1 /* 2 * Copyright (c) 1997, 2016, 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 "libadt/vectset.hpp" 27 #include "memory/allocation.inline.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "opto/block.hpp" 30 #include "opto/c2compiler.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/cfgnode.hpp" 33 #include "opto/machnode.hpp" 34 #include "opto/opcodes.hpp" 35 #include "opto/phaseX.hpp" 36 #include "opto/rootnode.hpp" 37 #include "opto/runtime.hpp" 38 #include "opto/chaitin.hpp" 39 #include "runtime/deoptimization.hpp" 40 41 // Portions of code courtesy of Clifford Click 42 43 // Optimization - Graph Style 44 45 // To avoid float value underflow 46 #define MIN_BLOCK_FREQUENCY 1.e-35f 47 48 //----------------------------schedule_node_into_block------------------------- 49 // Insert node n into block b. Look for projections of n and make sure they 50 // are in b also. 51 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) { 52 // Set basic block of n, Add n to b, 53 map_node_to_block(n, b); 54 b->add_inst(n); 55 56 // After Matching, nearly any old Node may have projections trailing it. 57 // These are usually machine-dependent flags. In any case, they might 58 // float to another block below this one. Move them up. 59 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 60 Node* use = n->fast_out(i); 61 if (use->is_Proj()) { 62 Block* buse = get_block_for_node(use); 63 if (buse != b) { // In wrong block? 64 if (buse != NULL) { 65 buse->find_remove(use); // Remove from wrong block 66 } 67 map_node_to_block(use, b); 68 b->add_inst(use); 69 } 70 } 71 } 72 } 73 74 //----------------------------replace_block_proj_ctrl------------------------- 75 // Nodes that have is_block_proj() nodes as their control need to use 76 // the appropriate Region for their actual block as their control since 77 // the projection will be in a predecessor block. 78 void PhaseCFG::replace_block_proj_ctrl( Node *n ) { 79 const Node *in0 = n->in(0); 80 assert(in0 != NULL, "Only control-dependent"); 81 const Node *p = in0->is_block_proj(); 82 if (p != NULL && p != n) { // Control from a block projection? 83 assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here"); 84 // Find trailing Region 85 Block *pb = get_block_for_node(in0); // Block-projection already has basic block 86 uint j = 0; 87 if (pb->_num_succs != 1) { // More then 1 successor? 88 // Search for successor 89 uint max = pb->number_of_nodes(); 90 assert( max > 1, "" ); 91 uint start = max - pb->_num_succs; 92 // Find which output path belongs to projection 93 for (j = start; j < max; j++) { 94 if( pb->get_node(j) == in0 ) 95 break; 96 } 97 assert( j < max, "must find" ); 98 // Change control to match head of successor basic block 99 j -= start; 100 } 101 n->set_req(0, pb->_succs[j]->head()); 102 } 103 } 104 105 bool PhaseCFG::is_dominator(Node* dom_node, Node* node) { 106 if (dom_node == node) { 107 return true; 108 } 109 Block* d = get_block_for_node(dom_node); 110 Block* n = get_block_for_node(node); 111 if (d == n) { 112 if (dom_node->is_block_start()) { 113 return true; 114 } 115 if (node->is_block_start()) { 116 return false; 117 } 118 if (dom_node->is_block_proj()) { 119 return false; 120 } 121 if (node->is_block_proj()) { 122 return true; 123 } 124 #ifdef ASSERT 125 node->dump(); 126 dom_node->dump(); 127 #endif 128 fatal("unhandled"); 129 return false; 130 } 131 return d->dom_lca(n) == d; 132 } 133 134 //------------------------------schedule_pinned_nodes-------------------------- 135 // Set the basic block for Nodes pinned into blocks 136 void PhaseCFG::schedule_pinned_nodes(VectorSet &visited) { 137 // Allocate node stack of size C->live_nodes()+8 to avoid frequent realloc 138 GrowableArray <Node *> spstack(C->live_nodes() + 8); 139 spstack.push(_root); 140 while (spstack.is_nonempty()) { 141 Node* node = spstack.pop(); 142 if (!visited.test_set(node->_idx)) { // Test node and flag it as visited 143 if (node->pinned() && !has_block(node)) { // Pinned? Nail it down! 144 assert(node->in(0), "pinned Node must have Control"); 145 // Before setting block replace block_proj control edge 146 replace_block_proj_ctrl(node); 147 Node* input = node->in(0); 148 while (!input->is_block_start()) { 149 input = input->in(0); 150 } 151 Block* block = get_block_for_node(input); // Basic block of controlling input 152 schedule_node_into_block(node, block); 153 } 154 155 // If the node has precedence edges (added when CastPP nodes are 156 // removed in final_graph_reshaping), fix the control of the 157 // node to cover the precedence edges and remove the 158 // dependencies. 159 Node* n = NULL; 160 for (uint i = node->len()-1; i >= node->req(); i--) { 161 Node* m = node->in(i); 162 if (m == NULL) continue; 163 // Skip the precedence edge if the test that guarded a CastPP: 164 // - was optimized out during escape analysis 165 // (OptimizePtrCompare): the CastPP's control isn't an end of 166 // block. 167 // - is moved in the branch of a dominating If: the control of 168 // the CastPP is then a Region. 169 if (m->is_block_proj() || m->is_block_start()) { 170 node->rm_prec(i); 171 if (n == NULL) { 172 n = m; 173 } else { 174 assert(is_dominator(n, m) || is_dominator(m, n), "one must dominate the other"); 175 n = is_dominator(n, m) ? m : n; 176 } 177 } 178 } 179 if (n != NULL) { 180 assert(node->in(0), "control should have been set"); 181 assert(is_dominator(n, node->in(0)) || is_dominator(node->in(0), n), "one must dominate the other"); 182 if (!is_dominator(n, node->in(0))) { 183 node->set_req(0, n); 184 } 185 } 186 187 // process all inputs that are non NULL 188 for (int i = node->req() - 1; i >= 0; --i) { 189 if (node->in(i) != NULL) { 190 spstack.push(node->in(i)); 191 } 192 } 193 } 194 } 195 } 196 197 #ifdef ASSERT 198 // Assert that new input b2 is dominated by all previous inputs. 199 // Check this by by seeing that it is dominated by b1, the deepest 200 // input observed until b2. 201 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) { 202 if (b1 == NULL) return; 203 assert(b1->_dom_depth < b2->_dom_depth, "sanity"); 204 Block* tmp = b2; 205 while (tmp != b1 && tmp != NULL) { 206 tmp = tmp->_idom; 207 } 208 if (tmp != b1) { 209 // Detected an unschedulable graph. Print some nice stuff and die. 210 tty->print_cr("!!! Unschedulable graph !!!"); 211 for (uint j=0; j<n->len(); j++) { // For all inputs 212 Node* inn = n->in(j); // Get input 213 if (inn == NULL) continue; // Ignore NULL, missing inputs 214 Block* inb = cfg->get_block_for_node(inn); 215 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order, 216 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth); 217 inn->dump(); 218 } 219 tty->print("Failing node: "); 220 n->dump(); 221 assert(false, "unscheduable graph"); 222 } 223 } 224 #endif 225 226 static Block* find_deepest_input(Node* n, const PhaseCFG* cfg) { 227 // Find the last input dominated by all other inputs. 228 Block* deepb = NULL; // Deepest block so far 229 int deepb_dom_depth = 0; 230 for (uint k = 0; k < n->len(); k++) { // For all inputs 231 Node* inn = n->in(k); // Get input 232 if (inn == NULL) continue; // Ignore NULL, missing inputs 233 Block* inb = cfg->get_block_for_node(inn); 234 assert(inb != NULL, "must already have scheduled this input"); 235 if (deepb_dom_depth < (int) inb->_dom_depth) { 236 // The new inb must be dominated by the previous deepb. 237 // The various inputs must be linearly ordered in the dom 238 // tree, or else there will not be a unique deepest block. 239 DEBUG_ONLY(assert_dom(deepb, inb, n, cfg)); 240 deepb = inb; // Save deepest block 241 deepb_dom_depth = deepb->_dom_depth; 242 } 243 } 244 assert(deepb != NULL, "must be at least one input to n"); 245 return deepb; 246 } 247 248 249 //------------------------------schedule_early--------------------------------- 250 // Find the earliest Block any instruction can be placed in. Some instructions 251 // are pinned into Blocks. Unpinned instructions can appear in last block in 252 // which all their inputs occur. 253 bool PhaseCFG::schedule_early(VectorSet &visited, Node_Stack &roots) { 254 // Allocate stack with enough space to avoid frequent realloc 255 Node_Stack nstack(roots.size() + 8); 256 // _root will be processed among C->top() inputs 257 roots.push(C->top(), 0); 258 visited.set(C->top()->_idx); 259 260 while (roots.size() != 0) { 261 // Use local variables nstack_top_n & nstack_top_i to cache values 262 // on stack's top. 263 Node* parent_node = roots.node(); 264 uint input_index = 0; 265 roots.pop(); 266 267 while (true) { 268 if (input_index == 0) { 269 // Fixup some control. Constants without control get attached 270 // to root and nodes that use is_block_proj() nodes should be attached 271 // to the region that starts their block. 272 const Node* control_input = parent_node->in(0); 273 if (control_input != NULL) { 274 replace_block_proj_ctrl(parent_node); 275 } else { 276 // Is a constant with NO inputs? 277 if (parent_node->req() == 1) { 278 parent_node->set_req(0, _root); 279 } 280 } 281 } 282 283 // First, visit all inputs and force them to get a block. If an 284 // input is already in a block we quit following inputs (to avoid 285 // cycles). Instead we put that Node on a worklist to be handled 286 // later (since IT'S inputs may not have a block yet). 287 288 // Assume all n's inputs will be processed 289 bool done = true; 290 291 while (input_index < parent_node->len()) { 292 Node* in = parent_node->in(input_index++); 293 if (in == NULL) { 294 continue; 295 } 296 297 int is_visited = visited.test_set(in->_idx); 298 if (!has_block(in)) { 299 if (is_visited) { 300 return false; 301 } 302 // Save parent node and next input's index. 303 nstack.push(parent_node, input_index); 304 // Process current input now. 305 parent_node = in; 306 input_index = 0; 307 // Not all n's inputs processed. 308 done = false; 309 break; 310 } else if (!is_visited) { 311 // Visit this guy later, using worklist 312 roots.push(in, 0); 313 } 314 } 315 316 if (done) { 317 // All of n's inputs have been processed, complete post-processing. 318 319 // Some instructions are pinned into a block. These include Region, 320 // Phi, Start, Return, and other control-dependent instructions and 321 // any projections which depend on them. 322 if (!parent_node->pinned()) { 323 // Set earliest legal block. 324 Block* earliest_block = find_deepest_input(parent_node, this); 325 map_node_to_block(parent_node, earliest_block); 326 } else { 327 assert(get_block_for_node(parent_node) == get_block_for_node(parent_node->in(0)), "Pinned Node should be at the same block as its control edge"); 328 } 329 330 if (nstack.is_empty()) { 331 // Finished all nodes on stack. 332 // Process next node on the worklist 'roots'. 333 break; 334 } 335 // Get saved parent node and next input's index. 336 parent_node = nstack.node(); 337 input_index = nstack.index(); 338 nstack.pop(); 339 } 340 } 341 } 342 return true; 343 } 344 345 //------------------------------dom_lca---------------------------------------- 346 // Find least common ancestor in dominator tree 347 // LCA is a current notion of LCA, to be raised above 'this'. 348 // As a convenient boundary condition, return 'this' if LCA is NULL. 349 // Find the LCA of those two nodes. 350 Block* Block::dom_lca(Block* LCA) { 351 if (LCA == NULL || LCA == this) return this; 352 353 Block* anc = this; 354 while (anc->_dom_depth > LCA->_dom_depth) 355 anc = anc->_idom; // Walk up till anc is as high as LCA 356 357 while (LCA->_dom_depth > anc->_dom_depth) 358 LCA = LCA->_idom; // Walk up till LCA is as high as anc 359 360 while (LCA != anc) { // Walk both up till they are the same 361 LCA = LCA->_idom; 362 anc = anc->_idom; 363 } 364 365 return LCA; 366 } 367 368 //--------------------------raise_LCA_above_use-------------------------------- 369 // We are placing a definition, and have been given a def->use edge. 370 // The definition must dominate the use, so move the LCA upward in the 371 // dominator tree to dominate the use. If the use is a phi, adjust 372 // the LCA only with the phi input paths which actually use this def. 373 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, const PhaseCFG* cfg) { 374 Block* buse = cfg->get_block_for_node(use); 375 if (buse == NULL) return LCA; // Unused killing Projs have no use block 376 if (!use->is_Phi()) return buse->dom_lca(LCA); 377 uint pmax = use->req(); // Number of Phi inputs 378 // Why does not this loop just break after finding the matching input to 379 // the Phi? Well...it's like this. I do not have true def-use/use-def 380 // chains. Means I cannot distinguish, from the def-use direction, which 381 // of many use-defs lead from the same use to the same def. That is, this 382 // Phi might have several uses of the same def. Each use appears in a 383 // different predecessor block. But when I enter here, I cannot distinguish 384 // which use-def edge I should find the predecessor block for. So I find 385 // them all. Means I do a little extra work if a Phi uses the same value 386 // more than once. 387 for (uint j=1; j<pmax; j++) { // For all inputs 388 if (use->in(j) == def) { // Found matching input? 389 Block* pred = cfg->get_block_for_node(buse->pred(j)); 390 LCA = pred->dom_lca(LCA); 391 } 392 } 393 return LCA; 394 } 395 396 //----------------------------raise_LCA_above_marks---------------------------- 397 // Return a new LCA that dominates LCA and any of its marked predecessors. 398 // Search all my parents up to 'early' (exclusive), looking for predecessors 399 // which are marked with the given index. Return the LCA (in the dom tree) 400 // of all marked blocks. If there are none marked, return the original 401 // LCA. 402 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, Block* early, const PhaseCFG* cfg) { 403 Block_List worklist; 404 worklist.push(LCA); 405 while (worklist.size() > 0) { 406 Block* mid = worklist.pop(); 407 if (mid == early) continue; // stop searching here 408 409 // Test and set the visited bit. 410 if (mid->raise_LCA_visited() == mark) continue; // already visited 411 412 // Don't process the current LCA, otherwise the search may terminate early 413 if (mid != LCA && mid->raise_LCA_mark() == mark) { 414 // Raise the LCA. 415 LCA = mid->dom_lca(LCA); 416 if (LCA == early) break; // stop searching everywhere 417 assert(early->dominates(LCA), "early is high enough"); 418 // Resume searching at that point, skipping intermediate levels. 419 worklist.push(LCA); 420 if (LCA == mid) 421 continue; // Don't mark as visited to avoid early termination. 422 } else { 423 // Keep searching through this block's predecessors. 424 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) { 425 Block* mid_parent = cfg->get_block_for_node(mid->pred(j)); 426 worklist.push(mid_parent); 427 } 428 } 429 mid->set_raise_LCA_visited(mark); 430 } 431 return LCA; 432 } 433 434 //--------------------------memory_early_block-------------------------------- 435 // This is a variation of find_deepest_input, the heart of schedule_early. 436 // Find the "early" block for a load, if we considered only memory and 437 // address inputs, that is, if other data inputs were ignored. 438 // 439 // Because a subset of edges are considered, the resulting block will 440 // be earlier (at a shallower dom_depth) than the true schedule_early 441 // point of the node. We compute this earlier block as a more permissive 442 // site for anti-dependency insertion, but only if subsume_loads is enabled. 443 static Block* memory_early_block(Node* load, Block* early, const PhaseCFG* cfg) { 444 Node* base; 445 Node* index; 446 Node* store = load->in(MemNode::Memory); 447 load->as_Mach()->memory_inputs(base, index); 448 449 assert(base != NodeSentinel && index != NodeSentinel, 450 "unexpected base/index inputs"); 451 452 Node* mem_inputs[4]; 453 int mem_inputs_length = 0; 454 if (base != NULL) mem_inputs[mem_inputs_length++] = base; 455 if (index != NULL) mem_inputs[mem_inputs_length++] = index; 456 if (store != NULL) mem_inputs[mem_inputs_length++] = store; 457 458 // In the comparision below, add one to account for the control input, 459 // which may be null, but always takes up a spot in the in array. 460 if (mem_inputs_length + 1 < (int) load->req()) { 461 // This "load" has more inputs than just the memory, base and index inputs. 462 // For purposes of checking anti-dependences, we need to start 463 // from the early block of only the address portion of the instruction, 464 // and ignore other blocks that may have factored into the wider 465 // schedule_early calculation. 466 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0); 467 468 Block* deepb = NULL; // Deepest block so far 469 int deepb_dom_depth = 0; 470 for (int i = 0; i < mem_inputs_length; i++) { 471 Block* inb = cfg->get_block_for_node(mem_inputs[i]); 472 if (deepb_dom_depth < (int) inb->_dom_depth) { 473 // The new inb must be dominated by the previous deepb. 474 // The various inputs must be linearly ordered in the dom 475 // tree, or else there will not be a unique deepest block. 476 DEBUG_ONLY(assert_dom(deepb, inb, load, cfg)); 477 deepb = inb; // Save deepest block 478 deepb_dom_depth = deepb->_dom_depth; 479 } 480 } 481 early = deepb; 482 } 483 484 return early; 485 } 486 487 //--------------------------insert_anti_dependences--------------------------- 488 // A load may need to witness memory that nearby stores can overwrite. 489 // For each nearby store, either insert an "anti-dependence" edge 490 // from the load to the store, or else move LCA upward to force the 491 // load to (eventually) be scheduled in a block above the store. 492 // 493 // Do not add edges to stores on distinct control-flow paths; 494 // only add edges to stores which might interfere. 495 // 496 // Return the (updated) LCA. There will not be any possibly interfering 497 // store between the load's "early block" and the updated LCA. 498 // Any stores in the updated LCA will have new precedence edges 499 // back to the load. The caller is expected to schedule the load 500 // in the LCA, in which case the precedence edges will make LCM 501 // preserve anti-dependences. The caller may also hoist the load 502 // above the LCA, if it is not the early block. 503 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) { 504 assert(load->needs_anti_dependence_check(), "must be a load of some sort"); 505 assert(LCA != NULL, ""); 506 DEBUG_ONLY(Block* LCA_orig = LCA); 507 508 // Compute the alias index. Loads and stores with different alias indices 509 // do not need anti-dependence edges. 510 int load_alias_idx = C->get_alias_index(load->adr_type()); 511 #ifdef ASSERT 512 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 && 513 (PrintOpto || VerifyAliases || 514 PrintMiscellaneous && (WizardMode || Verbose))) { 515 // Load nodes should not consume all of memory. 516 // Reporting a bottom type indicates a bug in adlc. 517 // If some particular type of node validly consumes all of memory, 518 // sharpen the preceding "if" to exclude it, so we can catch bugs here. 519 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory."); 520 load->dump(2); 521 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, ""); 522 } 523 #endif 524 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Opcodes::Op_StrComp), 525 "String compare is only known 'load' that does not conflict with any stores"); 526 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Opcodes::Op_StrEquals), 527 "String equals is a 'load' that does not conflict with any stores"); 528 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Opcodes::Op_StrIndexOf), 529 "String indexOf is a 'load' that does not conflict with any stores"); 530 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Opcodes::Op_StrIndexOfChar), 531 "String indexOfChar is a 'load' that does not conflict with any stores"); 532 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Opcodes::Op_AryEq), 533 "Arrays equals is a 'load' that does not conflict with any stores"); 534 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Opcodes::Op_HasNegatives), 535 "HasNegatives is a 'load' that does not conflict with any stores"); 536 537 if (!C->alias_type(load_alias_idx)->is_rewritable()) { 538 // It is impossible to spoil this load by putting stores before it, 539 // because we know that the stores will never update the value 540 // which 'load' must witness. 541 return LCA; 542 } 543 544 node_idx_t load_index = load->_idx; 545 546 // Note the earliest legal placement of 'load', as determined by 547 // by the unique point in the dom tree where all memory effects 548 // and other inputs are first available. (Computed by schedule_early.) 549 // For normal loads, 'early' is the shallowest place (dom graph wise) 550 // to look for anti-deps between this load and any store. 551 Block* early = get_block_for_node(load); 552 553 // If we are subsuming loads, compute an "early" block that only considers 554 // memory or address inputs. This block may be different than the 555 // schedule_early block in that it could be at an even shallower depth in the 556 // dominator tree, and allow for a broader discovery of anti-dependences. 557 if (C->subsume_loads()) { 558 early = memory_early_block(load, early, this); 559 } 560 561 ResourceArea *area = Thread::current()->resource_area(); 562 Node_List worklist_mem(area); // prior memory state to store 563 Node_List worklist_store(area); // possible-def to explore 564 Node_List worklist_visited(area); // visited mergemem nodes 565 Node_List non_early_stores(area); // all relevant stores outside of early 566 bool must_raise_LCA = false; 567 568 #ifdef TRACK_PHI_INPUTS 569 // %%% This extra checking fails because MergeMem nodes are not GVNed. 570 // Provide "phi_inputs" to check if every input to a PhiNode is from the 571 // original memory state. This indicates a PhiNode for which should not 572 // prevent the load from sinking. For such a block, set_raise_LCA_mark 573 // may be overly conservative. 574 // Mechanism: count inputs seen for each Phi encountered in worklist_store. 575 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0)); 576 #endif 577 578 // 'load' uses some memory state; look for users of the same state. 579 // Recurse through MergeMem nodes to the stores that use them. 580 581 // Each of these stores is a possible definition of memory 582 // that 'load' needs to use. We need to force 'load' 583 // to occur before each such store. When the store is in 584 // the same block as 'load', we insert an anti-dependence 585 // edge load->store. 586 587 // The relevant stores "nearby" the load consist of a tree rooted 588 // at initial_mem, with internal nodes of type MergeMem. 589 // Therefore, the branches visited by the worklist are of this form: 590 // initial_mem -> (MergeMem ->)* store 591 // The anti-dependence constraints apply only to the fringe of this tree. 592 593 Node* initial_mem = load->in(MemNode::Memory); 594 worklist_store.push(initial_mem); 595 worklist_visited.push(initial_mem); 596 worklist_mem.push(NULL); 597 while (worklist_store.size() > 0) { 598 // Examine a nearby store to see if it might interfere with our load. 599 Node* mem = worklist_mem.pop(); 600 Node* store = worklist_store.pop(); 601 Opcodes op = store->Opcode(); 602 603 // MergeMems do not directly have anti-deps. 604 // Treat them as internal nodes in a forward tree of memory states, 605 // the leaves of which are each a 'possible-def'. 606 if (store == initial_mem // root (exclusive) of tree we are searching 607 || op == Opcodes::Op_MergeMem // internal node of tree we are searching 608 ) { 609 mem = store; // It's not a possibly interfering store. 610 if (store == initial_mem) 611 initial_mem = NULL; // only process initial memory once 612 613 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 614 store = mem->fast_out(i); 615 if (store->is_MergeMem()) { 616 // Be sure we don't get into combinatorial problems. 617 // (Allow phis to be repeated; they can merge two relevant states.) 618 uint j = worklist_visited.size(); 619 for (; j > 0; j--) { 620 if (worklist_visited.at(j-1) == store) break; 621 } 622 if (j > 0) continue; // already on work list; do not repeat 623 worklist_visited.push(store); 624 } 625 worklist_mem.push(mem); 626 worklist_store.push(store); 627 } 628 continue; 629 } 630 631 if (op == Opcodes::Op_MachProj || op == Opcodes::Op_Catch) continue; 632 if (store->needs_anti_dependence_check()) continue; // not really a store 633 634 // Compute the alias index. Loads and stores with different alias 635 // indices do not need anti-dependence edges. Wide MemBar's are 636 // anti-dependent on everything (except immutable memories). 637 const TypePtr* adr_type = store->adr_type(); 638 if (!C->can_alias(adr_type, load_alias_idx)) continue; 639 640 // Most slow-path runtime calls do NOT modify Java memory, but 641 // they can block and so write Raw memory. 642 if (store->is_Mach()) { 643 MachNode* mstore = store->as_Mach(); 644 if (load_alias_idx != Compile::AliasIdxRaw) { 645 // Check for call into the runtime using the Java calling 646 // convention (and from there into a wrapper); it has no 647 // _method. Can't do this optimization for Native calls because 648 // they CAN write to Java memory. 649 if (mstore->ideal_Opcode() == Opcodes::Op_CallStaticJava) { 650 assert(mstore->is_MachSafePoint(), ""); 651 MachSafePointNode* ms = (MachSafePointNode*) mstore; 652 assert(ms->is_MachCallJava(), ""); 653 MachCallJavaNode* mcj = (MachCallJavaNode*) ms; 654 if (mcj->_method == NULL) { 655 // These runtime calls do not write to Java visible memory 656 // (other than Raw) and so do not require anti-dependence edges. 657 continue; 658 } 659 } 660 // Same for SafePoints: they read/write Raw but only read otherwise. 661 // This is basically a workaround for SafePoints only defining control 662 // instead of control + memory. 663 if (mstore->ideal_Opcode() == Opcodes::Op_SafePoint) 664 continue; 665 } else { 666 // Some raw memory, such as the load of "top" at an allocation, 667 // can be control dependent on the previous safepoint. See 668 // comments in GraphKit::allocate_heap() about control input. 669 // Inserting an anti-dep between such a safepoint and a use 670 // creates a cycle, and will cause a subsequent failure in 671 // local scheduling. (BugId 4919904) 672 // (%%% How can a control input be a safepoint and not a projection??) 673 if (mstore->ideal_Opcode() == Opcodes::Op_SafePoint && load->in(0) == mstore) 674 continue; 675 } 676 } 677 678 // Identify a block that the current load must be above, 679 // or else observe that 'store' is all the way up in the 680 // earliest legal block for 'load'. In the latter case, 681 // immediately insert an anti-dependence edge. 682 Block* store_block = get_block_for_node(store); 683 assert(store_block != NULL, "unused killing projections skipped above"); 684 685 if (store->is_Phi()) { 686 // 'load' uses memory which is one (or more) of the Phi's inputs. 687 // It must be scheduled not before the Phi, but rather before 688 // each of the relevant Phi inputs. 689 // 690 // Instead of finding the LCA of all inputs to a Phi that match 'mem', 691 // we mark each corresponding predecessor block and do a combined 692 // hoisting operation later (raise_LCA_above_marks). 693 // 694 // Do not assert(store_block != early, "Phi merging memory after access") 695 // PhiNode may be at start of block 'early' with backedge to 'early' 696 DEBUG_ONLY(bool found_match = false); 697 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) { 698 if (store->in(j) == mem) { // Found matching input? 699 DEBUG_ONLY(found_match = true); 700 Block* pred_block = get_block_for_node(store_block->pred(j)); 701 if (pred_block != early) { 702 // If any predecessor of the Phi matches the load's "early block", 703 // we do not need a precedence edge between the Phi and 'load' 704 // since the load will be forced into a block preceding the Phi. 705 pred_block->set_raise_LCA_mark(load_index); 706 assert(!LCA_orig->dominates(pred_block) || 707 early->dominates(pred_block), "early is high enough"); 708 must_raise_LCA = true; 709 } else { 710 // anti-dependent upon PHI pinned below 'early', no edge needed 711 LCA = early; // but can not schedule below 'early' 712 } 713 } 714 } 715 assert(found_match, "no worklist bug"); 716 #ifdef TRACK_PHI_INPUTS 717 #ifdef ASSERT 718 // This assert asks about correct handling of PhiNodes, which may not 719 // have all input edges directly from 'mem'. See BugId 4621264 720 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1; 721 // Increment by exactly one even if there are multiple copies of 'mem' 722 // coming into the phi, because we will run this block several times 723 // if there are several copies of 'mem'. (That's how DU iterators work.) 724 phi_inputs.at_put(store->_idx, num_mem_inputs); 725 assert(PhiNode::Input + num_mem_inputs < store->req(), 726 "Expect at least one phi input will not be from original memory state"); 727 #endif //ASSERT 728 #endif //TRACK_PHI_INPUTS 729 } else if (store_block != early) { 730 // 'store' is between the current LCA and earliest possible block. 731 // Label its block, and decide later on how to raise the LCA 732 // to include the effect on LCA of this store. 733 // If this store's block gets chosen as the raised LCA, we 734 // will find him on the non_early_stores list and stick him 735 // with a precedence edge. 736 // (But, don't bother if LCA is already raised all the way.) 737 if (LCA != early) { 738 store_block->set_raise_LCA_mark(load_index); 739 must_raise_LCA = true; 740 non_early_stores.push(store); 741 } 742 } else { 743 // Found a possibly-interfering store in the load's 'early' block. 744 // This means 'load' cannot sink at all in the dominator tree. 745 // Add an anti-dep edge, and squeeze 'load' into the highest block. 746 assert(store != load->in(0), "dependence cycle found"); 747 if (verify) { 748 assert(store->find_edge(load) != -1, "missing precedence edge"); 749 } else { 750 store->add_prec(load); 751 } 752 LCA = early; 753 // This turns off the process of gathering non_early_stores. 754 } 755 } 756 // (Worklist is now empty; all nearby stores have been visited.) 757 758 // Finished if 'load' must be scheduled in its 'early' block. 759 // If we found any stores there, they have already been given 760 // precedence edges. 761 if (LCA == early) return LCA; 762 763 // We get here only if there are no possibly-interfering stores 764 // in the load's 'early' block. Move LCA up above all predecessors 765 // which contain stores we have noted. 766 // 767 // The raised LCA block can be a home to such interfering stores, 768 // but its predecessors must not contain any such stores. 769 // 770 // The raised LCA will be a lower bound for placing the load, 771 // preventing the load from sinking past any block containing 772 // a store that may invalidate the memory state required by 'load'. 773 if (must_raise_LCA) 774 LCA = raise_LCA_above_marks(LCA, load->_idx, early, this); 775 if (LCA == early) return LCA; 776 777 // Insert anti-dependence edges from 'load' to each store 778 // in the non-early LCA block. 779 // Mine the non_early_stores list for such stores. 780 if (LCA->raise_LCA_mark() == load_index) { 781 while (non_early_stores.size() > 0) { 782 Node* store = non_early_stores.pop(); 783 Block* store_block = get_block_for_node(store); 784 if (store_block == LCA) { 785 // add anti_dependence from store to load in its own block 786 assert(store != load->in(0), "dependence cycle found"); 787 if (verify) { 788 assert(store->find_edge(load) != -1, "missing precedence edge"); 789 } else { 790 store->add_prec(load); 791 } 792 } else { 793 assert(store_block->raise_LCA_mark() == load_index, "block was marked"); 794 // Any other stores we found must be either inside the new LCA 795 // or else outside the original LCA. In the latter case, they 796 // did not interfere with any use of 'load'. 797 assert(LCA->dominates(store_block) 798 || !LCA_orig->dominates(store_block), "no stray stores"); 799 } 800 } 801 } 802 803 // Return the highest block containing stores; any stores 804 // within that block have been given anti-dependence edges. 805 return LCA; 806 } 807 808 // This class is used to iterate backwards over the nodes in the graph. 809 810 class Node_Backward_Iterator { 811 812 private: 813 Node_Backward_Iterator(); 814 815 public: 816 // Constructor for the iterator 817 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg); 818 819 // Postincrement operator to iterate over the nodes 820 Node *next(); 821 822 private: 823 VectorSet &_visited; 824 Node_Stack &_stack; 825 PhaseCFG &_cfg; 826 }; 827 828 // Constructor for the Node_Backward_Iterator 829 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_Stack &stack, PhaseCFG &cfg) 830 : _visited(visited), _stack(stack), _cfg(cfg) { 831 // The stack should contain exactly the root 832 stack.clear(); 833 stack.push(root, root->outcnt()); 834 835 // Clear the visited bits 836 visited.Clear(); 837 } 838 839 // Iterator for the Node_Backward_Iterator 840 Node *Node_Backward_Iterator::next() { 841 842 // If the _stack is empty, then just return NULL: finished. 843 if ( !_stack.size() ) 844 return NULL; 845 846 // I visit unvisited not-anti-dependence users first, then anti-dependent 847 // children next. I iterate backwards to support removal of nodes. 848 // The stack holds states consisting of 3 values: 849 // current Def node, flag which indicates 1st/2nd pass, index of current out edge 850 Node *self = (Node*)(((uintptr_t)_stack.node()) & ~1); 851 bool iterate_anti_dep = (((uintptr_t)_stack.node()) & 1); 852 uint idx = MIN2(_stack.index(), self->outcnt()); // Support removal of nodes. 853 _stack.pop(); 854 855 // I cycle here when I am entering a deeper level of recursion. 856 // The key variable 'self' was set prior to jumping here. 857 while( 1 ) { 858 859 _visited.set(self->_idx); 860 861 // Now schedule all uses as late as possible. 862 const Node* src = self->is_Proj() ? self->in(0) : self; 863 uint src_rpo = _cfg.get_block_for_node(src)->_rpo; 864 865 // Schedule all nodes in a post-order visit 866 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any 867 868 // Scan for unvisited nodes 869 while (idx > 0) { 870 // For all uses, schedule late 871 Node* n = self->raw_out(--idx); // Use 872 873 // Skip already visited children 874 if ( _visited.test(n->_idx) ) 875 continue; 876 877 // do not traverse backward control edges 878 Node *use = n->is_Proj() ? n->in(0) : n; 879 uint use_rpo = _cfg.get_block_for_node(use)->_rpo; 880 881 if ( use_rpo < src_rpo ) 882 continue; 883 884 // Phi nodes always precede uses in a basic block 885 if ( use_rpo == src_rpo && use->is_Phi() ) 886 continue; 887 888 unvisited = n; // Found unvisited 889 890 // Check for possible-anti-dependent 891 // 1st pass: No such nodes, 2nd pass: Only such nodes. 892 if (n->needs_anti_dependence_check() == iterate_anti_dep) { 893 unvisited = n; // Found unvisited 894 break; 895 } 896 } 897 898 // Did I find an unvisited not-anti-dependent Node? 899 if (!unvisited) { 900 if (!iterate_anti_dep) { 901 // 2nd pass: Iterate over nodes which needs_anti_dependence_check. 902 iterate_anti_dep = true; 903 idx = self->outcnt(); 904 continue; 905 } 906 break; // All done with children; post-visit 'self' 907 } 908 909 // Visit the unvisited Node. Contains the obvious push to 910 // indicate I'm entering a deeper level of recursion. I push the 911 // old state onto the _stack and set a new state and loop (recurse). 912 _stack.push((Node*)((uintptr_t)self | (uintptr_t)iterate_anti_dep), idx); 913 self = unvisited; 914 iterate_anti_dep = false; 915 idx = self->outcnt(); 916 } // End recursion loop 917 918 return self; 919 } 920 921 //------------------------------ComputeLatenciesBackwards---------------------- 922 // Compute the latency of all the instructions. 923 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_Stack &stack) { 924 #ifndef PRODUCT 925 if (trace_opto_pipelining()) 926 tty->print("\n#---- ComputeLatenciesBackwards ----\n"); 927 #endif 928 929 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this); 930 Node *n; 931 932 // Walk over all the nodes from last to first 933 while (n = iter.next()) { 934 // Set the latency for the definitions of this instruction 935 partial_latency_of_defs(n); 936 } 937 } // end ComputeLatenciesBackwards 938 939 //------------------------------partial_latency_of_defs------------------------ 940 // Compute the latency impact of this node on all defs. This computes 941 // a number that increases as we approach the beginning of the routine. 942 void PhaseCFG::partial_latency_of_defs(Node *n) { 943 // Set the latency for this instruction 944 #ifndef PRODUCT 945 if (trace_opto_pipelining()) { 946 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n)); 947 dump(); 948 } 949 #endif 950 951 if (n->is_Proj()) { 952 n = n->in(0); 953 } 954 955 if (n->is_Root()) { 956 return; 957 } 958 959 uint nlen = n->len(); 960 uint use_latency = get_latency_for_node(n); 961 uint use_pre_order = get_block_for_node(n)->_pre_order; 962 963 for (uint j = 0; j < nlen; j++) { 964 Node *def = n->in(j); 965 966 if (!def || def == n) { 967 continue; 968 } 969 970 // Walk backwards thru projections 971 if (def->is_Proj()) { 972 def = def->in(0); 973 } 974 975 #ifndef PRODUCT 976 if (trace_opto_pipelining()) { 977 tty->print("# in(%2d): ", j); 978 def->dump(); 979 } 980 #endif 981 982 // If the defining block is not known, assume it is ok 983 Block *def_block = get_block_for_node(def); 984 uint def_pre_order = def_block ? def_block->_pre_order : 0; 985 986 if ((use_pre_order < def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) { 987 continue; 988 } 989 990 uint delta_latency = n->latency(j); 991 uint current_latency = delta_latency + use_latency; 992 993 if (get_latency_for_node(def) < current_latency) { 994 set_latency_for_node(def, current_latency); 995 } 996 997 #ifndef PRODUCT 998 if (trace_opto_pipelining()) { 999 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, node_latency[%d] = %d", use_latency, j, delta_latency, current_latency, def->_idx, get_latency_for_node(def)); 1000 } 1001 #endif 1002 } 1003 } 1004 1005 //------------------------------latency_from_use------------------------------- 1006 // Compute the latency of a specific use 1007 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) { 1008 // If self-reference, return no latency 1009 if (use == n || use->is_Root()) { 1010 return 0; 1011 } 1012 1013 uint def_pre_order = get_block_for_node(def)->_pre_order; 1014 uint latency = 0; 1015 1016 // If the use is not a projection, then it is simple... 1017 if (!use->is_Proj()) { 1018 #ifndef PRODUCT 1019 if (trace_opto_pipelining()) { 1020 tty->print("# out(): "); 1021 use->dump(); 1022 } 1023 #endif 1024 1025 uint use_pre_order = get_block_for_node(use)->_pre_order; 1026 1027 if (use_pre_order < def_pre_order) 1028 return 0; 1029 1030 if (use_pre_order == def_pre_order && use->is_Phi()) 1031 return 0; 1032 1033 uint nlen = use->len(); 1034 uint nl = get_latency_for_node(use); 1035 1036 for ( uint j=0; j<nlen; j++ ) { 1037 if (use->in(j) == n) { 1038 // Change this if we want local latencies 1039 uint ul = use->latency(j); 1040 uint l = ul + nl; 1041 if (latency < l) latency = l; 1042 #ifndef PRODUCT 1043 if (trace_opto_pipelining()) { 1044 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d", 1045 nl, j, ul, l, latency); 1046 } 1047 #endif 1048 } 1049 } 1050 } else { 1051 // This is a projection, just grab the latency of the use(s) 1052 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) { 1053 uint l = latency_from_use(use, def, use->fast_out(j)); 1054 if (latency < l) latency = l; 1055 } 1056 } 1057 1058 return latency; 1059 } 1060 1061 //------------------------------latency_from_uses------------------------------ 1062 // Compute the latency of this instruction relative to all of it's uses. 1063 // This computes a number that increases as we approach the beginning of the 1064 // routine. 1065 void PhaseCFG::latency_from_uses(Node *n) { 1066 // Set the latency for this instruction 1067 #ifndef PRODUCT 1068 if (trace_opto_pipelining()) { 1069 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n)); 1070 dump(); 1071 } 1072 #endif 1073 uint latency=0; 1074 const Node *def = n->is_Proj() ? n->in(0): n; 1075 1076 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1077 uint l = latency_from_use(n, def, n->fast_out(i)); 1078 1079 if (latency < l) latency = l; 1080 } 1081 1082 set_latency_for_node(n, latency); 1083 } 1084 1085 //------------------------------hoist_to_cheaper_block------------------------- 1086 // Pick a block for node self, between early and LCA, that is a cheaper 1087 // alternative to LCA. 1088 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) { 1089 const double delta = 1+PROB_UNLIKELY_MAG(4); 1090 Block* least = LCA; 1091 double least_freq = least->_freq; 1092 uint target = get_latency_for_node(self); 1093 uint start_latency = get_latency_for_node(LCA->head()); 1094 uint end_latency = get_latency_for_node(LCA->get_node(LCA->end_idx())); 1095 bool in_latency = (target <= start_latency); 1096 const Block* root_block = get_block_for_node(_root); 1097 1098 // Turn off latency scheduling if scheduling is just plain off 1099 if (!C->do_scheduling()) 1100 in_latency = true; 1101 1102 // Do not hoist (to cover latency) instructions which target a 1103 // single register. Hoisting stretches the live range of the 1104 // single register and may force spilling. 1105 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL; 1106 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty()) 1107 in_latency = true; 1108 1109 #ifndef PRODUCT 1110 if (trace_opto_pipelining()) { 1111 tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self)); 1112 self->dump(); 1113 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g", 1114 LCA->_pre_order, 1115 LCA->head()->_idx, 1116 start_latency, 1117 LCA->get_node(LCA->end_idx())->_idx, 1118 end_latency, 1119 least_freq); 1120 } 1121 #endif 1122 1123 int cand_cnt = 0; // number of candidates tried 1124 1125 // Walk up the dominator tree from LCA (Lowest common ancestor) to 1126 // the earliest legal location. Capture the least execution frequency. 1127 while (LCA != early) { 1128 LCA = LCA->_idom; // Follow up the dominator tree 1129 1130 if (LCA == NULL) { 1131 // Bailout without retry 1132 C->record_method_not_compilable("late schedule failed: LCA == NULL"); 1133 return least; 1134 } 1135 1136 // Don't hoist machine instructions to the root basic block 1137 if (mach && LCA == root_block) 1138 break; 1139 1140 uint start_lat = get_latency_for_node(LCA->head()); 1141 uint end_idx = LCA->end_idx(); 1142 uint end_lat = get_latency_for_node(LCA->get_node(end_idx)); 1143 double LCA_freq = LCA->_freq; 1144 #ifndef PRODUCT 1145 if (trace_opto_pipelining()) { 1146 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g", 1147 LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq); 1148 } 1149 #endif 1150 cand_cnt++; 1151 if (LCA_freq < least_freq || // Better Frequency 1152 (StressGCM && Compile::randomized_select(cand_cnt)) || // Should be randomly accepted in stress mode 1153 (!StressGCM && // Otherwise, choose with latency 1154 !in_latency && // No block containing latency 1155 LCA_freq < least_freq * delta && // No worse frequency 1156 target >= end_lat && // within latency range 1157 !self->is_iteratively_computed() ) // But don't hoist IV increments 1158 // because they may end up above other uses of their phi forcing 1159 // their result register to be different from their input. 1160 ) { 1161 least = LCA; // Found cheaper block 1162 least_freq = LCA_freq; 1163 start_latency = start_lat; 1164 end_latency = end_lat; 1165 if (target <= start_lat) 1166 in_latency = true; 1167 } 1168 } 1169 1170 #ifndef PRODUCT 1171 if (trace_opto_pipelining()) { 1172 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g", 1173 least->_pre_order, start_latency, least_freq); 1174 } 1175 #endif 1176 1177 // See if the latency needs to be updated 1178 if (target < end_latency) { 1179 #ifndef PRODUCT 1180 if (trace_opto_pipelining()) { 1181 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency); 1182 } 1183 #endif 1184 set_latency_for_node(self, end_latency); 1185 partial_latency_of_defs(self); 1186 } 1187 1188 return least; 1189 } 1190 1191 1192 //------------------------------schedule_late----------------------------------- 1193 // Now schedule all codes as LATE as possible. This is the LCA in the 1194 // dominator tree of all USES of a value. Pick the block with the least 1195 // loop nesting depth that is lowest in the dominator tree. 1196 extern const char must_clone[]; 1197 void PhaseCFG::schedule_late(VectorSet &visited, Node_Stack &stack) { 1198 #ifndef PRODUCT 1199 if (trace_opto_pipelining()) 1200 tty->print("\n#---- schedule_late ----\n"); 1201 #endif 1202 1203 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this); 1204 Node *self; 1205 1206 // Walk over all the nodes from last to first 1207 while (self = iter.next()) { 1208 Block* early = get_block_for_node(self); // Earliest legal placement 1209 1210 if (self->is_top()) { 1211 // Top node goes in bb #2 with other constants. 1212 // It must be special-cased, because it has no out edges. 1213 early->add_inst(self); 1214 continue; 1215 } 1216 1217 // No uses, just terminate 1218 if (self->outcnt() == 0) { 1219 assert(self->is_MachProj(), "sanity"); 1220 continue; // Must be a dead machine projection 1221 } 1222 1223 // If node is pinned in the block, then no scheduling can be done. 1224 if( self->pinned() ) // Pinned in block? 1225 continue; 1226 1227 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL; 1228 if (mach) { 1229 switch (mach->ideal_Opcode()) { 1230 case Opcodes::Op_CreateEx: 1231 // Don't move exception creation 1232 early->add_inst(self); 1233 continue; 1234 break; 1235 case Opcodes::Op_CheckCastPP: 1236 // Don't move CheckCastPP nodes away from their input, if the input 1237 // is a rawptr (5071820). 1238 Node *def = self->in(1); 1239 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) { 1240 early->add_inst(self); 1241 #ifdef ASSERT 1242 _raw_oops.push(def); 1243 #endif 1244 continue; 1245 } 1246 break; 1247 } 1248 } 1249 1250 // Gather LCA of all uses 1251 Block *LCA = NULL; 1252 { 1253 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) { 1254 // For all uses, find LCA 1255 Node* use = self->fast_out(i); 1256 LCA = raise_LCA_above_use(LCA, use, self, this); 1257 } 1258 } // (Hide defs of imax, i from rest of block.) 1259 1260 // Place temps in the block of their use. This isn't a 1261 // requirement for correctness but it reduces useless 1262 // interference between temps and other nodes. 1263 if (mach != NULL && mach->is_MachTemp()) { 1264 map_node_to_block(self, LCA); 1265 LCA->add_inst(self); 1266 continue; 1267 } 1268 1269 // Check if 'self' could be anti-dependent on memory 1270 if (self->needs_anti_dependence_check()) { 1271 // Hoist LCA above possible-defs and insert anti-dependences to 1272 // defs in new LCA block. 1273 LCA = insert_anti_dependences(LCA, self); 1274 } 1275 1276 if (early->_dom_depth > LCA->_dom_depth) { 1277 // Somehow the LCA has moved above the earliest legal point. 1278 // (One way this can happen is via memory_early_block.) 1279 if (C->subsume_loads() == true && !C->failing()) { 1280 // Retry with subsume_loads == false 1281 // If this is the first failure, the sentinel string will "stick" 1282 // to the Compile object, and the C2Compiler will see it and retry. 1283 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 1284 } else { 1285 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth) 1286 C->record_method_not_compilable("late schedule failed: incorrect graph"); 1287 } 1288 return; 1289 } 1290 1291 // If there is no opportunity to hoist, then we're done. 1292 // In stress mode, try to hoist even the single operations. 1293 bool try_to_hoist = StressGCM || (LCA != early); 1294 1295 // Must clone guys stay next to use; no hoisting allowed. 1296 // Also cannot hoist guys that alter memory or are otherwise not 1297 // allocatable (hoisting can make a value live longer, leading to 1298 // anti and output dependency problems which are normally resolved 1299 // by the register allocator giving everyone a different register). 1300 if (mach != NULL && must_clone[static_cast<uint>(mach->ideal_Opcode())]) 1301 try_to_hoist = false; 1302 1303 Block* late = NULL; 1304 if (try_to_hoist) { 1305 // Now find the block with the least execution frequency. 1306 // Start at the latest schedule and work up to the earliest schedule 1307 // in the dominator tree. Thus the Node will dominate all its uses. 1308 late = hoist_to_cheaper_block(LCA, early, self); 1309 } else { 1310 // Just use the LCA of the uses. 1311 late = LCA; 1312 } 1313 1314 // Put the node into target block 1315 schedule_node_into_block(self, late); 1316 1317 #ifdef ASSERT 1318 if (self->needs_anti_dependence_check()) { 1319 // since precedence edges are only inserted when we're sure they 1320 // are needed make sure that after placement in a block we don't 1321 // need any new precedence edges. 1322 verify_anti_dependences(late, self); 1323 } 1324 #endif 1325 } // Loop until all nodes have been visited 1326 1327 } // end ScheduleLate 1328 1329 //------------------------------GlobalCodeMotion------------------------------- 1330 void PhaseCFG::global_code_motion() { 1331 ResourceMark rm; 1332 1333 #ifndef PRODUCT 1334 if (trace_opto_pipelining()) { 1335 tty->print("\n---- Start GlobalCodeMotion ----\n"); 1336 } 1337 #endif 1338 1339 // Initialize the node to block mapping for things on the proj_list 1340 for (uint i = 0; i < _matcher.number_of_projections(); i++) { 1341 unmap_node_from_block(_matcher.get_projection(i)); 1342 } 1343 1344 // Set the basic block for Nodes pinned into blocks 1345 Arena* arena = Thread::current()->resource_area(); 1346 VectorSet visited(arena); 1347 schedule_pinned_nodes(visited); 1348 1349 // Find the earliest Block any instruction can be placed in. Some 1350 // instructions are pinned into Blocks. Unpinned instructions can 1351 // appear in last block in which all their inputs occur. 1352 visited.Clear(); 1353 Node_Stack stack(arena, (C->live_nodes() >> 2) + 16); // pre-grow 1354 if (!schedule_early(visited, stack)) { 1355 // Bailout without retry 1356 C->record_method_not_compilable("early schedule failed"); 1357 return; 1358 } 1359 1360 // Build Def-Use edges. 1361 // Compute the latency information (via backwards walk) for all the 1362 // instructions in the graph 1363 _node_latency = new GrowableArray<uint>(); // resource_area allocation 1364 1365 if (C->do_scheduling()) { 1366 compute_latencies_backwards(visited, stack); 1367 } 1368 1369 // Now schedule all codes as LATE as possible. This is the LCA in the 1370 // dominator tree of all USES of a value. Pick the block with the least 1371 // loop nesting depth that is lowest in the dominator tree. 1372 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() ) 1373 schedule_late(visited, stack); 1374 if (C->failing()) { 1375 // schedule_late fails only when graph is incorrect. 1376 assert(!VerifyGraphEdges, "verification should have failed"); 1377 return; 1378 } 1379 1380 #ifndef PRODUCT 1381 if (trace_opto_pipelining()) { 1382 tty->print("\n---- Detect implicit null checks ----\n"); 1383 } 1384 #endif 1385 1386 // Detect implicit-null-check opportunities. Basically, find NULL checks 1387 // with suitable memory ops nearby. Use the memory op to do the NULL check. 1388 // I can generate a memory op if there is not one nearby. 1389 if (C->is_method_compilation()) { 1390 // By reversing the loop direction we get a very minor gain on mpegaudio. 1391 // Feel free to revert to a forward loop for clarity. 1392 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) { 1393 for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) { 1394 Node* proj = _matcher._null_check_tests[i]; 1395 Node* val = _matcher._null_check_tests[i + 1]; 1396 Block* block = get_block_for_node(proj); 1397 implicit_null_check(block, proj, val, C->allowed_deopt_reasons()); 1398 // The implicit_null_check will only perform the transformation 1399 // if the null branch is truly uncommon, *and* it leads to an 1400 // uncommon trap. Combined with the too_many_traps guards 1401 // above, this prevents SEGV storms reported in 6366351, 1402 // by recompiling offending methods without this optimization. 1403 } 1404 } 1405 1406 bool block_size_threshold_ok = false; 1407 intptr_t *recalc_pressure_nodes = NULL; 1408 if (OptoRegScheduling) { 1409 for (uint i = 0; i < number_of_blocks(); i++) { 1410 Block* block = get_block(i); 1411 if (block->number_of_nodes() > 10) { 1412 block_size_threshold_ok = true; 1413 break; 1414 } 1415 } 1416 } 1417 1418 // Enabling the scheduler for register pressure plus finding blocks of size to schedule for it 1419 // is key to enabling this feature. 1420 PhaseChaitin regalloc(C->unique(), *this, _matcher, true); 1421 ResourceArea live_arena; // Arena for liveness 1422 ResourceMark rm_live(&live_arena); 1423 PhaseLive live(*this, regalloc._lrg_map.names(), &live_arena, true); 1424 PhaseIFG ifg(&live_arena); 1425 if (OptoRegScheduling && block_size_threshold_ok) { 1426 regalloc.mark_ssa(); 1427 Compile::TracePhase tp("computeLive", &timers[_t_computeLive]); 1428 rm_live.reset_to_mark(); // Reclaim working storage 1429 IndexSet::reset_memory(C, &live_arena); 1430 uint node_size = regalloc._lrg_map.max_lrg_id(); 1431 ifg.init(node_size); // Empty IFG 1432 regalloc.set_ifg(ifg); 1433 regalloc.set_live(live); 1434 regalloc.gather_lrg_masks(false); // Collect LRG masks 1435 live.compute(node_size); // Compute liveness 1436 1437 recalc_pressure_nodes = NEW_RESOURCE_ARRAY(intptr_t, node_size); 1438 for (uint i = 0; i < node_size; i++) { 1439 recalc_pressure_nodes[i] = 0; 1440 } 1441 } 1442 _regalloc = ®alloc; 1443 1444 #ifndef PRODUCT 1445 if (trace_opto_pipelining()) { 1446 tty->print("\n---- Start Local Scheduling ----\n"); 1447 } 1448 #endif 1449 1450 // Schedule locally. Right now a simple topological sort. 1451 // Later, do a real latency aware scheduler. 1452 GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1); 1453 visited.Clear(); 1454 for (uint i = 0; i < number_of_blocks(); i++) { 1455 Block* block = get_block(i); 1456 if (!schedule_local(block, ready_cnt, visited, recalc_pressure_nodes)) { 1457 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 1458 C->record_method_not_compilable("local schedule failed"); 1459 } 1460 _regalloc = NULL; 1461 return; 1462 } 1463 } 1464 _regalloc = NULL; 1465 1466 // If we inserted any instructions between a Call and his CatchNode, 1467 // clone the instructions on all paths below the Catch. 1468 for (uint i = 0; i < number_of_blocks(); i++) { 1469 Block* block = get_block(i); 1470 call_catch_cleanup(block); 1471 } 1472 1473 #ifndef PRODUCT 1474 if (trace_opto_pipelining()) { 1475 tty->print("\n---- After GlobalCodeMotion ----\n"); 1476 for (uint i = 0; i < number_of_blocks(); i++) { 1477 Block* block = get_block(i); 1478 block->dump(); 1479 } 1480 } 1481 #endif 1482 // Dead. 1483 _node_latency = (GrowableArray<uint> *)0xdeadbeef; 1484 } 1485 1486 bool PhaseCFG::do_global_code_motion() { 1487 1488 build_dominator_tree(); 1489 if (C->failing()) { 1490 return false; 1491 } 1492 1493 NOT_PRODUCT( C->verify_graph_edges(); ) 1494 1495 estimate_block_frequency(); 1496 1497 global_code_motion(); 1498 1499 if (C->failing()) { 1500 return false; 1501 } 1502 1503 return true; 1504 } 1505 1506 //------------------------------Estimate_Block_Frequency----------------------- 1507 // Estimate block frequencies based on IfNode probabilities. 1508 void PhaseCFG::estimate_block_frequency() { 1509 1510 // Force conditional branches leading to uncommon traps to be unlikely, 1511 // not because we get to the uncommon_trap with less relative frequency, 1512 // but because an uncommon_trap typically causes a deopt, so we only get 1513 // there once. 1514 if (C->do_freq_based_layout()) { 1515 Block_List worklist; 1516 Block* root_blk = get_block(0); 1517 for (uint i = 1; i < root_blk->num_preds(); i++) { 1518 Block *pb = get_block_for_node(root_blk->pred(i)); 1519 if (pb->has_uncommon_code()) { 1520 worklist.push(pb); 1521 } 1522 } 1523 while (worklist.size() > 0) { 1524 Block* uct = worklist.pop(); 1525 if (uct == get_root_block()) { 1526 continue; 1527 } 1528 for (uint i = 1; i < uct->num_preds(); i++) { 1529 Block *pb = get_block_for_node(uct->pred(i)); 1530 if (pb->_num_succs == 1) { 1531 worklist.push(pb); 1532 } else if (pb->num_fall_throughs() == 2) { 1533 pb->update_uncommon_branch(uct); 1534 } 1535 } 1536 } 1537 } 1538 1539 // Create the loop tree and calculate loop depth. 1540 _root_loop = create_loop_tree(); 1541 _root_loop->compute_loop_depth(0); 1542 1543 // Compute block frequency of each block, relative to a single loop entry. 1544 _root_loop->compute_freq(); 1545 1546 // Adjust all frequencies to be relative to a single method entry 1547 _root_loop->_freq = 1.0; 1548 _root_loop->scale_freq(); 1549 1550 // Save outmost loop frequency for LRG frequency threshold 1551 _outer_loop_frequency = _root_loop->outer_loop_freq(); 1552 1553 // force paths ending at uncommon traps to be infrequent 1554 if (!C->do_freq_based_layout()) { 1555 Block_List worklist; 1556 Block* root_blk = get_block(0); 1557 for (uint i = 1; i < root_blk->num_preds(); i++) { 1558 Block *pb = get_block_for_node(root_blk->pred(i)); 1559 if (pb->has_uncommon_code()) { 1560 worklist.push(pb); 1561 } 1562 } 1563 while (worklist.size() > 0) { 1564 Block* uct = worklist.pop(); 1565 uct->_freq = PROB_MIN; 1566 for (uint i = 1; i < uct->num_preds(); i++) { 1567 Block *pb = get_block_for_node(uct->pred(i)); 1568 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) { 1569 worklist.push(pb); 1570 } 1571 } 1572 } 1573 } 1574 1575 #ifdef ASSERT 1576 for (uint i = 0; i < number_of_blocks(); i++) { 1577 Block* b = get_block(i); 1578 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency"); 1579 } 1580 #endif 1581 1582 #ifndef PRODUCT 1583 if (PrintCFGBlockFreq) { 1584 tty->print_cr("CFG Block Frequencies"); 1585 _root_loop->dump_tree(); 1586 if (Verbose) { 1587 tty->print_cr("PhaseCFG dump"); 1588 dump(); 1589 tty->print_cr("Node dump"); 1590 _root->dump(99999); 1591 } 1592 } 1593 #endif 1594 } 1595 1596 //----------------------------create_loop_tree-------------------------------- 1597 // Create a loop tree from the CFG 1598 CFGLoop* PhaseCFG::create_loop_tree() { 1599 1600 #ifdef ASSERT 1601 assert(get_block(0) == get_root_block(), "first block should be root block"); 1602 for (uint i = 0; i < number_of_blocks(); i++) { 1603 Block* block = get_block(i); 1604 // Check that _loop field are clear...we could clear them if not. 1605 assert(block->_loop == NULL, "clear _loop expected"); 1606 // Sanity check that the RPO numbering is reflected in the _blocks array. 1607 // It doesn't have to be for the loop tree to be built, but if it is not, 1608 // then the blocks have been reordered since dom graph building...which 1609 // may question the RPO numbering 1610 assert(block->_rpo == i, "unexpected reverse post order number"); 1611 } 1612 #endif 1613 1614 int idct = 0; 1615 CFGLoop* root_loop = new CFGLoop(idct++); 1616 1617 Block_List worklist; 1618 1619 // Assign blocks to loops 1620 for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block 1621 Block* block = get_block(i); 1622 1623 if (block->head()->is_Loop()) { 1624 Block* loop_head = block; 1625 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors"); 1626 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl); 1627 Block* tail = get_block_for_node(tail_n); 1628 1629 // Defensively filter out Loop nodes for non-single-entry loops. 1630 // For all reasonable loops, the head occurs before the tail in RPO. 1631 if (i <= tail->_rpo) { 1632 1633 // The tail and (recursive) predecessors of the tail 1634 // are made members of a new loop. 1635 1636 assert(worklist.size() == 0, "nonempty worklist"); 1637 CFGLoop* nloop = new CFGLoop(idct++); 1638 assert(loop_head->_loop == NULL, "just checking"); 1639 loop_head->_loop = nloop; 1640 // Add to nloop so push_pred() will skip over inner loops 1641 nloop->add_member(loop_head); 1642 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this); 1643 1644 while (worklist.size() > 0) { 1645 Block* member = worklist.pop(); 1646 if (member != loop_head) { 1647 for (uint j = 1; j < member->num_preds(); j++) { 1648 nloop->push_pred(member, j, worklist, this); 1649 } 1650 } 1651 } 1652 } 1653 } 1654 } 1655 1656 // Create a member list for each loop consisting 1657 // of both blocks and (immediate child) loops. 1658 for (uint i = 0; i < number_of_blocks(); i++) { 1659 Block* block = get_block(i); 1660 CFGLoop* lp = block->_loop; 1661 if (lp == NULL) { 1662 // Not assigned to a loop. Add it to the method's pseudo loop. 1663 block->_loop = root_loop; 1664 lp = root_loop; 1665 } 1666 if (lp == root_loop || block != lp->head()) { // loop heads are already members 1667 lp->add_member(block); 1668 } 1669 if (lp != root_loop) { 1670 if (lp->parent() == NULL) { 1671 // Not a nested loop. Make it a child of the method's pseudo loop. 1672 root_loop->add_nested_loop(lp); 1673 } 1674 if (block == lp->head()) { 1675 // Add nested loop to member list of parent loop. 1676 lp->parent()->add_member(lp); 1677 } 1678 } 1679 } 1680 1681 return root_loop; 1682 } 1683 1684 //------------------------------push_pred-------------------------------------- 1685 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) { 1686 Node* pred_n = blk->pred(i); 1687 Block* pred = cfg->get_block_for_node(pred_n); 1688 CFGLoop *pred_loop = pred->_loop; 1689 if (pred_loop == NULL) { 1690 // Filter out blocks for non-single-entry loops. 1691 // For all reasonable loops, the head occurs before the tail in RPO. 1692 if (pred->_rpo > head()->_rpo) { 1693 pred->_loop = this; 1694 worklist.push(pred); 1695 } 1696 } else if (pred_loop != this) { 1697 // Nested loop. 1698 while (pred_loop->_parent != NULL && pred_loop->_parent != this) { 1699 pred_loop = pred_loop->_parent; 1700 } 1701 // Make pred's loop be a child 1702 if (pred_loop->_parent == NULL) { 1703 add_nested_loop(pred_loop); 1704 // Continue with loop entry predecessor. 1705 Block* pred_head = pred_loop->head(); 1706 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors"); 1707 assert(pred_head != head(), "loop head in only one loop"); 1708 push_pred(pred_head, LoopNode::EntryControl, worklist, cfg); 1709 } else { 1710 assert(pred_loop->_parent == this && _parent == NULL, "just checking"); 1711 } 1712 } 1713 } 1714 1715 //------------------------------add_nested_loop-------------------------------- 1716 // Make cl a child of the current loop in the loop tree. 1717 void CFGLoop::add_nested_loop(CFGLoop* cl) { 1718 assert(_parent == NULL, "no parent yet"); 1719 assert(cl != this, "not my own parent"); 1720 cl->_parent = this; 1721 CFGLoop* ch = _child; 1722 if (ch == NULL) { 1723 _child = cl; 1724 } else { 1725 while (ch->_sibling != NULL) { ch = ch->_sibling; } 1726 ch->_sibling = cl; 1727 } 1728 } 1729 1730 //------------------------------compute_loop_depth----------------------------- 1731 // Store the loop depth in each CFGLoop object. 1732 // Recursively walk the children to do the same for them. 1733 void CFGLoop::compute_loop_depth(int depth) { 1734 _depth = depth; 1735 CFGLoop* ch = _child; 1736 while (ch != NULL) { 1737 ch->compute_loop_depth(depth + 1); 1738 ch = ch->_sibling; 1739 } 1740 } 1741 1742 //------------------------------compute_freq----------------------------------- 1743 // Compute the frequency of each block and loop, relative to a single entry 1744 // into the dominating loop head. 1745 void CFGLoop::compute_freq() { 1746 // Bottom up traversal of loop tree (visit inner loops first.) 1747 // Set loop head frequency to 1.0, then transitively 1748 // compute frequency for all successors in the loop, 1749 // as well as for each exit edge. Inner loops are 1750 // treated as single blocks with loop exit targets 1751 // as the successor blocks. 1752 1753 // Nested loops first 1754 CFGLoop* ch = _child; 1755 while (ch != NULL) { 1756 ch->compute_freq(); 1757 ch = ch->_sibling; 1758 } 1759 assert (_members.length() > 0, "no empty loops"); 1760 Block* hd = head(); 1761 hd->_freq = 1.0; 1762 for (int i = 0; i < _members.length(); i++) { 1763 CFGElement* s = _members.at(i); 1764 double freq = s->_freq; 1765 if (s->is_block()) { 1766 Block* b = s->as_Block(); 1767 for (uint j = 0; j < b->_num_succs; j++) { 1768 Block* sb = b->_succs[j]; 1769 update_succ_freq(sb, freq * b->succ_prob(j)); 1770 } 1771 } else { 1772 CFGLoop* lp = s->as_CFGLoop(); 1773 assert(lp->_parent == this, "immediate child"); 1774 for (int k = 0; k < lp->_exits.length(); k++) { 1775 Block* eb = lp->_exits.at(k).get_target(); 1776 double prob = lp->_exits.at(k).get_prob(); 1777 update_succ_freq(eb, freq * prob); 1778 } 1779 } 1780 } 1781 1782 // For all loops other than the outer, "method" loop, 1783 // sum and normalize the exit probability. The "method" loop 1784 // should keep the initial exit probability of 1, so that 1785 // inner blocks do not get erroneously scaled. 1786 if (_depth != 0) { 1787 // Total the exit probabilities for this loop. 1788 double exits_sum = 0.0f; 1789 for (int i = 0; i < _exits.length(); i++) { 1790 exits_sum += _exits.at(i).get_prob(); 1791 } 1792 1793 // Normalize the exit probabilities. Until now, the 1794 // probabilities estimate the possibility of exit per 1795 // a single loop iteration; afterward, they estimate 1796 // the probability of exit per loop entry. 1797 for (int i = 0; i < _exits.length(); i++) { 1798 Block* et = _exits.at(i).get_target(); 1799 float new_prob = 0.0f; 1800 if (_exits.at(i).get_prob() > 0.0f) { 1801 new_prob = _exits.at(i).get_prob() / exits_sum; 1802 } 1803 BlockProbPair bpp(et, new_prob); 1804 _exits.at_put(i, bpp); 1805 } 1806 1807 // Save the total, but guard against unreasonable probability, 1808 // as the value is used to estimate the loop trip count. 1809 // An infinite trip count would blur relative block 1810 // frequencies. 1811 if (exits_sum > 1.0f) exits_sum = 1.0; 1812 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN; 1813 _exit_prob = exits_sum; 1814 } 1815 } 1816 1817 //------------------------------succ_prob------------------------------------- 1818 // Determine the probability of reaching successor 'i' from the receiver block. 1819 float Block::succ_prob(uint i) { 1820 int eidx = end_idx(); 1821 Node *n = get_node(eidx); // Get ending Node 1822 1823 Opcodes op = n->Opcode(); 1824 if (n->is_Mach()) { 1825 if (n->is_MachNullCheck()) { 1826 // Can only reach here if called after lcm. The original Op_If is gone, 1827 // so we attempt to infer the probability from one or both of the 1828 // successor blocks. 1829 assert(_num_succs == 2, "expecting 2 successors of a null check"); 1830 // If either successor has only one predecessor, then the 1831 // probability estimate can be derived using the 1832 // relative frequency of the successor and this block. 1833 if (_succs[i]->num_preds() == 2) { 1834 return _succs[i]->_freq / _freq; 1835 } else if (_succs[1-i]->num_preds() == 2) { 1836 return 1 - (_succs[1-i]->_freq / _freq); 1837 } else { 1838 // Estimate using both successor frequencies 1839 float freq = _succs[i]->_freq; 1840 return freq / (freq + _succs[1-i]->_freq); 1841 } 1842 } 1843 op = n->as_Mach()->ideal_Opcode(); 1844 } 1845 1846 1847 // Switch on branch type 1848 switch( op ) { 1849 case Opcodes::Op_CountedLoopEnd: 1850 case Opcodes::Op_If: { 1851 assert (i < 2, "just checking"); 1852 // Conditionals pass on only part of their frequency 1853 float prob = n->as_MachIf()->_prob; 1854 assert(prob >= 0.0 && prob <= 1.0, "out of range probability"); 1855 // If succ[i] is the FALSE branch, invert path info 1856 if( get_node(i + eidx + 1)->Opcode() == Opcodes::Op_IfFalse ) { 1857 return 1.0f - prob; // not taken 1858 } else { 1859 return prob; // taken 1860 } 1861 } 1862 1863 case Opcodes::Op_Jump: 1864 // Divide the frequency between all successors evenly 1865 return 1.0f/_num_succs; 1866 1867 case Opcodes::Op_Catch: { 1868 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj(); 1869 if (ci->_con == CatchProjNode::fall_through_index) { 1870 // Fall-thru path gets the lion's share. 1871 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs; 1872 } else { 1873 // Presume exceptional paths are equally unlikely 1874 return PROB_UNLIKELY_MAG(5); 1875 } 1876 } 1877 1878 case Opcodes::Op_Root: 1879 case Opcodes::Op_Goto: 1880 // Pass frequency straight thru to target 1881 return 1.0f; 1882 1883 case Opcodes::Op_NeverBranch: 1884 return 0.0f; 1885 1886 case Opcodes::Op_TailCall: 1887 case Opcodes::Op_TailJump: 1888 case Opcodes::Op_Return: 1889 case Opcodes::Op_Halt: 1890 case Opcodes::Op_Rethrow: 1891 // Do not push out freq to root block 1892 return 0.0f; 1893 1894 default: 1895 ShouldNotReachHere(); 1896 } 1897 1898 return 0.0f; 1899 } 1900 1901 //------------------------------num_fall_throughs----------------------------- 1902 // Return the number of fall-through candidates for a block 1903 int Block::num_fall_throughs() { 1904 int eidx = end_idx(); 1905 Node *n = get_node(eidx); // Get ending Node 1906 1907 Opcodes op = n->Opcode(); 1908 if (n->is_Mach()) { 1909 if (n->is_MachNullCheck()) { 1910 // In theory, either side can fall-thru, for simplicity sake, 1911 // let's say only the false branch can now. 1912 return 1; 1913 } 1914 op = n->as_Mach()->ideal_Opcode(); 1915 } 1916 1917 // Switch on branch type 1918 switch( op ) { 1919 case Opcodes::Op_CountedLoopEnd: 1920 case Opcodes::Op_If: 1921 return 2; 1922 1923 case Opcodes::Op_Root: 1924 case Opcodes::Op_Goto: 1925 return 1; 1926 1927 case Opcodes::Op_Catch: { 1928 for (uint i = 0; i < _num_succs; i++) { 1929 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj(); 1930 if (ci->_con == CatchProjNode::fall_through_index) { 1931 return 1; 1932 } 1933 } 1934 return 0; 1935 } 1936 1937 case Opcodes::Op_Jump: 1938 case Opcodes::Op_NeverBranch: 1939 case Opcodes::Op_TailCall: 1940 case Opcodes::Op_TailJump: 1941 case Opcodes::Op_Return: 1942 case Opcodes::Op_Halt: 1943 case Opcodes::Op_Rethrow: 1944 return 0; 1945 1946 default: 1947 ShouldNotReachHere(); 1948 } 1949 1950 return 0; 1951 } 1952 1953 //------------------------------succ_fall_through----------------------------- 1954 // Return true if a specific successor could be fall-through target. 1955 bool Block::succ_fall_through(uint i) { 1956 int eidx = end_idx(); 1957 Node *n = get_node(eidx); // Get ending Node 1958 1959 Opcodes op = n->Opcode(); 1960 if (n->is_Mach()) { 1961 if (n->is_MachNullCheck()) { 1962 // In theory, either side can fall-thru, for simplicity sake, 1963 // let's say only the false branch can now. 1964 return get_node(i + eidx + 1)->Opcode() == Opcodes::Op_IfFalse; 1965 } 1966 op = n->as_Mach()->ideal_Opcode(); 1967 } 1968 1969 // Switch on branch type 1970 switch( op ) { 1971 case Opcodes::Op_CountedLoopEnd: 1972 case Opcodes::Op_If: 1973 case Opcodes::Op_Root: 1974 case Opcodes::Op_Goto: 1975 return true; 1976 1977 case Opcodes::Op_Catch: { 1978 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj(); 1979 return ci->_con == CatchProjNode::fall_through_index; 1980 } 1981 1982 case Opcodes::Op_Jump: 1983 case Opcodes::Op_NeverBranch: 1984 case Opcodes::Op_TailCall: 1985 case Opcodes::Op_TailJump: 1986 case Opcodes::Op_Return: 1987 case Opcodes::Op_Halt: 1988 case Opcodes::Op_Rethrow: 1989 return false; 1990 1991 default: 1992 ShouldNotReachHere(); 1993 } 1994 1995 return false; 1996 } 1997 1998 //------------------------------update_uncommon_branch------------------------ 1999 // Update the probability of a two-branch to be uncommon 2000 void Block::update_uncommon_branch(Block* ub) { 2001 int eidx = end_idx(); 2002 Node *n = get_node(eidx); // Get ending Node 2003 2004 Opcodes op = n->as_Mach()->ideal_Opcode(); 2005 2006 assert(op == Opcodes::Op_CountedLoopEnd || op == Opcodes::Op_If, "must be a If"); 2007 assert(num_fall_throughs() == 2, "must be a two way branch block"); 2008 2009 // Which successor is ub? 2010 uint s; 2011 for (s = 0; s <_num_succs; s++) { 2012 if (_succs[s] == ub) break; 2013 } 2014 assert(s < 2, "uncommon successor must be found"); 2015 2016 // If ub is the true path, make the proability small, else 2017 // ub is the false path, and make the probability large 2018 bool invert = (get_node(s + eidx + 1)->Opcode() == Opcodes::Op_IfFalse); 2019 2020 // Get existing probability 2021 float p = n->as_MachIf()->_prob; 2022 2023 if (invert) p = 1.0 - p; 2024 if (p > PROB_MIN) { 2025 p = PROB_MIN; 2026 } 2027 if (invert) p = 1.0 - p; 2028 2029 n->as_MachIf()->_prob = p; 2030 } 2031 2032 //------------------------------update_succ_freq------------------------------- 2033 // Update the appropriate frequency associated with block 'b', a successor of 2034 // a block in this loop. 2035 void CFGLoop::update_succ_freq(Block* b, double freq) { 2036 if (b->_loop == this) { 2037 if (b == head()) { 2038 // back branch within the loop 2039 // Do nothing now, the loop carried frequency will be 2040 // adjust later in scale_freq(). 2041 } else { 2042 // simple branch within the loop 2043 b->_freq += freq; 2044 } 2045 } else if (!in_loop_nest(b)) { 2046 // branch is exit from this loop 2047 BlockProbPair bpp(b, freq); 2048 _exits.append(bpp); 2049 } else { 2050 // branch into nested loop 2051 CFGLoop* ch = b->_loop; 2052 ch->_freq += freq; 2053 } 2054 } 2055 2056 //------------------------------in_loop_nest----------------------------------- 2057 // Determine if block b is in the receiver's loop nest. 2058 bool CFGLoop::in_loop_nest(Block* b) { 2059 int depth = _depth; 2060 CFGLoop* b_loop = b->_loop; 2061 int b_depth = b_loop->_depth; 2062 if (depth == b_depth) { 2063 return true; 2064 } 2065 while (b_depth > depth) { 2066 b_loop = b_loop->_parent; 2067 b_depth = b_loop->_depth; 2068 } 2069 return b_loop == this; 2070 } 2071 2072 //------------------------------scale_freq------------------------------------- 2073 // Scale frequency of loops and blocks by trip counts from outer loops 2074 // Do a top down traversal of loop tree (visit outer loops first.) 2075 void CFGLoop::scale_freq() { 2076 double loop_freq = _freq * trip_count(); 2077 _freq = loop_freq; 2078 for (int i = 0; i < _members.length(); i++) { 2079 CFGElement* s = _members.at(i); 2080 double block_freq = s->_freq * loop_freq; 2081 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY) 2082 block_freq = MIN_BLOCK_FREQUENCY; 2083 s->_freq = block_freq; 2084 } 2085 CFGLoop* ch = _child; 2086 while (ch != NULL) { 2087 ch->scale_freq(); 2088 ch = ch->_sibling; 2089 } 2090 } 2091 2092 // Frequency of outer loop 2093 double CFGLoop::outer_loop_freq() const { 2094 if (_child != NULL) { 2095 return _child->_freq; 2096 } 2097 return _freq; 2098 } 2099 2100 #ifndef PRODUCT 2101 //------------------------------dump_tree-------------------------------------- 2102 void CFGLoop::dump_tree() const { 2103 dump(); 2104 if (_child != NULL) _child->dump_tree(); 2105 if (_sibling != NULL) _sibling->dump_tree(); 2106 } 2107 2108 //------------------------------dump------------------------------------------- 2109 void CFGLoop::dump() const { 2110 for (int i = 0; i < _depth; i++) tty->print(" "); 2111 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n", 2112 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq); 2113 for (int i = 0; i < _depth; i++) tty->print(" "); 2114 tty->print(" members:"); 2115 int k = 0; 2116 for (int i = 0; i < _members.length(); i++) { 2117 if (k++ >= 6) { 2118 tty->print("\n "); 2119 for (int j = 0; j < _depth+1; j++) tty->print(" "); 2120 k = 0; 2121 } 2122 CFGElement *s = _members.at(i); 2123 if (s->is_block()) { 2124 Block *b = s->as_Block(); 2125 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq); 2126 } else { 2127 CFGLoop* lp = s->as_CFGLoop(); 2128 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq); 2129 } 2130 } 2131 tty->print("\n"); 2132 for (int i = 0; i < _depth; i++) tty->print(" "); 2133 tty->print(" exits: "); 2134 k = 0; 2135 for (int i = 0; i < _exits.length(); i++) { 2136 if (k++ >= 7) { 2137 tty->print("\n "); 2138 for (int j = 0; j < _depth+1; j++) tty->print(" "); 2139 k = 0; 2140 } 2141 Block *blk = _exits.at(i).get_target(); 2142 double prob = _exits.at(i).get_prob(); 2143 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100)); 2144 } 2145 tty->print("\n"); 2146 } 2147 #endif