1 /* 2 * Copyright (c) 1998, 2010, 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 "incls/_precompiled.incl" 26 #include "incls/_output.cpp.incl" 27 28 extern uint size_java_to_interp(); 29 extern uint reloc_java_to_interp(); 30 extern uint size_exception_handler(); 31 extern uint size_deopt_handler(); 32 33 #ifndef PRODUCT 34 #define DEBUG_ARG(x) , x 35 #else 36 #define DEBUG_ARG(x) 37 #endif 38 39 extern int emit_exception_handler(CodeBuffer &cbuf); 40 extern int emit_deopt_handler(CodeBuffer &cbuf); 41 42 //------------------------------Output----------------------------------------- 43 // Convert Nodes to instruction bits and pass off to the VM 44 void Compile::Output() { 45 // RootNode goes 46 assert( _cfg->_broot->_nodes.size() == 0, "" ); 47 48 // The number of new nodes (mostly MachNop) is proportional to 49 // the number of java calls and inner loops which are aligned. 50 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 51 C->inner_loops()*(OptoLoopAlignment-1)), 52 "out of nodes before code generation" ) ) { 53 return; 54 } 55 // Make sure I can find the Start Node 56 Block_Array& bbs = _cfg->_bbs; 57 Block *entry = _cfg->_blocks[1]; 58 Block *broot = _cfg->_broot; 59 60 const StartNode *start = entry->_nodes[0]->as_Start(); 61 62 // Replace StartNode with prolog 63 MachPrologNode *prolog = new (this) MachPrologNode(); 64 entry->_nodes.map( 0, prolog ); 65 bbs.map( prolog->_idx, entry ); 66 bbs.map( start->_idx, NULL ); // start is no longer in any block 67 68 // Virtual methods need an unverified entry point 69 70 if( is_osr_compilation() ) { 71 if( PoisonOSREntry ) { 72 // TODO: Should use a ShouldNotReachHereNode... 73 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 74 } 75 } else { 76 if( _method && !_method->flags().is_static() ) { 77 // Insert unvalidated entry point 78 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 79 } 80 81 } 82 83 84 // Break before main entry point 85 if( (_method && _method->break_at_execute()) 86 #ifndef PRODUCT 87 ||(OptoBreakpoint && is_method_compilation()) 88 ||(OptoBreakpointOSR && is_osr_compilation()) 89 ||(OptoBreakpointC2R && !_method) 90 #endif 91 ) { 92 // checking for _method means that OptoBreakpoint does not apply to 93 // runtime stubs or frame converters 94 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 95 } 96 97 // Insert epilogs before every return 98 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 99 Block *b = _cfg->_blocks[i]; 100 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point? 101 Node *m = b->end(); 102 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) { 103 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 104 b->add_inst( epilog ); 105 bbs.map(epilog->_idx, b); 106 //_regalloc->set_bad(epilog->_idx); // Already initialized this way. 107 } 108 } 109 } 110 111 # ifdef ENABLE_ZAP_DEAD_LOCALS 112 if ( ZapDeadCompiledLocals ) Insert_zap_nodes(); 113 # endif 114 115 ScheduleAndBundle(); 116 117 #ifndef PRODUCT 118 if (trace_opto_output()) { 119 tty->print("\n---- After ScheduleAndBundle ----\n"); 120 for (uint i = 0; i < _cfg->_num_blocks; i++) { 121 tty->print("\nBB#%03d:\n", i); 122 Block *bb = _cfg->_blocks[i]; 123 for (uint j = 0; j < bb->_nodes.size(); j++) { 124 Node *n = bb->_nodes[j]; 125 OptoReg::Name reg = _regalloc->get_reg_first(n); 126 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 127 n->dump(); 128 } 129 } 130 } 131 #endif 132 133 if (failing()) return; 134 135 BuildOopMaps(); 136 137 if (failing()) return; 138 139 Fill_buffer(); 140 } 141 142 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 143 // Determine if we need to generate a stack overflow check. 144 // Do it if the method is not a stub function and 145 // has java calls or has frame size > vm_page_size/8. 146 return (stub_function() == NULL && 147 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 148 } 149 150 bool Compile::need_register_stack_bang() const { 151 // Determine if we need to generate a register stack overflow check. 152 // This is only used on architectures which have split register 153 // and memory stacks (ie. IA64). 154 // Bang if the method is not a stub function and has java calls 155 return (stub_function() == NULL && has_java_calls()); 156 } 157 158 # ifdef ENABLE_ZAP_DEAD_LOCALS 159 160 161 // In order to catch compiler oop-map bugs, we have implemented 162 // a debugging mode called ZapDeadCompilerLocals. 163 // This mode causes the compiler to insert a call to a runtime routine, 164 // "zap_dead_locals", right before each place in compiled code 165 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 166 // The runtime routine checks that locations mapped as oops are really 167 // oops, that locations mapped as values do not look like oops, 168 // and that locations mapped as dead are not used later 169 // (by zapping them to an invalid address). 170 171 int Compile::_CompiledZap_count = 0; 172 173 void Compile::Insert_zap_nodes() { 174 bool skip = false; 175 176 177 // Dink with static counts because code code without the extra 178 // runtime calls is MUCH faster for debugging purposes 179 180 if ( CompileZapFirst == 0 ) ; // nothing special 181 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 182 else if ( CompileZapFirst == CompiledZap_count() ) 183 warning("starting zap compilation after skipping"); 184 185 if ( CompileZapLast == -1 ) ; // nothing special 186 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 187 else if ( CompileZapLast == CompiledZap_count() ) 188 warning("about to compile last zap"); 189 190 ++_CompiledZap_count; // counts skipped zaps, too 191 192 if ( skip ) return; 193 194 195 if ( _method == NULL ) 196 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 197 198 // Insert call to zap runtime stub before every node with an oop map 199 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 200 Block *b = _cfg->_blocks[i]; 201 for ( uint j = 0; j < b->_nodes.size(); ++j ) { 202 Node *n = b->_nodes[j]; 203 204 // Determining if we should insert a zap-a-lot node in output. 205 // We do that for all nodes that has oopmap info, except for calls 206 // to allocation. Calls to allocation passes in the old top-of-eden pointer 207 // and expect the C code to reset it. Hence, there can be no safepoints between 208 // the inlined-allocation and the call to new_Java, etc. 209 // We also cannot zap monitor calls, as they must hold the microlock 210 // during the call to Zap, which also wants to grab the microlock. 211 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 212 if ( insert ) { // it is MachSafePoint 213 if ( !n->is_MachCall() ) { 214 insert = false; 215 } else if ( n->is_MachCall() ) { 216 MachCallNode* call = n->as_MachCall(); 217 if (call->entry_point() == OptoRuntime::new_instance_Java() || 218 call->entry_point() == OptoRuntime::new_array_Java() || 219 call->entry_point() == OptoRuntime::multianewarray2_Java() || 220 call->entry_point() == OptoRuntime::multianewarray3_Java() || 221 call->entry_point() == OptoRuntime::multianewarray4_Java() || 222 call->entry_point() == OptoRuntime::multianewarray5_Java() || 223 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 224 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 225 ) { 226 insert = false; 227 } 228 } 229 if (insert) { 230 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 231 b->_nodes.insert( j, zap ); 232 _cfg->_bbs.map( zap->_idx, b ); 233 ++j; 234 } 235 } 236 } 237 } 238 } 239 240 241 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 242 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 243 CallStaticJavaNode* ideal_node = 244 new (this, tf->domain()->cnt()) CallStaticJavaNode( tf, 245 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 246 "call zap dead locals stub", 0, TypePtr::BOTTOM); 247 // We need to copy the OopMap from the site we're zapping at. 248 // We have to make a copy, because the zap site might not be 249 // a call site, and zap_dead is a call site. 250 OopMap* clone = node_to_check->oop_map()->deep_copy(); 251 252 // Add the cloned OopMap to the zap node 253 ideal_node->set_oop_map(clone); 254 return _matcher->match_sfpt(ideal_node); 255 } 256 257 //------------------------------is_node_getting_a_safepoint-------------------- 258 bool Compile::is_node_getting_a_safepoint( Node* n) { 259 // This code duplicates the logic prior to the call of add_safepoint 260 // below in this file. 261 if( n->is_MachSafePoint() ) return true; 262 return false; 263 } 264 265 # endif // ENABLE_ZAP_DEAD_LOCALS 266 267 //------------------------------compute_loop_first_inst_sizes------------------ 268 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 269 // of a loop. When aligning a loop we need to provide enough instructions 270 // in cpu's fetch buffer to feed decoders. The loop alignment could be 271 // avoided if we have enough instructions in fetch buffer at the head of a loop. 272 // By default, the size is set to 999999 by Block's constructor so that 273 // a loop will be aligned if the size is not reset here. 274 // 275 // Note: Mach instructions could contain several HW instructions 276 // so the size is estimated only. 277 // 278 void Compile::compute_loop_first_inst_sizes() { 279 // The next condition is used to gate the loop alignment optimization. 280 // Don't aligned a loop if there are enough instructions at the head of a loop 281 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 282 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 283 // equal to 11 bytes which is the largest address NOP instruction. 284 if( MaxLoopPad < OptoLoopAlignment-1 ) { 285 uint last_block = _cfg->_num_blocks-1; 286 for( uint i=1; i <= last_block; i++ ) { 287 Block *b = _cfg->_blocks[i]; 288 // Check the first loop's block which requires an alignment. 289 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) { 290 uint sum_size = 0; 291 uint inst_cnt = NumberOfLoopInstrToAlign; 292 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 293 294 // Check subsequent fallthrough blocks if the loop's first 295 // block(s) does not have enough instructions. 296 Block *nb = b; 297 while( inst_cnt > 0 && 298 i < last_block && 299 !_cfg->_blocks[i+1]->has_loop_alignment() && 300 !nb->has_successor(b) ) { 301 i++; 302 nb = _cfg->_blocks[i]; 303 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 304 } // while( inst_cnt > 0 && i < last_block ) 305 306 b->set_first_inst_size(sum_size); 307 } // f( b->head()->is_Loop() ) 308 } // for( i <= last_block ) 309 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 310 } 311 312 //----------------------Shorten_branches--------------------------------------- 313 // The architecture description provides short branch variants for some long 314 // branch instructions. Replace eligible long branches with short branches. 315 void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size) { 316 317 // fill in the nop array for bundling computations 318 MachNode *_nop_list[Bundle::_nop_count]; 319 Bundle::initialize_nops(_nop_list, this); 320 321 // ------------------ 322 // Compute size of each block, method size, and relocation information size 323 uint *jmp_end = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); 324 uint *blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1); 325 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) 326 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) 327 blk_starts[0] = 0; 328 329 // Initialize the sizes to 0 330 code_size = 0; // Size in bytes of generated code 331 stub_size = 0; // Size in bytes of all stub entries 332 // Size in bytes of all relocation entries, including those in local stubs. 333 // Start with 2-bytes of reloc info for the unvalidated entry point 334 reloc_size = 1; // Number of relocation entries 335 336 // Make three passes. The first computes pessimistic blk_starts, 337 // relative jmp_end and reloc_size information. The second performs 338 // short branch substitution using the pessimistic sizing. The 339 // third inserts nops where needed. 340 341 Node *nj; // tmp 342 343 // Step one, perform a pessimistic sizing pass. 344 uint i; 345 uint min_offset_from_last_call = 1; // init to a positive value 346 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 347 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 348 Block *b = _cfg->_blocks[i]; 349 350 // Sum all instruction sizes to compute block size 351 uint last_inst = b->_nodes.size(); 352 uint blk_size = 0; 353 for( uint j = 0; j<last_inst; j++ ) { 354 nj = b->_nodes[j]; 355 uint inst_size = nj->size(_regalloc); 356 blk_size += inst_size; 357 // Handle machine instruction nodes 358 if( nj->is_Mach() ) { 359 MachNode *mach = nj->as_Mach(); 360 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 361 reloc_size += mach->reloc(); 362 if( mach->is_MachCall() ) { 363 MachCallNode *mcall = mach->as_MachCall(); 364 // This destination address is NOT PC-relative 365 366 mcall->method_set((intptr_t)mcall->entry_point()); 367 368 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) { 369 stub_size += size_java_to_interp(); 370 reloc_size += reloc_java_to_interp(); 371 } 372 } else if (mach->is_MachSafePoint()) { 373 // If call/safepoint are adjacent, account for possible 374 // nop to disambiguate the two safepoints. 375 if (min_offset_from_last_call == 0) { 376 blk_size += nop_size; 377 } 378 } 379 } 380 min_offset_from_last_call += inst_size; 381 // Remember end of call offset 382 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 383 min_offset_from_last_call = 0; 384 } 385 } 386 387 // During short branch replacement, we store the relative (to blk_starts) 388 // end of jump in jmp_end, rather than the absolute end of jump. This 389 // is so that we do not need to recompute sizes of all nodes when we compute 390 // correct blk_starts in our next sizing pass. 391 jmp_end[i] = blk_size; 392 DEBUG_ONLY( jmp_target[i] = 0; ) 393 394 // When the next block starts a loop, we may insert pad NOP 395 // instructions. Since we cannot know our future alignment, 396 // assume the worst. 397 if( i<_cfg->_num_blocks-1 ) { 398 Block *nb = _cfg->_blocks[i+1]; 399 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 400 if( max_loop_pad > 0 ) { 401 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 402 blk_size += max_loop_pad; 403 } 404 } 405 406 // Save block size; update total method size 407 blk_starts[i+1] = blk_starts[i]+blk_size; 408 } 409 410 // Step two, replace eligible long jumps. 411 412 // Note: this will only get the long branches within short branch 413 // range. Another pass might detect more branches that became 414 // candidates because the shortening in the first pass exposed 415 // more opportunities. Unfortunately, this would require 416 // recomputing the starting and ending positions for the blocks 417 for( i=0; i<_cfg->_num_blocks; i++ ) { 418 Block *b = _cfg->_blocks[i]; 419 420 int j; 421 // Find the branch; ignore trailing NOPs. 422 for( j = b->_nodes.size()-1; j>=0; j-- ) { 423 nj = b->_nodes[j]; 424 if( !nj->is_Mach() || nj->as_Mach()->ideal_Opcode() != Op_Con ) 425 break; 426 } 427 428 if (j >= 0) { 429 if( nj->is_Mach() && nj->as_Mach()->may_be_short_branch() ) { 430 MachNode *mach = nj->as_Mach(); 431 // This requires the TRUE branch target be in succs[0] 432 uint bnum = b->non_connector_successor(0)->_pre_order; 433 uintptr_t target = blk_starts[bnum]; 434 if( mach->is_pc_relative() ) { 435 int offset = target-(blk_starts[i] + jmp_end[i]); 436 if (_matcher->is_short_branch_offset(mach->rule(), offset)) { 437 // We've got a winner. Replace this branch. 438 MachNode* replacement = mach->short_branch_version(this); 439 b->_nodes.map(j, replacement); 440 mach->subsume_by(replacement); 441 442 // Update the jmp_end size to save time in our 443 // next pass. 444 jmp_end[i] -= (mach->size(_regalloc) - replacement->size(_regalloc)); 445 DEBUG_ONLY( jmp_target[i] = bnum; ); 446 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 447 } 448 } else { 449 #ifndef PRODUCT 450 mach->dump(3); 451 #endif 452 Unimplemented(); 453 } 454 } 455 } 456 } 457 458 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 459 // of a loop. It is used to determine the padding for loop alignment. 460 compute_loop_first_inst_sizes(); 461 462 // Step 3, compute the offsets of all the labels 463 uint last_call_adr = max_uint; 464 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 465 // copy the offset of the beginning to the corresponding label 466 assert(labels[i].is_unused(), "cannot patch at this point"); 467 labels[i].bind_loc(blk_starts[i], CodeBuffer::SECT_INSTS); 468 469 // insert padding for any instructions that need it 470 Block *b = _cfg->_blocks[i]; 471 uint last_inst = b->_nodes.size(); 472 uint adr = blk_starts[i]; 473 for( uint j = 0; j<last_inst; j++ ) { 474 nj = b->_nodes[j]; 475 if( nj->is_Mach() ) { 476 int padding = nj->as_Mach()->compute_padding(adr); 477 // If call/safepoint are adjacent insert a nop (5010568) 478 if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() && 479 adr == last_call_adr ) { 480 padding = nop_size; 481 } 482 if(padding > 0) { 483 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 484 int nops_cnt = padding / nop_size; 485 MachNode *nop = new (this) MachNopNode(nops_cnt); 486 b->_nodes.insert(j++, nop); 487 _cfg->_bbs.map( nop->_idx, b ); 488 adr += padding; 489 last_inst++; 490 } 491 } 492 adr += nj->size(_regalloc); 493 494 // Remember end of call offset 495 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 496 last_call_adr = adr; 497 } 498 } 499 500 if ( i != _cfg->_num_blocks-1) { 501 // Get the size of the block 502 uint blk_size = adr - blk_starts[i]; 503 504 // When the next block is the top of a loop, we may insert pad NOP 505 // instructions. 506 Block *nb = _cfg->_blocks[i+1]; 507 int current_offset = blk_starts[i] + blk_size; 508 current_offset += nb->alignment_padding(current_offset); 509 // Save block size; update total method size 510 blk_starts[i+1] = current_offset; 511 } 512 } 513 514 #ifdef ASSERT 515 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 516 if( jmp_target[i] != 0 ) { 517 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_end[i]); 518 if (!_matcher->is_short_branch_offset(jmp_rule[i], offset)) { 519 tty->print_cr("target (%d) - jmp_end(%d) = offset (%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_end[i], offset, i, jmp_target[i]); 520 } 521 assert(_matcher->is_short_branch_offset(jmp_rule[i], offset), "Displacement too large for short jmp"); 522 } 523 } 524 #endif 525 526 // ------------------ 527 // Compute size for code buffer 528 code_size = blk_starts[i-1] + jmp_end[i-1]; 529 530 // Relocation records 531 reloc_size += 1; // Relo entry for exception handler 532 533 // Adjust reloc_size to number of record of relocation info 534 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 535 // a relocation index. 536 // The CodeBuffer will expand the locs array if this estimate is too low. 537 reloc_size *= 10 / sizeof(relocInfo); 538 } 539 540 //------------------------------FillLocArray----------------------------------- 541 // Create a bit of debug info and append it to the array. The mapping is from 542 // Java local or expression stack to constant, register or stack-slot. For 543 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 544 // entry has been taken care of and caller should skip it). 545 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 546 // This should never have accepted Bad before 547 assert(OptoReg::is_valid(regnum), "location must be valid"); 548 return (OptoReg::is_reg(regnum)) 549 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 550 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 551 } 552 553 554 ObjectValue* 555 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 556 for (int i = 0; i < objs->length(); i++) { 557 assert(objs->at(i)->is_object(), "corrupt object cache"); 558 ObjectValue* sv = (ObjectValue*) objs->at(i); 559 if (sv->id() == id) { 560 return sv; 561 } 562 } 563 // Otherwise.. 564 return NULL; 565 } 566 567 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 568 ObjectValue* sv ) { 569 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 570 objs->append(sv); 571 } 572 573 574 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 575 GrowableArray<ScopeValue*> *array, 576 GrowableArray<ScopeValue*> *objs ) { 577 assert( local, "use _top instead of null" ); 578 if (array->length() != idx) { 579 assert(array->length() == idx + 1, "Unexpected array count"); 580 // Old functionality: 581 // return 582 // New functionality: 583 // Assert if the local is not top. In product mode let the new node 584 // override the old entry. 585 assert(local == top(), "LocArray collision"); 586 if (local == top()) { 587 return; 588 } 589 array->pop(); 590 } 591 const Type *t = local->bottom_type(); 592 593 // Is it a safepoint scalar object node? 594 if (local->is_SafePointScalarObject()) { 595 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 596 597 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 598 if (sv == NULL) { 599 ciKlass* cik = t->is_oopptr()->klass(); 600 assert(cik->is_instance_klass() || 601 cik->is_array_klass(), "Not supported allocation."); 602 sv = new ObjectValue(spobj->_idx, 603 new ConstantOopWriteValue(cik->constant_encoding())); 604 Compile::set_sv_for_object_node(objs, sv); 605 606 uint first_ind = spobj->first_index(); 607 for (uint i = 0; i < spobj->n_fields(); i++) { 608 Node* fld_node = sfpt->in(first_ind+i); 609 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 610 } 611 } 612 array->append(sv); 613 return; 614 } 615 616 // Grab the register number for the local 617 OptoReg::Name regnum = _regalloc->get_reg_first(local); 618 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 619 // Record the double as two float registers. 620 // The register mask for such a value always specifies two adjacent 621 // float registers, with the lower register number even. 622 // Normally, the allocation of high and low words to these registers 623 // is irrelevant, because nearly all operations on register pairs 624 // (e.g., StoreD) treat them as a single unit. 625 // Here, we assume in addition that the words in these two registers 626 // stored "naturally" (by operations like StoreD and double stores 627 // within the interpreter) such that the lower-numbered register 628 // is written to the lower memory address. This may seem like 629 // a machine dependency, but it is not--it is a requirement on 630 // the author of the <arch>.ad file to ensure that, for every 631 // even/odd double-register pair to which a double may be allocated, 632 // the word in the even single-register is stored to the first 633 // memory word. (Note that register numbers are completely 634 // arbitrary, and are not tied to any machine-level encodings.) 635 #ifdef _LP64 636 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 637 array->append(new ConstantIntValue(0)); 638 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 639 } else if ( t->base() == Type::Long ) { 640 array->append(new ConstantIntValue(0)); 641 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 642 } else if ( t->base() == Type::RawPtr ) { 643 // jsr/ret return address which must be restored into a the full 644 // width 64-bit stack slot. 645 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 646 } 647 #else //_LP64 648 #ifdef SPARC 649 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 650 // For SPARC we have to swap high and low words for 651 // long values stored in a single-register (g0-g7). 652 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 653 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 654 } else 655 #endif //SPARC 656 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 657 // Repack the double/long as two jints. 658 // The convention the interpreter uses is that the second local 659 // holds the first raw word of the native double representation. 660 // This is actually reasonable, since locals and stack arrays 661 // grow downwards in all implementations. 662 // (If, on some machine, the interpreter's Java locals or stack 663 // were to grow upwards, the embedded doubles would be word-swapped.) 664 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 665 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 666 } 667 #endif //_LP64 668 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 669 OptoReg::is_reg(regnum) ) { 670 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 671 ? Location::float_in_dbl : Location::normal )); 672 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 673 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 674 ? Location::int_in_long : Location::normal )); 675 } else if( t->base() == Type::NarrowOop ) { 676 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 677 } else { 678 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 679 } 680 return; 681 } 682 683 // No register. It must be constant data. 684 switch (t->base()) { 685 case Type::Half: // Second half of a double 686 ShouldNotReachHere(); // Caller should skip 2nd halves 687 break; 688 case Type::AnyPtr: 689 array->append(new ConstantOopWriteValue(NULL)); 690 break; 691 case Type::AryPtr: 692 case Type::InstPtr: 693 case Type::KlassPtr: // fall through 694 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 695 break; 696 case Type::NarrowOop: 697 if (t == TypeNarrowOop::NULL_PTR) { 698 array->append(new ConstantOopWriteValue(NULL)); 699 } else { 700 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 701 } 702 break; 703 case Type::Int: 704 array->append(new ConstantIntValue(t->is_int()->get_con())); 705 break; 706 case Type::RawPtr: 707 // A return address (T_ADDRESS). 708 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 709 #ifdef _LP64 710 // Must be restored to the full-width 64-bit stack slot. 711 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 712 #else 713 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 714 #endif 715 break; 716 case Type::FloatCon: { 717 float f = t->is_float_constant()->getf(); 718 array->append(new ConstantIntValue(jint_cast(f))); 719 break; 720 } 721 case Type::DoubleCon: { 722 jdouble d = t->is_double_constant()->getd(); 723 #ifdef _LP64 724 array->append(new ConstantIntValue(0)); 725 array->append(new ConstantDoubleValue(d)); 726 #else 727 // Repack the double as two jints. 728 // The convention the interpreter uses is that the second local 729 // holds the first raw word of the native double representation. 730 // This is actually reasonable, since locals and stack arrays 731 // grow downwards in all implementations. 732 // (If, on some machine, the interpreter's Java locals or stack 733 // were to grow upwards, the embedded doubles would be word-swapped.) 734 jint *dp = (jint*)&d; 735 array->append(new ConstantIntValue(dp[1])); 736 array->append(new ConstantIntValue(dp[0])); 737 #endif 738 break; 739 } 740 case Type::Long: { 741 jlong d = t->is_long()->get_con(); 742 #ifdef _LP64 743 array->append(new ConstantIntValue(0)); 744 array->append(new ConstantLongValue(d)); 745 #else 746 // Repack the long as two jints. 747 // The convention the interpreter uses is that the second local 748 // holds the first raw word of the native double representation. 749 // This is actually reasonable, since locals and stack arrays 750 // grow downwards in all implementations. 751 // (If, on some machine, the interpreter's Java locals or stack 752 // were to grow upwards, the embedded doubles would be word-swapped.) 753 jint *dp = (jint*)&d; 754 array->append(new ConstantIntValue(dp[1])); 755 array->append(new ConstantIntValue(dp[0])); 756 #endif 757 break; 758 } 759 case Type::Top: // Add an illegal value here 760 array->append(new LocationValue(Location())); 761 break; 762 default: 763 ShouldNotReachHere(); 764 break; 765 } 766 } 767 768 // Determine if this node starts a bundle 769 bool Compile::starts_bundle(const Node *n) const { 770 return (_node_bundling_limit > n->_idx && 771 _node_bundling_base[n->_idx].starts_bundle()); 772 } 773 774 //--------------------------Process_OopMap_Node-------------------------------- 775 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 776 777 // Handle special safepoint nodes for synchronization 778 MachSafePointNode *sfn = mach->as_MachSafePoint(); 779 MachCallNode *mcall; 780 781 #ifdef ENABLE_ZAP_DEAD_LOCALS 782 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 783 #endif 784 785 int safepoint_pc_offset = current_offset; 786 bool is_method_handle_invoke = false; 787 bool return_oop = false; 788 789 // Add the safepoint in the DebugInfoRecorder 790 if( !mach->is_MachCall() ) { 791 mcall = NULL; 792 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 793 } else { 794 mcall = mach->as_MachCall(); 795 796 // Is the call a MethodHandle call? 797 if (mcall->is_MachCallJava()) { 798 if (mcall->as_MachCallJava()->_method_handle_invoke) { 799 assert(has_method_handle_invokes(), "must have been set during call generation"); 800 is_method_handle_invoke = true; 801 } 802 } 803 804 // Check if a call returns an object. 805 if (mcall->return_value_is_used() && 806 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 807 return_oop = true; 808 } 809 safepoint_pc_offset += mcall->ret_addr_offset(); 810 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 811 } 812 813 // Loop over the JVMState list to add scope information 814 // Do not skip safepoints with a NULL method, they need monitor info 815 JVMState* youngest_jvms = sfn->jvms(); 816 int max_depth = youngest_jvms->depth(); 817 818 // Allocate the object pool for scalar-replaced objects -- the map from 819 // small-integer keys (which can be recorded in the local and ostack 820 // arrays) to descriptions of the object state. 821 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 822 823 // Visit scopes from oldest to youngest. 824 for (int depth = 1; depth <= max_depth; depth++) { 825 JVMState* jvms = youngest_jvms->of_depth(depth); 826 int idx; 827 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 828 // Safepoints that do not have method() set only provide oop-map and monitor info 829 // to support GC; these do not support deoptimization. 830 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 831 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 832 int num_mon = jvms->nof_monitors(); 833 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 834 "JVMS local count must match that of the method"); 835 836 // Add Local and Expression Stack Information 837 838 // Insert locals into the locarray 839 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 840 for( idx = 0; idx < num_locs; idx++ ) { 841 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 842 } 843 844 // Insert expression stack entries into the exparray 845 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 846 for( idx = 0; idx < num_exps; idx++ ) { 847 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 848 } 849 850 // Add in mappings of the monitors 851 assert( !method || 852 !method->is_synchronized() || 853 method->is_native() || 854 num_mon > 0 || 855 !GenerateSynchronizationCode, 856 "monitors must always exist for synchronized methods"); 857 858 // Build the growable array of ScopeValues for exp stack 859 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 860 861 // Loop over monitors and insert into array 862 for(idx = 0; idx < num_mon; idx++) { 863 // Grab the node that defines this monitor 864 Node* box_node = sfn->monitor_box(jvms, idx); 865 Node* obj_node = sfn->monitor_obj(jvms, idx); 866 867 // Create ScopeValue for object 868 ScopeValue *scval = NULL; 869 870 if( obj_node->is_SafePointScalarObject() ) { 871 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 872 scval = Compile::sv_for_node_id(objs, spobj->_idx); 873 if (scval == NULL) { 874 const Type *t = obj_node->bottom_type(); 875 ciKlass* cik = t->is_oopptr()->klass(); 876 assert(cik->is_instance_klass() || 877 cik->is_array_klass(), "Not supported allocation."); 878 ObjectValue* sv = new ObjectValue(spobj->_idx, 879 new ConstantOopWriteValue(cik->constant_encoding())); 880 Compile::set_sv_for_object_node(objs, sv); 881 882 uint first_ind = spobj->first_index(); 883 for (uint i = 0; i < spobj->n_fields(); i++) { 884 Node* fld_node = sfn->in(first_ind+i); 885 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 886 } 887 scval = sv; 888 } 889 } else if( !obj_node->is_Con() ) { 890 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 891 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 892 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 893 } else { 894 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 895 } 896 } else { 897 const TypePtr *tp = obj_node->bottom_type()->make_ptr(); 898 scval = new ConstantOopWriteValue(tp->is_instptr()->const_oop()->constant_encoding()); 899 } 900 901 OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node); 902 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 903 while( !box_node->is_BoxLock() ) box_node = box_node->in(1); 904 monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated())); 905 } 906 907 // We dump the object pool first, since deoptimization reads it in first. 908 debug_info()->dump_object_pool(objs); 909 910 // Build first class objects to pass to scope 911 DebugToken *locvals = debug_info()->create_scope_values(locarray); 912 DebugToken *expvals = debug_info()->create_scope_values(exparray); 913 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 914 915 // Make method available for all Safepoints 916 ciMethod* scope_method = method ? method : _method; 917 // Describe the scope here 918 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 919 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 920 // Now we can describe the scope. 921 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 922 } // End jvms loop 923 924 // Mark the end of the scope set. 925 debug_info()->end_safepoint(safepoint_pc_offset); 926 } 927 928 929 930 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 931 class NonSafepointEmitter { 932 Compile* C; 933 JVMState* _pending_jvms; 934 int _pending_offset; 935 936 void emit_non_safepoint(); 937 938 public: 939 NonSafepointEmitter(Compile* compile) { 940 this->C = compile; 941 _pending_jvms = NULL; 942 _pending_offset = 0; 943 } 944 945 void observe_instruction(Node* n, int pc_offset) { 946 if (!C->debug_info()->recording_non_safepoints()) return; 947 948 Node_Notes* nn = C->node_notes_at(n->_idx); 949 if (nn == NULL || nn->jvms() == NULL) return; 950 if (_pending_jvms != NULL && 951 _pending_jvms->same_calls_as(nn->jvms())) { 952 // Repeated JVMS? Stretch it up here. 953 _pending_offset = pc_offset; 954 } else { 955 if (_pending_jvms != NULL && 956 _pending_offset < pc_offset) { 957 emit_non_safepoint(); 958 } 959 _pending_jvms = NULL; 960 if (pc_offset > C->debug_info()->last_pc_offset()) { 961 // This is the only way _pending_jvms can become non-NULL: 962 _pending_jvms = nn->jvms(); 963 _pending_offset = pc_offset; 964 } 965 } 966 } 967 968 // Stay out of the way of real safepoints: 969 void observe_safepoint(JVMState* jvms, int pc_offset) { 970 if (_pending_jvms != NULL && 971 !_pending_jvms->same_calls_as(jvms) && 972 _pending_offset < pc_offset) { 973 emit_non_safepoint(); 974 } 975 _pending_jvms = NULL; 976 } 977 978 void flush_at_end() { 979 if (_pending_jvms != NULL) { 980 emit_non_safepoint(); 981 } 982 _pending_jvms = NULL; 983 } 984 }; 985 986 void NonSafepointEmitter::emit_non_safepoint() { 987 JVMState* youngest_jvms = _pending_jvms; 988 int pc_offset = _pending_offset; 989 990 // Clear it now: 991 _pending_jvms = NULL; 992 993 DebugInformationRecorder* debug_info = C->debug_info(); 994 assert(debug_info->recording_non_safepoints(), "sanity"); 995 996 debug_info->add_non_safepoint(pc_offset); 997 int max_depth = youngest_jvms->depth(); 998 999 // Visit scopes from oldest to youngest. 1000 for (int depth = 1; depth <= max_depth; depth++) { 1001 JVMState* jvms = youngest_jvms->of_depth(depth); 1002 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1003 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1004 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1005 } 1006 1007 // Mark the end of the scope set. 1008 debug_info->end_non_safepoint(pc_offset); 1009 } 1010 1011 1012 1013 // helper for Fill_buffer bailout logic 1014 static void turn_off_compiler(Compile* C) { 1015 if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) { 1016 // Do not turn off compilation if a single giant method has 1017 // blown the code cache size. 1018 C->record_failure("excessive request to CodeCache"); 1019 } else { 1020 // Let CompilerBroker disable further compilations. 1021 C->record_failure("CodeCache is full"); 1022 } 1023 } 1024 1025 1026 //------------------------------Fill_buffer------------------------------------ 1027 void Compile::Fill_buffer() { 1028 1029 // Set the initially allocated size 1030 int code_req = initial_code_capacity; 1031 int locs_req = initial_locs_capacity; 1032 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1033 int const_req = initial_const_capacity; 1034 bool labels_not_set = true; 1035 1036 int pad_req = NativeCall::instruction_size; 1037 // The extra spacing after the code is necessary on some platforms. 1038 // Sometimes we need to patch in a jump after the last instruction, 1039 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1040 1041 uint i; 1042 // Compute the byte offset where we can store the deopt pc. 1043 if (fixed_slots() != 0) { 1044 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1045 } 1046 1047 // Compute prolog code size 1048 _method_size = 0; 1049 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1050 #ifdef IA64 1051 if (save_argument_registers()) { 1052 // 4815101: this is a stub with implicit and unknown precision fp args. 1053 // The usual spill mechanism can only generate stfd's in this case, which 1054 // doesn't work if the fp reg to spill contains a single-precision denorm. 1055 // Instead, we hack around the normal spill mechanism using stfspill's and 1056 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1057 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1058 // 1059 // If we ever implement 16-byte 'registers' == stack slots, we can 1060 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1061 // instead of stfd/stfs/ldfd/ldfs. 1062 _frame_slots += 8*(16/BytesPerInt); 1063 } 1064 #endif 1065 assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" ); 1066 1067 // Create an array of unused labels, one for each basic block 1068 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1); 1069 1070 for( i=0; i <= _cfg->_num_blocks; i++ ) { 1071 blk_labels[i].init(); 1072 } 1073 1074 if (has_mach_constant_base_node()) { 1075 // Fill the constant table. 1076 // Note: This must happen before Shorten_branches. 1077 for (i = 0; i < _cfg->_num_blocks; i++) { 1078 Block* b = _cfg->_blocks[i]; 1079 1080 for (uint j = 0; j < b->_nodes.size(); j++) { 1081 Node* n = b->_nodes[j]; 1082 1083 if (n->is_Mach()) { 1084 MachNode *mach = n->as_Mach(); 1085 1086 // If the MachNode is a MachConstantNode evaluate the 1087 // constant value section. 1088 if (mach->is_MachConstant()) { 1089 MachConstantNode* machcon = mach->as_MachConstant(); 1090 machcon->eval_constant(); 1091 } 1092 } 1093 } 1094 } 1095 1096 // Calculate the size of the constant table (including the padding 1097 // to the next section). 1098 const_req = mach_constant_base_node()->calculate_constant_table_size(); 1099 } 1100 1101 // Initialize the space for the BufferBlob used to find and verify 1102 // instruction size in MachNode::emit_size() 1103 init_scratch_buffer_blob(const_req); 1104 if (failing()) return; // Out of memory 1105 1106 // If this machine supports different size branch offsets, then pre-compute 1107 // the length of the blocks 1108 if( _matcher->is_short_branch_offset(-1, 0) ) { 1109 Shorten_branches(blk_labels, code_req, locs_req, stub_req); 1110 labels_not_set = false; 1111 } 1112 1113 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1114 int exception_handler_req = size_exception_handler(); 1115 int deopt_handler_req = size_deopt_handler(); 1116 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1117 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1118 stub_req += MAX_stubs_size; // ensure per-stub margin 1119 code_req += MAX_inst_size; // ensure per-instruction margin 1120 1121 if (StressCodeBuffers) 1122 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1123 1124 int total_req = 1125 const_req + 1126 code_req + 1127 pad_req + 1128 stub_req + 1129 exception_handler_req + 1130 deopt_handler_req; // deopt handler 1131 1132 if (has_method_handle_invokes()) 1133 total_req += deopt_handler_req; // deopt MH handler 1134 1135 CodeBuffer* cb = code_buffer(); 1136 cb->initialize(total_req, locs_req); 1137 1138 // Have we run out of code space? 1139 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1140 turn_off_compiler(this); 1141 return; 1142 } 1143 // Configure the code buffer. 1144 cb->initialize_consts_size(const_req); 1145 cb->initialize_stubs_size(stub_req); 1146 cb->initialize_oop_recorder(env()->oop_recorder()); 1147 1148 // fill in the nop array for bundling computations 1149 MachNode *_nop_list[Bundle::_nop_count]; 1150 Bundle::initialize_nops(_nop_list, this); 1151 1152 // Create oopmap set. 1153 _oop_map_set = new OopMapSet(); 1154 1155 // !!!!! This preserves old handling of oopmaps for now 1156 debug_info()->set_oopmaps(_oop_map_set); 1157 1158 // Count and start of implicit null check instructions 1159 uint inct_cnt = 0; 1160 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1161 1162 // Count and start of calls 1163 uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1164 1165 uint return_offset = 0; 1166 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1167 1168 int previous_offset = 0; 1169 int current_offset = 0; 1170 int last_call_offset = -1; 1171 1172 // Create an array of unused labels, one for each basic block, if printing is enabled 1173 #ifndef PRODUCT 1174 int *node_offsets = NULL; 1175 uint node_offset_limit = unique(); 1176 1177 1178 if ( print_assembly() ) 1179 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1180 #endif 1181 1182 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1183 1184 // ------------------ 1185 // Now fill in the code buffer 1186 Node *delay_slot = NULL; 1187 1188 for( i=0; i < _cfg->_num_blocks; i++ ) { 1189 Block *b = _cfg->_blocks[i]; 1190 1191 Node *head = b->head(); 1192 1193 // If this block needs to start aligned (i.e, can be reached other 1194 // than by falling-thru from the previous block), then force the 1195 // start of a new bundle. 1196 if( Pipeline::requires_bundling() && starts_bundle(head) ) 1197 cb->flush_bundle(true); 1198 1199 // Define the label at the beginning of the basic block 1200 if (labels_not_set) { 1201 MacroAssembler(cb).bind(blk_labels[b->_pre_order]); 1202 } else { 1203 assert(blk_labels[b->_pre_order].loc_pos() == cb->insts_size(), 1204 err_msg("label position does not match code offset: %d != %d", 1205 blk_labels[b->_pre_order].loc_pos(), cb->insts_size())); 1206 } 1207 1208 uint last_inst = b->_nodes.size(); 1209 1210 // Emit block normally, except for last instruction. 1211 // Emit means "dump code bits into code buffer". 1212 for( uint j = 0; j<last_inst; j++ ) { 1213 1214 // Get the node 1215 Node* n = b->_nodes[j]; 1216 1217 // See if delay slots are supported 1218 if (valid_bundle_info(n) && 1219 node_bundling(n)->used_in_unconditional_delay()) { 1220 assert(delay_slot == NULL, "no use of delay slot node"); 1221 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1222 1223 delay_slot = n; 1224 continue; 1225 } 1226 1227 // If this starts a new instruction group, then flush the current one 1228 // (but allow split bundles) 1229 if( Pipeline::requires_bundling() && starts_bundle(n) ) 1230 cb->flush_bundle(false); 1231 1232 // The following logic is duplicated in the code ifdeffed for 1233 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1234 // should be factored out. Or maybe dispersed to the nodes? 1235 1236 // Special handling for SafePoint/Call Nodes 1237 bool is_mcall = false; 1238 if( n->is_Mach() ) { 1239 MachNode *mach = n->as_Mach(); 1240 is_mcall = n->is_MachCall(); 1241 bool is_sfn = n->is_MachSafePoint(); 1242 1243 // If this requires all previous instructions be flushed, then do so 1244 if( is_sfn || is_mcall || mach->alignment_required() != 1) { 1245 cb->flush_bundle(true); 1246 current_offset = cb->insts_size(); 1247 } 1248 1249 // align the instruction if necessary 1250 int padding = mach->compute_padding(current_offset); 1251 // Make sure safepoint node for polling is distinct from a call's 1252 // return by adding a nop if needed. 1253 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) { 1254 padding = nop_size; 1255 } 1256 assert( labels_not_set || padding == 0, "instruction should already be aligned"); 1257 1258 if(padding > 0) { 1259 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1260 int nops_cnt = padding / nop_size; 1261 MachNode *nop = new (this) MachNopNode(nops_cnt); 1262 b->_nodes.insert(j++, nop); 1263 last_inst++; 1264 _cfg->_bbs.map( nop->_idx, b ); 1265 nop->emit(*cb, _regalloc); 1266 cb->flush_bundle(true); 1267 current_offset = cb->insts_size(); 1268 } 1269 1270 // Remember the start of the last call in a basic block 1271 if (is_mcall) { 1272 MachCallNode *mcall = mach->as_MachCall(); 1273 1274 // This destination address is NOT PC-relative 1275 mcall->method_set((intptr_t)mcall->entry_point()); 1276 1277 // Save the return address 1278 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1279 1280 if (!mcall->is_safepoint_node()) { 1281 is_mcall = false; 1282 is_sfn = false; 1283 } 1284 } 1285 1286 // sfn will be valid whenever mcall is valid now because of inheritance 1287 if( is_sfn || is_mcall ) { 1288 1289 // Handle special safepoint nodes for synchronization 1290 if( !is_mcall ) { 1291 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1292 // !!!!! Stubs only need an oopmap right now, so bail out 1293 if( sfn->jvms()->method() == NULL) { 1294 // Write the oopmap directly to the code blob??!! 1295 # ifdef ENABLE_ZAP_DEAD_LOCALS 1296 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1297 # endif 1298 continue; 1299 } 1300 } // End synchronization 1301 1302 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1303 current_offset); 1304 Process_OopMap_Node(mach, current_offset); 1305 } // End if safepoint 1306 1307 // If this is a null check, then add the start of the previous instruction to the list 1308 else if( mach->is_MachNullCheck() ) { 1309 inct_starts[inct_cnt++] = previous_offset; 1310 } 1311 1312 // If this is a branch, then fill in the label with the target BB's label 1313 else if ( mach->is_Branch() ) { 1314 1315 if ( mach->ideal_Opcode() == Op_Jump ) { 1316 for (uint h = 0; h < b->_num_succs; h++ ) { 1317 Block* succs_block = b->_succs[h]; 1318 for (uint j = 1; j < succs_block->num_preds(); j++) { 1319 Node* jpn = succs_block->pred(j); 1320 if ( jpn->is_JumpProj() && jpn->in(0) == mach ) { 1321 uint block_num = succs_block->non_connector()->_pre_order; 1322 Label *blkLabel = &blk_labels[block_num]; 1323 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1324 } 1325 } 1326 } 1327 } else { 1328 // For Branchs 1329 // This requires the TRUE branch target be in succs[0] 1330 uint block_num = b->non_connector_successor(0)->_pre_order; 1331 mach->label_set( blk_labels[block_num], block_num ); 1332 } 1333 } 1334 1335 #ifdef ASSERT 1336 // Check that oop-store precedes the card-mark 1337 else if( mach->ideal_Opcode() == Op_StoreCM ) { 1338 uint storeCM_idx = j; 1339 Node *oop_store = mach->in(mach->_cnt); // First precedence edge 1340 assert( oop_store != NULL, "storeCM expects a precedence edge"); 1341 uint i4; 1342 for( i4 = 0; i4 < last_inst; ++i4 ) { 1343 if( b->_nodes[i4] == oop_store ) break; 1344 } 1345 // Note: This test can provide a false failure if other precedence 1346 // edges have been added to the storeCMNode. 1347 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1348 } 1349 #endif 1350 1351 else if( !n->is_Proj() ) { 1352 // Remember the beginning of the previous instruction, in case 1353 // it's followed by a flag-kill and a null-check. Happens on 1354 // Intel all the time, with add-to-memory kind of opcodes. 1355 previous_offset = current_offset; 1356 } 1357 } 1358 1359 // Verify that there is sufficient space remaining 1360 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1361 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1362 turn_off_compiler(this); 1363 return; 1364 } 1365 1366 // Save the offset for the listing 1367 #ifndef PRODUCT 1368 if( node_offsets && n->_idx < node_offset_limit ) 1369 node_offsets[n->_idx] = cb->insts_size(); 1370 #endif 1371 1372 // "Normal" instruction case 1373 n->emit(*cb, _regalloc); 1374 current_offset = cb->insts_size(); 1375 non_safepoints.observe_instruction(n, current_offset); 1376 1377 // mcall is last "call" that can be a safepoint 1378 // record it so we can see if a poll will directly follow it 1379 // in which case we'll need a pad to make the PcDesc sites unique 1380 // see 5010568. This can be slightly inaccurate but conservative 1381 // in the case that return address is not actually at current_offset. 1382 // This is a small price to pay. 1383 1384 if (is_mcall) { 1385 last_call_offset = current_offset; 1386 } 1387 1388 // See if this instruction has a delay slot 1389 if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1390 assert(delay_slot != NULL, "expecting delay slot node"); 1391 1392 // Back up 1 instruction 1393 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1394 1395 // Save the offset for the listing 1396 #ifndef PRODUCT 1397 if( node_offsets && delay_slot->_idx < node_offset_limit ) 1398 node_offsets[delay_slot->_idx] = cb->insts_size(); 1399 #endif 1400 1401 // Support a SafePoint in the delay slot 1402 if( delay_slot->is_MachSafePoint() ) { 1403 MachNode *mach = delay_slot->as_Mach(); 1404 // !!!!! Stubs only need an oopmap right now, so bail out 1405 if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) { 1406 // Write the oopmap directly to the code blob??!! 1407 # ifdef ENABLE_ZAP_DEAD_LOCALS 1408 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1409 # endif 1410 delay_slot = NULL; 1411 continue; 1412 } 1413 1414 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1415 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1416 adjusted_offset); 1417 // Generate an OopMap entry 1418 Process_OopMap_Node(mach, adjusted_offset); 1419 } 1420 1421 // Insert the delay slot instruction 1422 delay_slot->emit(*cb, _regalloc); 1423 1424 // Don't reuse it 1425 delay_slot = NULL; 1426 } 1427 1428 } // End for all instructions in block 1429 1430 // If the next block is the top of a loop, pad this block out to align 1431 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1432 if( i<_cfg->_num_blocks-1 ) { 1433 Block *nb = _cfg->_blocks[i+1]; 1434 uint padding = nb->alignment_padding(current_offset); 1435 if( padding > 0 ) { 1436 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1437 b->_nodes.insert( b->_nodes.size(), nop ); 1438 _cfg->_bbs.map( nop->_idx, b ); 1439 nop->emit(*cb, _regalloc); 1440 current_offset = cb->insts_size(); 1441 } 1442 } 1443 1444 } // End of for all blocks 1445 1446 non_safepoints.flush_at_end(); 1447 1448 // Offset too large? 1449 if (failing()) return; 1450 1451 // Define a pseudo-label at the end of the code 1452 MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] ); 1453 1454 // Compute the size of the first block 1455 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1456 1457 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1458 1459 // ------------------ 1460 1461 #ifndef PRODUCT 1462 // Information on the size of the method, without the extraneous code 1463 Scheduling::increment_method_size(cb->insts_size()); 1464 #endif 1465 1466 // ------------------ 1467 // Fill in exception table entries. 1468 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1469 1470 // Only java methods have exception handlers and deopt handlers 1471 if (_method) { 1472 // Emit the exception handler code. 1473 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1474 // Emit the deopt handler code. 1475 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1476 1477 // Emit the MethodHandle deopt handler code (if required). 1478 if (has_method_handle_invokes()) { 1479 // We can use the same code as for the normal deopt handler, we 1480 // just need a different entry point address. 1481 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1482 } 1483 } 1484 1485 // One last check for failed CodeBuffer::expand: 1486 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1487 turn_off_compiler(this); 1488 return; 1489 } 1490 1491 #ifndef PRODUCT 1492 // Dump the assembly code, including basic-block numbers 1493 if (print_assembly()) { 1494 ttyLocker ttyl; // keep the following output all in one block 1495 if (!VMThread::should_terminate()) { // test this under the tty lock 1496 // This output goes directly to the tty, not the compiler log. 1497 // To enable tools to match it up with the compilation activity, 1498 // be sure to tag this tty output with the compile ID. 1499 if (xtty != NULL) { 1500 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1501 is_osr_compilation() ? " compile_kind='osr'" : 1502 ""); 1503 } 1504 if (method() != NULL) { 1505 method()->print_oop(); 1506 print_codes(); 1507 } 1508 dump_asm(node_offsets, node_offset_limit); 1509 if (xtty != NULL) { 1510 xtty->tail("opto_assembly"); 1511 } 1512 } 1513 } 1514 #endif 1515 1516 } 1517 1518 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1519 _inc_table.set_size(cnt); 1520 1521 uint inct_cnt = 0; 1522 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1523 Block *b = _cfg->_blocks[i]; 1524 Node *n = NULL; 1525 int j; 1526 1527 // Find the branch; ignore trailing NOPs. 1528 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1529 n = b->_nodes[j]; 1530 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1531 break; 1532 } 1533 1534 // If we didn't find anything, continue 1535 if( j < 0 ) continue; 1536 1537 // Compute ExceptionHandlerTable subtable entry and add it 1538 // (skip empty blocks) 1539 if( n->is_Catch() ) { 1540 1541 // Get the offset of the return from the call 1542 uint call_return = call_returns[b->_pre_order]; 1543 #ifdef ASSERT 1544 assert( call_return > 0, "no call seen for this basic block" ); 1545 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; 1546 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); 1547 #endif 1548 // last instruction is a CatchNode, find it's CatchProjNodes 1549 int nof_succs = b->_num_succs; 1550 // allocate space 1551 GrowableArray<intptr_t> handler_bcis(nof_succs); 1552 GrowableArray<intptr_t> handler_pcos(nof_succs); 1553 // iterate through all successors 1554 for (int j = 0; j < nof_succs; j++) { 1555 Block* s = b->_succs[j]; 1556 bool found_p = false; 1557 for( uint k = 1; k < s->num_preds(); k++ ) { 1558 Node *pk = s->pred(k); 1559 if( pk->is_CatchProj() && pk->in(0) == n ) { 1560 const CatchProjNode* p = pk->as_CatchProj(); 1561 found_p = true; 1562 // add the corresponding handler bci & pco information 1563 if( p->_con != CatchProjNode::fall_through_index ) { 1564 // p leads to an exception handler (and is not fall through) 1565 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1566 // no duplicates, please 1567 if( !handler_bcis.contains(p->handler_bci()) ) { 1568 uint block_num = s->non_connector()->_pre_order; 1569 handler_bcis.append(p->handler_bci()); 1570 handler_pcos.append(blk_labels[block_num].loc_pos()); 1571 } 1572 } 1573 } 1574 } 1575 assert(found_p, "no matching predecessor found"); 1576 // Note: Due to empty block removal, one block may have 1577 // several CatchProj inputs, from the same Catch. 1578 } 1579 1580 // Set the offset of the return from the call 1581 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1582 continue; 1583 } 1584 1585 // Handle implicit null exception table updates 1586 if( n->is_MachNullCheck() ) { 1587 uint block_num = b->non_connector_successor(0)->_pre_order; 1588 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1589 continue; 1590 } 1591 } // End of for all blocks fill in exception table entries 1592 } 1593 1594 // Static Variables 1595 #ifndef PRODUCT 1596 uint Scheduling::_total_nop_size = 0; 1597 uint Scheduling::_total_method_size = 0; 1598 uint Scheduling::_total_branches = 0; 1599 uint Scheduling::_total_unconditional_delays = 0; 1600 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1601 #endif 1602 1603 // Initializer for class Scheduling 1604 1605 Scheduling::Scheduling(Arena *arena, Compile &compile) 1606 : _arena(arena), 1607 _cfg(compile.cfg()), 1608 _bbs(compile.cfg()->_bbs), 1609 _regalloc(compile.regalloc()), 1610 _reg_node(arena), 1611 _bundle_instr_count(0), 1612 _bundle_cycle_number(0), 1613 _scheduled(arena), 1614 _available(arena), 1615 _next_node(NULL), 1616 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1617 _pinch_free_list(arena) 1618 #ifndef PRODUCT 1619 , _branches(0) 1620 , _unconditional_delays(0) 1621 #endif 1622 { 1623 // Create a MachNopNode 1624 _nop = new (&compile) MachNopNode(); 1625 1626 // Now that the nops are in the array, save the count 1627 // (but allow entries for the nops) 1628 _node_bundling_limit = compile.unique(); 1629 uint node_max = _regalloc->node_regs_max_index(); 1630 1631 compile.set_node_bundling_limit(_node_bundling_limit); 1632 1633 // This one is persistent within the Compile class 1634 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1635 1636 // Allocate space for fixed-size arrays 1637 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1638 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1639 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1640 1641 // Clear the arrays 1642 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1643 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1644 memset(_uses, 0, node_max * sizeof(short)); 1645 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1646 1647 // Clear the bundling information 1648 memcpy(_bundle_use_elements, 1649 Pipeline_Use::elaborated_elements, 1650 sizeof(Pipeline_Use::elaborated_elements)); 1651 1652 // Get the last node 1653 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1654 1655 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1656 } 1657 1658 #ifndef PRODUCT 1659 // Scheduling destructor 1660 Scheduling::~Scheduling() { 1661 _total_branches += _branches; 1662 _total_unconditional_delays += _unconditional_delays; 1663 } 1664 #endif 1665 1666 // Step ahead "i" cycles 1667 void Scheduling::step(uint i) { 1668 1669 Bundle *bundle = node_bundling(_next_node); 1670 bundle->set_starts_bundle(); 1671 1672 // Update the bundle record, but leave the flags information alone 1673 if (_bundle_instr_count > 0) { 1674 bundle->set_instr_count(_bundle_instr_count); 1675 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1676 } 1677 1678 // Update the state information 1679 _bundle_instr_count = 0; 1680 _bundle_cycle_number += i; 1681 _bundle_use.step(i); 1682 } 1683 1684 void Scheduling::step_and_clear() { 1685 Bundle *bundle = node_bundling(_next_node); 1686 bundle->set_starts_bundle(); 1687 1688 // Update the bundle record 1689 if (_bundle_instr_count > 0) { 1690 bundle->set_instr_count(_bundle_instr_count); 1691 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1692 1693 _bundle_cycle_number += 1; 1694 } 1695 1696 // Clear the bundling information 1697 _bundle_instr_count = 0; 1698 _bundle_use.reset(); 1699 1700 memcpy(_bundle_use_elements, 1701 Pipeline_Use::elaborated_elements, 1702 sizeof(Pipeline_Use::elaborated_elements)); 1703 } 1704 1705 //------------------------------ScheduleAndBundle------------------------------ 1706 // Perform instruction scheduling and bundling over the sequence of 1707 // instructions in backwards order. 1708 void Compile::ScheduleAndBundle() { 1709 1710 // Don't optimize this if it isn't a method 1711 if (!_method) 1712 return; 1713 1714 // Don't optimize this if scheduling is disabled 1715 if (!do_scheduling()) 1716 return; 1717 1718 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1719 1720 // Create a data structure for all the scheduling information 1721 Scheduling scheduling(Thread::current()->resource_area(), *this); 1722 1723 // Initialize the space for the BufferBlob used to find and verify 1724 // instruction size in MachNode::emit_size() 1725 init_scratch_buffer_blob(MAX_const_size); 1726 if (failing()) return; // Out of memory 1727 1728 // Walk backwards over each basic block, computing the needed alignment 1729 // Walk over all the basic blocks 1730 scheduling.DoScheduling(); 1731 1732 // Clear the BufferBlob used for scheduling. 1733 clear_scratch_buffer_blob(); 1734 } 1735 1736 //------------------------------ComputeLocalLatenciesForward------------------- 1737 // Compute the latency of all the instructions. This is fairly simple, 1738 // because we already have a legal ordering. Walk over the instructions 1739 // from first to last, and compute the latency of the instruction based 1740 // on the latency of the preceding instruction(s). 1741 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1742 #ifndef PRODUCT 1743 if (_cfg->C->trace_opto_output()) 1744 tty->print("# -> ComputeLocalLatenciesForward\n"); 1745 #endif 1746 1747 // Walk over all the schedulable instructions 1748 for( uint j=_bb_start; j < _bb_end; j++ ) { 1749 1750 // This is a kludge, forcing all latency calculations to start at 1. 1751 // Used to allow latency 0 to force an instruction to the beginning 1752 // of the bb 1753 uint latency = 1; 1754 Node *use = bb->_nodes[j]; 1755 uint nlen = use->len(); 1756 1757 // Walk over all the inputs 1758 for ( uint k=0; k < nlen; k++ ) { 1759 Node *def = use->in(k); 1760 if (!def) 1761 continue; 1762 1763 uint l = _node_latency[def->_idx] + use->latency(k); 1764 if (latency < l) 1765 latency = l; 1766 } 1767 1768 _node_latency[use->_idx] = latency; 1769 1770 #ifndef PRODUCT 1771 if (_cfg->C->trace_opto_output()) { 1772 tty->print("# latency %4d: ", latency); 1773 use->dump(); 1774 } 1775 #endif 1776 } 1777 1778 #ifndef PRODUCT 1779 if (_cfg->C->trace_opto_output()) 1780 tty->print("# <- ComputeLocalLatenciesForward\n"); 1781 #endif 1782 1783 } // end ComputeLocalLatenciesForward 1784 1785 // See if this node fits into the present instruction bundle 1786 bool Scheduling::NodeFitsInBundle(Node *n) { 1787 uint n_idx = n->_idx; 1788 1789 // If this is the unconditional delay instruction, then it fits 1790 if (n == _unconditional_delay_slot) { 1791 #ifndef PRODUCT 1792 if (_cfg->C->trace_opto_output()) 1793 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1794 #endif 1795 return (true); 1796 } 1797 1798 // If the node cannot be scheduled this cycle, skip it 1799 if (_current_latency[n_idx] > _bundle_cycle_number) { 1800 #ifndef PRODUCT 1801 if (_cfg->C->trace_opto_output()) 1802 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1803 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1804 #endif 1805 return (false); 1806 } 1807 1808 const Pipeline *node_pipeline = n->pipeline(); 1809 1810 uint instruction_count = node_pipeline->instructionCount(); 1811 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1812 instruction_count = 0; 1813 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1814 instruction_count++; 1815 1816 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1817 #ifndef PRODUCT 1818 if (_cfg->C->trace_opto_output()) 1819 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1820 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1821 #endif 1822 return (false); 1823 } 1824 1825 // Don't allow non-machine nodes to be handled this way 1826 if (!n->is_Mach() && instruction_count == 0) 1827 return (false); 1828 1829 // See if there is any overlap 1830 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1831 1832 if (delay > 0) { 1833 #ifndef PRODUCT 1834 if (_cfg->C->trace_opto_output()) 1835 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1836 #endif 1837 return false; 1838 } 1839 1840 #ifndef PRODUCT 1841 if (_cfg->C->trace_opto_output()) 1842 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1843 #endif 1844 1845 return true; 1846 } 1847 1848 Node * Scheduling::ChooseNodeToBundle() { 1849 uint siz = _available.size(); 1850 1851 if (siz == 0) { 1852 1853 #ifndef PRODUCT 1854 if (_cfg->C->trace_opto_output()) 1855 tty->print("# ChooseNodeToBundle: NULL\n"); 1856 #endif 1857 return (NULL); 1858 } 1859 1860 // Fast path, if only 1 instruction in the bundle 1861 if (siz == 1) { 1862 #ifndef PRODUCT 1863 if (_cfg->C->trace_opto_output()) { 1864 tty->print("# ChooseNodeToBundle (only 1): "); 1865 _available[0]->dump(); 1866 } 1867 #endif 1868 return (_available[0]); 1869 } 1870 1871 // Don't bother, if the bundle is already full 1872 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1873 for ( uint i = 0; i < siz; i++ ) { 1874 Node *n = _available[i]; 1875 1876 // Skip projections, we'll handle them another way 1877 if (n->is_Proj()) 1878 continue; 1879 1880 // This presupposed that instructions are inserted into the 1881 // available list in a legality order; i.e. instructions that 1882 // must be inserted first are at the head of the list 1883 if (NodeFitsInBundle(n)) { 1884 #ifndef PRODUCT 1885 if (_cfg->C->trace_opto_output()) { 1886 tty->print("# ChooseNodeToBundle: "); 1887 n->dump(); 1888 } 1889 #endif 1890 return (n); 1891 } 1892 } 1893 } 1894 1895 // Nothing fits in this bundle, choose the highest priority 1896 #ifndef PRODUCT 1897 if (_cfg->C->trace_opto_output()) { 1898 tty->print("# ChooseNodeToBundle: "); 1899 _available[0]->dump(); 1900 } 1901 #endif 1902 1903 return _available[0]; 1904 } 1905 1906 //------------------------------AddNodeToAvailableList------------------------- 1907 void Scheduling::AddNodeToAvailableList(Node *n) { 1908 assert( !n->is_Proj(), "projections never directly made available" ); 1909 #ifndef PRODUCT 1910 if (_cfg->C->trace_opto_output()) { 1911 tty->print("# AddNodeToAvailableList: "); 1912 n->dump(); 1913 } 1914 #endif 1915 1916 int latency = _current_latency[n->_idx]; 1917 1918 // Insert in latency order (insertion sort) 1919 uint i; 1920 for ( i=0; i < _available.size(); i++ ) 1921 if (_current_latency[_available[i]->_idx] > latency) 1922 break; 1923 1924 // Special Check for compares following branches 1925 if( n->is_Mach() && _scheduled.size() > 0 ) { 1926 int op = n->as_Mach()->ideal_Opcode(); 1927 Node *last = _scheduled[0]; 1928 if( last->is_MachIf() && last->in(1) == n && 1929 ( op == Op_CmpI || 1930 op == Op_CmpU || 1931 op == Op_CmpP || 1932 op == Op_CmpF || 1933 op == Op_CmpD || 1934 op == Op_CmpL ) ) { 1935 1936 // Recalculate position, moving to front of same latency 1937 for ( i=0 ; i < _available.size(); i++ ) 1938 if (_current_latency[_available[i]->_idx] >= latency) 1939 break; 1940 } 1941 } 1942 1943 // Insert the node in the available list 1944 _available.insert(i, n); 1945 1946 #ifndef PRODUCT 1947 if (_cfg->C->trace_opto_output()) 1948 dump_available(); 1949 #endif 1950 } 1951 1952 //------------------------------DecrementUseCounts----------------------------- 1953 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 1954 for ( uint i=0; i < n->len(); i++ ) { 1955 Node *def = n->in(i); 1956 if (!def) continue; 1957 if( def->is_Proj() ) // If this is a machine projection, then 1958 def = def->in(0); // propagate usage thru to the base instruction 1959 1960 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 1961 continue; 1962 1963 // Compute the latency 1964 uint l = _bundle_cycle_number + n->latency(i); 1965 if (_current_latency[def->_idx] < l) 1966 _current_latency[def->_idx] = l; 1967 1968 // If this does not have uses then schedule it 1969 if ((--_uses[def->_idx]) == 0) 1970 AddNodeToAvailableList(def); 1971 } 1972 } 1973 1974 //------------------------------AddNodeToBundle-------------------------------- 1975 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 1976 #ifndef PRODUCT 1977 if (_cfg->C->trace_opto_output()) { 1978 tty->print("# AddNodeToBundle: "); 1979 n->dump(); 1980 } 1981 #endif 1982 1983 // Remove this from the available list 1984 uint i; 1985 for (i = 0; i < _available.size(); i++) 1986 if (_available[i] == n) 1987 break; 1988 assert(i < _available.size(), "entry in _available list not found"); 1989 _available.remove(i); 1990 1991 // See if this fits in the current bundle 1992 const Pipeline *node_pipeline = n->pipeline(); 1993 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 1994 1995 // Check for instructions to be placed in the delay slot. We 1996 // do this before we actually schedule the current instruction, 1997 // because the delay slot follows the current instruction. 1998 if (Pipeline::_branch_has_delay_slot && 1999 node_pipeline->hasBranchDelay() && 2000 !_unconditional_delay_slot) { 2001 2002 uint siz = _available.size(); 2003 2004 // Conditional branches can support an instruction that 2005 // is unconditionally executed and not dependent by the 2006 // branch, OR a conditionally executed instruction if 2007 // the branch is taken. In practice, this means that 2008 // the first instruction at the branch target is 2009 // copied to the delay slot, and the branch goes to 2010 // the instruction after that at the branch target 2011 if ( n->is_Mach() && n->is_Branch() ) { 2012 2013 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2014 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2015 2016 #ifndef PRODUCT 2017 _branches++; 2018 #endif 2019 2020 // At least 1 instruction is on the available list 2021 // that is not dependent on the branch 2022 for (uint i = 0; i < siz; i++) { 2023 Node *d = _available[i]; 2024 const Pipeline *avail_pipeline = d->pipeline(); 2025 2026 // Don't allow safepoints in the branch shadow, that will 2027 // cause a number of difficulties 2028 if ( avail_pipeline->instructionCount() == 1 && 2029 !avail_pipeline->hasMultipleBundles() && 2030 !avail_pipeline->hasBranchDelay() && 2031 Pipeline::instr_has_unit_size() && 2032 d->size(_regalloc) == Pipeline::instr_unit_size() && 2033 NodeFitsInBundle(d) && 2034 !node_bundling(d)->used_in_delay()) { 2035 2036 if (d->is_Mach() && !d->is_MachSafePoint()) { 2037 // A node that fits in the delay slot was found, so we need to 2038 // set the appropriate bits in the bundle pipeline information so 2039 // that it correctly indicates resource usage. Later, when we 2040 // attempt to add this instruction to the bundle, we will skip 2041 // setting the resource usage. 2042 _unconditional_delay_slot = d; 2043 node_bundling(n)->set_use_unconditional_delay(); 2044 node_bundling(d)->set_used_in_unconditional_delay(); 2045 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2046 _current_latency[d->_idx] = _bundle_cycle_number; 2047 _next_node = d; 2048 ++_bundle_instr_count; 2049 #ifndef PRODUCT 2050 _unconditional_delays++; 2051 #endif 2052 break; 2053 } 2054 } 2055 } 2056 } 2057 2058 // No delay slot, add a nop to the usage 2059 if (!_unconditional_delay_slot) { 2060 // See if adding an instruction in the delay slot will overflow 2061 // the bundle. 2062 if (!NodeFitsInBundle(_nop)) { 2063 #ifndef PRODUCT 2064 if (_cfg->C->trace_opto_output()) 2065 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2066 #endif 2067 step(1); 2068 } 2069 2070 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2071 _next_node = _nop; 2072 ++_bundle_instr_count; 2073 } 2074 2075 // See if the instruction in the delay slot requires a 2076 // step of the bundles 2077 if (!NodeFitsInBundle(n)) { 2078 #ifndef PRODUCT 2079 if (_cfg->C->trace_opto_output()) 2080 tty->print("# *** STEP(branch won't fit) ***\n"); 2081 #endif 2082 // Update the state information 2083 _bundle_instr_count = 0; 2084 _bundle_cycle_number += 1; 2085 _bundle_use.step(1); 2086 } 2087 } 2088 2089 // Get the number of instructions 2090 uint instruction_count = node_pipeline->instructionCount(); 2091 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2092 instruction_count = 0; 2093 2094 // Compute the latency information 2095 uint delay = 0; 2096 2097 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2098 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2099 if (relative_latency < 0) 2100 relative_latency = 0; 2101 2102 delay = _bundle_use.full_latency(relative_latency, node_usage); 2103 2104 // Does not fit in this bundle, start a new one 2105 if (delay > 0) { 2106 step(delay); 2107 2108 #ifndef PRODUCT 2109 if (_cfg->C->trace_opto_output()) 2110 tty->print("# *** STEP(%d) ***\n", delay); 2111 #endif 2112 } 2113 } 2114 2115 // If this was placed in the delay slot, ignore it 2116 if (n != _unconditional_delay_slot) { 2117 2118 if (delay == 0) { 2119 if (node_pipeline->hasMultipleBundles()) { 2120 #ifndef PRODUCT 2121 if (_cfg->C->trace_opto_output()) 2122 tty->print("# *** STEP(multiple instructions) ***\n"); 2123 #endif 2124 step(1); 2125 } 2126 2127 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2128 #ifndef PRODUCT 2129 if (_cfg->C->trace_opto_output()) 2130 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2131 instruction_count + _bundle_instr_count, 2132 Pipeline::_max_instrs_per_cycle); 2133 #endif 2134 step(1); 2135 } 2136 } 2137 2138 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2139 _bundle_instr_count++; 2140 2141 // Set the node's latency 2142 _current_latency[n->_idx] = _bundle_cycle_number; 2143 2144 // Now merge the functional unit information 2145 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2146 _bundle_use.add_usage(node_usage); 2147 2148 // Increment the number of instructions in this bundle 2149 _bundle_instr_count += instruction_count; 2150 2151 // Remember this node for later 2152 if (n->is_Mach()) 2153 _next_node = n; 2154 } 2155 2156 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2157 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2158 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2159 // into the block. All other scheduled nodes get put in the schedule here. 2160 int op = n->Opcode(); 2161 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2162 (op != Op_Node && // Not an unused antidepedence node and 2163 // not an unallocated boxlock 2164 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2165 2166 // Push any trailing projections 2167 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2168 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2169 Node *foi = n->fast_out(i); 2170 if( foi->is_Proj() ) 2171 _scheduled.push(foi); 2172 } 2173 } 2174 2175 // Put the instruction in the schedule list 2176 _scheduled.push(n); 2177 } 2178 2179 #ifndef PRODUCT 2180 if (_cfg->C->trace_opto_output()) 2181 dump_available(); 2182 #endif 2183 2184 // Walk all the definitions, decrementing use counts, and 2185 // if a definition has a 0 use count, place it in the available list. 2186 DecrementUseCounts(n,bb); 2187 } 2188 2189 //------------------------------ComputeUseCount-------------------------------- 2190 // This method sets the use count within a basic block. We will ignore all 2191 // uses outside the current basic block. As we are doing a backwards walk, 2192 // any node we reach that has a use count of 0 may be scheduled. This also 2193 // avoids the problem of cyclic references from phi nodes, as long as phi 2194 // nodes are at the front of the basic block. This method also initializes 2195 // the available list to the set of instructions that have no uses within this 2196 // basic block. 2197 void Scheduling::ComputeUseCount(const Block *bb) { 2198 #ifndef PRODUCT 2199 if (_cfg->C->trace_opto_output()) 2200 tty->print("# -> ComputeUseCount\n"); 2201 #endif 2202 2203 // Clear the list of available and scheduled instructions, just in case 2204 _available.clear(); 2205 _scheduled.clear(); 2206 2207 // No delay slot specified 2208 _unconditional_delay_slot = NULL; 2209 2210 #ifdef ASSERT 2211 for( uint i=0; i < bb->_nodes.size(); i++ ) 2212 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2213 #endif 2214 2215 // Force the _uses count to never go to zero for unscheduable pieces 2216 // of the block 2217 for( uint k = 0; k < _bb_start; k++ ) 2218 _uses[bb->_nodes[k]->_idx] = 1; 2219 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2220 _uses[bb->_nodes[l]->_idx] = 1; 2221 2222 // Iterate backwards over the instructions in the block. Don't count the 2223 // branch projections at end or the block header instructions. 2224 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2225 Node *n = bb->_nodes[j]; 2226 if( n->is_Proj() ) continue; // Projections handled another way 2227 2228 // Account for all uses 2229 for ( uint k = 0; k < n->len(); k++ ) { 2230 Node *inp = n->in(k); 2231 if (!inp) continue; 2232 assert(inp != n, "no cycles allowed" ); 2233 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2234 if( inp->is_Proj() ) // Skip through Proj's 2235 inp = inp->in(0); 2236 ++_uses[inp->_idx]; // Count 1 block-local use 2237 } 2238 } 2239 2240 // If this instruction has a 0 use count, then it is available 2241 if (!_uses[n->_idx]) { 2242 _current_latency[n->_idx] = _bundle_cycle_number; 2243 AddNodeToAvailableList(n); 2244 } 2245 2246 #ifndef PRODUCT 2247 if (_cfg->C->trace_opto_output()) { 2248 tty->print("# uses: %3d: ", _uses[n->_idx]); 2249 n->dump(); 2250 } 2251 #endif 2252 } 2253 2254 #ifndef PRODUCT 2255 if (_cfg->C->trace_opto_output()) 2256 tty->print("# <- ComputeUseCount\n"); 2257 #endif 2258 } 2259 2260 // This routine performs scheduling on each basic block in reverse order, 2261 // using instruction latencies and taking into account function unit 2262 // availability. 2263 void Scheduling::DoScheduling() { 2264 #ifndef PRODUCT 2265 if (_cfg->C->trace_opto_output()) 2266 tty->print("# -> DoScheduling\n"); 2267 #endif 2268 2269 Block *succ_bb = NULL; 2270 Block *bb; 2271 2272 // Walk over all the basic blocks in reverse order 2273 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2274 bb = _cfg->_blocks[i]; 2275 2276 #ifndef PRODUCT 2277 if (_cfg->C->trace_opto_output()) { 2278 tty->print("# Schedule BB#%03d (initial)\n", i); 2279 for (uint j = 0; j < bb->_nodes.size(); j++) 2280 bb->_nodes[j]->dump(); 2281 } 2282 #endif 2283 2284 // On the head node, skip processing 2285 if( bb == _cfg->_broot ) 2286 continue; 2287 2288 // Skip empty, connector blocks 2289 if (bb->is_connector()) 2290 continue; 2291 2292 // If the following block is not the sole successor of 2293 // this one, then reset the pipeline information 2294 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2295 #ifndef PRODUCT 2296 if (_cfg->C->trace_opto_output()) { 2297 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2298 _next_node->_idx, _bundle_instr_count); 2299 } 2300 #endif 2301 step_and_clear(); 2302 } 2303 2304 // Leave untouched the starting instruction, any Phis, a CreateEx node 2305 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2306 _bb_end = bb->_nodes.size()-1; 2307 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2308 Node *n = bb->_nodes[_bb_start]; 2309 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2310 // Also, MachIdealNodes do not get scheduled 2311 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2312 MachNode *mach = n->as_Mach(); 2313 int iop = mach->ideal_Opcode(); 2314 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2315 if( iop == Op_Con ) continue; // Do not schedule Top 2316 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2317 mach->pipeline() == MachNode::pipeline_class() && 2318 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2319 continue; 2320 break; // Funny loop structure to be sure... 2321 } 2322 // Compute last "interesting" instruction in block - last instruction we 2323 // might schedule. _bb_end points just after last schedulable inst. We 2324 // normally schedule conditional branches (despite them being forced last 2325 // in the block), because they have delay slots we can fill. Calls all 2326 // have their delay slots filled in the template expansions, so we don't 2327 // bother scheduling them. 2328 Node *last = bb->_nodes[_bb_end]; 2329 if( last->is_Catch() || 2330 // Exclude unreachable path case when Halt node is in a separate block. 2331 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2332 // There must be a prior call. Skip it. 2333 while( !bb->_nodes[--_bb_end]->is_Call() ) { 2334 assert( bb->_nodes[_bb_end]->is_Proj(), "skipping projections after expected call" ); 2335 } 2336 } else if( last->is_MachNullCheck() ) { 2337 // Backup so the last null-checked memory instruction is 2338 // outside the schedulable range. Skip over the nullcheck, 2339 // projection, and the memory nodes. 2340 Node *mem = last->in(1); 2341 do { 2342 _bb_end--; 2343 } while (mem != bb->_nodes[_bb_end]); 2344 } else { 2345 // Set _bb_end to point after last schedulable inst. 2346 _bb_end++; 2347 } 2348 2349 assert( _bb_start <= _bb_end, "inverted block ends" ); 2350 2351 // Compute the register antidependencies for the basic block 2352 ComputeRegisterAntidependencies(bb); 2353 if (_cfg->C->failing()) return; // too many D-U pinch points 2354 2355 // Compute intra-bb latencies for the nodes 2356 ComputeLocalLatenciesForward(bb); 2357 2358 // Compute the usage within the block, and set the list of all nodes 2359 // in the block that have no uses within the block. 2360 ComputeUseCount(bb); 2361 2362 // Schedule the remaining instructions in the block 2363 while ( _available.size() > 0 ) { 2364 Node *n = ChooseNodeToBundle(); 2365 AddNodeToBundle(n,bb); 2366 } 2367 2368 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2369 #ifdef ASSERT 2370 for( uint l = _bb_start; l < _bb_end; l++ ) { 2371 Node *n = bb->_nodes[l]; 2372 uint m; 2373 for( m = 0; m < _bb_end-_bb_start; m++ ) 2374 if( _scheduled[m] == n ) 2375 break; 2376 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2377 } 2378 #endif 2379 2380 // Now copy the instructions (in reverse order) back to the block 2381 for ( uint k = _bb_start; k < _bb_end; k++ ) 2382 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2383 2384 #ifndef PRODUCT 2385 if (_cfg->C->trace_opto_output()) { 2386 tty->print("# Schedule BB#%03d (final)\n", i); 2387 uint current = 0; 2388 for (uint j = 0; j < bb->_nodes.size(); j++) { 2389 Node *n = bb->_nodes[j]; 2390 if( valid_bundle_info(n) ) { 2391 Bundle *bundle = node_bundling(n); 2392 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2393 tty->print("*** Bundle: "); 2394 bundle->dump(); 2395 } 2396 n->dump(); 2397 } 2398 } 2399 } 2400 #endif 2401 #ifdef ASSERT 2402 verify_good_schedule(bb,"after block local scheduling"); 2403 #endif 2404 } 2405 2406 #ifndef PRODUCT 2407 if (_cfg->C->trace_opto_output()) 2408 tty->print("# <- DoScheduling\n"); 2409 #endif 2410 2411 // Record final node-bundling array location 2412 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2413 2414 } // end DoScheduling 2415 2416 //------------------------------verify_good_schedule--------------------------- 2417 // Verify that no live-range used in the block is killed in the block by a 2418 // wrong DEF. This doesn't verify live-ranges that span blocks. 2419 2420 // Check for edge existence. Used to avoid adding redundant precedence edges. 2421 static bool edge_from_to( Node *from, Node *to ) { 2422 for( uint i=0; i<from->len(); i++ ) 2423 if( from->in(i) == to ) 2424 return true; 2425 return false; 2426 } 2427 2428 #ifdef ASSERT 2429 //------------------------------verify_do_def---------------------------------- 2430 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2431 // Check for bad kills 2432 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2433 Node *prior_use = _reg_node[def]; 2434 if( prior_use && !edge_from_to(prior_use,n) ) { 2435 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2436 n->dump(); 2437 tty->print_cr("..."); 2438 prior_use->dump(); 2439 assert(edge_from_to(prior_use,n),msg); 2440 } 2441 _reg_node.map(def,NULL); // Kill live USEs 2442 } 2443 } 2444 2445 //------------------------------verify_good_schedule--------------------------- 2446 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2447 2448 // Zap to something reasonable for the verify code 2449 _reg_node.clear(); 2450 2451 // Walk over the block backwards. Check to make sure each DEF doesn't 2452 // kill a live value (other than the one it's supposed to). Add each 2453 // USE to the live set. 2454 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2455 Node *n = b->_nodes[i]; 2456 int n_op = n->Opcode(); 2457 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2458 // Fat-proj kills a slew of registers 2459 RegMask rm = n->out_RegMask();// Make local copy 2460 while( rm.is_NotEmpty() ) { 2461 OptoReg::Name kill = rm.find_first_elem(); 2462 rm.Remove(kill); 2463 verify_do_def( n, kill, msg ); 2464 } 2465 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2466 // Get DEF'd registers the normal way 2467 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2468 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2469 } 2470 2471 // Now make all USEs live 2472 for( uint i=1; i<n->req(); i++ ) { 2473 Node *def = n->in(i); 2474 assert(def != 0, "input edge required"); 2475 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2476 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2477 if( OptoReg::is_valid(reg_lo) ) { 2478 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2479 _reg_node.map(reg_lo,n); 2480 } 2481 if( OptoReg::is_valid(reg_hi) ) { 2482 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2483 _reg_node.map(reg_hi,n); 2484 } 2485 } 2486 2487 } 2488 2489 // Zap to something reasonable for the Antidependence code 2490 _reg_node.clear(); 2491 } 2492 #endif 2493 2494 // Conditionally add precedence edges. Avoid putting edges on Projs. 2495 static void add_prec_edge_from_to( Node *from, Node *to ) { 2496 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2497 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2498 from = from->in(0); 2499 } 2500 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2501 !edge_from_to( from, to ) ) // Avoid duplicate edge 2502 from->add_prec(to); 2503 } 2504 2505 //------------------------------anti_do_def------------------------------------ 2506 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2507 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2508 return; 2509 2510 Node *pinch = _reg_node[def_reg]; // Get pinch point 2511 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2512 is_def ) { // Check for a true def (not a kill) 2513 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2514 return; 2515 } 2516 2517 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2518 debug_only( def = (Node*)0xdeadbeef; ) 2519 2520 // After some number of kills there _may_ be a later def 2521 Node *later_def = NULL; 2522 2523 // Finding a kill requires a real pinch-point. 2524 // Check for not already having a pinch-point. 2525 // Pinch points are Op_Node's. 2526 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2527 later_def = pinch; // Must be def/kill as optimistic pinch-point 2528 if ( _pinch_free_list.size() > 0) { 2529 pinch = _pinch_free_list.pop(); 2530 } else { 2531 pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be 2532 } 2533 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2534 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2535 return; 2536 } 2537 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2538 _reg_node.map(def_reg,pinch); // Record pinch-point 2539 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2540 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2541 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2542 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2543 later_def = NULL; // and no later def 2544 } 2545 pinch->set_req(0,later_def); // Hook later def so we can find it 2546 } else { // Else have valid pinch point 2547 if( pinch->in(0) ) // If there is a later-def 2548 later_def = pinch->in(0); // Get it 2549 } 2550 2551 // Add output-dependence edge from later def to kill 2552 if( later_def ) // If there is some original def 2553 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2554 2555 // See if current kill is also a use, and so is forced to be the pinch-point. 2556 if( pinch->Opcode() == Op_Node ) { 2557 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2558 for( uint i=1; i<uses->req(); i++ ) { 2559 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2560 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2561 // Yes, found a use/kill pinch-point 2562 pinch->set_req(0,NULL); // 2563 pinch->replace_by(kill); // Move anti-dep edges up 2564 pinch = kill; 2565 _reg_node.map(def_reg,pinch); 2566 return; 2567 } 2568 } 2569 } 2570 2571 // Add edge from kill to pinch-point 2572 add_prec_edge_from_to(kill,pinch); 2573 } 2574 2575 //------------------------------anti_do_use------------------------------------ 2576 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2577 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2578 return; 2579 Node *pinch = _reg_node[use_reg]; // Get pinch point 2580 // Check for no later def_reg/kill in block 2581 if( pinch && _bbs[pinch->_idx] == b && 2582 // Use has to be block-local as well 2583 _bbs[use->_idx] == b ) { 2584 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2585 pinch->req() == 1 ) { // pinch not yet in block? 2586 pinch->del_req(0); // yank pointer to later-def, also set flag 2587 // Insert the pinch-point in the block just after the last use 2588 b->_nodes.insert(b->find_node(use)+1,pinch); 2589 _bb_end++; // Increase size scheduled region in block 2590 } 2591 2592 add_prec_edge_from_to(pinch,use); 2593 } 2594 } 2595 2596 //------------------------------ComputeRegisterAntidependences----------------- 2597 // We insert antidependences between the reads and following write of 2598 // allocated registers to prevent illegal code motion. Hopefully, the 2599 // number of added references should be fairly small, especially as we 2600 // are only adding references within the current basic block. 2601 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2602 2603 #ifdef ASSERT 2604 verify_good_schedule(b,"before block local scheduling"); 2605 #endif 2606 2607 // A valid schedule, for each register independently, is an endless cycle 2608 // of: a def, then some uses (connected to the def by true dependencies), 2609 // then some kills (defs with no uses), finally the cycle repeats with a new 2610 // def. The uses are allowed to float relative to each other, as are the 2611 // kills. No use is allowed to slide past a kill (or def). This requires 2612 // antidependencies between all uses of a single def and all kills that 2613 // follow, up to the next def. More edges are redundant, because later defs 2614 // & kills are already serialized with true or antidependencies. To keep 2615 // the edge count down, we add a 'pinch point' node if there's more than 2616 // one use or more than one kill/def. 2617 2618 // We add dependencies in one bottom-up pass. 2619 2620 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2621 2622 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2623 // register. If not, we record the DEF/KILL in _reg_node, the 2624 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2625 // "pinch point", a new Node that's in the graph but not in the block. 2626 // We put edges from the prior and current DEF/KILLs to the pinch point. 2627 // We put the pinch point in _reg_node. If there's already a pinch point 2628 // we merely add an edge from the current DEF/KILL to the pinch point. 2629 2630 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2631 // put an edge from the pinch point to the USE. 2632 2633 // To be expedient, the _reg_node array is pre-allocated for the whole 2634 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2635 // or a valid def/kill/pinch-point, or a leftover node from some prior 2636 // block. Leftover node from some prior block is treated like a NULL (no 2637 // prior def, so no anti-dependence needed). Valid def is distinguished by 2638 // it being in the current block. 2639 bool fat_proj_seen = false; 2640 uint last_safept = _bb_end-1; 2641 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2642 Node* last_safept_node = end_node; 2643 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2644 Node *n = b->_nodes[i]; 2645 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2646 if( n->Opcode() == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2647 // Fat-proj kills a slew of registers 2648 // This can add edges to 'n' and obscure whether or not it was a def, 2649 // hence the is_def flag. 2650 fat_proj_seen = true; 2651 RegMask rm = n->out_RegMask();// Make local copy 2652 while( rm.is_NotEmpty() ) { 2653 OptoReg::Name kill = rm.find_first_elem(); 2654 rm.Remove(kill); 2655 anti_do_def( b, n, kill, is_def ); 2656 } 2657 } else { 2658 // Get DEF'd registers the normal way 2659 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2660 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2661 } 2662 2663 // Check each register used by this instruction for a following DEF/KILL 2664 // that must occur afterward and requires an anti-dependence edge. 2665 for( uint j=0; j<n->req(); j++ ) { 2666 Node *def = n->in(j); 2667 if( def ) { 2668 assert( def->Opcode() != Op_MachProj || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2669 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2670 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2671 } 2672 } 2673 // Do not allow defs of new derived values to float above GC 2674 // points unless the base is definitely available at the GC point. 2675 2676 Node *m = b->_nodes[i]; 2677 2678 // Add precedence edge from following safepoint to use of derived pointer 2679 if( last_safept_node != end_node && 2680 m != last_safept_node) { 2681 for (uint k = 1; k < m->req(); k++) { 2682 const Type *t = m->in(k)->bottom_type(); 2683 if( t->isa_oop_ptr() && 2684 t->is_ptr()->offset() != 0 ) { 2685 last_safept_node->add_prec( m ); 2686 break; 2687 } 2688 } 2689 } 2690 2691 if( n->jvms() ) { // Precedence edge from derived to safept 2692 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2693 if( b->_nodes[last_safept] != last_safept_node ) { 2694 last_safept = b->find_node(last_safept_node); 2695 } 2696 for( uint j=last_safept; j > i; j-- ) { 2697 Node *mach = b->_nodes[j]; 2698 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2699 mach->add_prec( n ); 2700 } 2701 last_safept = i; 2702 last_safept_node = m; 2703 } 2704 } 2705 2706 if (fat_proj_seen) { 2707 // Garbage collect pinch nodes that were not consumed. 2708 // They are usually created by a fat kill MachProj for a call. 2709 garbage_collect_pinch_nodes(); 2710 } 2711 } 2712 2713 //------------------------------garbage_collect_pinch_nodes------------------------------- 2714 2715 // Garbage collect pinch nodes for reuse by other blocks. 2716 // 2717 // The block scheduler's insertion of anti-dependence 2718 // edges creates many pinch nodes when the block contains 2719 // 2 or more Calls. A pinch node is used to prevent a 2720 // combinatorial explosion of edges. If a set of kills for a 2721 // register is anti-dependent on a set of uses (or defs), rather 2722 // than adding an edge in the graph between each pair of kill 2723 // and use (or def), a pinch is inserted between them: 2724 // 2725 // use1 use2 use3 2726 // \ | / 2727 // \ | / 2728 // pinch 2729 // / | \ 2730 // / | \ 2731 // kill1 kill2 kill3 2732 // 2733 // One pinch node is created per register killed when 2734 // the second call is encountered during a backwards pass 2735 // over the block. Most of these pinch nodes are never 2736 // wired into the graph because the register is never 2737 // used or def'ed in the block. 2738 // 2739 void Scheduling::garbage_collect_pinch_nodes() { 2740 #ifndef PRODUCT 2741 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2742 #endif 2743 int trace_cnt = 0; 2744 for (uint k = 0; k < _reg_node.Size(); k++) { 2745 Node* pinch = _reg_node[k]; 2746 if (pinch != NULL && pinch->Opcode() == Op_Node && 2747 // no predecence input edges 2748 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2749 cleanup_pinch(pinch); 2750 _pinch_free_list.push(pinch); 2751 _reg_node.map(k, NULL); 2752 #ifndef PRODUCT 2753 if (_cfg->C->trace_opto_output()) { 2754 trace_cnt++; 2755 if (trace_cnt > 40) { 2756 tty->print("\n"); 2757 trace_cnt = 0; 2758 } 2759 tty->print(" %d", pinch->_idx); 2760 } 2761 #endif 2762 } 2763 } 2764 #ifndef PRODUCT 2765 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2766 #endif 2767 } 2768 2769 // Clean up a pinch node for reuse. 2770 void Scheduling::cleanup_pinch( Node *pinch ) { 2771 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2772 2773 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2774 Node* use = pinch->last_out(i); 2775 uint uses_found = 0; 2776 for (uint j = use->req(); j < use->len(); j++) { 2777 if (use->in(j) == pinch) { 2778 use->rm_prec(j); 2779 uses_found++; 2780 } 2781 } 2782 assert(uses_found > 0, "must be a precedence edge"); 2783 i -= uses_found; // we deleted 1 or more copies of this edge 2784 } 2785 // May have a later_def entry 2786 pinch->set_req(0, NULL); 2787 } 2788 2789 //------------------------------print_statistics------------------------------- 2790 #ifndef PRODUCT 2791 2792 void Scheduling::dump_available() const { 2793 tty->print("#Availist "); 2794 for (uint i = 0; i < _available.size(); i++) 2795 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2796 tty->cr(); 2797 } 2798 2799 // Print Scheduling Statistics 2800 void Scheduling::print_statistics() { 2801 // Print the size added by nops for bundling 2802 tty->print("Nops added %d bytes to total of %d bytes", 2803 _total_nop_size, _total_method_size); 2804 if (_total_method_size > 0) 2805 tty->print(", for %.2f%%", 2806 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2807 tty->print("\n"); 2808 2809 // Print the number of branch shadows filled 2810 if (Pipeline::_branch_has_delay_slot) { 2811 tty->print("Of %d branches, %d had unconditional delay slots filled", 2812 _total_branches, _total_unconditional_delays); 2813 if (_total_branches > 0) 2814 tty->print(", for %.2f%%", 2815 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2816 tty->print("\n"); 2817 } 2818 2819 uint total_instructions = 0, total_bundles = 0; 2820 2821 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2822 uint bundle_count = _total_instructions_per_bundle[i]; 2823 total_instructions += bundle_count * i; 2824 total_bundles += bundle_count; 2825 } 2826 2827 if (total_bundles > 0) 2828 tty->print("Average ILP (excluding nops) is %.2f\n", 2829 ((double)total_instructions) / ((double)total_bundles)); 2830 } 2831 #endif