _loop field are clear...we could clear them if not. 1438 assert(b->_loop == NULL, "clear _loop expected"); 1439 // Sanity check that the RPO numbering is reflected in the _blocks array. 1440 // It doesn't have to be for the loop tree to be built, but if it is not, 1441 // then the blocks have been reordered since dom graph building...which 1442 // may question the RPO numbering 1443 assert(b->_rpo == i, "unexpected reverse post order number"); 1444 } 1445 #endif 1446 1447 int idct = 0; 1448 CFGLoop* root_loop = new CFGLoop(idct++); 1449 1450 Block_List worklist; 1451 1452 // Assign blocks to loops 1453 for(uint i = _num_blocks - 1; i > 0; i-- ) { // skip Root block 1454 Block *b = _blocks[i]; 1455 1456 if (b->head()->is_Loop()) { 1457 Block* loop_head = b; 1458 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors"); 1459 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl); 1460 Block* tail = _bbs[tail_n->_idx]; 1461 1462 // Defensively filter out Loop nodes for non-single-entry loops. 1463 // For all reasonable loops, the head occurs before the tail in RPO. 1464 if (i <= tail->_rpo) { 1465 1466 // The tail and (recursive) predecessors of the tail 1467 // are made members of a new loop. 1468 1469 assert(worklist.size() == 0, "nonempty worklist"); 1470 CFGLoop* nloop = new CFGLoop(idct++); 1471 assert(loop_head->_loop == NULL, "just checking"); 1472 loop_head->_loop = nloop; 1473 // Add to nloop so push_pred() will skip over inner loops 1474 nloop->add_member(loop_head); 1475 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, _bbs); 1476 1477 while (worklist.size() > 0) { 1478 Block* member = worklist.pop(); 1479 if (member != loop_head) { 1480 for (uint j = 1; j < member->num_preds(); j++) { 1481 nloop->push_pred(member, j, worklist, _bbs); 1482 } 1483 } 1484 } 1485 } 1486 } 1487 } 1488 1489 // Create a member list for each loop consisting 1490 // of both blocks and (immediate child) loops. 1491 for (uint i = 0; i < _num_blocks; i++) { 1492 Block *b = _blocks[i]; 1493 CFGLoop* lp = b->_loop; 1494 if (lp == NULL) { 1495 // Not assigned to a loop. Add it to the method's pseudo loop. 1496 b->_loop = root_loop; 1497 lp = root_loop; 1498 } 1499 if (lp == root_loop || b != lp->head()) { // loop heads are already members 1500 lp->add_member(b); 1501 } 1502 if (lp != root_loop) { 1503 if (lp->parent() == NULL) { 1504 // Not a nested loop. Make it a child of the method's pseudo loop. 1505 root_loop->add_nested_loop(lp); 1506 } 1507 if (b == lp->head()) { 1508 // Add nested loop to member list of parent loop. 1509 lp->parent()->add_member(lp); 1510 } 1511 } 1512 } 1513 1514 return root_loop; 1515 } 1516 1517 //------------------------------push_pred-------------------------------------- 1518 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, Block_Array& node_to_blk) { 1519 Node* pred_n = blk->pred(i); 1520 Block* pred = node_to_blk[pred_n->_idx]; 1521 CFGLoop *pred_loop = pred->_loop; 1522 if (pred_loop == NULL) { 1523 // Filter out blocks for non-single-entry loops. 1524 // For all reasonable loops, the head occurs before the tail in RPO. 1525 if (pred->_rpo > head()->_rpo) { 1526 pred->_loop = this; 1527 worklist.push(pred); 1528 } 1529 } else if (pred_loop != this) { 1530 // Nested loop. 1531 while (pred_loop->_parent != NULL && pred_loop->_parent != this) { 1532 pred_loop = pred_loop->_parent; 1533 } 1534 // Make pred's loop be a child 1535 if (pred_loop->_parent == NULL) { 1536 add_nested_loop(pred_loop); 1537 // Continue with loop entry predecessor. 1538 Block* pred_head = pred_loop->head(); 1539 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors"); 1540 assert(pred_head != head(), "loop head in only one loop"); 1541 push_pred(pred_head, LoopNode::EntryControl, worklist, node_to_blk); 1542 } else { 1543 assert(pred_loop->_parent == this && _parent == NULL, "just checking"); 1544 } 1545 } 1546 } 1547 1548 //------------------------------add_nested_loop-------------------------------- 1549 // Make cl a child of the current loop in the loop tree. 1550 void CFGLoop::add_nested_loop(CFGLoop* cl) { 1551 assert(_parent == NULL, "no parent yet"); 1552 assert(cl != this, "not my own parent"); 1553 cl->_parent = this; 1554 CFGLoop* ch = _child; 1555 if (ch == NULL) { 1556 _child = cl; 1557 } else { 1558 while (ch->_sibling != NULL) { ch = ch->_sibling; } 1559 ch->_sibling = cl; 1560 } 1561 } 1562 1563 //------------------------------compute_loop_depth----------------------------- 1564 // Store the loop depth in each CFGLoop object. 1565 // Recursively walk the children to do the same for them. 1566 void CFGLoop::compute_loop_depth(int depth) { 1567 _depth = depth; 1568 CFGLoop* ch = _child; 1569 while (ch != NULL) { 1570 ch->compute_loop_depth(depth + 1); 1571 ch = ch->_sibling; 1572 } 1573 } 1574 1575 //------------------------------compute_freq----------------------------------- 1576 // Compute the frequency of each block and loop, relative to a single entry 1577 // into the dominating loop head. 1578 void CFGLoop::compute_freq() { 1579 // Bottom up traversal of loop tree (visit inner loops first.) 1580 // Set loop head frequency to 1.0, then transitively 1581 // compute frequency for all successors in the loop, 1582 // as well as for each exit edge. Inner loops are 1583 // treated as single blocks with loop exit targets 1584 // as the successor blocks. 1585 1586 // Nested loops first 1587 CFGLoop* ch = _child; 1588 while (ch != NULL) { 1589 ch->compute_freq(); 1590 ch = ch->_sibling; 1591 } 1592 assert (_members.length() > 0, "no empty loops"); 1593 Block* hd = head(); 1594 hd->_freq = 1.0f; 1595 for (int i = 0; i < _members.length(); i++) { 1596 CFGElement* s = _members.at(i); 1597 float freq = s->_freq; 1598 if (s->is_block()) { 1599 Block* b = s->as_Block(); 1600 for (uint j = 0; j < b->_num_succs; j++) { 1601 Block* sb = b->_succs[j]; 1602 update_succ_freq(sb, freq * b->succ_prob(j)); 1603 } 1604 } else { 1605 CFGLoop* lp = s->as_CFGLoop(); 1606 assert(lp->_parent == this, "immediate child"); 1607 for (int k = 0; k < lp->_exits.length(); k++) { 1608 Block* eb = lp->_exits.at(k).get_target(); 1609 float prob = lp->_exits.at(k).get_prob(); 1610 update_succ_freq(eb, freq * prob); 1611 } 1612 } 1613 } 1614 1615 // For all loops other than the outer, "method" loop, 1616 // sum and normalize the exit probability. The "method" loop 1617 // should keep the initial exit probability of 1, so that 1618 // inner blocks do not get erroneously scaled. 1619 if (_depth != 0) { 1620 // Total the exit probabilities for this loop. 1621 float exits_sum = 0.0f; 1622 for (int i = 0; i < _exits.length(); i++) { 1623 exits_sum += _exits.at(i).get_prob(); 1624 } 1625 1626 // Normalize the exit probabilities. Until now, the 1627 // probabilities estimate the possibility of exit per 1628 // a single loop iteration; afterward, they estimate 1629 // the probability of exit per loop entry. 1630 for (int i = 0; i < _exits.length(); i++) { 1631 Block* et = _exits.at(i).get_target(); 1632 float new_prob = 0.0f; 1633 if (_exits.at(i).get_prob() > 0.0f) { 1634 new_prob = _exits.at(i).get_prob() / exits_sum; 1635 } 1636 BlockProbPair bpp(et, new_prob); 1637 _exits.at_put(i, bpp); 1638 } 1639 1640 // Save the total, but guard against unreasonable probability, 1641 // as the value is used to estimate the loop trip count. 1642 // An infinite trip count would blur relative block 1643 // frequencies. 1644 if (exits_sum > 1.0f) exits_sum = 1.0; 1645 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN; 1646 _exit_prob = exits_sum; 1647 } 1648 } 1649 1650 //------------------------------succ_prob------------------------------------- 1651 // Determine the probability of reaching successor 'i' from the receiver block. 1652 float Block::succ_prob(uint i) { 1653 int eidx = end_idx(); 1654 Node *n = _nodes[eidx]; // Get ending Node 1655 1656 int op = n->Opcode(); 1657 if (n->is_Mach()) { 1658 if (n->is_MachNullCheck()) { 1659 // Can only reach here if called after lcm. The original Op_If is gone, 1660 // so we attempt to infer the probability from one or both of the 1661 // successor blocks. 1662 assert(_num_succs == 2, "expecting 2 successors of a null check"); 1663 // If either successor has only one predecessor, then the 1664 // probability estimate can be derived using the 1665 // relative frequency of the successor and this block. 1666 if (_succs[i]->num_preds() == 2) { 1667 return _succs[i]->_freq / _freq; 1668 } else if (_succs[1-i]->num_preds() == 2) { 1669 return 1 - (_succs[1-i]->_freq / _freq); 1670 } else { 1671 // Estimate using both successor frequencies 1672 float freq = _succs[i]->_freq; 1673 return freq / (freq + _succs[1-i]->_freq); 1674 } 1675 } 1676 op = n->as_Mach()->ideal_Opcode(); 1677 } 1678 1679 1680 // Switch on branch type 1681 switch( op ) { 1682 case Op_CountedLoopEnd: 1683 case Op_If: { 1684 assert (i < 2, "just checking"); 1685 // Conditionals pass on only part of their frequency 1686 float prob = n->as_MachIf()->_prob; 1687 assert(prob >= 0.0 && prob <= 1.0, "out of range probability"); 1688 // If succ[i] is the FALSE branch, invert path info 1689 if( _nodes[i + eidx + 1]->Opcode() == Op_IfFalse ) { 1690 return 1.0f - prob; // not taken 1691 } else { 1692 return prob; // taken 1693 } 1694 } 1695 1696 case Op_Jump: 1697 // Divide the frequency between all successors evenly 1698 return 1.0f/_num_succs; 1699 1700 case Op_Catch: { 1701 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj(); 1702 if (ci->_con == CatchProjNode::fall_through_index) { 1703 // Fall-thru path gets the lion's share. 1704 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs; 1705 } else { 1706 // Presume exceptional paths are equally unlikely 1707 return PROB_UNLIKELY_MAG(5); 1708 } 1709 } 1710 1711 case Op_Root: 1712 case Op_Goto: 1713 // Pass frequency straight thru to target 1714 return 1.0f; 1715 1716 case Op_NeverBranch: 1717 return 0.0f; 1718 1719 case Op_TailCall: 1720 case Op_TailJump: 1721 case Op_Return: 1722 case Op_Halt: 1723 case Op_Rethrow: 1724 // Do not push out freq to root block 1725 return 0.0f; 1726 1727 default: 1728 ShouldNotReachHere(); 1729 } 1730 1731 return 0.0f; 1732 } 1733 1734 //------------------------------num_fall_throughs----------------------------- 1735 // Return the number of fall-through candidates for a block 1736 int Block::num_fall_throughs() { 1737 int eidx = end_idx(); 1738 Node *n = _nodes[eidx]; // Get ending Node 1739 1740 int op = n->Opcode(); 1741 if (n->is_Mach()) { 1742 if (n->is_MachNullCheck()) { 1743 // In theory, either side can fall-thru, for simplicity sake, 1744 // let's say only the false branch can now. 1745 return 1; 1746 } 1747 op = n->as_Mach()->ideal_Opcode(); 1748 } 1749 1750 // Switch on branch type 1751 switch( op ) { 1752 case Op_CountedLoopEnd: 1753 case Op_If: 1754 return 2; 1755 1756 case Op_Root: 1757 case Op_Goto: 1758 return 1; 1759 1760 case Op_Catch: { 1761 for (uint i = 0; i < _num_succs; i++) { 1762 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj(); 1763 if (ci->_con == CatchProjNode::fall_through_index) { 1764 return 1; 1765 } 1766 } 1767 return 0; 1768 } 1769 1770 case Op_Jump: 1771 case Op_NeverBranch: 1772 case Op_TailCall: 1773 case Op_TailJump: 1774 case Op_Return: 1775 case Op_Halt: 1776 case Op_Rethrow: 1777 return 0; 1778 1779 default: 1780 ShouldNotReachHere(); 1781 } 1782 1783 return 0; 1784 } 1785 1786 //------------------------------succ_fall_through----------------------------- 1787 // Return true if a specific successor could be fall-through target. 1788 bool Block::succ_fall_through(uint i) { 1789 int eidx = end_idx(); 1790 Node *n = _nodes[eidx]; // Get ending Node 1791 1792 int op = n->Opcode(); 1793 if (n->is_Mach()) { 1794 if (n->is_MachNullCheck()) { 1795 // In theory, either side can fall-thru, for simplicity sake, 1796 // let's say only the false branch can now. 1797 return _nodes[i + eidx + 1]->Opcode() == Op_IfFalse; 1798 } 1799 op = n->as_Mach()->ideal_Opcode(); 1800 } 1801 1802 // Switch on branch type 1803 switch( op ) { 1804 case Op_CountedLoopEnd: 1805 case Op_If: 1806 case Op_Root: 1807 case Op_Goto: 1808 return true; 1809 1810 case Op_Catch: { 1811 const CatchProjNode *ci = _nodes[i + eidx + 1]->as_CatchProj(); 1812 return ci->_con == CatchProjNode::fall_through_index; 1813 } 1814 1815 case Op_Jump: 1816 case Op_NeverBranch: 1817 case Op_TailCall: 1818 case Op_TailJump: 1819 case Op_Return: 1820 case Op_Halt: 1821 case Op_Rethrow: 1822 return false; 1823 1824 default: 1825 ShouldNotReachHere(); 1826 } 1827 1828 return false; 1829 } 1830 1831 //------------------------------update_uncommon_branch------------------------ 1832 // Update the probability of a two-branch to be uncommon 1833 void Block::update_uncommon_branch(Block* ub) { 1834 int eidx = end_idx(); 1835 Node *n = _nodes[eidx]; // Get ending Node 1836 1837 int op = n->as_Mach()->ideal_Opcode(); 1838 1839 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If"); 1840 assert(num_fall_throughs() == 2, "must be a two way branch block"); 1841 1842 // Which successor is ub? 1843 uint s; 1844 for (s = 0; s <_num_succs; s++) { 1845 if (_succs[s] == ub) break; 1846 } 1847 assert(s < 2, "uncommon successor must be found"); 1848 1849 // If ub is the true path, make the proability small, else 1850 // ub is the false path, and make the probability large 1851 bool invert = (_nodes[s + eidx + 1]->Opcode() == Op_IfFalse); 1852 1853 // Get existing probability 1854 float p = n->as_MachIf()->_prob; 1855 1856 if (invert) p = 1.0 - p; 1857 if (p > PROB_MIN) { 1858 p = PROB_MIN; 1859 } 1860 if (invert) p = 1.0 - p; 1861 1862 n->as_MachIf()->_prob = p; 1863 } 1864 1865 //------------------------------update_succ_freq------------------------------- 1866 // Update the appropriate frequency associated with block 'b', a successor of 1867 // a block in this loop. 1868 void CFGLoop::update_succ_freq(Block* b, float freq) { 1869 if (b->_loop == this) { 1870 if (b == head()) { 1871 // back branch within the loop 1872 // Do nothing now, the loop carried frequency will be 1873 // adjust later in scale_freq(). 1874 } else { 1875 // simple branch within the loop 1876 b->_freq += freq; 1877 } 1878 } else if (!in_loop_nest(b)) { 1879 // branch is exit from this loop 1880 BlockProbPair bpp(b, freq); 1881 _exits.append(bpp); 1882 } else { 1883 // branch into nested loop 1884 CFGLoop* ch = b->_loop; 1885 ch->_freq += freq; 1886 } 1887 } 1888 1889 //------------------------------in_loop_nest----------------------------------- 1890 // Determine if block b is in the receiver's loop nest. 1891 bool CFGLoop::in_loop_nest(Block* b) { 1892 int depth = _depth; 1893 CFGLoop* b_loop = b->_loop; 1894 int b_depth = b_loop->_depth; 1895 if (depth == b_depth) { 1896 return true; 1897 } 1898 while (b_depth > depth) { 1899 b_loop = b_loop->_parent; 1900 b_depth = b_loop->_depth; 1901 } 1902 return b_loop == this; 1903 } 1904 1905 //------------------------------scale_freq------------------------------------- 1906 // Scale frequency of loops and blocks by trip counts from outer loops 1907 // Do a top down traversal of loop tree (visit outer loops first.) 1908 void CFGLoop::scale_freq() { 1909 float loop_freq = _freq * trip_count(); 1910 _freq = loop_freq; 1911 for (int i = 0; i < _members.length(); i++) { 1912 CFGElement* s = _members.at(i); 1913 float block_freq = s->_freq * loop_freq; 1914 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY) 1915 block_freq = MIN_BLOCK_FREQUENCY; 1916 s->_freq = block_freq; 1917 } 1918 CFGLoop* ch = _child; 1919 while (ch != NULL) { 1920 ch->scale_freq(); 1921 ch = ch->_sibling; 1922 } 1923 } 1924 1925 // Frequency of outer loop 1926 float CFGLoop::outer_loop_freq() const { 1927 if (_child != NULL) { 1928 return _child->_freq; 1929 } 1930 return _freq; 1931 } 1932 1933 #ifndef PRODUCT 1934 //------------------------------dump_tree-------------------------------------- 1935 void CFGLoop::dump_tree() const { 1936 dump(); 1937 if (_child != NULL) _child->dump_tree(); 1938 if (_sibling != NULL) _sibling->dump_tree(); 1939 } 1940 1941 //------------------------------dump------------------------------------------- 1942 void CFGLoop::dump() const { 1943 for (int i = 0; i < _depth; i++) tty->print(" "); 1944 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n", 1945 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq); 1946 for (int i = 0; i < _depth; i++) tty->print(" "); 1947 tty->print(" members:", _id); 1948 int k = 0; 1949 for (int i = 0; i < _members.length(); i++) { 1950 if (k++ >= 6) { 1951 tty->print("\n "); 1952 for (int j = 0; j < _depth+1; j++) tty->print(" "); 1953 k = 0; 1954 } 1955 CFGElement *s = _members.at(i); 1956 if (s->is_block()) { 1957 Block *b = s->as_Block(); 1958 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq); 1959 } else { 1960 CFGLoop* lp = s->as_CFGLoop(); 1961 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq); 1962 } 1963 } 1964 tty->print("\n"); 1965 for (int i = 0; i < _depth; i++) tty->print(" "); 1966 tty->print(" exits: "); 1967 k = 0; 1968 for (int i = 0; i < _exits.length(); i++) { 1969 if (k++ >= 7) { 1970 tty->print("\n "); 1971 for (int j = 0; j < _depth+1; j++) tty->print(" "); 1972 k = 0; 1973 } 1974 Block *blk = _exits.at(i).get_target(); 1975 float prob = _exits.at(i).get_prob(); 1976 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100)); 1977 } 1978 tty->print("\n"); 1979 } 1980 #endif