1 /* 2 * Copyright 1998-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #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 // Initialize the space for the BufferBlob used to find and verify 49 // instruction size in MachNode::emit_size() 50 init_scratch_buffer_blob(); 51 if (failing()) return; // Out of memory 52 53 // The number of new nodes (mostly MachNop) is proportional to 54 // the number of java calls and inner loops which are aligned. 55 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 56 C->inner_loops()*(OptoLoopAlignment-1)), 57 "out of nodes before code generation" ) ) { 58 return; 59 } 60 // Make sure I can find the Start Node 61 Block_Array& bbs = _cfg->_bbs; 62 Block *entry = _cfg->_blocks[1]; 63 Block *broot = _cfg->_broot; 64 65 const StartNode *start = entry->_nodes[0]->as_Start(); 66 67 // Replace StartNode with prolog 68 MachPrologNode *prolog = new (this) MachPrologNode(); 69 entry->_nodes.map( 0, prolog ); 70 bbs.map( prolog->_idx, entry ); 71 bbs.map( start->_idx, NULL ); // start is no longer in any block 72 73 // Virtual methods need an unverified entry point 74 75 if( is_osr_compilation() ) { 76 if( PoisonOSREntry ) { 77 // TODO: Should use a ShouldNotReachHereNode... 78 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 79 } 80 } else { 81 if( _method && !_method->flags().is_static() ) { 82 // Insert unvalidated entry point 83 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 84 } 85 86 } 87 88 89 // Break before main entry point 90 if( (_method && _method->break_at_execute()) 91 #ifndef PRODUCT 92 ||(OptoBreakpoint && is_method_compilation()) 93 ||(OptoBreakpointOSR && is_osr_compilation()) 94 ||(OptoBreakpointC2R && !_method) 95 #endif 96 ) { 97 // checking for _method means that OptoBreakpoint does not apply to 98 // runtime stubs or frame converters 99 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 100 } 101 102 // Insert epilogs before every return 103 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 104 Block *b = _cfg->_blocks[i]; 105 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point? 106 Node *m = b->end(); 107 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) { 108 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 109 b->add_inst( epilog ); 110 bbs.map(epilog->_idx, b); 111 //_regalloc->set_bad(epilog->_idx); // Already initialized this way. 112 } 113 } 114 } 115 116 # ifdef ENABLE_ZAP_DEAD_LOCALS 117 if ( ZapDeadCompiledLocals ) Insert_zap_nodes(); 118 # endif 119 120 ScheduleAndBundle(); 121 122 #ifndef PRODUCT 123 if (trace_opto_output()) { 124 tty->print("\n---- After ScheduleAndBundle ----\n"); 125 for (uint i = 0; i < _cfg->_num_blocks; i++) { 126 tty->print("\nBB#%03d:\n", i); 127 Block *bb = _cfg->_blocks[i]; 128 for (uint j = 0; j < bb->_nodes.size(); j++) { 129 Node *n = bb->_nodes[j]; 130 OptoReg::Name reg = _regalloc->get_reg_first(n); 131 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 132 n->dump(); 133 } 134 } 135 } 136 #endif 137 138 if (failing()) return; 139 140 BuildOopMaps(); 141 142 if (failing()) return; 143 144 Fill_buffer(); 145 } 146 147 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 148 // Determine if we need to generate a stack overflow check. 149 // Do it if the method is not a stub function and 150 // has java calls or has frame size > vm_page_size/8. 151 return (stub_function() == NULL && 152 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 153 } 154 155 bool Compile::need_register_stack_bang() const { 156 // Determine if we need to generate a register stack overflow check. 157 // This is only used on architectures which have split register 158 // and memory stacks (ie. IA64). 159 // Bang if the method is not a stub function and has java calls 160 return (stub_function() == NULL && has_java_calls()); 161 } 162 163 # ifdef ENABLE_ZAP_DEAD_LOCALS 164 165 166 // In order to catch compiler oop-map bugs, we have implemented 167 // a debugging mode called ZapDeadCompilerLocals. 168 // This mode causes the compiler to insert a call to a runtime routine, 169 // "zap_dead_locals", right before each place in compiled code 170 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 171 // The runtime routine checks that locations mapped as oops are really 172 // oops, that locations mapped as values do not look like oops, 173 // and that locations mapped as dead are not used later 174 // (by zapping them to an invalid address). 175 176 int Compile::_CompiledZap_count = 0; 177 178 void Compile::Insert_zap_nodes() { 179 bool skip = false; 180 181 182 // Dink with static counts because code code without the extra 183 // runtime calls is MUCH faster for debugging purposes 184 185 if ( CompileZapFirst == 0 ) ; // nothing special 186 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 187 else if ( CompileZapFirst == CompiledZap_count() ) 188 warning("starting zap compilation after skipping"); 189 190 if ( CompileZapLast == -1 ) ; // nothing special 191 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 192 else if ( CompileZapLast == CompiledZap_count() ) 193 warning("about to compile last zap"); 194 195 ++_CompiledZap_count; // counts skipped zaps, too 196 197 if ( skip ) return; 198 199 200 if ( _method == NULL ) 201 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 202 203 // Insert call to zap runtime stub before every node with an oop map 204 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 205 Block *b = _cfg->_blocks[i]; 206 for ( uint j = 0; j < b->_nodes.size(); ++j ) { 207 Node *n = b->_nodes[j]; 208 209 // Determining if we should insert a zap-a-lot node in output. 210 // We do that for all nodes that has oopmap info, except for calls 211 // to allocation. Calls to allocation passes in the old top-of-eden pointer 212 // and expect the C code to reset it. Hence, there can be no safepoints between 213 // the inlined-allocation and the call to new_Java, etc. 214 // We also cannot zap monitor calls, as they must hold the microlock 215 // during the call to Zap, which also wants to grab the microlock. 216 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 217 if ( insert ) { // it is MachSafePoint 218 if ( !n->is_MachCall() ) { 219 insert = false; 220 } else if ( n->is_MachCall() ) { 221 MachCallNode* call = n->as_MachCall(); 222 if (call->entry_point() == OptoRuntime::new_instance_Java() || 223 call->entry_point() == OptoRuntime::new_array_Java() || 224 call->entry_point() == OptoRuntime::multianewarray2_Java() || 225 call->entry_point() == OptoRuntime::multianewarray3_Java() || 226 call->entry_point() == OptoRuntime::multianewarray4_Java() || 227 call->entry_point() == OptoRuntime::multianewarray5_Java() || 228 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 229 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 230 ) { 231 insert = false; 232 } 233 } 234 if (insert) { 235 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 236 b->_nodes.insert( j, zap ); 237 _cfg->_bbs.map( zap->_idx, b ); 238 ++j; 239 } 240 } 241 } 242 } 243 } 244 245 246 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 247 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 248 CallStaticJavaNode* ideal_node = 249 new (this, tf->domain()->cnt()) CallStaticJavaNode( tf, 250 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 251 "call zap dead locals stub", 0, TypePtr::BOTTOM); 252 // We need to copy the OopMap from the site we're zapping at. 253 // We have to make a copy, because the zap site might not be 254 // a call site, and zap_dead is a call site. 255 OopMap* clone = node_to_check->oop_map()->deep_copy(); 256 257 // Add the cloned OopMap to the zap node 258 ideal_node->set_oop_map(clone); 259 return _matcher->match_sfpt(ideal_node); 260 } 261 262 //------------------------------is_node_getting_a_safepoint-------------------- 263 bool Compile::is_node_getting_a_safepoint( Node* n) { 264 // This code duplicates the logic prior to the call of add_safepoint 265 // below in this file. 266 if( n->is_MachSafePoint() ) return true; 267 return false; 268 } 269 270 # endif // ENABLE_ZAP_DEAD_LOCALS 271 272 //------------------------------compute_loop_first_inst_sizes------------------ 273 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 274 // of a loop. When aligning a loop we need to provide enough instructions 275 // in cpu's fetch buffer to feed decoders. The loop alignment could be 276 // avoided if we have enough instructions in fetch buffer at the head of a loop. 277 // By default, the size is set to 999999 by Block's constructor so that 278 // a loop will be aligned if the size is not reset here. 279 // 280 // Note: Mach instructions could contain several HW instructions 281 // so the size is estimated only. 282 // 283 void Compile::compute_loop_first_inst_sizes() { 284 // The next condition is used to gate the loop alignment optimization. 285 // Don't aligned a loop if there are enough instructions at the head of a loop 286 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 287 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 288 // equal to 11 bytes which is the largest address NOP instruction. 289 if( MaxLoopPad < OptoLoopAlignment-1 ) { 290 uint last_block = _cfg->_num_blocks-1; 291 for( uint i=1; i <= last_block; i++ ) { 292 Block *b = _cfg->_blocks[i]; 293 // Check the first loop's block which requires an alignment. 294 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) { 295 uint sum_size = 0; 296 uint inst_cnt = NumberOfLoopInstrToAlign; 297 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 298 299 // Check subsequent fallthrough blocks if the loop's first 300 // block(s) does not have enough instructions. 301 Block *nb = b; 302 while( inst_cnt > 0 && 303 i < last_block && 304 !_cfg->_blocks[i+1]->has_loop_alignment() && 305 !nb->has_successor(b) ) { 306 i++; 307 nb = _cfg->_blocks[i]; 308 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 309 } // while( inst_cnt > 0 && i < last_block ) 310 311 b->set_first_inst_size(sum_size); 312 } // f( b->head()->is_Loop() ) 313 } // for( i <= last_block ) 314 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 315 } 316 317 //----------------------Shorten_branches--------------------------------------- 318 // The architecture description provides short branch variants for some long 319 // branch instructions. Replace eligible long branches with short branches. 320 void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size, int& const_size) { 321 322 // fill in the nop array for bundling computations 323 MachNode *_nop_list[Bundle::_nop_count]; 324 Bundle::initialize_nops(_nop_list, this); 325 326 // ------------------ 327 // Compute size of each block, method size, and relocation information size 328 uint *jmp_end = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); 329 uint *blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1); 330 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) 331 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) 332 blk_starts[0] = 0; 333 334 // Initialize the sizes to 0 335 code_size = 0; // Size in bytes of generated code 336 stub_size = 0; // Size in bytes of all stub entries 337 // Size in bytes of all relocation entries, including those in local stubs. 338 // Start with 2-bytes of reloc info for the unvalidated entry point 339 reloc_size = 1; // Number of relocation entries 340 const_size = 0; // size of fp constants in words 341 342 // Make three passes. The first computes pessimistic blk_starts, 343 // relative jmp_end, reloc_size and const_size information. 344 // The second performs short branch substitution using the pessimistic 345 // sizing. The third inserts nops where needed. 346 347 Node *nj; // tmp 348 349 // Step one, perform a pessimistic sizing pass. 350 uint i; 351 uint min_offset_from_last_call = 1; // init to a positive value 352 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 353 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 354 Block *b = _cfg->_blocks[i]; 355 356 // Sum all instruction sizes to compute block size 357 uint last_inst = b->_nodes.size(); 358 uint blk_size = 0; 359 for( uint j = 0; j<last_inst; j++ ) { 360 nj = b->_nodes[j]; 361 uint inst_size = nj->size(_regalloc); 362 blk_size += inst_size; 363 // Handle machine instruction nodes 364 if( nj->is_Mach() ) { 365 MachNode *mach = nj->as_Mach(); 366 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 367 reloc_size += mach->reloc(); 368 const_size += mach->const_size(); 369 if( mach->is_MachCall() ) { 370 MachCallNode *mcall = mach->as_MachCall(); 371 // This destination address is NOT PC-relative 372 373 mcall->method_set((intptr_t)mcall->entry_point()); 374 375 if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) { 376 stub_size += size_java_to_interp(); 377 reloc_size += reloc_java_to_interp(); 378 } 379 } else if (mach->is_MachSafePoint()) { 380 // If call/safepoint are adjacent, account for possible 381 // nop to disambiguate the two safepoints. 382 if (min_offset_from_last_call == 0) { 383 blk_size += nop_size; 384 } 385 } 386 } 387 min_offset_from_last_call += inst_size; 388 // Remember end of call offset 389 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 390 min_offset_from_last_call = 0; 391 } 392 } 393 394 // During short branch replacement, we store the relative (to blk_starts) 395 // end of jump in jmp_end, rather than the absolute end of jump. This 396 // is so that we do not need to recompute sizes of all nodes when we compute 397 // correct blk_starts in our next sizing pass. 398 jmp_end[i] = blk_size; 399 DEBUG_ONLY( jmp_target[i] = 0; ) 400 401 // When the next block starts a loop, we may insert pad NOP 402 // instructions. Since we cannot know our future alignment, 403 // assume the worst. 404 if( i<_cfg->_num_blocks-1 ) { 405 Block *nb = _cfg->_blocks[i+1]; 406 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 407 if( max_loop_pad > 0 ) { 408 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 409 blk_size += max_loop_pad; 410 } 411 } 412 413 // Save block size; update total method size 414 blk_starts[i+1] = blk_starts[i]+blk_size; 415 } 416 417 // Step two, replace eligible long jumps. 418 419 // Note: this will only get the long branches within short branch 420 // range. Another pass might detect more branches that became 421 // candidates because the shortening in the first pass exposed 422 // more opportunities. Unfortunately, this would require 423 // recomputing the starting and ending positions for the blocks 424 for( i=0; i<_cfg->_num_blocks; i++ ) { 425 Block *b = _cfg->_blocks[i]; 426 427 int j; 428 // Find the branch; ignore trailing NOPs. 429 for( j = b->_nodes.size()-1; j>=0; j-- ) { 430 nj = b->_nodes[j]; 431 if( !nj->is_Mach() || nj->as_Mach()->ideal_Opcode() != Op_Con ) 432 break; 433 } 434 435 if (j >= 0) { 436 if( nj->is_Mach() && nj->as_Mach()->may_be_short_branch() ) { 437 MachNode *mach = nj->as_Mach(); 438 // This requires the TRUE branch target be in succs[0] 439 uint bnum = b->non_connector_successor(0)->_pre_order; 440 uintptr_t target = blk_starts[bnum]; 441 if( mach->is_pc_relative() ) { 442 int offset = target-(blk_starts[i] + jmp_end[i]); 443 if (_matcher->is_short_branch_offset(mach->rule(), offset)) { 444 // We've got a winner. Replace this branch. 445 MachNode* replacement = mach->short_branch_version(this); 446 b->_nodes.map(j, replacement); 447 mach->subsume_by(replacement); 448 449 // Update the jmp_end size to save time in our 450 // next pass. 451 jmp_end[i] -= (mach->size(_regalloc) - replacement->size(_regalloc)); 452 DEBUG_ONLY( jmp_target[i] = bnum; ); 453 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 454 } 455 } else { 456 #ifndef PRODUCT 457 mach->dump(3); 458 #endif 459 Unimplemented(); 460 } 461 } 462 } 463 } 464 465 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 466 // of a loop. It is used to determine the padding for loop alignment. 467 compute_loop_first_inst_sizes(); 468 469 // Step 3, compute the offsets of all the labels 470 uint last_call_adr = max_uint; 471 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 472 // copy the offset of the beginning to the corresponding label 473 assert(labels[i].is_unused(), "cannot patch at this point"); 474 labels[i].bind_loc(blk_starts[i], CodeBuffer::SECT_INSTS); 475 476 // insert padding for any instructions that need it 477 Block *b = _cfg->_blocks[i]; 478 uint last_inst = b->_nodes.size(); 479 uint adr = blk_starts[i]; 480 for( uint j = 0; j<last_inst; j++ ) { 481 nj = b->_nodes[j]; 482 if( nj->is_Mach() ) { 483 int padding = nj->as_Mach()->compute_padding(adr); 484 // If call/safepoint are adjacent insert a nop (5010568) 485 if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() && 486 adr == last_call_adr ) { 487 padding = nop_size; 488 } 489 if(padding > 0) { 490 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 491 int nops_cnt = padding / nop_size; 492 MachNode *nop = new (this) MachNopNode(nops_cnt); 493 b->_nodes.insert(j++, nop); 494 _cfg->_bbs.map( nop->_idx, b ); 495 adr += padding; 496 last_inst++; 497 } 498 } 499 adr += nj->size(_regalloc); 500 501 // Remember end of call offset 502 if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { 503 last_call_adr = adr; 504 } 505 } 506 507 if ( i != _cfg->_num_blocks-1) { 508 // Get the size of the block 509 uint blk_size = adr - blk_starts[i]; 510 511 // When the next block is the top of a loop, we may insert pad NOP 512 // instructions. 513 Block *nb = _cfg->_blocks[i+1]; 514 int current_offset = blk_starts[i] + blk_size; 515 current_offset += nb->alignment_padding(current_offset); 516 // Save block size; update total method size 517 blk_starts[i+1] = current_offset; 518 } 519 } 520 521 #ifdef ASSERT 522 for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks 523 if( jmp_target[i] != 0 ) { 524 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_end[i]); 525 if (!_matcher->is_short_branch_offset(jmp_rule[i], offset)) { 526 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]); 527 } 528 assert(_matcher->is_short_branch_offset(jmp_rule[i], offset), "Displacement too large for short jmp"); 529 } 530 } 531 #endif 532 533 // ------------------ 534 // Compute size for code buffer 535 code_size = blk_starts[i-1] + jmp_end[i-1]; 536 537 // Relocation records 538 reloc_size += 1; // Relo entry for exception handler 539 540 // Adjust reloc_size to number of record of relocation info 541 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 542 // a relocation index. 543 // The CodeBuffer will expand the locs array if this estimate is too low. 544 reloc_size *= 10 / sizeof(relocInfo); 545 546 // Adjust const_size to number of bytes 547 const_size *= 2*jintSize; // both float and double take two words per entry 548 549 } 550 551 //------------------------------FillLocArray----------------------------------- 552 // Create a bit of debug info and append it to the array. The mapping is from 553 // Java local or expression stack to constant, register or stack-slot. For 554 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 555 // entry has been taken care of and caller should skip it). 556 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 557 // This should never have accepted Bad before 558 assert(OptoReg::is_valid(regnum), "location must be valid"); 559 return (OptoReg::is_reg(regnum)) 560 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 561 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 562 } 563 564 565 ObjectValue* 566 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 567 for (int i = 0; i < objs->length(); i++) { 568 assert(objs->at(i)->is_object(), "corrupt object cache"); 569 ObjectValue* sv = (ObjectValue*) objs->at(i); 570 if (sv->id() == id) { 571 return sv; 572 } 573 } 574 // Otherwise.. 575 return NULL; 576 } 577 578 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 579 ObjectValue* sv ) { 580 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 581 objs->append(sv); 582 } 583 584 585 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 586 GrowableArray<ScopeValue*> *array, 587 GrowableArray<ScopeValue*> *objs ) { 588 assert( local, "use _top instead of null" ); 589 if (array->length() != idx) { 590 assert(array->length() == idx + 1, "Unexpected array count"); 591 // Old functionality: 592 // return 593 // New functionality: 594 // Assert if the local is not top. In product mode let the new node 595 // override the old entry. 596 assert(local == top(), "LocArray collision"); 597 if (local == top()) { 598 return; 599 } 600 array->pop(); 601 } 602 const Type *t = local->bottom_type(); 603 604 // Is it a safepoint scalar object node? 605 if (local->is_SafePointScalarObject()) { 606 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 607 608 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 609 if (sv == NULL) { 610 ciKlass* cik = t->is_oopptr()->klass(); 611 assert(cik->is_instance_klass() || 612 cik->is_array_klass(), "Not supported allocation."); 613 sv = new ObjectValue(spobj->_idx, 614 new ConstantOopWriteValue(cik->constant_encoding())); 615 Compile::set_sv_for_object_node(objs, sv); 616 617 uint first_ind = spobj->first_index(); 618 for (uint i = 0; i < spobj->n_fields(); i++) { 619 Node* fld_node = sfpt->in(first_ind+i); 620 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 621 } 622 } 623 array->append(sv); 624 return; 625 } 626 627 // Grab the register number for the local 628 OptoReg::Name regnum = _regalloc->get_reg_first(local); 629 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 630 // Record the double as two float registers. 631 // The register mask for such a value always specifies two adjacent 632 // float registers, with the lower register number even. 633 // Normally, the allocation of high and low words to these registers 634 // is irrelevant, because nearly all operations on register pairs 635 // (e.g., StoreD) treat them as a single unit. 636 // Here, we assume in addition that the words in these two registers 637 // stored "naturally" (by operations like StoreD and double stores 638 // within the interpreter) such that the lower-numbered register 639 // is written to the lower memory address. This may seem like 640 // a machine dependency, but it is not--it is a requirement on 641 // the author of the <arch>.ad file to ensure that, for every 642 // even/odd double-register pair to which a double may be allocated, 643 // the word in the even single-register is stored to the first 644 // memory word. (Note that register numbers are completely 645 // arbitrary, and are not tied to any machine-level encodings.) 646 #ifdef _LP64 647 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 648 array->append(new ConstantIntValue(0)); 649 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 650 } else if ( t->base() == Type::Long ) { 651 array->append(new ConstantIntValue(0)); 652 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 653 } else if ( t->base() == Type::RawPtr ) { 654 // jsr/ret return address which must be restored into a the full 655 // width 64-bit stack slot. 656 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 657 } 658 #else //_LP64 659 #ifdef SPARC 660 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 661 // For SPARC we have to swap high and low words for 662 // long values stored in a single-register (g0-g7). 663 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 664 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 665 } else 666 #endif //SPARC 667 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 668 // Repack the double/long as two jints. 669 // The convention the interpreter uses is that the second local 670 // holds the first raw word of the native double representation. 671 // This is actually reasonable, since locals and stack arrays 672 // grow downwards in all implementations. 673 // (If, on some machine, the interpreter's Java locals or stack 674 // were to grow upwards, the embedded doubles would be word-swapped.) 675 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 676 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 677 } 678 #endif //_LP64 679 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 680 OptoReg::is_reg(regnum) ) { 681 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double 682 ? Location::float_in_dbl : Location::normal )); 683 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 684 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 685 ? Location::int_in_long : Location::normal )); 686 } else if( t->base() == Type::NarrowOop ) { 687 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 688 } else { 689 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 690 } 691 return; 692 } 693 694 // No register. It must be constant data. 695 switch (t->base()) { 696 case Type::Half: // Second half of a double 697 ShouldNotReachHere(); // Caller should skip 2nd halves 698 break; 699 case Type::AnyPtr: 700 array->append(new ConstantOopWriteValue(NULL)); 701 break; 702 case Type::AryPtr: 703 case Type::InstPtr: 704 case Type::KlassPtr: // fall through 705 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 706 break; 707 case Type::NarrowOop: 708 if (t == TypeNarrowOop::NULL_PTR) { 709 array->append(new ConstantOopWriteValue(NULL)); 710 } else { 711 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 712 } 713 break; 714 case Type::Int: 715 array->append(new ConstantIntValue(t->is_int()->get_con())); 716 break; 717 case Type::RawPtr: 718 // A return address (T_ADDRESS). 719 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 720 #ifdef _LP64 721 // Must be restored to the full-width 64-bit stack slot. 722 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 723 #else 724 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 725 #endif 726 break; 727 case Type::FloatCon: { 728 float f = t->is_float_constant()->getf(); 729 array->append(new ConstantIntValue(jint_cast(f))); 730 break; 731 } 732 case Type::DoubleCon: { 733 jdouble d = t->is_double_constant()->getd(); 734 #ifdef _LP64 735 array->append(new ConstantIntValue(0)); 736 array->append(new ConstantDoubleValue(d)); 737 #else 738 // Repack the double as two jints. 739 // The convention the interpreter uses is that the second local 740 // holds the first raw word of the native double representation. 741 // This is actually reasonable, since locals and stack arrays 742 // grow downwards in all implementations. 743 // (If, on some machine, the interpreter's Java locals or stack 744 // were to grow upwards, the embedded doubles would be word-swapped.) 745 jint *dp = (jint*)&d; 746 array->append(new ConstantIntValue(dp[1])); 747 array->append(new ConstantIntValue(dp[0])); 748 #endif 749 break; 750 } 751 case Type::Long: { 752 jlong d = t->is_long()->get_con(); 753 #ifdef _LP64 754 array->append(new ConstantIntValue(0)); 755 array->append(new ConstantLongValue(d)); 756 #else 757 // Repack the long as two jints. 758 // The convention the interpreter uses is that the second local 759 // holds the first raw word of the native double representation. 760 // This is actually reasonable, since locals and stack arrays 761 // grow downwards in all implementations. 762 // (If, on some machine, the interpreter's Java locals or stack 763 // were to grow upwards, the embedded doubles would be word-swapped.) 764 jint *dp = (jint*)&d; 765 array->append(new ConstantIntValue(dp[1])); 766 array->append(new ConstantIntValue(dp[0])); 767 #endif 768 break; 769 } 770 case Type::Top: // Add an illegal value here 771 array->append(new LocationValue(Location())); 772 break; 773 default: 774 ShouldNotReachHere(); 775 break; 776 } 777 } 778 779 // Determine if this node starts a bundle 780 bool Compile::starts_bundle(const Node *n) const { 781 return (_node_bundling_limit > n->_idx && 782 _node_bundling_base[n->_idx].starts_bundle()); 783 } 784 785 //--------------------------Process_OopMap_Node-------------------------------- 786 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 787 788 // Handle special safepoint nodes for synchronization 789 MachSafePointNode *sfn = mach->as_MachSafePoint(); 790 MachCallNode *mcall; 791 792 #ifdef ENABLE_ZAP_DEAD_LOCALS 793 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 794 #endif 795 796 int safepoint_pc_offset = current_offset; 797 798 // Add the safepoint in the DebugInfoRecorder 799 if( !mach->is_MachCall() ) { 800 mcall = NULL; 801 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 802 } else { 803 mcall = mach->as_MachCall(); 804 safepoint_pc_offset += mcall->ret_addr_offset(); 805 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 806 } 807 808 // Loop over the JVMState list to add scope information 809 // Do not skip safepoints with a NULL method, they need monitor info 810 JVMState* youngest_jvms = sfn->jvms(); 811 int max_depth = youngest_jvms->depth(); 812 813 // Allocate the object pool for scalar-replaced objects -- the map from 814 // small-integer keys (which can be recorded in the local and ostack 815 // arrays) to descriptions of the object state. 816 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 817 818 // Visit scopes from oldest to youngest. 819 for (int depth = 1; depth <= max_depth; depth++) { 820 JVMState* jvms = youngest_jvms->of_depth(depth); 821 int idx; 822 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 823 // Safepoints that do not have method() set only provide oop-map and monitor info 824 // to support GC; these do not support deoptimization. 825 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 826 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 827 int num_mon = jvms->nof_monitors(); 828 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 829 "JVMS local count must match that of the method"); 830 831 // Add Local and Expression Stack Information 832 833 // Insert locals into the locarray 834 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 835 for( idx = 0; idx < num_locs; idx++ ) { 836 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 837 } 838 839 // Insert expression stack entries into the exparray 840 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 841 for( idx = 0; idx < num_exps; idx++ ) { 842 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 843 } 844 845 // Add in mappings of the monitors 846 assert( !method || 847 !method->is_synchronized() || 848 method->is_native() || 849 num_mon > 0 || 850 !GenerateSynchronizationCode, 851 "monitors must always exist for synchronized methods"); 852 853 // Build the growable array of ScopeValues for exp stack 854 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 855 856 // Loop over monitors and insert into array 857 for(idx = 0; idx < num_mon; idx++) { 858 // Grab the node that defines this monitor 859 Node* box_node = sfn->monitor_box(jvms, idx); 860 Node* obj_node = sfn->monitor_obj(jvms, idx); 861 862 // Create ScopeValue for object 863 ScopeValue *scval = NULL; 864 865 if( obj_node->is_SafePointScalarObject() ) { 866 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 867 scval = Compile::sv_for_node_id(objs, spobj->_idx); 868 if (scval == NULL) { 869 const Type *t = obj_node->bottom_type(); 870 ciKlass* cik = t->is_oopptr()->klass(); 871 assert(cik->is_instance_klass() || 872 cik->is_array_klass(), "Not supported allocation."); 873 ObjectValue* sv = new ObjectValue(spobj->_idx, 874 new ConstantOopWriteValue(cik->constant_encoding())); 875 Compile::set_sv_for_object_node(objs, sv); 876 877 uint first_ind = spobj->first_index(); 878 for (uint i = 0; i < spobj->n_fields(); i++) { 879 Node* fld_node = sfn->in(first_ind+i); 880 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 881 } 882 scval = sv; 883 } 884 } else if( !obj_node->is_Con() ) { 885 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 886 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 887 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 888 } else { 889 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 890 } 891 } else { 892 const TypePtr *tp = obj_node->bottom_type()->make_ptr(); 893 scval = new ConstantOopWriteValue(tp->is_instptr()->const_oop()->constant_encoding()); 894 } 895 896 OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node); 897 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 898 while( !box_node->is_BoxLock() ) box_node = box_node->in(1); 899 monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated())); 900 } 901 902 // We dump the object pool first, since deoptimization reads it in first. 903 debug_info()->dump_object_pool(objs); 904 905 // Build first class objects to pass to scope 906 DebugToken *locvals = debug_info()->create_scope_values(locarray); 907 DebugToken *expvals = debug_info()->create_scope_values(exparray); 908 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 909 910 // Make method available for all Safepoints 911 ciMethod* scope_method = method ? method : _method; 912 // Describe the scope here 913 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 914 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 915 // Now we can describe the scope. 916 bool is_method_handle_invoke = false; 917 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, locvals, expvals, monvals); 918 } // End jvms loop 919 920 // Mark the end of the scope set. 921 debug_info()->end_safepoint(safepoint_pc_offset); 922 } 923 924 925 926 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 927 class NonSafepointEmitter { 928 Compile* C; 929 JVMState* _pending_jvms; 930 int _pending_offset; 931 932 void emit_non_safepoint(); 933 934 public: 935 NonSafepointEmitter(Compile* compile) { 936 this->C = compile; 937 _pending_jvms = NULL; 938 _pending_offset = 0; 939 } 940 941 void observe_instruction(Node* n, int pc_offset) { 942 if (!C->debug_info()->recording_non_safepoints()) return; 943 944 Node_Notes* nn = C->node_notes_at(n->_idx); 945 if (nn == NULL || nn->jvms() == NULL) return; 946 if (_pending_jvms != NULL && 947 _pending_jvms->same_calls_as(nn->jvms())) { 948 // Repeated JVMS? Stretch it up here. 949 _pending_offset = pc_offset; 950 } else { 951 if (_pending_jvms != NULL && 952 _pending_offset < pc_offset) { 953 emit_non_safepoint(); 954 } 955 _pending_jvms = NULL; 956 if (pc_offset > C->debug_info()->last_pc_offset()) { 957 // This is the only way _pending_jvms can become non-NULL: 958 _pending_jvms = nn->jvms(); 959 _pending_offset = pc_offset; 960 } 961 } 962 } 963 964 // Stay out of the way of real safepoints: 965 void observe_safepoint(JVMState* jvms, int pc_offset) { 966 if (_pending_jvms != NULL && 967 !_pending_jvms->same_calls_as(jvms) && 968 _pending_offset < pc_offset) { 969 emit_non_safepoint(); 970 } 971 _pending_jvms = NULL; 972 } 973 974 void flush_at_end() { 975 if (_pending_jvms != NULL) { 976 emit_non_safepoint(); 977 } 978 _pending_jvms = NULL; 979 } 980 }; 981 982 void NonSafepointEmitter::emit_non_safepoint() { 983 JVMState* youngest_jvms = _pending_jvms; 984 int pc_offset = _pending_offset; 985 986 // Clear it now: 987 _pending_jvms = NULL; 988 989 DebugInformationRecorder* debug_info = C->debug_info(); 990 assert(debug_info->recording_non_safepoints(), "sanity"); 991 992 debug_info->add_non_safepoint(pc_offset); 993 int max_depth = youngest_jvms->depth(); 994 995 // Visit scopes from oldest to youngest. 996 for (int depth = 1; depth <= max_depth; depth++) { 997 JVMState* jvms = youngest_jvms->of_depth(depth); 998 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 999 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1000 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1001 } 1002 1003 // Mark the end of the scope set. 1004 debug_info->end_non_safepoint(pc_offset); 1005 } 1006 1007 1008 1009 // helper for Fill_buffer bailout logic 1010 static void turn_off_compiler(Compile* C) { 1011 if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) { 1012 // Do not turn off compilation if a single giant method has 1013 // blown the code cache size. 1014 C->record_failure("excessive request to CodeCache"); 1015 } else { 1016 // Let CompilerBroker disable further compilations. 1017 C->record_failure("CodeCache is full"); 1018 } 1019 } 1020 1021 1022 //------------------------------Fill_buffer------------------------------------ 1023 void Compile::Fill_buffer() { 1024 1025 // Set the initially allocated size 1026 int code_req = initial_code_capacity; 1027 int locs_req = initial_locs_capacity; 1028 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1029 int const_req = initial_const_capacity; 1030 bool labels_not_set = true; 1031 1032 int pad_req = NativeCall::instruction_size; 1033 // The extra spacing after the code is necessary on some platforms. 1034 // Sometimes we need to patch in a jump after the last instruction, 1035 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1036 1037 uint i; 1038 // Compute the byte offset where we can store the deopt pc. 1039 if (fixed_slots() != 0) { 1040 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1041 } 1042 1043 // Compute prolog code size 1044 _method_size = 0; 1045 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1046 #ifdef IA64 1047 if (save_argument_registers()) { 1048 // 4815101: this is a stub with implicit and unknown precision fp args. 1049 // The usual spill mechanism can only generate stfd's in this case, which 1050 // doesn't work if the fp reg to spill contains a single-precision denorm. 1051 // Instead, we hack around the normal spill mechanism using stfspill's and 1052 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1053 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1054 // 1055 // If we ever implement 16-byte 'registers' == stack slots, we can 1056 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1057 // instead of stfd/stfs/ldfd/ldfs. 1058 _frame_slots += 8*(16/BytesPerInt); 1059 } 1060 #endif 1061 assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" ); 1062 1063 // Create an array of unused labels, one for each basic block 1064 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1); 1065 1066 for( i=0; i <= _cfg->_num_blocks; i++ ) { 1067 blk_labels[i].init(); 1068 } 1069 1070 // If this machine supports different size branch offsets, then pre-compute 1071 // the length of the blocks 1072 if( _matcher->is_short_branch_offset(-1, 0) ) { 1073 Shorten_branches(blk_labels, code_req, locs_req, stub_req, const_req); 1074 labels_not_set = false; 1075 } 1076 1077 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1078 int exception_handler_req = size_exception_handler(); 1079 int deopt_handler_req = size_deopt_handler(); 1080 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1081 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1082 stub_req += MAX_stubs_size; // ensure per-stub margin 1083 code_req += MAX_inst_size; // ensure per-instruction margin 1084 if (StressCodeBuffers) 1085 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1086 int total_req = code_req + pad_req + stub_req + exception_handler_req + deopt_handler_req + const_req; 1087 CodeBuffer* cb = code_buffer(); 1088 cb->initialize(total_req, locs_req); 1089 1090 // Have we run out of code space? 1091 if (cb->blob() == NULL) { 1092 turn_off_compiler(this); 1093 return; 1094 } 1095 // Configure the code buffer. 1096 cb->initialize_consts_size(const_req); 1097 cb->initialize_stubs_size(stub_req); 1098 cb->initialize_oop_recorder(env()->oop_recorder()); 1099 1100 // fill in the nop array for bundling computations 1101 MachNode *_nop_list[Bundle::_nop_count]; 1102 Bundle::initialize_nops(_nop_list, this); 1103 1104 // Create oopmap set. 1105 _oop_map_set = new OopMapSet(); 1106 1107 // !!!!! This preserves old handling of oopmaps for now 1108 debug_info()->set_oopmaps(_oop_map_set); 1109 1110 // Count and start of implicit null check instructions 1111 uint inct_cnt = 0; 1112 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1113 1114 // Count and start of calls 1115 uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1116 1117 uint return_offset = 0; 1118 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1119 1120 int previous_offset = 0; 1121 int current_offset = 0; 1122 int last_call_offset = -1; 1123 1124 // Create an array of unused labels, one for each basic block, if printing is enabled 1125 #ifndef PRODUCT 1126 int *node_offsets = NULL; 1127 uint node_offset_limit = unique(); 1128 1129 1130 if ( print_assembly() ) 1131 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1132 #endif 1133 1134 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1135 1136 // ------------------ 1137 // Now fill in the code buffer 1138 Node *delay_slot = NULL; 1139 1140 for( i=0; i < _cfg->_num_blocks; i++ ) { 1141 Block *b = _cfg->_blocks[i]; 1142 1143 Node *head = b->head(); 1144 1145 // If this block needs to start aligned (i.e, can be reached other 1146 // than by falling-thru from the previous block), then force the 1147 // start of a new bundle. 1148 if( Pipeline::requires_bundling() && starts_bundle(head) ) 1149 cb->flush_bundle(true); 1150 1151 // Define the label at the beginning of the basic block 1152 if( labels_not_set ) 1153 MacroAssembler(cb).bind( blk_labels[b->_pre_order] ); 1154 1155 else 1156 assert( blk_labels[b->_pre_order].loc_pos() == cb->code_size(), 1157 "label position does not match code offset" ); 1158 1159 uint last_inst = b->_nodes.size(); 1160 1161 // Emit block normally, except for last instruction. 1162 // Emit means "dump code bits into code buffer". 1163 for( uint j = 0; j<last_inst; j++ ) { 1164 1165 // Get the node 1166 Node* n = b->_nodes[j]; 1167 1168 // See if delay slots are supported 1169 if (valid_bundle_info(n) && 1170 node_bundling(n)->used_in_unconditional_delay()) { 1171 assert(delay_slot == NULL, "no use of delay slot node"); 1172 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1173 1174 delay_slot = n; 1175 continue; 1176 } 1177 1178 // If this starts a new instruction group, then flush the current one 1179 // (but allow split bundles) 1180 if( Pipeline::requires_bundling() && starts_bundle(n) ) 1181 cb->flush_bundle(false); 1182 1183 // The following logic is duplicated in the code ifdeffed for 1184 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1185 // should be factored out. Or maybe dispersed to the nodes? 1186 1187 // Special handling for SafePoint/Call Nodes 1188 bool is_mcall = false; 1189 if( n->is_Mach() ) { 1190 MachNode *mach = n->as_Mach(); 1191 is_mcall = n->is_MachCall(); 1192 bool is_sfn = n->is_MachSafePoint(); 1193 1194 // If this requires all previous instructions be flushed, then do so 1195 if( is_sfn || is_mcall || mach->alignment_required() != 1) { 1196 cb->flush_bundle(true); 1197 current_offset = cb->code_size(); 1198 } 1199 1200 // align the instruction if necessary 1201 int padding = mach->compute_padding(current_offset); 1202 // Make sure safepoint node for polling is distinct from a call's 1203 // return by adding a nop if needed. 1204 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) { 1205 padding = nop_size; 1206 } 1207 assert( labels_not_set || padding == 0, "instruction should already be aligned") 1208 1209 if(padding > 0) { 1210 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1211 int nops_cnt = padding / nop_size; 1212 MachNode *nop = new (this) MachNopNode(nops_cnt); 1213 b->_nodes.insert(j++, nop); 1214 last_inst++; 1215 _cfg->_bbs.map( nop->_idx, b ); 1216 nop->emit(*cb, _regalloc); 1217 cb->flush_bundle(true); 1218 current_offset = cb->code_size(); 1219 } 1220 1221 // Remember the start of the last call in a basic block 1222 if (is_mcall) { 1223 MachCallNode *mcall = mach->as_MachCall(); 1224 1225 // This destination address is NOT PC-relative 1226 mcall->method_set((intptr_t)mcall->entry_point()); 1227 1228 // Save the return address 1229 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1230 1231 if (!mcall->is_safepoint_node()) { 1232 is_mcall = false; 1233 is_sfn = false; 1234 } 1235 } 1236 1237 // sfn will be valid whenever mcall is valid now because of inheritance 1238 if( is_sfn || is_mcall ) { 1239 1240 // Handle special safepoint nodes for synchronization 1241 if( !is_mcall ) { 1242 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1243 // !!!!! Stubs only need an oopmap right now, so bail out 1244 if( sfn->jvms()->method() == NULL) { 1245 // Write the oopmap directly to the code blob??!! 1246 # ifdef ENABLE_ZAP_DEAD_LOCALS 1247 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1248 # endif 1249 continue; 1250 } 1251 } // End synchronization 1252 1253 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1254 current_offset); 1255 Process_OopMap_Node(mach, current_offset); 1256 } // End if safepoint 1257 1258 // If this is a null check, then add the start of the previous instruction to the list 1259 else if( mach->is_MachNullCheck() ) { 1260 inct_starts[inct_cnt++] = previous_offset; 1261 } 1262 1263 // If this is a branch, then fill in the label with the target BB's label 1264 else if ( mach->is_Branch() ) { 1265 1266 if ( mach->ideal_Opcode() == Op_Jump ) { 1267 for (uint h = 0; h < b->_num_succs; h++ ) { 1268 Block* succs_block = b->_succs[h]; 1269 for (uint j = 1; j < succs_block->num_preds(); j++) { 1270 Node* jpn = succs_block->pred(j); 1271 if ( jpn->is_JumpProj() && jpn->in(0) == mach ) { 1272 uint block_num = succs_block->non_connector()->_pre_order; 1273 Label *blkLabel = &blk_labels[block_num]; 1274 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1275 } 1276 } 1277 } 1278 } else { 1279 // For Branchs 1280 // This requires the TRUE branch target be in succs[0] 1281 uint block_num = b->non_connector_successor(0)->_pre_order; 1282 mach->label_set( blk_labels[block_num], block_num ); 1283 } 1284 } 1285 1286 #ifdef ASSERT 1287 // Check that oop-store precedes the card-mark 1288 else if( mach->ideal_Opcode() == Op_StoreCM ) { 1289 uint storeCM_idx = j; 1290 Node *oop_store = mach->in(mach->_cnt); // First precedence edge 1291 assert( oop_store != NULL, "storeCM expects a precedence edge"); 1292 uint i4; 1293 for( i4 = 0; i4 < last_inst; ++i4 ) { 1294 if( b->_nodes[i4] == oop_store ) break; 1295 } 1296 // Note: This test can provide a false failure if other precedence 1297 // edges have been added to the storeCMNode. 1298 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1299 } 1300 #endif 1301 1302 else if( !n->is_Proj() ) { 1303 // Remember the beginning of the previous instruction, in case 1304 // it's followed by a flag-kill and a null-check. Happens on 1305 // Intel all the time, with add-to-memory kind of opcodes. 1306 previous_offset = current_offset; 1307 } 1308 } 1309 1310 // Verify that there is sufficient space remaining 1311 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1312 if (cb->blob() == NULL) { 1313 turn_off_compiler(this); 1314 return; 1315 } 1316 1317 // Save the offset for the listing 1318 #ifndef PRODUCT 1319 if( node_offsets && n->_idx < node_offset_limit ) 1320 node_offsets[n->_idx] = cb->code_size(); 1321 #endif 1322 1323 // "Normal" instruction case 1324 n->emit(*cb, _regalloc); 1325 current_offset = cb->code_size(); 1326 non_safepoints.observe_instruction(n, current_offset); 1327 1328 // mcall is last "call" that can be a safepoint 1329 // record it so we can see if a poll will directly follow it 1330 // in which case we'll need a pad to make the PcDesc sites unique 1331 // see 5010568. This can be slightly inaccurate but conservative 1332 // in the case that return address is not actually at current_offset. 1333 // This is a small price to pay. 1334 1335 if (is_mcall) { 1336 last_call_offset = current_offset; 1337 } 1338 1339 // See if this instruction has a delay slot 1340 if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1341 assert(delay_slot != NULL, "expecting delay slot node"); 1342 1343 // Back up 1 instruction 1344 cb->set_code_end( 1345 cb->code_end()-Pipeline::instr_unit_size()); 1346 1347 // Save the offset for the listing 1348 #ifndef PRODUCT 1349 if( node_offsets && delay_slot->_idx < node_offset_limit ) 1350 node_offsets[delay_slot->_idx] = cb->code_size(); 1351 #endif 1352 1353 // Support a SafePoint in the delay slot 1354 if( delay_slot->is_MachSafePoint() ) { 1355 MachNode *mach = delay_slot->as_Mach(); 1356 // !!!!! Stubs only need an oopmap right now, so bail out 1357 if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) { 1358 // Write the oopmap directly to the code blob??!! 1359 # ifdef ENABLE_ZAP_DEAD_LOCALS 1360 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1361 # endif 1362 delay_slot = NULL; 1363 continue; 1364 } 1365 1366 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1367 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1368 adjusted_offset); 1369 // Generate an OopMap entry 1370 Process_OopMap_Node(mach, adjusted_offset); 1371 } 1372 1373 // Insert the delay slot instruction 1374 delay_slot->emit(*cb, _regalloc); 1375 1376 // Don't reuse it 1377 delay_slot = NULL; 1378 } 1379 1380 } // End for all instructions in block 1381 1382 // If the next block is the top of a loop, pad this block out to align 1383 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1384 if( i<_cfg->_num_blocks-1 ) { 1385 Block *nb = _cfg->_blocks[i+1]; 1386 uint padding = nb->alignment_padding(current_offset); 1387 if( padding > 0 ) { 1388 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1389 b->_nodes.insert( b->_nodes.size(), nop ); 1390 _cfg->_bbs.map( nop->_idx, b ); 1391 nop->emit(*cb, _regalloc); 1392 current_offset = cb->code_size(); 1393 } 1394 } 1395 1396 } // End of for all blocks 1397 1398 non_safepoints.flush_at_end(); 1399 1400 // Offset too large? 1401 if (failing()) return; 1402 1403 // Define a pseudo-label at the end of the code 1404 MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] ); 1405 1406 // Compute the size of the first block 1407 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1408 1409 assert(cb->code_size() < 500000, "method is unreasonably large"); 1410 1411 // ------------------ 1412 1413 #ifndef PRODUCT 1414 // Information on the size of the method, without the extraneous code 1415 Scheduling::increment_method_size(cb->code_size()); 1416 #endif 1417 1418 // ------------------ 1419 // Fill in exception table entries. 1420 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1421 1422 // Only java methods have exception handlers and deopt handlers 1423 if (_method) { 1424 // Emit the exception handler code. 1425 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1426 // Emit the deopt handler code. 1427 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1428 } 1429 1430 // One last check for failed CodeBuffer::expand: 1431 if (cb->blob() == NULL) { 1432 turn_off_compiler(this); 1433 return; 1434 } 1435 1436 #ifndef PRODUCT 1437 // Dump the assembly code, including basic-block numbers 1438 if (print_assembly()) { 1439 ttyLocker ttyl; // keep the following output all in one block 1440 if (!VMThread::should_terminate()) { // test this under the tty lock 1441 // This output goes directly to the tty, not the compiler log. 1442 // To enable tools to match it up with the compilation activity, 1443 // be sure to tag this tty output with the compile ID. 1444 if (xtty != NULL) { 1445 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1446 is_osr_compilation() ? " compile_kind='osr'" : 1447 ""); 1448 } 1449 if (method() != NULL) { 1450 method()->print_oop(); 1451 print_codes(); 1452 } 1453 dump_asm(node_offsets, node_offset_limit); 1454 if (xtty != NULL) { 1455 xtty->tail("opto_assembly"); 1456 } 1457 } 1458 } 1459 #endif 1460 1461 } 1462 1463 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1464 _inc_table.set_size(cnt); 1465 1466 uint inct_cnt = 0; 1467 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1468 Block *b = _cfg->_blocks[i]; 1469 Node *n = NULL; 1470 int j; 1471 1472 // Find the branch; ignore trailing NOPs. 1473 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1474 n = b->_nodes[j]; 1475 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1476 break; 1477 } 1478 1479 // If we didn't find anything, continue 1480 if( j < 0 ) continue; 1481 1482 // Compute ExceptionHandlerTable subtable entry and add it 1483 // (skip empty blocks) 1484 if( n->is_Catch() ) { 1485 1486 // Get the offset of the return from the call 1487 uint call_return = call_returns[b->_pre_order]; 1488 #ifdef ASSERT 1489 assert( call_return > 0, "no call seen for this basic block" ); 1490 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; 1491 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); 1492 #endif 1493 // last instruction is a CatchNode, find it's CatchProjNodes 1494 int nof_succs = b->_num_succs; 1495 // allocate space 1496 GrowableArray<intptr_t> handler_bcis(nof_succs); 1497 GrowableArray<intptr_t> handler_pcos(nof_succs); 1498 // iterate through all successors 1499 for (int j = 0; j < nof_succs; j++) { 1500 Block* s = b->_succs[j]; 1501 bool found_p = false; 1502 for( uint k = 1; k < s->num_preds(); k++ ) { 1503 Node *pk = s->pred(k); 1504 if( pk->is_CatchProj() && pk->in(0) == n ) { 1505 const CatchProjNode* p = pk->as_CatchProj(); 1506 found_p = true; 1507 // add the corresponding handler bci & pco information 1508 if( p->_con != CatchProjNode::fall_through_index ) { 1509 // p leads to an exception handler (and is not fall through) 1510 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1511 // no duplicates, please 1512 if( !handler_bcis.contains(p->handler_bci()) ) { 1513 uint block_num = s->non_connector()->_pre_order; 1514 handler_bcis.append(p->handler_bci()); 1515 handler_pcos.append(blk_labels[block_num].loc_pos()); 1516 } 1517 } 1518 } 1519 } 1520 assert(found_p, "no matching predecessor found"); 1521 // Note: Due to empty block removal, one block may have 1522 // several CatchProj inputs, from the same Catch. 1523 } 1524 1525 // Set the offset of the return from the call 1526 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1527 continue; 1528 } 1529 1530 // Handle implicit null exception table updates 1531 if( n->is_MachNullCheck() ) { 1532 uint block_num = b->non_connector_successor(0)->_pre_order; 1533 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1534 continue; 1535 } 1536 } // End of for all blocks fill in exception table entries 1537 } 1538 1539 // Static Variables 1540 #ifndef PRODUCT 1541 uint Scheduling::_total_nop_size = 0; 1542 uint Scheduling::_total_method_size = 0; 1543 uint Scheduling::_total_branches = 0; 1544 uint Scheduling::_total_unconditional_delays = 0; 1545 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1546 #endif 1547 1548 // Initializer for class Scheduling 1549 1550 Scheduling::Scheduling(Arena *arena, Compile &compile) 1551 : _arena(arena), 1552 _cfg(compile.cfg()), 1553 _bbs(compile.cfg()->_bbs), 1554 _regalloc(compile.regalloc()), 1555 _reg_node(arena), 1556 _bundle_instr_count(0), 1557 _bundle_cycle_number(0), 1558 _scheduled(arena), 1559 _available(arena), 1560 _next_node(NULL), 1561 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1562 _pinch_free_list(arena) 1563 #ifndef PRODUCT 1564 , _branches(0) 1565 , _unconditional_delays(0) 1566 #endif 1567 { 1568 // Create a MachNopNode 1569 _nop = new (&compile) MachNopNode(); 1570 1571 // Now that the nops are in the array, save the count 1572 // (but allow entries for the nops) 1573 _node_bundling_limit = compile.unique(); 1574 uint node_max = _regalloc->node_regs_max_index(); 1575 1576 compile.set_node_bundling_limit(_node_bundling_limit); 1577 1578 // This one is persistent within the Compile class 1579 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1580 1581 // Allocate space for fixed-size arrays 1582 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1583 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1584 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1585 1586 // Clear the arrays 1587 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1588 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1589 memset(_uses, 0, node_max * sizeof(short)); 1590 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1591 1592 // Clear the bundling information 1593 memcpy(_bundle_use_elements, 1594 Pipeline_Use::elaborated_elements, 1595 sizeof(Pipeline_Use::elaborated_elements)); 1596 1597 // Get the last node 1598 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1599 1600 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1601 } 1602 1603 #ifndef PRODUCT 1604 // Scheduling destructor 1605 Scheduling::~Scheduling() { 1606 _total_branches += _branches; 1607 _total_unconditional_delays += _unconditional_delays; 1608 } 1609 #endif 1610 1611 // Step ahead "i" cycles 1612 void Scheduling::step(uint i) { 1613 1614 Bundle *bundle = node_bundling(_next_node); 1615 bundle->set_starts_bundle(); 1616 1617 // Update the bundle record, but leave the flags information alone 1618 if (_bundle_instr_count > 0) { 1619 bundle->set_instr_count(_bundle_instr_count); 1620 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1621 } 1622 1623 // Update the state information 1624 _bundle_instr_count = 0; 1625 _bundle_cycle_number += i; 1626 _bundle_use.step(i); 1627 } 1628 1629 void Scheduling::step_and_clear() { 1630 Bundle *bundle = node_bundling(_next_node); 1631 bundle->set_starts_bundle(); 1632 1633 // Update the bundle record 1634 if (_bundle_instr_count > 0) { 1635 bundle->set_instr_count(_bundle_instr_count); 1636 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1637 1638 _bundle_cycle_number += 1; 1639 } 1640 1641 // Clear the bundling information 1642 _bundle_instr_count = 0; 1643 _bundle_use.reset(); 1644 1645 memcpy(_bundle_use_elements, 1646 Pipeline_Use::elaborated_elements, 1647 sizeof(Pipeline_Use::elaborated_elements)); 1648 } 1649 1650 //------------------------------ScheduleAndBundle------------------------------ 1651 // Perform instruction scheduling and bundling over the sequence of 1652 // instructions in backwards order. 1653 void Compile::ScheduleAndBundle() { 1654 1655 // Don't optimize this if it isn't a method 1656 if (!_method) 1657 return; 1658 1659 // Don't optimize this if scheduling is disabled 1660 if (!do_scheduling()) 1661 return; 1662 1663 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1664 1665 // Create a data structure for all the scheduling information 1666 Scheduling scheduling(Thread::current()->resource_area(), *this); 1667 1668 // Walk backwards over each basic block, computing the needed alignment 1669 // Walk over all the basic blocks 1670 scheduling.DoScheduling(); 1671 } 1672 1673 //------------------------------ComputeLocalLatenciesForward------------------- 1674 // Compute the latency of all the instructions. This is fairly simple, 1675 // because we already have a legal ordering. Walk over the instructions 1676 // from first to last, and compute the latency of the instruction based 1677 // on the latency of the preceding instruction(s). 1678 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1679 #ifndef PRODUCT 1680 if (_cfg->C->trace_opto_output()) 1681 tty->print("# -> ComputeLocalLatenciesForward\n"); 1682 #endif 1683 1684 // Walk over all the schedulable instructions 1685 for( uint j=_bb_start; j < _bb_end; j++ ) { 1686 1687 // This is a kludge, forcing all latency calculations to start at 1. 1688 // Used to allow latency 0 to force an instruction to the beginning 1689 // of the bb 1690 uint latency = 1; 1691 Node *use = bb->_nodes[j]; 1692 uint nlen = use->len(); 1693 1694 // Walk over all the inputs 1695 for ( uint k=0; k < nlen; k++ ) { 1696 Node *def = use->in(k); 1697 if (!def) 1698 continue; 1699 1700 uint l = _node_latency[def->_idx] + use->latency(k); 1701 if (latency < l) 1702 latency = l; 1703 } 1704 1705 _node_latency[use->_idx] = latency; 1706 1707 #ifndef PRODUCT 1708 if (_cfg->C->trace_opto_output()) { 1709 tty->print("# latency %4d: ", latency); 1710 use->dump(); 1711 } 1712 #endif 1713 } 1714 1715 #ifndef PRODUCT 1716 if (_cfg->C->trace_opto_output()) 1717 tty->print("# <- ComputeLocalLatenciesForward\n"); 1718 #endif 1719 1720 } // end ComputeLocalLatenciesForward 1721 1722 // See if this node fits into the present instruction bundle 1723 bool Scheduling::NodeFitsInBundle(Node *n) { 1724 uint n_idx = n->_idx; 1725 1726 // If this is the unconditional delay instruction, then it fits 1727 if (n == _unconditional_delay_slot) { 1728 #ifndef PRODUCT 1729 if (_cfg->C->trace_opto_output()) 1730 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1731 #endif 1732 return (true); 1733 } 1734 1735 // If the node cannot be scheduled this cycle, skip it 1736 if (_current_latency[n_idx] > _bundle_cycle_number) { 1737 #ifndef PRODUCT 1738 if (_cfg->C->trace_opto_output()) 1739 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1740 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1741 #endif 1742 return (false); 1743 } 1744 1745 const Pipeline *node_pipeline = n->pipeline(); 1746 1747 uint instruction_count = node_pipeline->instructionCount(); 1748 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1749 instruction_count = 0; 1750 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1751 instruction_count++; 1752 1753 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1754 #ifndef PRODUCT 1755 if (_cfg->C->trace_opto_output()) 1756 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1757 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1758 #endif 1759 return (false); 1760 } 1761 1762 // Don't allow non-machine nodes to be handled this way 1763 if (!n->is_Mach() && instruction_count == 0) 1764 return (false); 1765 1766 // See if there is any overlap 1767 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1768 1769 if (delay > 0) { 1770 #ifndef PRODUCT 1771 if (_cfg->C->trace_opto_output()) 1772 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1773 #endif 1774 return false; 1775 } 1776 1777 #ifndef PRODUCT 1778 if (_cfg->C->trace_opto_output()) 1779 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1780 #endif 1781 1782 return true; 1783 } 1784 1785 Node * Scheduling::ChooseNodeToBundle() { 1786 uint siz = _available.size(); 1787 1788 if (siz == 0) { 1789 1790 #ifndef PRODUCT 1791 if (_cfg->C->trace_opto_output()) 1792 tty->print("# ChooseNodeToBundle: NULL\n"); 1793 #endif 1794 return (NULL); 1795 } 1796 1797 // Fast path, if only 1 instruction in the bundle 1798 if (siz == 1) { 1799 #ifndef PRODUCT 1800 if (_cfg->C->trace_opto_output()) { 1801 tty->print("# ChooseNodeToBundle (only 1): "); 1802 _available[0]->dump(); 1803 } 1804 #endif 1805 return (_available[0]); 1806 } 1807 1808 // Don't bother, if the bundle is already full 1809 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1810 for ( uint i = 0; i < siz; i++ ) { 1811 Node *n = _available[i]; 1812 1813 // Skip projections, we'll handle them another way 1814 if (n->is_Proj()) 1815 continue; 1816 1817 // This presupposed that instructions are inserted into the 1818 // available list in a legality order; i.e. instructions that 1819 // must be inserted first are at the head of the list 1820 if (NodeFitsInBundle(n)) { 1821 #ifndef PRODUCT 1822 if (_cfg->C->trace_opto_output()) { 1823 tty->print("# ChooseNodeToBundle: "); 1824 n->dump(); 1825 } 1826 #endif 1827 return (n); 1828 } 1829 } 1830 } 1831 1832 // Nothing fits in this bundle, choose the highest priority 1833 #ifndef PRODUCT 1834 if (_cfg->C->trace_opto_output()) { 1835 tty->print("# ChooseNodeToBundle: "); 1836 _available[0]->dump(); 1837 } 1838 #endif 1839 1840 return _available[0]; 1841 } 1842 1843 //------------------------------AddNodeToAvailableList------------------------- 1844 void Scheduling::AddNodeToAvailableList(Node *n) { 1845 assert( !n->is_Proj(), "projections never directly made available" ); 1846 #ifndef PRODUCT 1847 if (_cfg->C->trace_opto_output()) { 1848 tty->print("# AddNodeToAvailableList: "); 1849 n->dump(); 1850 } 1851 #endif 1852 1853 int latency = _current_latency[n->_idx]; 1854 1855 // Insert in latency order (insertion sort) 1856 uint i; 1857 for ( i=0; i < _available.size(); i++ ) 1858 if (_current_latency[_available[i]->_idx] > latency) 1859 break; 1860 1861 // Special Check for compares following branches 1862 if( n->is_Mach() && _scheduled.size() > 0 ) { 1863 int op = n->as_Mach()->ideal_Opcode(); 1864 Node *last = _scheduled[0]; 1865 if( last->is_MachIf() && last->in(1) == n && 1866 ( op == Op_CmpI || 1867 op == Op_CmpU || 1868 op == Op_CmpP || 1869 op == Op_CmpF || 1870 op == Op_CmpD || 1871 op == Op_CmpL ) ) { 1872 1873 // Recalculate position, moving to front of same latency 1874 for ( i=0 ; i < _available.size(); i++ ) 1875 if (_current_latency[_available[i]->_idx] >= latency) 1876 break; 1877 } 1878 } 1879 1880 // Insert the node in the available list 1881 _available.insert(i, n); 1882 1883 #ifndef PRODUCT 1884 if (_cfg->C->trace_opto_output()) 1885 dump_available(); 1886 #endif 1887 } 1888 1889 //------------------------------DecrementUseCounts----------------------------- 1890 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 1891 for ( uint i=0; i < n->len(); i++ ) { 1892 Node *def = n->in(i); 1893 if (!def) continue; 1894 if( def->is_Proj() ) // If this is a machine projection, then 1895 def = def->in(0); // propagate usage thru to the base instruction 1896 1897 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 1898 continue; 1899 1900 // Compute the latency 1901 uint l = _bundle_cycle_number + n->latency(i); 1902 if (_current_latency[def->_idx] < l) 1903 _current_latency[def->_idx] = l; 1904 1905 // If this does not have uses then schedule it 1906 if ((--_uses[def->_idx]) == 0) 1907 AddNodeToAvailableList(def); 1908 } 1909 } 1910 1911 //------------------------------AddNodeToBundle-------------------------------- 1912 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 1913 #ifndef PRODUCT 1914 if (_cfg->C->trace_opto_output()) { 1915 tty->print("# AddNodeToBundle: "); 1916 n->dump(); 1917 } 1918 #endif 1919 1920 // Remove this from the available list 1921 uint i; 1922 for (i = 0; i < _available.size(); i++) 1923 if (_available[i] == n) 1924 break; 1925 assert(i < _available.size(), "entry in _available list not found"); 1926 _available.remove(i); 1927 1928 // See if this fits in the current bundle 1929 const Pipeline *node_pipeline = n->pipeline(); 1930 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 1931 1932 // Check for instructions to be placed in the delay slot. We 1933 // do this before we actually schedule the current instruction, 1934 // because the delay slot follows the current instruction. 1935 if (Pipeline::_branch_has_delay_slot && 1936 node_pipeline->hasBranchDelay() && 1937 !_unconditional_delay_slot) { 1938 1939 uint siz = _available.size(); 1940 1941 // Conditional branches can support an instruction that 1942 // is unconditionally executed and not dependent by the 1943 // branch, OR a conditionally executed instruction if 1944 // the branch is taken. In practice, this means that 1945 // the first instruction at the branch target is 1946 // copied to the delay slot, and the branch goes to 1947 // the instruction after that at the branch target 1948 if ( n->is_Mach() && n->is_Branch() ) { 1949 1950 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 1951 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 1952 1953 #ifndef PRODUCT 1954 _branches++; 1955 #endif 1956 1957 // At least 1 instruction is on the available list 1958 // that is not dependent on the branch 1959 for (uint i = 0; i < siz; i++) { 1960 Node *d = _available[i]; 1961 const Pipeline *avail_pipeline = d->pipeline(); 1962 1963 // Don't allow safepoints in the branch shadow, that will 1964 // cause a number of difficulties 1965 if ( avail_pipeline->instructionCount() == 1 && 1966 !avail_pipeline->hasMultipleBundles() && 1967 !avail_pipeline->hasBranchDelay() && 1968 Pipeline::instr_has_unit_size() && 1969 d->size(_regalloc) == Pipeline::instr_unit_size() && 1970 NodeFitsInBundle(d) && 1971 !node_bundling(d)->used_in_delay()) { 1972 1973 if (d->is_Mach() && !d->is_MachSafePoint()) { 1974 // A node that fits in the delay slot was found, so we need to 1975 // set the appropriate bits in the bundle pipeline information so 1976 // that it correctly indicates resource usage. Later, when we 1977 // attempt to add this instruction to the bundle, we will skip 1978 // setting the resource usage. 1979 _unconditional_delay_slot = d; 1980 node_bundling(n)->set_use_unconditional_delay(); 1981 node_bundling(d)->set_used_in_unconditional_delay(); 1982 _bundle_use.add_usage(avail_pipeline->resourceUse()); 1983 _current_latency[d->_idx] = _bundle_cycle_number; 1984 _next_node = d; 1985 ++_bundle_instr_count; 1986 #ifndef PRODUCT 1987 _unconditional_delays++; 1988 #endif 1989 break; 1990 } 1991 } 1992 } 1993 } 1994 1995 // No delay slot, add a nop to the usage 1996 if (!_unconditional_delay_slot) { 1997 // See if adding an instruction in the delay slot will overflow 1998 // the bundle. 1999 if (!NodeFitsInBundle(_nop)) { 2000 #ifndef PRODUCT 2001 if (_cfg->C->trace_opto_output()) 2002 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2003 #endif 2004 step(1); 2005 } 2006 2007 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2008 _next_node = _nop; 2009 ++_bundle_instr_count; 2010 } 2011 2012 // See if the instruction in the delay slot requires a 2013 // step of the bundles 2014 if (!NodeFitsInBundle(n)) { 2015 #ifndef PRODUCT 2016 if (_cfg->C->trace_opto_output()) 2017 tty->print("# *** STEP(branch won't fit) ***\n"); 2018 #endif 2019 // Update the state information 2020 _bundle_instr_count = 0; 2021 _bundle_cycle_number += 1; 2022 _bundle_use.step(1); 2023 } 2024 } 2025 2026 // Get the number of instructions 2027 uint instruction_count = node_pipeline->instructionCount(); 2028 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2029 instruction_count = 0; 2030 2031 // Compute the latency information 2032 uint delay = 0; 2033 2034 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2035 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2036 if (relative_latency < 0) 2037 relative_latency = 0; 2038 2039 delay = _bundle_use.full_latency(relative_latency, node_usage); 2040 2041 // Does not fit in this bundle, start a new one 2042 if (delay > 0) { 2043 step(delay); 2044 2045 #ifndef PRODUCT 2046 if (_cfg->C->trace_opto_output()) 2047 tty->print("# *** STEP(%d) ***\n", delay); 2048 #endif 2049 } 2050 } 2051 2052 // If this was placed in the delay slot, ignore it 2053 if (n != _unconditional_delay_slot) { 2054 2055 if (delay == 0) { 2056 if (node_pipeline->hasMultipleBundles()) { 2057 #ifndef PRODUCT 2058 if (_cfg->C->trace_opto_output()) 2059 tty->print("# *** STEP(multiple instructions) ***\n"); 2060 #endif 2061 step(1); 2062 } 2063 2064 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2065 #ifndef PRODUCT 2066 if (_cfg->C->trace_opto_output()) 2067 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2068 instruction_count + _bundle_instr_count, 2069 Pipeline::_max_instrs_per_cycle); 2070 #endif 2071 step(1); 2072 } 2073 } 2074 2075 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2076 _bundle_instr_count++; 2077 2078 // Set the node's latency 2079 _current_latency[n->_idx] = _bundle_cycle_number; 2080 2081 // Now merge the functional unit information 2082 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2083 _bundle_use.add_usage(node_usage); 2084 2085 // Increment the number of instructions in this bundle 2086 _bundle_instr_count += instruction_count; 2087 2088 // Remember this node for later 2089 if (n->is_Mach()) 2090 _next_node = n; 2091 } 2092 2093 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2094 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2095 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2096 // into the block. All other scheduled nodes get put in the schedule here. 2097 int op = n->Opcode(); 2098 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2099 (op != Op_Node && // Not an unused antidepedence node and 2100 // not an unallocated boxlock 2101 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2102 2103 // Push any trailing projections 2104 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2105 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2106 Node *foi = n->fast_out(i); 2107 if( foi->is_Proj() ) 2108 _scheduled.push(foi); 2109 } 2110 } 2111 2112 // Put the instruction in the schedule list 2113 _scheduled.push(n); 2114 } 2115 2116 #ifndef PRODUCT 2117 if (_cfg->C->trace_opto_output()) 2118 dump_available(); 2119 #endif 2120 2121 // Walk all the definitions, decrementing use counts, and 2122 // if a definition has a 0 use count, place it in the available list. 2123 DecrementUseCounts(n,bb); 2124 } 2125 2126 //------------------------------ComputeUseCount-------------------------------- 2127 // This method sets the use count within a basic block. We will ignore all 2128 // uses outside the current basic block. As we are doing a backwards walk, 2129 // any node we reach that has a use count of 0 may be scheduled. This also 2130 // avoids the problem of cyclic references from phi nodes, as long as phi 2131 // nodes are at the front of the basic block. This method also initializes 2132 // the available list to the set of instructions that have no uses within this 2133 // basic block. 2134 void Scheduling::ComputeUseCount(const Block *bb) { 2135 #ifndef PRODUCT 2136 if (_cfg->C->trace_opto_output()) 2137 tty->print("# -> ComputeUseCount\n"); 2138 #endif 2139 2140 // Clear the list of available and scheduled instructions, just in case 2141 _available.clear(); 2142 _scheduled.clear(); 2143 2144 // No delay slot specified 2145 _unconditional_delay_slot = NULL; 2146 2147 #ifdef ASSERT 2148 for( uint i=0; i < bb->_nodes.size(); i++ ) 2149 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2150 #endif 2151 2152 // Force the _uses count to never go to zero for unscheduable pieces 2153 // of the block 2154 for( uint k = 0; k < _bb_start; k++ ) 2155 _uses[bb->_nodes[k]->_idx] = 1; 2156 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2157 _uses[bb->_nodes[l]->_idx] = 1; 2158 2159 // Iterate backwards over the instructions in the block. Don't count the 2160 // branch projections at end or the block header instructions. 2161 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2162 Node *n = bb->_nodes[j]; 2163 if( n->is_Proj() ) continue; // Projections handled another way 2164 2165 // Account for all uses 2166 for ( uint k = 0; k < n->len(); k++ ) { 2167 Node *inp = n->in(k); 2168 if (!inp) continue; 2169 assert(inp != n, "no cycles allowed" ); 2170 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2171 if( inp->is_Proj() ) // Skip through Proj's 2172 inp = inp->in(0); 2173 ++_uses[inp->_idx]; // Count 1 block-local use 2174 } 2175 } 2176 2177 // If this instruction has a 0 use count, then it is available 2178 if (!_uses[n->_idx]) { 2179 _current_latency[n->_idx] = _bundle_cycle_number; 2180 AddNodeToAvailableList(n); 2181 } 2182 2183 #ifndef PRODUCT 2184 if (_cfg->C->trace_opto_output()) { 2185 tty->print("# uses: %3d: ", _uses[n->_idx]); 2186 n->dump(); 2187 } 2188 #endif 2189 } 2190 2191 #ifndef PRODUCT 2192 if (_cfg->C->trace_opto_output()) 2193 tty->print("# <- ComputeUseCount\n"); 2194 #endif 2195 } 2196 2197 // This routine performs scheduling on each basic block in reverse order, 2198 // using instruction latencies and taking into account function unit 2199 // availability. 2200 void Scheduling::DoScheduling() { 2201 #ifndef PRODUCT 2202 if (_cfg->C->trace_opto_output()) 2203 tty->print("# -> DoScheduling\n"); 2204 #endif 2205 2206 Block *succ_bb = NULL; 2207 Block *bb; 2208 2209 // Walk over all the basic blocks in reverse order 2210 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2211 bb = _cfg->_blocks[i]; 2212 2213 #ifndef PRODUCT 2214 if (_cfg->C->trace_opto_output()) { 2215 tty->print("# Schedule BB#%03d (initial)\n", i); 2216 for (uint j = 0; j < bb->_nodes.size(); j++) 2217 bb->_nodes[j]->dump(); 2218 } 2219 #endif 2220 2221 // On the head node, skip processing 2222 if( bb == _cfg->_broot ) 2223 continue; 2224 2225 // Skip empty, connector blocks 2226 if (bb->is_connector()) 2227 continue; 2228 2229 // If the following block is not the sole successor of 2230 // this one, then reset the pipeline information 2231 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2232 #ifndef PRODUCT 2233 if (_cfg->C->trace_opto_output()) { 2234 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2235 _next_node->_idx, _bundle_instr_count); 2236 } 2237 #endif 2238 step_and_clear(); 2239 } 2240 2241 // Leave untouched the starting instruction, any Phis, a CreateEx node 2242 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2243 _bb_end = bb->_nodes.size()-1; 2244 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2245 Node *n = bb->_nodes[_bb_start]; 2246 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2247 // Also, MachIdealNodes do not get scheduled 2248 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2249 MachNode *mach = n->as_Mach(); 2250 int iop = mach->ideal_Opcode(); 2251 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2252 if( iop == Op_Con ) continue; // Do not schedule Top 2253 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2254 mach->pipeline() == MachNode::pipeline_class() && 2255 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2256 continue; 2257 break; // Funny loop structure to be sure... 2258 } 2259 // Compute last "interesting" instruction in block - last instruction we 2260 // might schedule. _bb_end points just after last schedulable inst. We 2261 // normally schedule conditional branches (despite them being forced last 2262 // in the block), because they have delay slots we can fill. Calls all 2263 // have their delay slots filled in the template expansions, so we don't 2264 // bother scheduling them. 2265 Node *last = bb->_nodes[_bb_end]; 2266 if( last->is_Catch() || 2267 // Exclude unreachable path case when Halt node is in a separate block. 2268 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2269 // There must be a prior call. Skip it. 2270 while( !bb->_nodes[--_bb_end]->is_Call() ) { 2271 assert( bb->_nodes[_bb_end]->is_Proj(), "skipping projections after expected call" ); 2272 } 2273 } else if( last->is_MachNullCheck() ) { 2274 // Backup so the last null-checked memory instruction is 2275 // outside the schedulable range. Skip over the nullcheck, 2276 // projection, and the memory nodes. 2277 Node *mem = last->in(1); 2278 do { 2279 _bb_end--; 2280 } while (mem != bb->_nodes[_bb_end]); 2281 } else { 2282 // Set _bb_end to point after last schedulable inst. 2283 _bb_end++; 2284 } 2285 2286 assert( _bb_start <= _bb_end, "inverted block ends" ); 2287 2288 // Compute the register antidependencies for the basic block 2289 ComputeRegisterAntidependencies(bb); 2290 if (_cfg->C->failing()) return; // too many D-U pinch points 2291 2292 // Compute intra-bb latencies for the nodes 2293 ComputeLocalLatenciesForward(bb); 2294 2295 // Compute the usage within the block, and set the list of all nodes 2296 // in the block that have no uses within the block. 2297 ComputeUseCount(bb); 2298 2299 // Schedule the remaining instructions in the block 2300 while ( _available.size() > 0 ) { 2301 Node *n = ChooseNodeToBundle(); 2302 AddNodeToBundle(n,bb); 2303 } 2304 2305 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2306 #ifdef ASSERT 2307 for( uint l = _bb_start; l < _bb_end; l++ ) { 2308 Node *n = bb->_nodes[l]; 2309 uint m; 2310 for( m = 0; m < _bb_end-_bb_start; m++ ) 2311 if( _scheduled[m] == n ) 2312 break; 2313 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2314 } 2315 #endif 2316 2317 // Now copy the instructions (in reverse order) back to the block 2318 for ( uint k = _bb_start; k < _bb_end; k++ ) 2319 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2320 2321 #ifndef PRODUCT 2322 if (_cfg->C->trace_opto_output()) { 2323 tty->print("# Schedule BB#%03d (final)\n", i); 2324 uint current = 0; 2325 for (uint j = 0; j < bb->_nodes.size(); j++) { 2326 Node *n = bb->_nodes[j]; 2327 if( valid_bundle_info(n) ) { 2328 Bundle *bundle = node_bundling(n); 2329 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2330 tty->print("*** Bundle: "); 2331 bundle->dump(); 2332 } 2333 n->dump(); 2334 } 2335 } 2336 } 2337 #endif 2338 #ifdef ASSERT 2339 verify_good_schedule(bb,"after block local scheduling"); 2340 #endif 2341 } 2342 2343 #ifndef PRODUCT 2344 if (_cfg->C->trace_opto_output()) 2345 tty->print("# <- DoScheduling\n"); 2346 #endif 2347 2348 // Record final node-bundling array location 2349 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2350 2351 } // end DoScheduling 2352 2353 //------------------------------verify_good_schedule--------------------------- 2354 // Verify that no live-range used in the block is killed in the block by a 2355 // wrong DEF. This doesn't verify live-ranges that span blocks. 2356 2357 // Check for edge existence. Used to avoid adding redundant precedence edges. 2358 static bool edge_from_to( Node *from, Node *to ) { 2359 for( uint i=0; i<from->len(); i++ ) 2360 if( from->in(i) == to ) 2361 return true; 2362 return false; 2363 } 2364 2365 #ifdef ASSERT 2366 //------------------------------verify_do_def---------------------------------- 2367 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2368 // Check for bad kills 2369 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2370 Node *prior_use = _reg_node[def]; 2371 if( prior_use && !edge_from_to(prior_use,n) ) { 2372 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2373 n->dump(); 2374 tty->print_cr("..."); 2375 prior_use->dump(); 2376 assert_msg(edge_from_to(prior_use,n),msg); 2377 } 2378 _reg_node.map(def,NULL); // Kill live USEs 2379 } 2380 } 2381 2382 //------------------------------verify_good_schedule--------------------------- 2383 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2384 2385 // Zap to something reasonable for the verify code 2386 _reg_node.clear(); 2387 2388 // Walk over the block backwards. Check to make sure each DEF doesn't 2389 // kill a live value (other than the one it's supposed to). Add each 2390 // USE to the live set. 2391 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2392 Node *n = b->_nodes[i]; 2393 int n_op = n->Opcode(); 2394 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2395 // Fat-proj kills a slew of registers 2396 RegMask rm = n->out_RegMask();// Make local copy 2397 while( rm.is_NotEmpty() ) { 2398 OptoReg::Name kill = rm.find_first_elem(); 2399 rm.Remove(kill); 2400 verify_do_def( n, kill, msg ); 2401 } 2402 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2403 // Get DEF'd registers the normal way 2404 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2405 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2406 } 2407 2408 // Now make all USEs live 2409 for( uint i=1; i<n->req(); i++ ) { 2410 Node *def = n->in(i); 2411 assert(def != 0, "input edge required"); 2412 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2413 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2414 if( OptoReg::is_valid(reg_lo) ) { 2415 assert_msg(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg ); 2416 _reg_node.map(reg_lo,n); 2417 } 2418 if( OptoReg::is_valid(reg_hi) ) { 2419 assert_msg(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg ); 2420 _reg_node.map(reg_hi,n); 2421 } 2422 } 2423 2424 } 2425 2426 // Zap to something reasonable for the Antidependence code 2427 _reg_node.clear(); 2428 } 2429 #endif 2430 2431 // Conditionally add precedence edges. Avoid putting edges on Projs. 2432 static void add_prec_edge_from_to( Node *from, Node *to ) { 2433 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2434 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2435 from = from->in(0); 2436 } 2437 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2438 !edge_from_to( from, to ) ) // Avoid duplicate edge 2439 from->add_prec(to); 2440 } 2441 2442 //------------------------------anti_do_def------------------------------------ 2443 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2444 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2445 return; 2446 2447 Node *pinch = _reg_node[def_reg]; // Get pinch point 2448 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2449 is_def ) { // Check for a true def (not a kill) 2450 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2451 return; 2452 } 2453 2454 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2455 debug_only( def = (Node*)0xdeadbeef; ) 2456 2457 // After some number of kills there _may_ be a later def 2458 Node *later_def = NULL; 2459 2460 // Finding a kill requires a real pinch-point. 2461 // Check for not already having a pinch-point. 2462 // Pinch points are Op_Node's. 2463 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2464 later_def = pinch; // Must be def/kill as optimistic pinch-point 2465 if ( _pinch_free_list.size() > 0) { 2466 pinch = _pinch_free_list.pop(); 2467 } else { 2468 pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be 2469 } 2470 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2471 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2472 return; 2473 } 2474 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2475 _reg_node.map(def_reg,pinch); // Record pinch-point 2476 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2477 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2478 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2479 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2480 later_def = NULL; // and no later def 2481 } 2482 pinch->set_req(0,later_def); // Hook later def so we can find it 2483 } else { // Else have valid pinch point 2484 if( pinch->in(0) ) // If there is a later-def 2485 later_def = pinch->in(0); // Get it 2486 } 2487 2488 // Add output-dependence edge from later def to kill 2489 if( later_def ) // If there is some original def 2490 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2491 2492 // See if current kill is also a use, and so is forced to be the pinch-point. 2493 if( pinch->Opcode() == Op_Node ) { 2494 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2495 for( uint i=1; i<uses->req(); i++ ) { 2496 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2497 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2498 // Yes, found a use/kill pinch-point 2499 pinch->set_req(0,NULL); // 2500 pinch->replace_by(kill); // Move anti-dep edges up 2501 pinch = kill; 2502 _reg_node.map(def_reg,pinch); 2503 return; 2504 } 2505 } 2506 } 2507 2508 // Add edge from kill to pinch-point 2509 add_prec_edge_from_to(kill,pinch); 2510 } 2511 2512 //------------------------------anti_do_use------------------------------------ 2513 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2514 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2515 return; 2516 Node *pinch = _reg_node[use_reg]; // Get pinch point 2517 // Check for no later def_reg/kill in block 2518 if( pinch && _bbs[pinch->_idx] == b && 2519 // Use has to be block-local as well 2520 _bbs[use->_idx] == b ) { 2521 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2522 pinch->req() == 1 ) { // pinch not yet in block? 2523 pinch->del_req(0); // yank pointer to later-def, also set flag 2524 // Insert the pinch-point in the block just after the last use 2525 b->_nodes.insert(b->find_node(use)+1,pinch); 2526 _bb_end++; // Increase size scheduled region in block 2527 } 2528 2529 add_prec_edge_from_to(pinch,use); 2530 } 2531 } 2532 2533 //------------------------------ComputeRegisterAntidependences----------------- 2534 // We insert antidependences between the reads and following write of 2535 // allocated registers to prevent illegal code motion. Hopefully, the 2536 // number of added references should be fairly small, especially as we 2537 // are only adding references within the current basic block. 2538 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2539 2540 #ifdef ASSERT 2541 verify_good_schedule(b,"before block local scheduling"); 2542 #endif 2543 2544 // A valid schedule, for each register independently, is an endless cycle 2545 // of: a def, then some uses (connected to the def by true dependencies), 2546 // then some kills (defs with no uses), finally the cycle repeats with a new 2547 // def. The uses are allowed to float relative to each other, as are the 2548 // kills. No use is allowed to slide past a kill (or def). This requires 2549 // antidependencies between all uses of a single def and all kills that 2550 // follow, up to the next def. More edges are redundant, because later defs 2551 // & kills are already serialized with true or antidependencies. To keep 2552 // the edge count down, we add a 'pinch point' node if there's more than 2553 // one use or more than one kill/def. 2554 2555 // We add dependencies in one bottom-up pass. 2556 2557 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2558 2559 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2560 // register. If not, we record the DEF/KILL in _reg_node, the 2561 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2562 // "pinch point", a new Node that's in the graph but not in the block. 2563 // We put edges from the prior and current DEF/KILLs to the pinch point. 2564 // We put the pinch point in _reg_node. If there's already a pinch point 2565 // we merely add an edge from the current DEF/KILL to the pinch point. 2566 2567 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2568 // put an edge from the pinch point to the USE. 2569 2570 // To be expedient, the _reg_node array is pre-allocated for the whole 2571 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2572 // or a valid def/kill/pinch-point, or a leftover node from some prior 2573 // block. Leftover node from some prior block is treated like a NULL (no 2574 // prior def, so no anti-dependence needed). Valid def is distinguished by 2575 // it being in the current block. 2576 bool fat_proj_seen = false; 2577 uint last_safept = _bb_end-1; 2578 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2579 Node* last_safept_node = end_node; 2580 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2581 Node *n = b->_nodes[i]; 2582 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2583 if( n->Opcode() == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2584 // Fat-proj kills a slew of registers 2585 // This can add edges to 'n' and obscure whether or not it was a def, 2586 // hence the is_def flag. 2587 fat_proj_seen = true; 2588 RegMask rm = n->out_RegMask();// Make local copy 2589 while( rm.is_NotEmpty() ) { 2590 OptoReg::Name kill = rm.find_first_elem(); 2591 rm.Remove(kill); 2592 anti_do_def( b, n, kill, is_def ); 2593 } 2594 } else { 2595 // Get DEF'd registers the normal way 2596 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2597 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2598 } 2599 2600 // Check each register used by this instruction for a following DEF/KILL 2601 // that must occur afterward and requires an anti-dependence edge. 2602 for( uint j=0; j<n->req(); j++ ) { 2603 Node *def = n->in(j); 2604 if( def ) { 2605 assert( def->Opcode() != Op_MachProj || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2606 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2607 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2608 } 2609 } 2610 // Do not allow defs of new derived values to float above GC 2611 // points unless the base is definitely available at the GC point. 2612 2613 Node *m = b->_nodes[i]; 2614 2615 // Add precedence edge from following safepoint to use of derived pointer 2616 if( last_safept_node != end_node && 2617 m != last_safept_node) { 2618 for (uint k = 1; k < m->req(); k++) { 2619 const Type *t = m->in(k)->bottom_type(); 2620 if( t->isa_oop_ptr() && 2621 t->is_ptr()->offset() != 0 ) { 2622 last_safept_node->add_prec( m ); 2623 break; 2624 } 2625 } 2626 } 2627 2628 if( n->jvms() ) { // Precedence edge from derived to safept 2629 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2630 if( b->_nodes[last_safept] != last_safept_node ) { 2631 last_safept = b->find_node(last_safept_node); 2632 } 2633 for( uint j=last_safept; j > i; j-- ) { 2634 Node *mach = b->_nodes[j]; 2635 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2636 mach->add_prec( n ); 2637 } 2638 last_safept = i; 2639 last_safept_node = m; 2640 } 2641 } 2642 2643 if (fat_proj_seen) { 2644 // Garbage collect pinch nodes that were not consumed. 2645 // They are usually created by a fat kill MachProj for a call. 2646 garbage_collect_pinch_nodes(); 2647 } 2648 } 2649 2650 //------------------------------garbage_collect_pinch_nodes------------------------------- 2651 2652 // Garbage collect pinch nodes for reuse by other blocks. 2653 // 2654 // The block scheduler's insertion of anti-dependence 2655 // edges creates many pinch nodes when the block contains 2656 // 2 or more Calls. A pinch node is used to prevent a 2657 // combinatorial explosion of edges. If a set of kills for a 2658 // register is anti-dependent on a set of uses (or defs), rather 2659 // than adding an edge in the graph between each pair of kill 2660 // and use (or def), a pinch is inserted between them: 2661 // 2662 // use1 use2 use3 2663 // \ | / 2664 // \ | / 2665 // pinch 2666 // / | \ 2667 // / | \ 2668 // kill1 kill2 kill3 2669 // 2670 // One pinch node is created per register killed when 2671 // the second call is encountered during a backwards pass 2672 // over the block. Most of these pinch nodes are never 2673 // wired into the graph because the register is never 2674 // used or def'ed in the block. 2675 // 2676 void Scheduling::garbage_collect_pinch_nodes() { 2677 #ifndef PRODUCT 2678 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2679 #endif 2680 int trace_cnt = 0; 2681 for (uint k = 0; k < _reg_node.Size(); k++) { 2682 Node* pinch = _reg_node[k]; 2683 if (pinch != NULL && pinch->Opcode() == Op_Node && 2684 // no predecence input edges 2685 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2686 cleanup_pinch(pinch); 2687 _pinch_free_list.push(pinch); 2688 _reg_node.map(k, NULL); 2689 #ifndef PRODUCT 2690 if (_cfg->C->trace_opto_output()) { 2691 trace_cnt++; 2692 if (trace_cnt > 40) { 2693 tty->print("\n"); 2694 trace_cnt = 0; 2695 } 2696 tty->print(" %d", pinch->_idx); 2697 } 2698 #endif 2699 } 2700 } 2701 #ifndef PRODUCT 2702 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2703 #endif 2704 } 2705 2706 // Clean up a pinch node for reuse. 2707 void Scheduling::cleanup_pinch( Node *pinch ) { 2708 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2709 2710 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2711 Node* use = pinch->last_out(i); 2712 uint uses_found = 0; 2713 for (uint j = use->req(); j < use->len(); j++) { 2714 if (use->in(j) == pinch) { 2715 use->rm_prec(j); 2716 uses_found++; 2717 } 2718 } 2719 assert(uses_found > 0, "must be a precedence edge"); 2720 i -= uses_found; // we deleted 1 or more copies of this edge 2721 } 2722 // May have a later_def entry 2723 pinch->set_req(0, NULL); 2724 } 2725 2726 //------------------------------print_statistics------------------------------- 2727 #ifndef PRODUCT 2728 2729 void Scheduling::dump_available() const { 2730 tty->print("#Availist "); 2731 for (uint i = 0; i < _available.size(); i++) 2732 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2733 tty->cr(); 2734 } 2735 2736 // Print Scheduling Statistics 2737 void Scheduling::print_statistics() { 2738 // Print the size added by nops for bundling 2739 tty->print("Nops added %d bytes to total of %d bytes", 2740 _total_nop_size, _total_method_size); 2741 if (_total_method_size > 0) 2742 tty->print(", for %.2f%%", 2743 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2744 tty->print("\n"); 2745 2746 // Print the number of branch shadows filled 2747 if (Pipeline::_branch_has_delay_slot) { 2748 tty->print("Of %d branches, %d had unconditional delay slots filled", 2749 _total_branches, _total_unconditional_delays); 2750 if (_total_branches > 0) 2751 tty->print(", for %.2f%%", 2752 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2753 tty->print("\n"); 2754 } 2755 2756 uint total_instructions = 0, total_bundles = 0; 2757 2758 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2759 uint bundle_count = _total_instructions_per_bundle[i]; 2760 total_instructions += bundle_count * i; 2761 total_bundles += bundle_count; 2762 } 2763 2764 if (total_bundles > 0) 2765 tty->print("Average ILP (excluding nops) is %.2f\n", 2766 ((double)total_instructions) / ((double)total_bundles)); 2767 } 2768 #endif