1 /* 2 * Copyright (c) 1997, 2013, 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 "opto/block.hpp" 29 #include "opto/c2compiler.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/cfgnode.hpp" 32 #include "opto/machnode.hpp" 33 #include "opto/opcodes.hpp" 34 #include "opto/phaseX.hpp" 35 #include "opto/rootnode.hpp" 36 #include "opto/runtime.hpp" 37 #include "runtime/deoptimization.hpp" 38 #ifdef TARGET_ARCH_MODEL_x86_32 39 # include "adfiles/ad_x86_32.hpp" 40 #endif 41 #ifdef TARGET_ARCH_MODEL_x86_64 42 # include "adfiles/ad_x86_64.hpp" 43 #endif 44 #ifdef TARGET_ARCH_MODEL_sparc 45 # include "adfiles/ad_sparc.hpp" 46 #endif 47 #ifdef TARGET_ARCH_MODEL_zero 48 # include "adfiles/ad_zero.hpp" 49 #endif 50 #ifdef TARGET_ARCH_MODEL_arm 51 # include "adfiles/ad_arm.hpp" 52 #endif 53 #ifdef TARGET_ARCH_MODEL_ppc_32 54 # include "adfiles/ad_ppc_32.hpp" 55 #endif 56 #ifdef TARGET_ARCH_MODEL_ppc_64 57 # include "adfiles/ad_ppc_64.hpp" 58 #endif 59 60 61 // Portions of code courtesy of Clifford Click 62 63 // Optimization - Graph Style 64 65 // To avoid float value underflow 66 #define MIN_BLOCK_FREQUENCY 1.e-35f 67 68 //----------------------------schedule_node_into_block------------------------- 69 // Insert node n into block b. Look for projections of n and make sure they 70 // are in b also. 71 void PhaseCFG::schedule_node_into_block( Node *n, Block *b ) { 72 // Set basic block of n, Add n to b, 73 map_node_to_block(n, b); 74 b->add_inst(n); 75 76 // After Matching, nearly any old Node may have projections trailing it. 77 // These are usually machine-dependent flags. In any case, they might 78 // float to another block below this one. Move them up. 79 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 80 Node* use = n->fast_out(i); 81 if (use->is_Proj()) { 82 Block* buse = get_block_for_node(use); 83 if (buse != b) { // In wrong block? 84 if (buse != NULL) { 85 buse->find_remove(use); // Remove from wrong block 86 } 87 map_node_to_block(use, b); 88 b->add_inst(use); 89 } 90 } 91 } 92 } 93 94 //----------------------------replace_block_proj_ctrl------------------------- 95 // Nodes that have is_block_proj() nodes as their control need to use 96 // the appropriate Region for their actual block as their control since 97 // the projection will be in a predecessor block. 98 void PhaseCFG::replace_block_proj_ctrl( Node *n ) { 99 const Node *in0 = n->in(0); 100 assert(in0 != NULL, "Only control-dependent"); 101 const Node *p = in0->is_block_proj(); 102 if (p != NULL && p != n) { // Control from a block projection? 103 assert(!n->pinned() || n->is_MachConstantBase(), "only pinned MachConstantBase node is expected here"); 104 // Find trailing Region 105 Block *pb = get_block_for_node(in0); // Block-projection already has basic block 106 uint j = 0; 107 if (pb->_num_succs != 1) { // More then 1 successor? 108 // Search for successor 109 uint max = pb->number_of_nodes(); 110 assert( max > 1, "" ); 111 uint start = max - pb->_num_succs; 112 // Find which output path belongs to projection 113 for (j = start; j < max; j++) { 114 if( pb->get_node(j) == in0 ) 115 break; 116 } 117 assert( j < max, "must find" ); 118 // Change control to match head of successor basic block 119 j -= start; 120 } 121 n->set_req(0, pb->_succs[j]->head()); 122 } 123 } 124 125 126 //------------------------------schedule_pinned_nodes-------------------------- 127 // Set the basic block for Nodes pinned into blocks 128 void PhaseCFG::schedule_pinned_nodes(VectorSet &visited) { 129 // Allocate node stack of size C->unique()+8 to avoid frequent realloc 130 GrowableArray <Node *> spstack(C->unique() + 8); 131 spstack.push(_root); 132 while (spstack.is_nonempty()) { 133 Node* node = spstack.pop(); 134 if (!visited.test_set(node->_idx)) { // Test node and flag it as visited 135 if (node->pinned() && !has_block(node)) { // Pinned? Nail it down! 136 assert(node->in(0), "pinned Node must have Control"); 137 // Before setting block replace block_proj control edge 138 replace_block_proj_ctrl(node); 139 Node* input = node->in(0); 140 while (!input->is_block_start()) { 141 input = input->in(0); 142 } 143 Block* block = get_block_for_node(input); // Basic block of controlling input 144 schedule_node_into_block(node, block); 145 } 146 147 // process all inputs that are non NULL 148 for (int i = node->req() - 1; i >= 0; --i) { 149 if (node->in(i) != NULL) { 150 spstack.push(node->in(i)); 151 } 152 } 153 } 154 } 155 } 156 157 #ifdef ASSERT 158 // Assert that new input b2 is dominated by all previous inputs. 159 // Check this by by seeing that it is dominated by b1, the deepest 160 // input observed until b2. 161 static void assert_dom(Block* b1, Block* b2, Node* n, const PhaseCFG* cfg) { 162 if (b1 == NULL) return; 163 assert(b1->_dom_depth < b2->_dom_depth, "sanity"); 164 Block* tmp = b2; 165 while (tmp != b1 && tmp != NULL) { 166 tmp = tmp->_idom; 167 } 168 if (tmp != b1) { 169 // Detected an unschedulable graph. Print some nice stuff and die. 170 tty->print_cr("!!! Unschedulable graph !!!"); 171 for (uint j=0; j<n->len(); j++) { // For all inputs 172 Node* inn = n->in(j); // Get input 173 if (inn == NULL) continue; // Ignore NULL, missing inputs 174 Block* inb = cfg->get_block_for_node(inn); 175 tty->print("B%d idom=B%d depth=%2d ",inb->_pre_order, 176 inb->_idom ? inb->_idom->_pre_order : 0, inb->_dom_depth); 177 inn->dump(); 178 } 179 tty->print("Failing node: "); 180 n->dump(); 181 assert(false, "unscheduable graph"); 182 } 183 } 184 #endif 185 186 static Block* find_deepest_input(Node* n, const PhaseCFG* cfg) { 187 // Find the last input dominated by all other inputs. 188 Block* deepb = NULL; // Deepest block so far 189 int deepb_dom_depth = 0; 190 for (uint k = 0; k < n->len(); k++) { // For all inputs 191 Node* inn = n->in(k); // Get input 192 if (inn == NULL) continue; // Ignore NULL, missing inputs 193 Block* inb = cfg->get_block_for_node(inn); 194 assert(inb != NULL, "must already have scheduled this input"); 195 if (deepb_dom_depth < (int) inb->_dom_depth) { 196 // The new inb must be dominated by the previous deepb. 197 // The various inputs must be linearly ordered in the dom 198 // tree, or else there will not be a unique deepest block. 199 DEBUG_ONLY(assert_dom(deepb, inb, n, cfg)); 200 deepb = inb; // Save deepest block 201 deepb_dom_depth = deepb->_dom_depth; 202 } 203 } 204 assert(deepb != NULL, "must be at least one input to n"); 205 return deepb; 206 } 207 208 209 //------------------------------schedule_early--------------------------------- 210 // Find the earliest Block any instruction can be placed in. Some instructions 211 // are pinned into Blocks. Unpinned instructions can appear in last block in 212 // which all their inputs occur. 213 bool PhaseCFG::schedule_early(VectorSet &visited, Node_List &roots) { 214 // Allocate stack with enough space to avoid frequent realloc 215 Node_Stack nstack(roots.Size() + 8); 216 // _root will be processed among C->top() inputs 217 roots.push(C->top()); 218 visited.set(C->top()->_idx); 219 220 while (roots.size() != 0) { 221 // Use local variables nstack_top_n & nstack_top_i to cache values 222 // on stack's top. 223 Node* parent_node = roots.pop(); 224 uint input_index = 0; 225 226 while (true) { 227 if (input_index == 0) { 228 // Fixup some control. Constants without control get attached 229 // to root and nodes that use is_block_proj() nodes should be attached 230 // to the region that starts their block. 231 const Node* control_input = parent_node->in(0); 232 if (control_input != NULL) { 233 replace_block_proj_ctrl(parent_node); 234 } else { 235 // Is a constant with NO inputs? 236 if (parent_node->req() == 1) { 237 parent_node->set_req(0, _root); 238 } 239 } 240 } 241 242 // First, visit all inputs and force them to get a block. If an 243 // input is already in a block we quit following inputs (to avoid 244 // cycles). Instead we put that Node on a worklist to be handled 245 // later (since IT'S inputs may not have a block yet). 246 247 // Assume all n's inputs will be processed 248 bool done = true; 249 250 while (input_index < parent_node->len()) { 251 Node* in = parent_node->in(input_index++); 252 if (in == NULL) { 253 continue; 254 } 255 256 int is_visited = visited.test_set(in->_idx); 257 if (!has_block(in)) { 258 if (is_visited) { 259 return false; 260 } 261 // Save parent node and next input's index. 262 nstack.push(parent_node, input_index); 263 // Process current input now. 264 parent_node = in; 265 input_index = 0; 266 // Not all n's inputs processed. 267 done = false; 268 break; 269 } else if (!is_visited) { 270 // Visit this guy later, using worklist 271 roots.push(in); 272 } 273 } 274 275 if (done) { 276 // All of n's inputs have been processed, complete post-processing. 277 278 // Some instructions are pinned into a block. These include Region, 279 // Phi, Start, Return, and other control-dependent instructions and 280 // any projections which depend on them. 281 if (!parent_node->pinned()) { 282 // Set earliest legal block. 283 Block* earliest_block = find_deepest_input(parent_node, this); 284 map_node_to_block(parent_node, earliest_block); 285 } else { 286 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"); 287 } 288 289 if (nstack.is_empty()) { 290 // Finished all nodes on stack. 291 // Process next node on the worklist 'roots'. 292 break; 293 } 294 // Get saved parent node and next input's index. 295 parent_node = nstack.node(); 296 input_index = nstack.index(); 297 nstack.pop(); 298 } 299 } 300 } 301 return true; 302 } 303 304 //------------------------------dom_lca---------------------------------------- 305 // Find least common ancestor in dominator tree 306 // LCA is a current notion of LCA, to be raised above 'this'. 307 // As a convenient boundary condition, return 'this' if LCA is NULL. 308 // Find the LCA of those two nodes. 309 Block* Block::dom_lca(Block* LCA) { 310 if (LCA == NULL || LCA == this) return this; 311 312 Block* anc = this; 313 while (anc->_dom_depth > LCA->_dom_depth) 314 anc = anc->_idom; // Walk up till anc is as high as LCA 315 316 while (LCA->_dom_depth > anc->_dom_depth) 317 LCA = LCA->_idom; // Walk up till LCA is as high as anc 318 319 while (LCA != anc) { // Walk both up till they are the same 320 LCA = LCA->_idom; 321 anc = anc->_idom; 322 } 323 324 return LCA; 325 } 326 327 //--------------------------raise_LCA_above_use-------------------------------- 328 // We are placing a definition, and have been given a def->use edge. 329 // The definition must dominate the use, so move the LCA upward in the 330 // dominator tree to dominate the use. If the use is a phi, adjust 331 // the LCA only with the phi input paths which actually use this def. 332 static Block* raise_LCA_above_use(Block* LCA, Node* use, Node* def, const PhaseCFG* cfg) { 333 Block* buse = cfg->get_block_for_node(use); 334 if (buse == NULL) return LCA; // Unused killing Projs have no use block 335 if (!use->is_Phi()) return buse->dom_lca(LCA); 336 uint pmax = use->req(); // Number of Phi inputs 337 // Why does not this loop just break after finding the matching input to 338 // the Phi? Well...it's like this. I do not have true def-use/use-def 339 // chains. Means I cannot distinguish, from the def-use direction, which 340 // of many use-defs lead from the same use to the same def. That is, this 341 // Phi might have several uses of the same def. Each use appears in a 342 // different predecessor block. But when I enter here, I cannot distinguish 343 // which use-def edge I should find the predecessor block for. So I find 344 // them all. Means I do a little extra work if a Phi uses the same value 345 // more than once. 346 for (uint j=1; j<pmax; j++) { // For all inputs 347 if (use->in(j) == def) { // Found matching input? 348 Block* pred = cfg->get_block_for_node(buse->pred(j)); 349 LCA = pred->dom_lca(LCA); 350 } 351 } 352 return LCA; 353 } 354 355 //----------------------------raise_LCA_above_marks---------------------------- 356 // Return a new LCA that dominates LCA and any of its marked predecessors. 357 // Search all my parents up to 'early' (exclusive), looking for predecessors 358 // which are marked with the given index. Return the LCA (in the dom tree) 359 // of all marked blocks. If there are none marked, return the original 360 // LCA. 361 static Block* raise_LCA_above_marks(Block* LCA, node_idx_t mark, Block* early, const PhaseCFG* cfg) { 362 Block_List worklist; 363 worklist.push(LCA); 364 while (worklist.size() > 0) { 365 Block* mid = worklist.pop(); 366 if (mid == early) continue; // stop searching here 367 368 // Test and set the visited bit. 369 if (mid->raise_LCA_visited() == mark) continue; // already visited 370 371 // Don't process the current LCA, otherwise the search may terminate early 372 if (mid != LCA && mid->raise_LCA_mark() == mark) { 373 // Raise the LCA. 374 LCA = mid->dom_lca(LCA); 375 if (LCA == early) break; // stop searching everywhere 376 assert(early->dominates(LCA), "early is high enough"); 377 // Resume searching at that point, skipping intermediate levels. 378 worklist.push(LCA); 379 if (LCA == mid) 380 continue; // Don't mark as visited to avoid early termination. 381 } else { 382 // Keep searching through this block's predecessors. 383 for (uint j = 1, jmax = mid->num_preds(); j < jmax; j++) { 384 Block* mid_parent = cfg->get_block_for_node(mid->pred(j)); 385 worklist.push(mid_parent); 386 } 387 } 388 mid->set_raise_LCA_visited(mark); 389 } 390 return LCA; 391 } 392 393 //--------------------------memory_early_block-------------------------------- 394 // This is a variation of find_deepest_input, the heart of schedule_early. 395 // Find the "early" block for a load, if we considered only memory and 396 // address inputs, that is, if other data inputs were ignored. 397 // 398 // Because a subset of edges are considered, the resulting block will 399 // be earlier (at a shallower dom_depth) than the true schedule_early 400 // point of the node. We compute this earlier block as a more permissive 401 // site for anti-dependency insertion, but only if subsume_loads is enabled. 402 static Block* memory_early_block(Node* load, Block* early, const PhaseCFG* cfg) { 403 Node* base; 404 Node* index; 405 Node* store = load->in(MemNode::Memory); 406 load->as_Mach()->memory_inputs(base, index); 407 408 assert(base != NodeSentinel && index != NodeSentinel, 409 "unexpected base/index inputs"); 410 411 Node* mem_inputs[4]; 412 int mem_inputs_length = 0; 413 if (base != NULL) mem_inputs[mem_inputs_length++] = base; 414 if (index != NULL) mem_inputs[mem_inputs_length++] = index; 415 if (store != NULL) mem_inputs[mem_inputs_length++] = store; 416 417 // In the comparision below, add one to account for the control input, 418 // which may be null, but always takes up a spot in the in array. 419 if (mem_inputs_length + 1 < (int) load->req()) { 420 // This "load" has more inputs than just the memory, base and index inputs. 421 // For purposes of checking anti-dependences, we need to start 422 // from the early block of only the address portion of the instruction, 423 // and ignore other blocks that may have factored into the wider 424 // schedule_early calculation. 425 if (load->in(0) != NULL) mem_inputs[mem_inputs_length++] = load->in(0); 426 427 Block* deepb = NULL; // Deepest block so far 428 int deepb_dom_depth = 0; 429 for (int i = 0; i < mem_inputs_length; i++) { 430 Block* inb = cfg->get_block_for_node(mem_inputs[i]); 431 if (deepb_dom_depth < (int) inb->_dom_depth) { 432 // The new inb must be dominated by the previous deepb. 433 // The various inputs must be linearly ordered in the dom 434 // tree, or else there will not be a unique deepest block. 435 DEBUG_ONLY(assert_dom(deepb, inb, load, cfg)); 436 deepb = inb; // Save deepest block 437 deepb_dom_depth = deepb->_dom_depth; 438 } 439 } 440 early = deepb; 441 } 442 443 return early; 444 } 445 446 //--------------------------insert_anti_dependences--------------------------- 447 // A load may need to witness memory that nearby stores can overwrite. 448 // For each nearby store, either insert an "anti-dependence" edge 449 // from the load to the store, or else move LCA upward to force the 450 // load to (eventually) be scheduled in a block above the store. 451 // 452 // Do not add edges to stores on distinct control-flow paths; 453 // only add edges to stores which might interfere. 454 // 455 // Return the (updated) LCA. There will not be any possibly interfering 456 // store between the load's "early block" and the updated LCA. 457 // Any stores in the updated LCA will have new precedence edges 458 // back to the load. The caller is expected to schedule the load 459 // in the LCA, in which case the precedence edges will make LCM 460 // preserve anti-dependences. The caller may also hoist the load 461 // above the LCA, if it is not the early block. 462 Block* PhaseCFG::insert_anti_dependences(Block* LCA, Node* load, bool verify) { 463 assert(load->needs_anti_dependence_check(), "must be a load of some sort"); 464 assert(LCA != NULL, ""); 465 DEBUG_ONLY(Block* LCA_orig = LCA); 466 467 // Compute the alias index. Loads and stores with different alias indices 468 // do not need anti-dependence edges. 469 uint load_alias_idx = C->get_alias_index(load->adr_type()); 470 #ifdef ASSERT 471 if (load_alias_idx == Compile::AliasIdxBot && C->AliasLevel() > 0 && 472 (PrintOpto || VerifyAliases || 473 PrintMiscellaneous && (WizardMode || Verbose))) { 474 // Load nodes should not consume all of memory. 475 // Reporting a bottom type indicates a bug in adlc. 476 // If some particular type of node validly consumes all of memory, 477 // sharpen the preceding "if" to exclude it, so we can catch bugs here. 478 tty->print_cr("*** Possible Anti-Dependence Bug: Load consumes all of memory."); 479 load->dump(2); 480 if (VerifyAliases) assert(load_alias_idx != Compile::AliasIdxBot, ""); 481 } 482 #endif 483 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrComp), 484 "String compare is only known 'load' that does not conflict with any stores"); 485 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrEquals), 486 "String equals is a 'load' that does not conflict with any stores"); 487 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_StrIndexOf), 488 "String indexOf is a 'load' that does not conflict with any stores"); 489 assert(load_alias_idx || (load->is_Mach() && load->as_Mach()->ideal_Opcode() == Op_AryEq), 490 "Arrays equals is a 'load' that do not conflict with any stores"); 491 492 if (!C->alias_type(load_alias_idx)->is_rewritable()) { 493 // It is impossible to spoil this load by putting stores before it, 494 // because we know that the stores will never update the value 495 // which 'load' must witness. 496 return LCA; 497 } 498 499 node_idx_t load_index = load->_idx; 500 501 // Note the earliest legal placement of 'load', as determined by 502 // by the unique point in the dom tree where all memory effects 503 // and other inputs are first available. (Computed by schedule_early.) 504 // For normal loads, 'early' is the shallowest place (dom graph wise) 505 // to look for anti-deps between this load and any store. 506 Block* early = get_block_for_node(load); 507 508 // If we are subsuming loads, compute an "early" block that only considers 509 // memory or address inputs. This block may be different than the 510 // schedule_early block in that it could be at an even shallower depth in the 511 // dominator tree, and allow for a broader discovery of anti-dependences. 512 if (C->subsume_loads()) { 513 early = memory_early_block(load, early, this); 514 } 515 516 ResourceArea *area = Thread::current()->resource_area(); 517 Node_List worklist_mem(area); // prior memory state to store 518 Node_List worklist_store(area); // possible-def to explore 519 Node_List worklist_visited(area); // visited mergemem nodes 520 Node_List non_early_stores(area); // all relevant stores outside of early 521 bool must_raise_LCA = false; 522 523 #ifdef TRACK_PHI_INPUTS 524 // %%% This extra checking fails because MergeMem nodes are not GVNed. 525 // Provide "phi_inputs" to check if every input to a PhiNode is from the 526 // original memory state. This indicates a PhiNode for which should not 527 // prevent the load from sinking. For such a block, set_raise_LCA_mark 528 // may be overly conservative. 529 // Mechanism: count inputs seen for each Phi encountered in worklist_store. 530 DEBUG_ONLY(GrowableArray<uint> phi_inputs(area, C->unique(),0,0)); 531 #endif 532 533 // 'load' uses some memory state; look for users of the same state. 534 // Recurse through MergeMem nodes to the stores that use them. 535 536 // Each of these stores is a possible definition of memory 537 // that 'load' needs to use. We need to force 'load' 538 // to occur before each such store. When the store is in 539 // the same block as 'load', we insert an anti-dependence 540 // edge load->store. 541 542 // The relevant stores "nearby" the load consist of a tree rooted 543 // at initial_mem, with internal nodes of type MergeMem. 544 // Therefore, the branches visited by the worklist are of this form: 545 // initial_mem -> (MergeMem ->)* store 546 // The anti-dependence constraints apply only to the fringe of this tree. 547 548 Node* initial_mem = load->in(MemNode::Memory); 549 worklist_store.push(initial_mem); 550 worklist_visited.push(initial_mem); 551 worklist_mem.push(NULL); 552 while (worklist_store.size() > 0) { 553 // Examine a nearby store to see if it might interfere with our load. 554 Node* mem = worklist_mem.pop(); 555 Node* store = worklist_store.pop(); 556 uint op = store->Opcode(); 557 558 // MergeMems do not directly have anti-deps. 559 // Treat them as internal nodes in a forward tree of memory states, 560 // the leaves of which are each a 'possible-def'. 561 if (store == initial_mem // root (exclusive) of tree we are searching 562 || op == Op_MergeMem // internal node of tree we are searching 563 ) { 564 mem = store; // It's not a possibly interfering store. 565 if (store == initial_mem) 566 initial_mem = NULL; // only process initial memory once 567 568 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 569 store = mem->fast_out(i); 570 if (store->is_MergeMem()) { 571 // Be sure we don't get into combinatorial problems. 572 // (Allow phis to be repeated; they can merge two relevant states.) 573 uint j = worklist_visited.size(); 574 for (; j > 0; j--) { 575 if (worklist_visited.at(j-1) == store) break; 576 } 577 if (j > 0) continue; // already on work list; do not repeat 578 worklist_visited.push(store); 579 } 580 worklist_mem.push(mem); 581 worklist_store.push(store); 582 } 583 continue; 584 } 585 586 if (op == Op_MachProj || op == Op_Catch) continue; 587 if (store->needs_anti_dependence_check()) continue; // not really a store 588 589 // Compute the alias index. Loads and stores with different alias 590 // indices do not need anti-dependence edges. Wide MemBar's are 591 // anti-dependent on everything (except immutable memories). 592 const TypePtr* adr_type = store->adr_type(); 593 if (!C->can_alias(adr_type, load_alias_idx)) continue; 594 595 // Most slow-path runtime calls do NOT modify Java memory, but 596 // they can block and so write Raw memory. 597 if (store->is_Mach()) { 598 MachNode* mstore = store->as_Mach(); 599 if (load_alias_idx != Compile::AliasIdxRaw) { 600 // Check for call into the runtime using the Java calling 601 // convention (and from there into a wrapper); it has no 602 // _method. Can't do this optimization for Native calls because 603 // they CAN write to Java memory. 604 if (mstore->ideal_Opcode() == Op_CallStaticJava) { 605 assert(mstore->is_MachSafePoint(), ""); 606 MachSafePointNode* ms = (MachSafePointNode*) mstore; 607 assert(ms->is_MachCallJava(), ""); 608 MachCallJavaNode* mcj = (MachCallJavaNode*) ms; 609 if (mcj->_method == NULL) { 610 // These runtime calls do not write to Java visible memory 611 // (other than Raw) and so do not require anti-dependence edges. 612 continue; 613 } 614 } 615 // Same for SafePoints: they read/write Raw but only read otherwise. 616 // This is basically a workaround for SafePoints only defining control 617 // instead of control + memory. 618 if (mstore->ideal_Opcode() == Op_SafePoint) 619 continue; 620 } else { 621 // Some raw memory, such as the load of "top" at an allocation, 622 // can be control dependent on the previous safepoint. See 623 // comments in GraphKit::allocate_heap() about control input. 624 // Inserting an anti-dep between such a safepoint and a use 625 // creates a cycle, and will cause a subsequent failure in 626 // local scheduling. (BugId 4919904) 627 // (%%% How can a control input be a safepoint and not a projection??) 628 if (mstore->ideal_Opcode() == Op_SafePoint && load->in(0) == mstore) 629 continue; 630 } 631 } 632 633 // Identify a block that the current load must be above, 634 // or else observe that 'store' is all the way up in the 635 // earliest legal block for 'load'. In the latter case, 636 // immediately insert an anti-dependence edge. 637 Block* store_block = get_block_for_node(store); 638 assert(store_block != NULL, "unused killing projections skipped above"); 639 640 if (store->is_Phi()) { 641 // 'load' uses memory which is one (or more) of the Phi's inputs. 642 // It must be scheduled not before the Phi, but rather before 643 // each of the relevant Phi inputs. 644 // 645 // Instead of finding the LCA of all inputs to a Phi that match 'mem', 646 // we mark each corresponding predecessor block and do a combined 647 // hoisting operation later (raise_LCA_above_marks). 648 // 649 // Do not assert(store_block != early, "Phi merging memory after access") 650 // PhiNode may be at start of block 'early' with backedge to 'early' 651 DEBUG_ONLY(bool found_match = false); 652 for (uint j = PhiNode::Input, jmax = store->req(); j < jmax; j++) { 653 if (store->in(j) == mem) { // Found matching input? 654 DEBUG_ONLY(found_match = true); 655 Block* pred_block = get_block_for_node(store_block->pred(j)); 656 if (pred_block != early) { 657 // If any predecessor of the Phi matches the load's "early block", 658 // we do not need a precedence edge between the Phi and 'load' 659 // since the load will be forced into a block preceding the Phi. 660 pred_block->set_raise_LCA_mark(load_index); 661 assert(!LCA_orig->dominates(pred_block) || 662 early->dominates(pred_block), "early is high enough"); 663 must_raise_LCA = true; 664 } else { 665 // anti-dependent upon PHI pinned below 'early', no edge needed 666 LCA = early; // but can not schedule below 'early' 667 } 668 } 669 } 670 assert(found_match, "no worklist bug"); 671 #ifdef TRACK_PHI_INPUTS 672 #ifdef ASSERT 673 // This assert asks about correct handling of PhiNodes, which may not 674 // have all input edges directly from 'mem'. See BugId 4621264 675 int num_mem_inputs = phi_inputs.at_grow(store->_idx,0) + 1; 676 // Increment by exactly one even if there are multiple copies of 'mem' 677 // coming into the phi, because we will run this block several times 678 // if there are several copies of 'mem'. (That's how DU iterators work.) 679 phi_inputs.at_put(store->_idx, num_mem_inputs); 680 assert(PhiNode::Input + num_mem_inputs < store->req(), 681 "Expect at least one phi input will not be from original memory state"); 682 #endif //ASSERT 683 #endif //TRACK_PHI_INPUTS 684 } else if (store_block != early) { 685 // 'store' is between the current LCA and earliest possible block. 686 // Label its block, and decide later on how to raise the LCA 687 // to include the effect on LCA of this store. 688 // If this store's block gets chosen as the raised LCA, we 689 // will find him on the non_early_stores list and stick him 690 // with a precedence edge. 691 // (But, don't bother if LCA is already raised all the way.) 692 if (LCA != early) { 693 store_block->set_raise_LCA_mark(load_index); 694 must_raise_LCA = true; 695 non_early_stores.push(store); 696 } 697 } else { 698 // Found a possibly-interfering store in the load's 'early' block. 699 // This means 'load' cannot sink at all in the dominator tree. 700 // Add an anti-dep edge, and squeeze 'load' into the highest block. 701 assert(store != load->in(0), "dependence cycle found"); 702 if (verify) { 703 assert(store->find_edge(load) != -1, "missing precedence edge"); 704 } else { 705 store->add_prec(load); 706 } 707 LCA = early; 708 // This turns off the process of gathering non_early_stores. 709 } 710 } 711 // (Worklist is now empty; all nearby stores have been visited.) 712 713 // Finished if 'load' must be scheduled in its 'early' block. 714 // If we found any stores there, they have already been given 715 // precedence edges. 716 if (LCA == early) return LCA; 717 718 // We get here only if there are no possibly-interfering stores 719 // in the load's 'early' block. Move LCA up above all predecessors 720 // which contain stores we have noted. 721 // 722 // The raised LCA block can be a home to such interfering stores, 723 // but its predecessors must not contain any such stores. 724 // 725 // The raised LCA will be a lower bound for placing the load, 726 // preventing the load from sinking past any block containing 727 // a store that may invalidate the memory state required by 'load'. 728 if (must_raise_LCA) 729 LCA = raise_LCA_above_marks(LCA, load->_idx, early, this); 730 if (LCA == early) return LCA; 731 732 // Insert anti-dependence edges from 'load' to each store 733 // in the non-early LCA block. 734 // Mine the non_early_stores list for such stores. 735 if (LCA->raise_LCA_mark() == load_index) { 736 while (non_early_stores.size() > 0) { 737 Node* store = non_early_stores.pop(); 738 Block* store_block = get_block_for_node(store); 739 if (store_block == LCA) { 740 // add anti_dependence from store to load in its own block 741 assert(store != load->in(0), "dependence cycle found"); 742 if (verify) { 743 assert(store->find_edge(load) != -1, "missing precedence edge"); 744 } else { 745 store->add_prec(load); 746 } 747 } else { 748 assert(store_block->raise_LCA_mark() == load_index, "block was marked"); 749 // Any other stores we found must be either inside the new LCA 750 // or else outside the original LCA. In the latter case, they 751 // did not interfere with any use of 'load'. 752 assert(LCA->dominates(store_block) 753 || !LCA_orig->dominates(store_block), "no stray stores"); 754 } 755 } 756 } 757 758 // Return the highest block containing stores; any stores 759 // within that block have been given anti-dependence edges. 760 return LCA; 761 } 762 763 // This class is used to iterate backwards over the nodes in the graph. 764 765 class Node_Backward_Iterator { 766 767 private: 768 Node_Backward_Iterator(); 769 770 public: 771 // Constructor for the iterator 772 Node_Backward_Iterator(Node *root, VectorSet &visited, Node_List &stack, PhaseCFG &cfg); 773 774 // Postincrement operator to iterate over the nodes 775 Node *next(); 776 777 private: 778 VectorSet &_visited; 779 Node_List &_stack; 780 PhaseCFG &_cfg; 781 }; 782 783 // Constructor for the Node_Backward_Iterator 784 Node_Backward_Iterator::Node_Backward_Iterator( Node *root, VectorSet &visited, Node_List &stack, PhaseCFG &cfg) 785 : _visited(visited), _stack(stack), _cfg(cfg) { 786 // The stack should contain exactly the root 787 stack.clear(); 788 stack.push(root); 789 790 // Clear the visited bits 791 visited.Clear(); 792 } 793 794 // Iterator for the Node_Backward_Iterator 795 Node *Node_Backward_Iterator::next() { 796 797 // If the _stack is empty, then just return NULL: finished. 798 if ( !_stack.size() ) 799 return NULL; 800 801 // '_stack' is emulating a real _stack. The 'visit-all-users' loop has been 802 // made stateless, so I do not need to record the index 'i' on my _stack. 803 // Instead I visit all users each time, scanning for unvisited users. 804 // I visit unvisited not-anti-dependence users first, then anti-dependent 805 // children next. 806 Node *self = _stack.pop(); 807 808 // I cycle here when I am entering a deeper level of recursion. 809 // The key variable 'self' was set prior to jumping here. 810 while( 1 ) { 811 812 _visited.set(self->_idx); 813 814 // Now schedule all uses as late as possible. 815 const Node* src = self->is_Proj() ? self->in(0) : self; 816 uint src_rpo = _cfg.get_block_for_node(src)->_rpo; 817 818 // Schedule all nodes in a post-order visit 819 Node *unvisited = NULL; // Unvisited anti-dependent Node, if any 820 821 // Scan for unvisited nodes 822 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) { 823 // For all uses, schedule late 824 Node* n = self->fast_out(i); // Use 825 826 // Skip already visited children 827 if ( _visited.test(n->_idx) ) 828 continue; 829 830 // do not traverse backward control edges 831 Node *use = n->is_Proj() ? n->in(0) : n; 832 uint use_rpo = _cfg.get_block_for_node(use)->_rpo; 833 834 if ( use_rpo < src_rpo ) 835 continue; 836 837 // Phi nodes always precede uses in a basic block 838 if ( use_rpo == src_rpo && use->is_Phi() ) 839 continue; 840 841 unvisited = n; // Found unvisited 842 843 // Check for possible-anti-dependent 844 if( !n->needs_anti_dependence_check() ) 845 break; // Not visited, not anti-dep; schedule it NOW 846 } 847 848 // Did I find an unvisited not-anti-dependent Node? 849 if ( !unvisited ) 850 break; // All done with children; post-visit 'self' 851 852 // Visit the unvisited Node. Contains the obvious push to 853 // indicate I'm entering a deeper level of recursion. I push the 854 // old state onto the _stack and set a new state and loop (recurse). 855 _stack.push(self); 856 self = unvisited; 857 } // End recursion loop 858 859 return self; 860 } 861 862 //------------------------------ComputeLatenciesBackwards---------------------- 863 // Compute the latency of all the instructions. 864 void PhaseCFG::compute_latencies_backwards(VectorSet &visited, Node_List &stack) { 865 #ifndef PRODUCT 866 if (trace_opto_pipelining()) 867 tty->print("\n#---- ComputeLatenciesBackwards ----\n"); 868 #endif 869 870 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this); 871 Node *n; 872 873 // Walk over all the nodes from last to first 874 while (n = iter.next()) { 875 // Set the latency for the definitions of this instruction 876 partial_latency_of_defs(n); 877 } 878 } // end ComputeLatenciesBackwards 879 880 //------------------------------partial_latency_of_defs------------------------ 881 // Compute the latency impact of this node on all defs. This computes 882 // a number that increases as we approach the beginning of the routine. 883 void PhaseCFG::partial_latency_of_defs(Node *n) { 884 // Set the latency for this instruction 885 #ifndef PRODUCT 886 if (trace_opto_pipelining()) { 887 tty->print("# latency_to_inputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n)); 888 dump(); 889 } 890 #endif 891 892 if (n->is_Proj()) { 893 n = n->in(0); 894 } 895 896 if (n->is_Root()) { 897 return; 898 } 899 900 uint nlen = n->len(); 901 uint use_latency = get_latency_for_node(n); 902 uint use_pre_order = get_block_for_node(n)->_pre_order; 903 904 for (uint j = 0; j < nlen; j++) { 905 Node *def = n->in(j); 906 907 if (!def || def == n) { 908 continue; 909 } 910 911 // Walk backwards thru projections 912 if (def->is_Proj()) { 913 def = def->in(0); 914 } 915 916 #ifndef PRODUCT 917 if (trace_opto_pipelining()) { 918 tty->print("# in(%2d): ", j); 919 def->dump(); 920 } 921 #endif 922 923 // If the defining block is not known, assume it is ok 924 Block *def_block = get_block_for_node(def); 925 uint def_pre_order = def_block ? def_block->_pre_order : 0; 926 927 if ((use_pre_order < def_pre_order) || (use_pre_order == def_pre_order && n->is_Phi())) { 928 continue; 929 } 930 931 uint delta_latency = n->latency(j); 932 uint current_latency = delta_latency + use_latency; 933 934 if (get_latency_for_node(def) < current_latency) { 935 set_latency_for_node(def, current_latency); 936 } 937 938 #ifndef PRODUCT 939 if (trace_opto_pipelining()) { 940 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)); 941 } 942 #endif 943 } 944 } 945 946 //------------------------------latency_from_use------------------------------- 947 // Compute the latency of a specific use 948 int PhaseCFG::latency_from_use(Node *n, const Node *def, Node *use) { 949 // If self-reference, return no latency 950 if (use == n || use->is_Root()) { 951 return 0; 952 } 953 954 uint def_pre_order = get_block_for_node(def)->_pre_order; 955 uint latency = 0; 956 957 // If the use is not a projection, then it is simple... 958 if (!use->is_Proj()) { 959 #ifndef PRODUCT 960 if (trace_opto_pipelining()) { 961 tty->print("# out(): "); 962 use->dump(); 963 } 964 #endif 965 966 uint use_pre_order = get_block_for_node(use)->_pre_order; 967 968 if (use_pre_order < def_pre_order) 969 return 0; 970 971 if (use_pre_order == def_pre_order && use->is_Phi()) 972 return 0; 973 974 uint nlen = use->len(); 975 uint nl = get_latency_for_node(use); 976 977 for ( uint j=0; j<nlen; j++ ) { 978 if (use->in(j) == n) { 979 // Change this if we want local latencies 980 uint ul = use->latency(j); 981 uint l = ul + nl; 982 if (latency < l) latency = l; 983 #ifndef PRODUCT 984 if (trace_opto_pipelining()) { 985 tty->print_cr("# %d + edge_latency(%d) == %d -> %d, latency = %d", 986 nl, j, ul, l, latency); 987 } 988 #endif 989 } 990 } 991 } else { 992 // This is a projection, just grab the latency of the use(s) 993 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) { 994 uint l = latency_from_use(use, def, use->fast_out(j)); 995 if (latency < l) latency = l; 996 } 997 } 998 999 return latency; 1000 } 1001 1002 //------------------------------latency_from_uses------------------------------ 1003 // Compute the latency of this instruction relative to all of it's uses. 1004 // This computes a number that increases as we approach the beginning of the 1005 // routine. 1006 void PhaseCFG::latency_from_uses(Node *n) { 1007 // Set the latency for this instruction 1008 #ifndef PRODUCT 1009 if (trace_opto_pipelining()) { 1010 tty->print("# latency_from_outputs: node_latency[%d] = %d for node", n->_idx, get_latency_for_node(n)); 1011 dump(); 1012 } 1013 #endif 1014 uint latency=0; 1015 const Node *def = n->is_Proj() ? n->in(0): n; 1016 1017 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1018 uint l = latency_from_use(n, def, n->fast_out(i)); 1019 1020 if (latency < l) latency = l; 1021 } 1022 1023 set_latency_for_node(n, latency); 1024 } 1025 1026 //------------------------------hoist_to_cheaper_block------------------------- 1027 // Pick a block for node self, between early and LCA, that is a cheaper 1028 // alternative to LCA. 1029 Block* PhaseCFG::hoist_to_cheaper_block(Block* LCA, Block* early, Node* self) { 1030 const double delta = 1+PROB_UNLIKELY_MAG(4); 1031 Block* least = LCA; 1032 double least_freq = least->_freq; 1033 uint target = get_latency_for_node(self); 1034 uint start_latency = get_latency_for_node(LCA->head()); 1035 uint end_latency = get_latency_for_node(LCA->get_node(LCA->end_idx())); 1036 bool in_latency = (target <= start_latency); 1037 const Block* root_block = get_block_for_node(_root); 1038 1039 // Turn off latency scheduling if scheduling is just plain off 1040 if (!C->do_scheduling()) 1041 in_latency = true; 1042 1043 // Do not hoist (to cover latency) instructions which target a 1044 // single register. Hoisting stretches the live range of the 1045 // single register and may force spilling. 1046 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL; 1047 if (mach && mach->out_RegMask().is_bound1() && mach->out_RegMask().is_NotEmpty()) 1048 in_latency = true; 1049 1050 #ifndef PRODUCT 1051 if (trace_opto_pipelining()) { 1052 tty->print("# Find cheaper block for latency %d: ", get_latency_for_node(self)); 1053 self->dump(); 1054 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g", 1055 LCA->_pre_order, 1056 LCA->head()->_idx, 1057 start_latency, 1058 LCA->get_node(LCA->end_idx())->_idx, 1059 end_latency, 1060 least_freq); 1061 } 1062 #endif 1063 1064 int cand_cnt = 0; // number of candidates tried 1065 1066 // Walk up the dominator tree from LCA (Lowest common ancestor) to 1067 // the earliest legal location. Capture the least execution frequency. 1068 while (LCA != early) { 1069 LCA = LCA->_idom; // Follow up the dominator tree 1070 1071 if (LCA == NULL) { 1072 // Bailout without retry 1073 C->record_method_not_compilable("late schedule failed: LCA == NULL"); 1074 return least; 1075 } 1076 1077 // Don't hoist machine instructions to the root basic block 1078 if (mach && LCA == root_block) 1079 break; 1080 1081 uint start_lat = get_latency_for_node(LCA->head()); 1082 uint end_idx = LCA->end_idx(); 1083 uint end_lat = get_latency_for_node(LCA->get_node(end_idx)); 1084 double LCA_freq = LCA->_freq; 1085 #ifndef PRODUCT 1086 if (trace_opto_pipelining()) { 1087 tty->print_cr("# B%d: start latency for [%4d]=%d, end latency for [%4d]=%d, freq=%g", 1088 LCA->_pre_order, LCA->head()->_idx, start_lat, end_idx, end_lat, LCA_freq); 1089 } 1090 #endif 1091 cand_cnt++; 1092 if (LCA_freq < least_freq || // Better Frequency 1093 (StressGCM && Compile::randomized_select(cand_cnt)) || // Should be randomly accepted in stress mode 1094 (!StressGCM && // Otherwise, choose with latency 1095 !in_latency && // No block containing latency 1096 LCA_freq < least_freq * delta && // No worse frequency 1097 target >= end_lat && // within latency range 1098 !self->is_iteratively_computed() ) // But don't hoist IV increments 1099 // because they may end up above other uses of their phi forcing 1100 // their result register to be different from their input. 1101 ) { 1102 least = LCA; // Found cheaper block 1103 least_freq = LCA_freq; 1104 start_latency = start_lat; 1105 end_latency = end_lat; 1106 if (target <= start_lat) 1107 in_latency = true; 1108 } 1109 } 1110 1111 #ifndef PRODUCT 1112 if (trace_opto_pipelining()) { 1113 tty->print_cr("# Choose block B%d with start latency=%d and freq=%g", 1114 least->_pre_order, start_latency, least_freq); 1115 } 1116 #endif 1117 1118 // See if the latency needs to be updated 1119 if (target < end_latency) { 1120 #ifndef PRODUCT 1121 if (trace_opto_pipelining()) { 1122 tty->print_cr("# Change latency for [%4d] from %d to %d", self->_idx, target, end_latency); 1123 } 1124 #endif 1125 set_latency_for_node(self, end_latency); 1126 partial_latency_of_defs(self); 1127 } 1128 1129 return least; 1130 } 1131 1132 1133 //------------------------------schedule_late----------------------------------- 1134 // Now schedule all codes as LATE as possible. This is the LCA in the 1135 // dominator tree of all USES of a value. Pick the block with the least 1136 // loop nesting depth that is lowest in the dominator tree. 1137 extern const char must_clone[]; 1138 void PhaseCFG::schedule_late(VectorSet &visited, Node_List &stack) { 1139 #ifndef PRODUCT 1140 if (trace_opto_pipelining()) 1141 tty->print("\n#---- schedule_late ----\n"); 1142 #endif 1143 1144 Node_Backward_Iterator iter((Node *)_root, visited, stack, *this); 1145 Node *self; 1146 1147 // Walk over all the nodes from last to first 1148 while (self = iter.next()) { 1149 Block* early = get_block_for_node(self); // Earliest legal placement 1150 1151 if (self->is_top()) { 1152 // Top node goes in bb #2 with other constants. 1153 // It must be special-cased, because it has no out edges. 1154 early->add_inst(self); 1155 continue; 1156 } 1157 1158 // No uses, just terminate 1159 if (self->outcnt() == 0) { 1160 assert(self->is_MachProj(), "sanity"); 1161 continue; // Must be a dead machine projection 1162 } 1163 1164 // If node is pinned in the block, then no scheduling can be done. 1165 if( self->pinned() ) // Pinned in block? 1166 continue; 1167 1168 MachNode* mach = self->is_Mach() ? self->as_Mach() : NULL; 1169 if (mach) { 1170 switch (mach->ideal_Opcode()) { 1171 case Op_CreateEx: 1172 // Don't move exception creation 1173 early->add_inst(self); 1174 continue; 1175 break; 1176 case Op_CheckCastPP: 1177 // Don't move CheckCastPP nodes away from their input, if the input 1178 // is a rawptr (5071820). 1179 Node *def = self->in(1); 1180 if (def != NULL && def->bottom_type()->base() == Type::RawPtr) { 1181 early->add_inst(self); 1182 #ifdef ASSERT 1183 _raw_oops.push(def); 1184 #endif 1185 continue; 1186 } 1187 break; 1188 } 1189 } 1190 1191 // Gather LCA of all uses 1192 Block *LCA = NULL; 1193 { 1194 for (DUIterator_Fast imax, i = self->fast_outs(imax); i < imax; i++) { 1195 // For all uses, find LCA 1196 Node* use = self->fast_out(i); 1197 LCA = raise_LCA_above_use(LCA, use, self, this); 1198 } 1199 } // (Hide defs of imax, i from rest of block.) 1200 1201 // Place temps in the block of their use. This isn't a 1202 // requirement for correctness but it reduces useless 1203 // interference between temps and other nodes. 1204 if (mach != NULL && mach->is_MachTemp()) { 1205 map_node_to_block(self, LCA); 1206 LCA->add_inst(self); 1207 continue; 1208 } 1209 1210 // Check if 'self' could be anti-dependent on memory 1211 if (self->needs_anti_dependence_check()) { 1212 // Hoist LCA above possible-defs and insert anti-dependences to 1213 // defs in new LCA block. 1214 LCA = insert_anti_dependences(LCA, self); 1215 } 1216 1217 if (early->_dom_depth > LCA->_dom_depth) { 1218 // Somehow the LCA has moved above the earliest legal point. 1219 // (One way this can happen is via memory_early_block.) 1220 if (C->subsume_loads() == true && !C->failing()) { 1221 // Retry with subsume_loads == false 1222 // If this is the first failure, the sentinel string will "stick" 1223 // to the Compile object, and the C2Compiler will see it and retry. 1224 C->record_failure(C2Compiler::retry_no_subsuming_loads()); 1225 } else { 1226 // Bailout without retry when (early->_dom_depth > LCA->_dom_depth) 1227 C->record_method_not_compilable("late schedule failed: incorrect graph"); 1228 } 1229 return; 1230 } 1231 1232 // If there is no opportunity to hoist, then we're done. 1233 // In stress mode, try to hoist even the single operations. 1234 bool try_to_hoist = StressGCM || (LCA != early); 1235 1236 // Must clone guys stay next to use; no hoisting allowed. 1237 // Also cannot hoist guys that alter memory or are otherwise not 1238 // allocatable (hoisting can make a value live longer, leading to 1239 // anti and output dependency problems which are normally resolved 1240 // by the register allocator giving everyone a different register). 1241 if (mach != NULL && must_clone[mach->ideal_Opcode()]) 1242 try_to_hoist = false; 1243 1244 Block* late = NULL; 1245 if (try_to_hoist) { 1246 // Now find the block with the least execution frequency. 1247 // Start at the latest schedule and work up to the earliest schedule 1248 // in the dominator tree. Thus the Node will dominate all its uses. 1249 late = hoist_to_cheaper_block(LCA, early, self); 1250 } else { 1251 // Just use the LCA of the uses. 1252 late = LCA; 1253 } 1254 1255 // Put the node into target block 1256 schedule_node_into_block(self, late); 1257 1258 #ifdef ASSERT 1259 if (self->needs_anti_dependence_check()) { 1260 // since precedence edges are only inserted when we're sure they 1261 // are needed make sure that after placement in a block we don't 1262 // need any new precedence edges. 1263 verify_anti_dependences(late, self); 1264 } 1265 #endif 1266 } // Loop until all nodes have been visited 1267 1268 } // end ScheduleLate 1269 1270 //------------------------------GlobalCodeMotion------------------------------- 1271 void PhaseCFG::global_code_motion() { 1272 ResourceMark rm; 1273 1274 #ifndef PRODUCT 1275 if (trace_opto_pipelining()) { 1276 tty->print("\n---- Start GlobalCodeMotion ----\n"); 1277 } 1278 #endif 1279 1280 // Initialize the node to block mapping for things on the proj_list 1281 for (uint i = 0; i < _matcher.number_of_projections(); i++) { 1282 unmap_node_from_block(_matcher.get_projection(i)); 1283 } 1284 1285 // Set the basic block for Nodes pinned into blocks 1286 Arena* arena = Thread::current()->resource_area(); 1287 VectorSet visited(arena); 1288 schedule_pinned_nodes(visited); 1289 1290 // Find the earliest Block any instruction can be placed in. Some 1291 // instructions are pinned into Blocks. Unpinned instructions can 1292 // appear in last block in which all their inputs occur. 1293 visited.Clear(); 1294 Node_List stack(arena); 1295 // Pre-grow the list 1296 stack.map((C->unique() >> 1) + 16, NULL); 1297 if (!schedule_early(visited, stack)) { 1298 // Bailout without retry 1299 C->record_method_not_compilable("early schedule failed"); 1300 return; 1301 } 1302 1303 // Build Def-Use edges. 1304 // Compute the latency information (via backwards walk) for all the 1305 // instructions in the graph 1306 _node_latency = new GrowableArray<uint>(); // resource_area allocation 1307 1308 if (C->do_scheduling()) { 1309 compute_latencies_backwards(visited, stack); 1310 } 1311 1312 // Now schedule all codes as LATE as possible. This is the LCA in the 1313 // dominator tree of all USES of a value. Pick the block with the least 1314 // loop nesting depth that is lowest in the dominator tree. 1315 // ( visited.Clear() called in schedule_late()->Node_Backward_Iterator() ) 1316 schedule_late(visited, stack); 1317 if (C->failing()) { 1318 // schedule_late fails only when graph is incorrect. 1319 assert(!VerifyGraphEdges, "verification should have failed"); 1320 return; 1321 } 1322 1323 #ifndef PRODUCT 1324 if (trace_opto_pipelining()) { 1325 tty->print("\n---- Detect implicit null checks ----\n"); 1326 } 1327 #endif 1328 1329 // Detect implicit-null-check opportunities. Basically, find NULL checks 1330 // with suitable memory ops nearby. Use the memory op to do the NULL check. 1331 // I can generate a memory op if there is not one nearby. 1332 if (C->is_method_compilation()) { 1333 // By reversing the loop direction we get a very minor gain on mpegaudio. 1334 // Feel free to revert to a forward loop for clarity. 1335 // for( int i=0; i < (int)matcher._null_check_tests.size(); i+=2 ) { 1336 for (int i = _matcher._null_check_tests.size() - 2; i >= 0; i -= 2) { 1337 Node* proj = _matcher._null_check_tests[i]; 1338 Node* val = _matcher._null_check_tests[i + 1]; 1339 Block* block = get_block_for_node(proj); 1340 implicit_null_check(block, proj, val, C->allowed_deopt_reasons()); 1341 // The implicit_null_check will only perform the transformation 1342 // if the null branch is truly uncommon, *and* it leads to an 1343 // uncommon trap. Combined with the too_many_traps guards 1344 // above, this prevents SEGV storms reported in 6366351, 1345 // by recompiling offending methods without this optimization. 1346 } 1347 } 1348 1349 #ifndef PRODUCT 1350 if (trace_opto_pipelining()) { 1351 tty->print("\n---- Start Local Scheduling ----\n"); 1352 } 1353 #endif 1354 1355 // Schedule locally. Right now a simple topological sort. 1356 // Later, do a real latency aware scheduler. 1357 GrowableArray<int> ready_cnt(C->unique(), C->unique(), -1); 1358 visited.Clear(); 1359 for (uint i = 0; i < number_of_blocks(); i++) { 1360 Block* block = get_block(i); 1361 if (!schedule_local(block, ready_cnt, visited)) { 1362 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 1363 C->record_method_not_compilable("local schedule failed"); 1364 } 1365 return; 1366 } 1367 } 1368 1369 // If we inserted any instructions between a Call and his CatchNode, 1370 // clone the instructions on all paths below the Catch. 1371 for (uint i = 0; i < number_of_blocks(); i++) { 1372 Block* block = get_block(i); 1373 call_catch_cleanup(block); 1374 } 1375 1376 #ifndef PRODUCT 1377 if (trace_opto_pipelining()) { 1378 tty->print("\n---- After GlobalCodeMotion ----\n"); 1379 for (uint i = 0; i < number_of_blocks(); i++) { 1380 Block* block = get_block(i); 1381 block->dump(); 1382 } 1383 } 1384 #endif 1385 // Dead. 1386 _node_latency = (GrowableArray<uint> *)0xdeadbeef; 1387 } 1388 1389 bool PhaseCFG::do_global_code_motion() { 1390 1391 build_dominator_tree(); 1392 if (C->failing()) { 1393 return false; 1394 } 1395 1396 NOT_PRODUCT( C->verify_graph_edges(); ) 1397 1398 estimate_block_frequency(); 1399 1400 global_code_motion(); 1401 1402 if (C->failing()) { 1403 return false; 1404 } 1405 1406 return true; 1407 } 1408 1409 //------------------------------Estimate_Block_Frequency----------------------- 1410 // Estimate block frequencies based on IfNode probabilities. 1411 void PhaseCFG::estimate_block_frequency() { 1412 1413 // Force conditional branches leading to uncommon traps to be unlikely, 1414 // not because we get to the uncommon_trap with less relative frequency, 1415 // but because an uncommon_trap typically causes a deopt, so we only get 1416 // there once. 1417 if (C->do_freq_based_layout()) { 1418 Block_List worklist; 1419 Block* root_blk = get_block(0); 1420 for (uint i = 1; i < root_blk->num_preds(); i++) { 1421 Block *pb = get_block_for_node(root_blk->pred(i)); 1422 if (pb->has_uncommon_code()) { 1423 worklist.push(pb); 1424 } 1425 } 1426 while (worklist.size() > 0) { 1427 Block* uct = worklist.pop(); 1428 if (uct == get_root_block()) { 1429 continue; 1430 } 1431 for (uint i = 1; i < uct->num_preds(); i++) { 1432 Block *pb = get_block_for_node(uct->pred(i)); 1433 if (pb->_num_succs == 1) { 1434 worklist.push(pb); 1435 } else if (pb->num_fall_throughs() == 2) { 1436 pb->update_uncommon_branch(uct); 1437 } 1438 } 1439 } 1440 } 1441 1442 // Create the loop tree and calculate loop depth. 1443 _root_loop = create_loop_tree(); 1444 _root_loop->compute_loop_depth(0); 1445 1446 // Compute block frequency of each block, relative to a single loop entry. 1447 _root_loop->compute_freq(); 1448 1449 // Adjust all frequencies to be relative to a single method entry 1450 _root_loop->_freq = 1.0; 1451 _root_loop->scale_freq(); 1452 1453 // Save outmost loop frequency for LRG frequency threshold 1454 _outer_loop_frequency = _root_loop->outer_loop_freq(); 1455 1456 // force paths ending at uncommon traps to be infrequent 1457 if (!C->do_freq_based_layout()) { 1458 Block_List worklist; 1459 Block* root_blk = get_block(0); 1460 for (uint i = 1; i < root_blk->num_preds(); i++) { 1461 Block *pb = get_block_for_node(root_blk->pred(i)); 1462 if (pb->has_uncommon_code()) { 1463 worklist.push(pb); 1464 } 1465 } 1466 while (worklist.size() > 0) { 1467 Block* uct = worklist.pop(); 1468 uct->_freq = PROB_MIN; 1469 for (uint i = 1; i < uct->num_preds(); i++) { 1470 Block *pb = get_block_for_node(uct->pred(i)); 1471 if (pb->_num_succs == 1 && pb->_freq > PROB_MIN) { 1472 worklist.push(pb); 1473 } 1474 } 1475 } 1476 } 1477 1478 #ifdef ASSERT 1479 for (uint i = 0; i < number_of_blocks(); i++) { 1480 Block* b = get_block(i); 1481 assert(b->_freq >= MIN_BLOCK_FREQUENCY, "Register Allocator requires meaningful block frequency"); 1482 } 1483 #endif 1484 1485 #ifndef PRODUCT 1486 if (PrintCFGBlockFreq) { 1487 tty->print_cr("CFG Block Frequencies"); 1488 _root_loop->dump_tree(); 1489 if (Verbose) { 1490 tty->print_cr("PhaseCFG dump"); 1491 dump(); 1492 tty->print_cr("Node dump"); 1493 _root->dump(99999); 1494 } 1495 } 1496 #endif 1497 } 1498 1499 //----------------------------create_loop_tree-------------------------------- 1500 // Create a loop tree from the CFG 1501 CFGLoop* PhaseCFG::create_loop_tree() { 1502 1503 #ifdef ASSERT 1504 assert(get_block(0) == get_root_block(), "first block should be root block"); 1505 for (uint i = 0; i < number_of_blocks(); i++) { 1506 Block* block = get_block(i); 1507 // Check that _loop field are clear...we could clear them if not. 1508 assert(block->_loop == NULL, "clear _loop expected"); 1509 // Sanity check that the RPO numbering is reflected in the _blocks array. 1510 // It doesn't have to be for the loop tree to be built, but if it is not, 1511 // then the blocks have been reordered since dom graph building...which 1512 // may question the RPO numbering 1513 assert(block->_rpo == i, "unexpected reverse post order number"); 1514 } 1515 #endif 1516 1517 int idct = 0; 1518 CFGLoop* root_loop = new CFGLoop(idct++); 1519 1520 Block_List worklist; 1521 1522 // Assign blocks to loops 1523 for(uint i = number_of_blocks() - 1; i > 0; i-- ) { // skip Root block 1524 Block* block = get_block(i); 1525 1526 if (block->head()->is_Loop()) { 1527 Block* loop_head = block; 1528 assert(loop_head->num_preds() - 1 == 2, "loop must have 2 predecessors"); 1529 Node* tail_n = loop_head->pred(LoopNode::LoopBackControl); 1530 Block* tail = get_block_for_node(tail_n); 1531 1532 // Defensively filter out Loop nodes for non-single-entry loops. 1533 // For all reasonable loops, the head occurs before the tail in RPO. 1534 if (i <= tail->_rpo) { 1535 1536 // The tail and (recursive) predecessors of the tail 1537 // are made members of a new loop. 1538 1539 assert(worklist.size() == 0, "nonempty worklist"); 1540 CFGLoop* nloop = new CFGLoop(idct++); 1541 assert(loop_head->_loop == NULL, "just checking"); 1542 loop_head->_loop = nloop; 1543 // Add to nloop so push_pred() will skip over inner loops 1544 nloop->add_member(loop_head); 1545 nloop->push_pred(loop_head, LoopNode::LoopBackControl, worklist, this); 1546 1547 while (worklist.size() > 0) { 1548 Block* member = worklist.pop(); 1549 if (member != loop_head) { 1550 for (uint j = 1; j < member->num_preds(); j++) { 1551 nloop->push_pred(member, j, worklist, this); 1552 } 1553 } 1554 } 1555 } 1556 } 1557 } 1558 1559 // Create a member list for each loop consisting 1560 // of both blocks and (immediate child) loops. 1561 for (uint i = 0; i < number_of_blocks(); i++) { 1562 Block* block = get_block(i); 1563 CFGLoop* lp = block->_loop; 1564 if (lp == NULL) { 1565 // Not assigned to a loop. Add it to the method's pseudo loop. 1566 block->_loop = root_loop; 1567 lp = root_loop; 1568 } 1569 if (lp == root_loop || block != lp->head()) { // loop heads are already members 1570 lp->add_member(block); 1571 } 1572 if (lp != root_loop) { 1573 if (lp->parent() == NULL) { 1574 // Not a nested loop. Make it a child of the method's pseudo loop. 1575 root_loop->add_nested_loop(lp); 1576 } 1577 if (block == lp->head()) { 1578 // Add nested loop to member list of parent loop. 1579 lp->parent()->add_member(lp); 1580 } 1581 } 1582 } 1583 1584 return root_loop; 1585 } 1586 1587 //------------------------------push_pred-------------------------------------- 1588 void CFGLoop::push_pred(Block* blk, int i, Block_List& worklist, PhaseCFG* cfg) { 1589 Node* pred_n = blk->pred(i); 1590 Block* pred = cfg->get_block_for_node(pred_n); 1591 CFGLoop *pred_loop = pred->_loop; 1592 if (pred_loop == NULL) { 1593 // Filter out blocks for non-single-entry loops. 1594 // For all reasonable loops, the head occurs before the tail in RPO. 1595 if (pred->_rpo > head()->_rpo) { 1596 pred->_loop = this; 1597 worklist.push(pred); 1598 } 1599 } else if (pred_loop != this) { 1600 // Nested loop. 1601 while (pred_loop->_parent != NULL && pred_loop->_parent != this) { 1602 pred_loop = pred_loop->_parent; 1603 } 1604 // Make pred's loop be a child 1605 if (pred_loop->_parent == NULL) { 1606 add_nested_loop(pred_loop); 1607 // Continue with loop entry predecessor. 1608 Block* pred_head = pred_loop->head(); 1609 assert(pred_head->num_preds() - 1 == 2, "loop must have 2 predecessors"); 1610 assert(pred_head != head(), "loop head in only one loop"); 1611 push_pred(pred_head, LoopNode::EntryControl, worklist, cfg); 1612 } else { 1613 assert(pred_loop->_parent == this && _parent == NULL, "just checking"); 1614 } 1615 } 1616 } 1617 1618 //------------------------------add_nested_loop-------------------------------- 1619 // Make cl a child of the current loop in the loop tree. 1620 void CFGLoop::add_nested_loop(CFGLoop* cl) { 1621 assert(_parent == NULL, "no parent yet"); 1622 assert(cl != this, "not my own parent"); 1623 cl->_parent = this; 1624 CFGLoop* ch = _child; 1625 if (ch == NULL) { 1626 _child = cl; 1627 } else { 1628 while (ch->_sibling != NULL) { ch = ch->_sibling; } 1629 ch->_sibling = cl; 1630 } 1631 } 1632 1633 //------------------------------compute_loop_depth----------------------------- 1634 // Store the loop depth in each CFGLoop object. 1635 // Recursively walk the children to do the same for them. 1636 void CFGLoop::compute_loop_depth(int depth) { 1637 _depth = depth; 1638 CFGLoop* ch = _child; 1639 while (ch != NULL) { 1640 ch->compute_loop_depth(depth + 1); 1641 ch = ch->_sibling; 1642 } 1643 } 1644 1645 //------------------------------compute_freq----------------------------------- 1646 // Compute the frequency of each block and loop, relative to a single entry 1647 // into the dominating loop head. 1648 void CFGLoop::compute_freq() { 1649 // Bottom up traversal of loop tree (visit inner loops first.) 1650 // Set loop head frequency to 1.0, then transitively 1651 // compute frequency for all successors in the loop, 1652 // as well as for each exit edge. Inner loops are 1653 // treated as single blocks with loop exit targets 1654 // as the successor blocks. 1655 1656 // Nested loops first 1657 CFGLoop* ch = _child; 1658 while (ch != NULL) { 1659 ch->compute_freq(); 1660 ch = ch->_sibling; 1661 } 1662 assert (_members.length() > 0, "no empty loops"); 1663 Block* hd = head(); 1664 hd->_freq = 1.0f; 1665 for (int i = 0; i < _members.length(); i++) { 1666 CFGElement* s = _members.at(i); 1667 float freq = s->_freq; 1668 if (s->is_block()) { 1669 Block* b = s->as_Block(); 1670 for (uint j = 0; j < b->_num_succs; j++) { 1671 Block* sb = b->_succs[j]; 1672 update_succ_freq(sb, freq * b->succ_prob(j)); 1673 } 1674 } else { 1675 CFGLoop* lp = s->as_CFGLoop(); 1676 assert(lp->_parent == this, "immediate child"); 1677 for (int k = 0; k < lp->_exits.length(); k++) { 1678 Block* eb = lp->_exits.at(k).get_target(); 1679 float prob = lp->_exits.at(k).get_prob(); 1680 update_succ_freq(eb, freq * prob); 1681 } 1682 } 1683 } 1684 1685 // For all loops other than the outer, "method" loop, 1686 // sum and normalize the exit probability. The "method" loop 1687 // should keep the initial exit probability of 1, so that 1688 // inner blocks do not get erroneously scaled. 1689 if (_depth != 0) { 1690 // Total the exit probabilities for this loop. 1691 float exits_sum = 0.0f; 1692 for (int i = 0; i < _exits.length(); i++) { 1693 exits_sum += _exits.at(i).get_prob(); 1694 } 1695 1696 // Normalize the exit probabilities. Until now, the 1697 // probabilities estimate the possibility of exit per 1698 // a single loop iteration; afterward, they estimate 1699 // the probability of exit per loop entry. 1700 for (int i = 0; i < _exits.length(); i++) { 1701 Block* et = _exits.at(i).get_target(); 1702 float new_prob = 0.0f; 1703 if (_exits.at(i).get_prob() > 0.0f) { 1704 new_prob = _exits.at(i).get_prob() / exits_sum; 1705 } 1706 BlockProbPair bpp(et, new_prob); 1707 _exits.at_put(i, bpp); 1708 } 1709 1710 // Save the total, but guard against unreasonable probability, 1711 // as the value is used to estimate the loop trip count. 1712 // An infinite trip count would blur relative block 1713 // frequencies. 1714 if (exits_sum > 1.0f) exits_sum = 1.0; 1715 if (exits_sum < PROB_MIN) exits_sum = PROB_MIN; 1716 _exit_prob = exits_sum; 1717 } 1718 } 1719 1720 //------------------------------succ_prob------------------------------------- 1721 // Determine the probability of reaching successor 'i' from the receiver block. 1722 float Block::succ_prob(uint i) { 1723 int eidx = end_idx(); 1724 Node *n = get_node(eidx); // Get ending Node 1725 1726 int op = n->Opcode(); 1727 if (n->is_Mach()) { 1728 if (n->is_MachNullCheck()) { 1729 // Can only reach here if called after lcm. The original Op_If is gone, 1730 // so we attempt to infer the probability from one or both of the 1731 // successor blocks. 1732 assert(_num_succs == 2, "expecting 2 successors of a null check"); 1733 // If either successor has only one predecessor, then the 1734 // probability estimate can be derived using the 1735 // relative frequency of the successor and this block. 1736 if (_succs[i]->num_preds() == 2) { 1737 return _succs[i]->_freq / _freq; 1738 } else if (_succs[1-i]->num_preds() == 2) { 1739 return 1 - (_succs[1-i]->_freq / _freq); 1740 } else { 1741 // Estimate using both successor frequencies 1742 float freq = _succs[i]->_freq; 1743 return freq / (freq + _succs[1-i]->_freq); 1744 } 1745 } 1746 op = n->as_Mach()->ideal_Opcode(); 1747 } 1748 1749 1750 // Switch on branch type 1751 switch( op ) { 1752 case Op_CountedLoopEnd: 1753 case Op_If: { 1754 assert (i < 2, "just checking"); 1755 // Conditionals pass on only part of their frequency 1756 float prob = n->as_MachIf()->_prob; 1757 assert(prob >= 0.0 && prob <= 1.0, "out of range probability"); 1758 // If succ[i] is the FALSE branch, invert path info 1759 if( get_node(i + eidx + 1)->Opcode() == Op_IfFalse ) { 1760 return 1.0f - prob; // not taken 1761 } else { 1762 return prob; // taken 1763 } 1764 } 1765 1766 case Op_Jump: 1767 // Divide the frequency between all successors evenly 1768 return 1.0f/_num_succs; 1769 1770 case Op_Catch: { 1771 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj(); 1772 if (ci->_con == CatchProjNode::fall_through_index) { 1773 // Fall-thru path gets the lion's share. 1774 return 1.0f - PROB_UNLIKELY_MAG(5)*_num_succs; 1775 } else { 1776 // Presume exceptional paths are equally unlikely 1777 return PROB_UNLIKELY_MAG(5); 1778 } 1779 } 1780 1781 case Op_Root: 1782 case Op_Goto: 1783 // Pass frequency straight thru to target 1784 return 1.0f; 1785 1786 case Op_NeverBranch: 1787 return 0.0f; 1788 1789 case Op_TailCall: 1790 case Op_TailJump: 1791 case Op_Return: 1792 case Op_Halt: 1793 case Op_Rethrow: 1794 // Do not push out freq to root block 1795 return 0.0f; 1796 1797 default: 1798 ShouldNotReachHere(); 1799 } 1800 1801 return 0.0f; 1802 } 1803 1804 //------------------------------num_fall_throughs----------------------------- 1805 // Return the number of fall-through candidates for a block 1806 int Block::num_fall_throughs() { 1807 int eidx = end_idx(); 1808 Node *n = get_node(eidx); // Get ending Node 1809 1810 int op = n->Opcode(); 1811 if (n->is_Mach()) { 1812 if (n->is_MachNullCheck()) { 1813 // In theory, either side can fall-thru, for simplicity sake, 1814 // let's say only the false branch can now. 1815 return 1; 1816 } 1817 op = n->as_Mach()->ideal_Opcode(); 1818 } 1819 1820 // Switch on branch type 1821 switch( op ) { 1822 case Op_CountedLoopEnd: 1823 case Op_If: 1824 return 2; 1825 1826 case Op_Root: 1827 case Op_Goto: 1828 return 1; 1829 1830 case Op_Catch: { 1831 for (uint i = 0; i < _num_succs; i++) { 1832 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj(); 1833 if (ci->_con == CatchProjNode::fall_through_index) { 1834 return 1; 1835 } 1836 } 1837 return 0; 1838 } 1839 1840 case Op_Jump: 1841 case Op_NeverBranch: 1842 case Op_TailCall: 1843 case Op_TailJump: 1844 case Op_Return: 1845 case Op_Halt: 1846 case Op_Rethrow: 1847 return 0; 1848 1849 default: 1850 ShouldNotReachHere(); 1851 } 1852 1853 return 0; 1854 } 1855 1856 //------------------------------succ_fall_through----------------------------- 1857 // Return true if a specific successor could be fall-through target. 1858 bool Block::succ_fall_through(uint i) { 1859 int eidx = end_idx(); 1860 Node *n = get_node(eidx); // Get ending Node 1861 1862 int op = n->Opcode(); 1863 if (n->is_Mach()) { 1864 if (n->is_MachNullCheck()) { 1865 // In theory, either side can fall-thru, for simplicity sake, 1866 // let's say only the false branch can now. 1867 return get_node(i + eidx + 1)->Opcode() == Op_IfFalse; 1868 } 1869 op = n->as_Mach()->ideal_Opcode(); 1870 } 1871 1872 // Switch on branch type 1873 switch( op ) { 1874 case Op_CountedLoopEnd: 1875 case Op_If: 1876 case Op_Root: 1877 case Op_Goto: 1878 return true; 1879 1880 case Op_Catch: { 1881 const CatchProjNode *ci = get_node(i + eidx + 1)->as_CatchProj(); 1882 return ci->_con == CatchProjNode::fall_through_index; 1883 } 1884 1885 case Op_Jump: 1886 case Op_NeverBranch: 1887 case Op_TailCall: 1888 case Op_TailJump: 1889 case Op_Return: 1890 case Op_Halt: 1891 case Op_Rethrow: 1892 return false; 1893 1894 default: 1895 ShouldNotReachHere(); 1896 } 1897 1898 return false; 1899 } 1900 1901 //------------------------------update_uncommon_branch------------------------ 1902 // Update the probability of a two-branch to be uncommon 1903 void Block::update_uncommon_branch(Block* ub) { 1904 int eidx = end_idx(); 1905 Node *n = get_node(eidx); // Get ending Node 1906 1907 int op = n->as_Mach()->ideal_Opcode(); 1908 1909 assert(op == Op_CountedLoopEnd || op == Op_If, "must be a If"); 1910 assert(num_fall_throughs() == 2, "must be a two way branch block"); 1911 1912 // Which successor is ub? 1913 uint s; 1914 for (s = 0; s <_num_succs; s++) { 1915 if (_succs[s] == ub) break; 1916 } 1917 assert(s < 2, "uncommon successor must be found"); 1918 1919 // If ub is the true path, make the proability small, else 1920 // ub is the false path, and make the probability large 1921 bool invert = (get_node(s + eidx + 1)->Opcode() == Op_IfFalse); 1922 1923 // Get existing probability 1924 float p = n->as_MachIf()->_prob; 1925 1926 if (invert) p = 1.0 - p; 1927 if (p > PROB_MIN) { 1928 p = PROB_MIN; 1929 } 1930 if (invert) p = 1.0 - p; 1931 1932 n->as_MachIf()->_prob = p; 1933 } 1934 1935 //------------------------------update_succ_freq------------------------------- 1936 // Update the appropriate frequency associated with block 'b', a successor of 1937 // a block in this loop. 1938 void CFGLoop::update_succ_freq(Block* b, float freq) { 1939 if (b->_loop == this) { 1940 if (b == head()) { 1941 // back branch within the loop 1942 // Do nothing now, the loop carried frequency will be 1943 // adjust later in scale_freq(). 1944 } else { 1945 // simple branch within the loop 1946 b->_freq += freq; 1947 } 1948 } else if (!in_loop_nest(b)) { 1949 // branch is exit from this loop 1950 BlockProbPair bpp(b, freq); 1951 _exits.append(bpp); 1952 } else { 1953 // branch into nested loop 1954 CFGLoop* ch = b->_loop; 1955 ch->_freq += freq; 1956 } 1957 } 1958 1959 //------------------------------in_loop_nest----------------------------------- 1960 // Determine if block b is in the receiver's loop nest. 1961 bool CFGLoop::in_loop_nest(Block* b) { 1962 int depth = _depth; 1963 CFGLoop* b_loop = b->_loop; 1964 int b_depth = b_loop->_depth; 1965 if (depth == b_depth) { 1966 return true; 1967 } 1968 while (b_depth > depth) { 1969 b_loop = b_loop->_parent; 1970 b_depth = b_loop->_depth; 1971 } 1972 return b_loop == this; 1973 } 1974 1975 //------------------------------scale_freq------------------------------------- 1976 // Scale frequency of loops and blocks by trip counts from outer loops 1977 // Do a top down traversal of loop tree (visit outer loops first.) 1978 void CFGLoop::scale_freq() { 1979 float loop_freq = _freq * trip_count(); 1980 _freq = loop_freq; 1981 for (int i = 0; i < _members.length(); i++) { 1982 CFGElement* s = _members.at(i); 1983 float block_freq = s->_freq * loop_freq; 1984 if (g_isnan(block_freq) || block_freq < MIN_BLOCK_FREQUENCY) 1985 block_freq = MIN_BLOCK_FREQUENCY; 1986 s->_freq = block_freq; 1987 } 1988 CFGLoop* ch = _child; 1989 while (ch != NULL) { 1990 ch->scale_freq(); 1991 ch = ch->_sibling; 1992 } 1993 } 1994 1995 // Frequency of outer loop 1996 float CFGLoop::outer_loop_freq() const { 1997 if (_child != NULL) { 1998 return _child->_freq; 1999 } 2000 return _freq; 2001 } 2002 2003 #ifndef PRODUCT 2004 //------------------------------dump_tree-------------------------------------- 2005 void CFGLoop::dump_tree() const { 2006 dump(); 2007 if (_child != NULL) _child->dump_tree(); 2008 if (_sibling != NULL) _sibling->dump_tree(); 2009 } 2010 2011 //------------------------------dump------------------------------------------- 2012 void CFGLoop::dump() const { 2013 for (int i = 0; i < _depth; i++) tty->print(" "); 2014 tty->print("%s: %d trip_count: %6.0f freq: %6.0f\n", 2015 _depth == 0 ? "Method" : "Loop", _id, trip_count(), _freq); 2016 for (int i = 0; i < _depth; i++) tty->print(" "); 2017 tty->print(" members:", _id); 2018 int k = 0; 2019 for (int i = 0; i < _members.length(); i++) { 2020 if (k++ >= 6) { 2021 tty->print("\n "); 2022 for (int j = 0; j < _depth+1; j++) tty->print(" "); 2023 k = 0; 2024 } 2025 CFGElement *s = _members.at(i); 2026 if (s->is_block()) { 2027 Block *b = s->as_Block(); 2028 tty->print(" B%d(%6.3f)", b->_pre_order, b->_freq); 2029 } else { 2030 CFGLoop* lp = s->as_CFGLoop(); 2031 tty->print(" L%d(%6.3f)", lp->_id, lp->_freq); 2032 } 2033 } 2034 tty->print("\n"); 2035 for (int i = 0; i < _depth; i++) tty->print(" "); 2036 tty->print(" exits: "); 2037 k = 0; 2038 for (int i = 0; i < _exits.length(); i++) { 2039 if (k++ >= 7) { 2040 tty->print("\n "); 2041 for (int j = 0; j < _depth+1; j++) tty->print(" "); 2042 k = 0; 2043 } 2044 Block *blk = _exits.at(i).get_target(); 2045 float prob = _exits.at(i).get_prob(); 2046 tty->print(" ->%d@%d%%", blk->_pre_order, (int)(prob*100)); 2047 } 2048 tty->print("\n"); 2049 } 2050 #endif