output.cpp revision 1472:c18cbe5936b8
1/* 2 * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25#include "incls/_precompiled.incl" 26#include "incls/_output.cpp.incl" 27 28extern uint size_java_to_interp(); 29extern uint reloc_java_to_interp(); 30extern uint size_exception_handler(); 31extern uint size_deopt_handler(); 32 33#ifndef PRODUCT 34#define DEBUG_ARG(x) , x 35#else 36#define DEBUG_ARG(x) 37#endif 38 39extern int emit_exception_handler(CodeBuffer &cbuf); 40extern int emit_deopt_handler(CodeBuffer &cbuf); 41 42//------------------------------Output----------------------------------------- 43// Convert Nodes to instruction bits and pass off to the VM 44void 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 147bool 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 155bool 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 176int Compile::_CompiledZap_count = 0; 177 178void 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 246Node* 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-------------------- 263bool 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// 283void 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. 320void 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). 556static 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 565ObjectValue* 566Compile::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 578void 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 585void 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 780bool 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-------------------------------- 786void 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 bool is_method_handle_invoke = false; 798 bool return_oop = false; 799 800 // Add the safepoint in the DebugInfoRecorder 801 if( !mach->is_MachCall() ) { 802 mcall = NULL; 803 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 804 } else { 805 mcall = mach->as_MachCall(); 806 807 // Is the call a MethodHandle call? 808 if (mcall->is_MachCallJava()) { 809 if (mcall->as_MachCallJava()->_method_handle_invoke) { 810 assert(has_method_handle_invokes(), "must have been set during call generation"); 811 is_method_handle_invoke = true; 812 } 813 } 814 815 // Check if a call returns an object. 816 if (mcall->return_value_is_used() && 817 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 818 return_oop = true; 819 } 820 safepoint_pc_offset += mcall->ret_addr_offset(); 821 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 822 } 823 824 // Loop over the JVMState list to add scope information 825 // Do not skip safepoints with a NULL method, they need monitor info 826 JVMState* youngest_jvms = sfn->jvms(); 827 int max_depth = youngest_jvms->depth(); 828 829 // Allocate the object pool for scalar-replaced objects -- the map from 830 // small-integer keys (which can be recorded in the local and ostack 831 // arrays) to descriptions of the object state. 832 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 833 834 // Visit scopes from oldest to youngest. 835 for (int depth = 1; depth <= max_depth; depth++) { 836 JVMState* jvms = youngest_jvms->of_depth(depth); 837 int idx; 838 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 839 // Safepoints that do not have method() set only provide oop-map and monitor info 840 // to support GC; these do not support deoptimization. 841 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 842 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 843 int num_mon = jvms->nof_monitors(); 844 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 845 "JVMS local count must match that of the method"); 846 847 // Add Local and Expression Stack Information 848 849 // Insert locals into the locarray 850 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 851 for( idx = 0; idx < num_locs; idx++ ) { 852 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 853 } 854 855 // Insert expression stack entries into the exparray 856 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 857 for( idx = 0; idx < num_exps; idx++ ) { 858 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 859 } 860 861 // Add in mappings of the monitors 862 assert( !method || 863 !method->is_synchronized() || 864 method->is_native() || 865 num_mon > 0 || 866 !GenerateSynchronizationCode, 867 "monitors must always exist for synchronized methods"); 868 869 // Build the growable array of ScopeValues for exp stack 870 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 871 872 // Loop over monitors and insert into array 873 for(idx = 0; idx < num_mon; idx++) { 874 // Grab the node that defines this monitor 875 Node* box_node = sfn->monitor_box(jvms, idx); 876 Node* obj_node = sfn->monitor_obj(jvms, idx); 877 878 // Create ScopeValue for object 879 ScopeValue *scval = NULL; 880 881 if( obj_node->is_SafePointScalarObject() ) { 882 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 883 scval = Compile::sv_for_node_id(objs, spobj->_idx); 884 if (scval == NULL) { 885 const Type *t = obj_node->bottom_type(); 886 ciKlass* cik = t->is_oopptr()->klass(); 887 assert(cik->is_instance_klass() || 888 cik->is_array_klass(), "Not supported allocation."); 889 ObjectValue* sv = new ObjectValue(spobj->_idx, 890 new ConstantOopWriteValue(cik->constant_encoding())); 891 Compile::set_sv_for_object_node(objs, sv); 892 893 uint first_ind = spobj->first_index(); 894 for (uint i = 0; i < spobj->n_fields(); i++) { 895 Node* fld_node = sfn->in(first_ind+i); 896 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 897 } 898 scval = sv; 899 } 900 } else if( !obj_node->is_Con() ) { 901 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 902 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 903 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 904 } else { 905 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 906 } 907 } else { 908 const TypePtr *tp = obj_node->bottom_type()->make_ptr(); 909 scval = new ConstantOopWriteValue(tp->is_instptr()->const_oop()->constant_encoding()); 910 } 911 912 OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node); 913 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 914 while( !box_node->is_BoxLock() ) box_node = box_node->in(1); 915 monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated())); 916 } 917 918 // We dump the object pool first, since deoptimization reads it in first. 919 debug_info()->dump_object_pool(objs); 920 921 // Build first class objects to pass to scope 922 DebugToken *locvals = debug_info()->create_scope_values(locarray); 923 DebugToken *expvals = debug_info()->create_scope_values(exparray); 924 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 925 926 // Make method available for all Safepoints 927 ciMethod* scope_method = method ? method : _method; 928 // Describe the scope here 929 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 930 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 931 // Now we can describe the scope. 932 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 933 } // End jvms loop 934 935 // Mark the end of the scope set. 936 debug_info()->end_safepoint(safepoint_pc_offset); 937} 938 939 940 941// A simplified version of Process_OopMap_Node, to handle non-safepoints. 942class NonSafepointEmitter { 943 Compile* C; 944 JVMState* _pending_jvms; 945 int _pending_offset; 946 947 void emit_non_safepoint(); 948 949 public: 950 NonSafepointEmitter(Compile* compile) { 951 this->C = compile; 952 _pending_jvms = NULL; 953 _pending_offset = 0; 954 } 955 956 void observe_instruction(Node* n, int pc_offset) { 957 if (!C->debug_info()->recording_non_safepoints()) return; 958 959 Node_Notes* nn = C->node_notes_at(n->_idx); 960 if (nn == NULL || nn->jvms() == NULL) return; 961 if (_pending_jvms != NULL && 962 _pending_jvms->same_calls_as(nn->jvms())) { 963 // Repeated JVMS? Stretch it up here. 964 _pending_offset = pc_offset; 965 } else { 966 if (_pending_jvms != NULL && 967 _pending_offset < pc_offset) { 968 emit_non_safepoint(); 969 } 970 _pending_jvms = NULL; 971 if (pc_offset > C->debug_info()->last_pc_offset()) { 972 // This is the only way _pending_jvms can become non-NULL: 973 _pending_jvms = nn->jvms(); 974 _pending_offset = pc_offset; 975 } 976 } 977 } 978 979 // Stay out of the way of real safepoints: 980 void observe_safepoint(JVMState* jvms, int pc_offset) { 981 if (_pending_jvms != NULL && 982 !_pending_jvms->same_calls_as(jvms) && 983 _pending_offset < pc_offset) { 984 emit_non_safepoint(); 985 } 986 _pending_jvms = NULL; 987 } 988 989 void flush_at_end() { 990 if (_pending_jvms != NULL) { 991 emit_non_safepoint(); 992 } 993 _pending_jvms = NULL; 994 } 995}; 996 997void NonSafepointEmitter::emit_non_safepoint() { 998 JVMState* youngest_jvms = _pending_jvms; 999 int pc_offset = _pending_offset; 1000 1001 // Clear it now: 1002 _pending_jvms = NULL; 1003 1004 DebugInformationRecorder* debug_info = C->debug_info(); 1005 assert(debug_info->recording_non_safepoints(), "sanity"); 1006 1007 debug_info->add_non_safepoint(pc_offset); 1008 int max_depth = youngest_jvms->depth(); 1009 1010 // Visit scopes from oldest to youngest. 1011 for (int depth = 1; depth <= max_depth; depth++) { 1012 JVMState* jvms = youngest_jvms->of_depth(depth); 1013 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1014 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1015 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1016 } 1017 1018 // Mark the end of the scope set. 1019 debug_info->end_non_safepoint(pc_offset); 1020} 1021 1022 1023 1024// helper for Fill_buffer bailout logic 1025static void turn_off_compiler(Compile* C) { 1026 if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) { 1027 // Do not turn off compilation if a single giant method has 1028 // blown the code cache size. 1029 C->record_failure("excessive request to CodeCache"); 1030 } else { 1031 // Let CompilerBroker disable further compilations. 1032 C->record_failure("CodeCache is full"); 1033 } 1034} 1035 1036 1037//------------------------------Fill_buffer------------------------------------ 1038void Compile::Fill_buffer() { 1039 1040 // Set the initially allocated size 1041 int code_req = initial_code_capacity; 1042 int locs_req = initial_locs_capacity; 1043 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1044 int const_req = initial_const_capacity; 1045 bool labels_not_set = true; 1046 1047 int pad_req = NativeCall::instruction_size; 1048 // The extra spacing after the code is necessary on some platforms. 1049 // Sometimes we need to patch in a jump after the last instruction, 1050 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1051 1052 uint i; 1053 // Compute the byte offset where we can store the deopt pc. 1054 if (fixed_slots() != 0) { 1055 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1056 } 1057 1058 // Compute prolog code size 1059 _method_size = 0; 1060 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1061#ifdef IA64 1062 if (save_argument_registers()) { 1063 // 4815101: this is a stub with implicit and unknown precision fp args. 1064 // The usual spill mechanism can only generate stfd's in this case, which 1065 // doesn't work if the fp reg to spill contains a single-precision denorm. 1066 // Instead, we hack around the normal spill mechanism using stfspill's and 1067 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1068 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1069 // 1070 // If we ever implement 16-byte 'registers' == stack slots, we can 1071 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1072 // instead of stfd/stfs/ldfd/ldfs. 1073 _frame_slots += 8*(16/BytesPerInt); 1074 } 1075#endif 1076 assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" ); 1077 1078 // Create an array of unused labels, one for each basic block 1079 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1); 1080 1081 for( i=0; i <= _cfg->_num_blocks; i++ ) { 1082 blk_labels[i].init(); 1083 } 1084 1085 // If this machine supports different size branch offsets, then pre-compute 1086 // the length of the blocks 1087 if( _matcher->is_short_branch_offset(-1, 0) ) { 1088 Shorten_branches(blk_labels, code_req, locs_req, stub_req, const_req); 1089 labels_not_set = false; 1090 } 1091 1092 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1093 int exception_handler_req = size_exception_handler(); 1094 int deopt_handler_req = size_deopt_handler(); 1095 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1096 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1097 stub_req += MAX_stubs_size; // ensure per-stub margin 1098 code_req += MAX_inst_size; // ensure per-instruction margin 1099 1100 if (StressCodeBuffers) 1101 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1102 1103 int total_req = 1104 code_req + 1105 pad_req + 1106 stub_req + 1107 exception_handler_req + 1108 deopt_handler_req + // deopt handler 1109 const_req; 1110 1111 if (has_method_handle_invokes()) 1112 total_req += deopt_handler_req; // deopt MH handler 1113 1114 CodeBuffer* cb = code_buffer(); 1115 cb->initialize(total_req, locs_req); 1116 1117 // Have we run out of code space? 1118 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1119 turn_off_compiler(this); 1120 return; 1121 } 1122 // Configure the code buffer. 1123 cb->initialize_consts_size(const_req); 1124 cb->initialize_stubs_size(stub_req); 1125 cb->initialize_oop_recorder(env()->oop_recorder()); 1126 1127 // fill in the nop array for bundling computations 1128 MachNode *_nop_list[Bundle::_nop_count]; 1129 Bundle::initialize_nops(_nop_list, this); 1130 1131 // Create oopmap set. 1132 _oop_map_set = new OopMapSet(); 1133 1134 // !!!!! This preserves old handling of oopmaps for now 1135 debug_info()->set_oopmaps(_oop_map_set); 1136 1137 // Count and start of implicit null check instructions 1138 uint inct_cnt = 0; 1139 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1140 1141 // Count and start of calls 1142 uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); 1143 1144 uint return_offset = 0; 1145 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1146 1147 int previous_offset = 0; 1148 int current_offset = 0; 1149 int last_call_offset = -1; 1150 1151 // Create an array of unused labels, one for each basic block, if printing is enabled 1152#ifndef PRODUCT 1153 int *node_offsets = NULL; 1154 uint node_offset_limit = unique(); 1155 1156 1157 if ( print_assembly() ) 1158 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1159#endif 1160 1161 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1162 1163 // ------------------ 1164 // Now fill in the code buffer 1165 Node *delay_slot = NULL; 1166 1167 for( i=0; i < _cfg->_num_blocks; i++ ) { 1168 Block *b = _cfg->_blocks[i]; 1169 1170 Node *head = b->head(); 1171 1172 // If this block needs to start aligned (i.e, can be reached other 1173 // than by falling-thru from the previous block), then force the 1174 // start of a new bundle. 1175 if( Pipeline::requires_bundling() && starts_bundle(head) ) 1176 cb->flush_bundle(true); 1177 1178 // Define the label at the beginning of the basic block 1179 if( labels_not_set ) 1180 MacroAssembler(cb).bind( blk_labels[b->_pre_order] ); 1181 1182 else 1183 assert( blk_labels[b->_pre_order].loc_pos() == cb->code_size(), 1184 "label position does not match code offset" ); 1185 1186 uint last_inst = b->_nodes.size(); 1187 1188 // Emit block normally, except for last instruction. 1189 // Emit means "dump code bits into code buffer". 1190 for( uint j = 0; j<last_inst; j++ ) { 1191 1192 // Get the node 1193 Node* n = b->_nodes[j]; 1194 1195 // See if delay slots are supported 1196 if (valid_bundle_info(n) && 1197 node_bundling(n)->used_in_unconditional_delay()) { 1198 assert(delay_slot == NULL, "no use of delay slot node"); 1199 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1200 1201 delay_slot = n; 1202 continue; 1203 } 1204 1205 // If this starts a new instruction group, then flush the current one 1206 // (but allow split bundles) 1207 if( Pipeline::requires_bundling() && starts_bundle(n) ) 1208 cb->flush_bundle(false); 1209 1210 // The following logic is duplicated in the code ifdeffed for 1211 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1212 // should be factored out. Or maybe dispersed to the nodes? 1213 1214 // Special handling for SafePoint/Call Nodes 1215 bool is_mcall = false; 1216 if( n->is_Mach() ) { 1217 MachNode *mach = n->as_Mach(); 1218 is_mcall = n->is_MachCall(); 1219 bool is_sfn = n->is_MachSafePoint(); 1220 1221 // If this requires all previous instructions be flushed, then do so 1222 if( is_sfn || is_mcall || mach->alignment_required() != 1) { 1223 cb->flush_bundle(true); 1224 current_offset = cb->code_size(); 1225 } 1226 1227 // align the instruction if necessary 1228 int padding = mach->compute_padding(current_offset); 1229 // Make sure safepoint node for polling is distinct from a call's 1230 // return by adding a nop if needed. 1231 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) { 1232 padding = nop_size; 1233 } 1234 assert( labels_not_set || padding == 0, "instruction should already be aligned"); 1235 1236 if(padding > 0) { 1237 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1238 int nops_cnt = padding / nop_size; 1239 MachNode *nop = new (this) MachNopNode(nops_cnt); 1240 b->_nodes.insert(j++, nop); 1241 last_inst++; 1242 _cfg->_bbs.map( nop->_idx, b ); 1243 nop->emit(*cb, _regalloc); 1244 cb->flush_bundle(true); 1245 current_offset = cb->code_size(); 1246 } 1247 1248 // Remember the start of the last call in a basic block 1249 if (is_mcall) { 1250 MachCallNode *mcall = mach->as_MachCall(); 1251 1252 // This destination address is NOT PC-relative 1253 mcall->method_set((intptr_t)mcall->entry_point()); 1254 1255 // Save the return address 1256 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1257 1258 if (!mcall->is_safepoint_node()) { 1259 is_mcall = false; 1260 is_sfn = false; 1261 } 1262 } 1263 1264 // sfn will be valid whenever mcall is valid now because of inheritance 1265 if( is_sfn || is_mcall ) { 1266 1267 // Handle special safepoint nodes for synchronization 1268 if( !is_mcall ) { 1269 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1270 // !!!!! Stubs only need an oopmap right now, so bail out 1271 if( sfn->jvms()->method() == NULL) { 1272 // Write the oopmap directly to the code blob??!! 1273# ifdef ENABLE_ZAP_DEAD_LOCALS 1274 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1275# endif 1276 continue; 1277 } 1278 } // End synchronization 1279 1280 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1281 current_offset); 1282 Process_OopMap_Node(mach, current_offset); 1283 } // End if safepoint 1284 1285 // If this is a null check, then add the start of the previous instruction to the list 1286 else if( mach->is_MachNullCheck() ) { 1287 inct_starts[inct_cnt++] = previous_offset; 1288 } 1289 1290 // If this is a branch, then fill in the label with the target BB's label 1291 else if ( mach->is_Branch() ) { 1292 1293 if ( mach->ideal_Opcode() == Op_Jump ) { 1294 for (uint h = 0; h < b->_num_succs; h++ ) { 1295 Block* succs_block = b->_succs[h]; 1296 for (uint j = 1; j < succs_block->num_preds(); j++) { 1297 Node* jpn = succs_block->pred(j); 1298 if ( jpn->is_JumpProj() && jpn->in(0) == mach ) { 1299 uint block_num = succs_block->non_connector()->_pre_order; 1300 Label *blkLabel = &blk_labels[block_num]; 1301 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1302 } 1303 } 1304 } 1305 } else { 1306 // For Branchs 1307 // This requires the TRUE branch target be in succs[0] 1308 uint block_num = b->non_connector_successor(0)->_pre_order; 1309 mach->label_set( blk_labels[block_num], block_num ); 1310 } 1311 } 1312 1313#ifdef ASSERT 1314 // Check that oop-store precedes the card-mark 1315 else if( mach->ideal_Opcode() == Op_StoreCM ) { 1316 uint storeCM_idx = j; 1317 Node *oop_store = mach->in(mach->_cnt); // First precedence edge 1318 assert( oop_store != NULL, "storeCM expects a precedence edge"); 1319 uint i4; 1320 for( i4 = 0; i4 < last_inst; ++i4 ) { 1321 if( b->_nodes[i4] == oop_store ) break; 1322 } 1323 // Note: This test can provide a false failure if other precedence 1324 // edges have been added to the storeCMNode. 1325 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1326 } 1327#endif 1328 1329 else if( !n->is_Proj() ) { 1330 // Remember the beginning of the previous instruction, in case 1331 // it's followed by a flag-kill and a null-check. Happens on 1332 // Intel all the time, with add-to-memory kind of opcodes. 1333 previous_offset = current_offset; 1334 } 1335 } 1336 1337 // Verify that there is sufficient space remaining 1338 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1339 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1340 turn_off_compiler(this); 1341 return; 1342 } 1343 1344 // Save the offset for the listing 1345#ifndef PRODUCT 1346 if( node_offsets && n->_idx < node_offset_limit ) 1347 node_offsets[n->_idx] = cb->code_size(); 1348#endif 1349 1350 // "Normal" instruction case 1351 n->emit(*cb, _regalloc); 1352 current_offset = cb->code_size(); 1353 non_safepoints.observe_instruction(n, current_offset); 1354 1355 // mcall is last "call" that can be a safepoint 1356 // record it so we can see if a poll will directly follow it 1357 // in which case we'll need a pad to make the PcDesc sites unique 1358 // see 5010568. This can be slightly inaccurate but conservative 1359 // in the case that return address is not actually at current_offset. 1360 // This is a small price to pay. 1361 1362 if (is_mcall) { 1363 last_call_offset = current_offset; 1364 } 1365 1366 // See if this instruction has a delay slot 1367 if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1368 assert(delay_slot != NULL, "expecting delay slot node"); 1369 1370 // Back up 1 instruction 1371 cb->set_code_end( 1372 cb->code_end()-Pipeline::instr_unit_size()); 1373 1374 // Save the offset for the listing 1375#ifndef PRODUCT 1376 if( node_offsets && delay_slot->_idx < node_offset_limit ) 1377 node_offsets[delay_slot->_idx] = cb->code_size(); 1378#endif 1379 1380 // Support a SafePoint in the delay slot 1381 if( delay_slot->is_MachSafePoint() ) { 1382 MachNode *mach = delay_slot->as_Mach(); 1383 // !!!!! Stubs only need an oopmap right now, so bail out 1384 if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) { 1385 // Write the oopmap directly to the code blob??!! 1386# ifdef ENABLE_ZAP_DEAD_LOCALS 1387 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1388# endif 1389 delay_slot = NULL; 1390 continue; 1391 } 1392 1393 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1394 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1395 adjusted_offset); 1396 // Generate an OopMap entry 1397 Process_OopMap_Node(mach, adjusted_offset); 1398 } 1399 1400 // Insert the delay slot instruction 1401 delay_slot->emit(*cb, _regalloc); 1402 1403 // Don't reuse it 1404 delay_slot = NULL; 1405 } 1406 1407 } // End for all instructions in block 1408 1409 // If the next block is the top of a loop, pad this block out to align 1410 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1411 if( i<_cfg->_num_blocks-1 ) { 1412 Block *nb = _cfg->_blocks[i+1]; 1413 uint padding = nb->alignment_padding(current_offset); 1414 if( padding > 0 ) { 1415 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1416 b->_nodes.insert( b->_nodes.size(), nop ); 1417 _cfg->_bbs.map( nop->_idx, b ); 1418 nop->emit(*cb, _regalloc); 1419 current_offset = cb->code_size(); 1420 } 1421 } 1422 1423 } // End of for all blocks 1424 1425 non_safepoints.flush_at_end(); 1426 1427 // Offset too large? 1428 if (failing()) return; 1429 1430 // Define a pseudo-label at the end of the code 1431 MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] ); 1432 1433 // Compute the size of the first block 1434 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1435 1436 assert(cb->code_size() < 500000, "method is unreasonably large"); 1437 1438 // ------------------ 1439 1440#ifndef PRODUCT 1441 // Information on the size of the method, without the extraneous code 1442 Scheduling::increment_method_size(cb->code_size()); 1443#endif 1444 1445 // ------------------ 1446 // Fill in exception table entries. 1447 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1448 1449 // Only java methods have exception handlers and deopt handlers 1450 if (_method) { 1451 // Emit the exception handler code. 1452 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1453 // Emit the deopt handler code. 1454 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1455 1456 // Emit the MethodHandle deopt handler code (if required). 1457 if (has_method_handle_invokes()) { 1458 // We can use the same code as for the normal deopt handler, we 1459 // just need a different entry point address. 1460 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1461 } 1462 } 1463 1464 // One last check for failed CodeBuffer::expand: 1465 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1466 turn_off_compiler(this); 1467 return; 1468 } 1469 1470#ifndef PRODUCT 1471 // Dump the assembly code, including basic-block numbers 1472 if (print_assembly()) { 1473 ttyLocker ttyl; // keep the following output all in one block 1474 if (!VMThread::should_terminate()) { // test this under the tty lock 1475 // This output goes directly to the tty, not the compiler log. 1476 // To enable tools to match it up with the compilation activity, 1477 // be sure to tag this tty output with the compile ID. 1478 if (xtty != NULL) { 1479 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1480 is_osr_compilation() ? " compile_kind='osr'" : 1481 ""); 1482 } 1483 if (method() != NULL) { 1484 method()->print_oop(); 1485 print_codes(); 1486 } 1487 dump_asm(node_offsets, node_offset_limit); 1488 if (xtty != NULL) { 1489 xtty->tail("opto_assembly"); 1490 } 1491 } 1492 } 1493#endif 1494 1495} 1496 1497void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1498 _inc_table.set_size(cnt); 1499 1500 uint inct_cnt = 0; 1501 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1502 Block *b = _cfg->_blocks[i]; 1503 Node *n = NULL; 1504 int j; 1505 1506 // Find the branch; ignore trailing NOPs. 1507 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1508 n = b->_nodes[j]; 1509 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1510 break; 1511 } 1512 1513 // If we didn't find anything, continue 1514 if( j < 0 ) continue; 1515 1516 // Compute ExceptionHandlerTable subtable entry and add it 1517 // (skip empty blocks) 1518 if( n->is_Catch() ) { 1519 1520 // Get the offset of the return from the call 1521 uint call_return = call_returns[b->_pre_order]; 1522#ifdef ASSERT 1523 assert( call_return > 0, "no call seen for this basic block" ); 1524 while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; 1525 assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); 1526#endif 1527 // last instruction is a CatchNode, find it's CatchProjNodes 1528 int nof_succs = b->_num_succs; 1529 // allocate space 1530 GrowableArray<intptr_t> handler_bcis(nof_succs); 1531 GrowableArray<intptr_t> handler_pcos(nof_succs); 1532 // iterate through all successors 1533 for (int j = 0; j < nof_succs; j++) { 1534 Block* s = b->_succs[j]; 1535 bool found_p = false; 1536 for( uint k = 1; k < s->num_preds(); k++ ) { 1537 Node *pk = s->pred(k); 1538 if( pk->is_CatchProj() && pk->in(0) == n ) { 1539 const CatchProjNode* p = pk->as_CatchProj(); 1540 found_p = true; 1541 // add the corresponding handler bci & pco information 1542 if( p->_con != CatchProjNode::fall_through_index ) { 1543 // p leads to an exception handler (and is not fall through) 1544 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1545 // no duplicates, please 1546 if( !handler_bcis.contains(p->handler_bci()) ) { 1547 uint block_num = s->non_connector()->_pre_order; 1548 handler_bcis.append(p->handler_bci()); 1549 handler_pcos.append(blk_labels[block_num].loc_pos()); 1550 } 1551 } 1552 } 1553 } 1554 assert(found_p, "no matching predecessor found"); 1555 // Note: Due to empty block removal, one block may have 1556 // several CatchProj inputs, from the same Catch. 1557 } 1558 1559 // Set the offset of the return from the call 1560 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1561 continue; 1562 } 1563 1564 // Handle implicit null exception table updates 1565 if( n->is_MachNullCheck() ) { 1566 uint block_num = b->non_connector_successor(0)->_pre_order; 1567 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1568 continue; 1569 } 1570 } // End of for all blocks fill in exception table entries 1571} 1572 1573// Static Variables 1574#ifndef PRODUCT 1575uint Scheduling::_total_nop_size = 0; 1576uint Scheduling::_total_method_size = 0; 1577uint Scheduling::_total_branches = 0; 1578uint Scheduling::_total_unconditional_delays = 0; 1579uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1580#endif 1581 1582// Initializer for class Scheduling 1583 1584Scheduling::Scheduling(Arena *arena, Compile &compile) 1585 : _arena(arena), 1586 _cfg(compile.cfg()), 1587 _bbs(compile.cfg()->_bbs), 1588 _regalloc(compile.regalloc()), 1589 _reg_node(arena), 1590 _bundle_instr_count(0), 1591 _bundle_cycle_number(0), 1592 _scheduled(arena), 1593 _available(arena), 1594 _next_node(NULL), 1595 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1596 _pinch_free_list(arena) 1597#ifndef PRODUCT 1598 , _branches(0) 1599 , _unconditional_delays(0) 1600#endif 1601{ 1602 // Create a MachNopNode 1603 _nop = new (&compile) MachNopNode(); 1604 1605 // Now that the nops are in the array, save the count 1606 // (but allow entries for the nops) 1607 _node_bundling_limit = compile.unique(); 1608 uint node_max = _regalloc->node_regs_max_index(); 1609 1610 compile.set_node_bundling_limit(_node_bundling_limit); 1611 1612 // This one is persistent within the Compile class 1613 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1614 1615 // Allocate space for fixed-size arrays 1616 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1617 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1618 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1619 1620 // Clear the arrays 1621 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1622 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1623 memset(_uses, 0, node_max * sizeof(short)); 1624 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1625 1626 // Clear the bundling information 1627 memcpy(_bundle_use_elements, 1628 Pipeline_Use::elaborated_elements, 1629 sizeof(Pipeline_Use::elaborated_elements)); 1630 1631 // Get the last node 1632 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1633 1634 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1635} 1636 1637#ifndef PRODUCT 1638// Scheduling destructor 1639Scheduling::~Scheduling() { 1640 _total_branches += _branches; 1641 _total_unconditional_delays += _unconditional_delays; 1642} 1643#endif 1644 1645// Step ahead "i" cycles 1646void Scheduling::step(uint i) { 1647 1648 Bundle *bundle = node_bundling(_next_node); 1649 bundle->set_starts_bundle(); 1650 1651 // Update the bundle record, but leave the flags information alone 1652 if (_bundle_instr_count > 0) { 1653 bundle->set_instr_count(_bundle_instr_count); 1654 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1655 } 1656 1657 // Update the state information 1658 _bundle_instr_count = 0; 1659 _bundle_cycle_number += i; 1660 _bundle_use.step(i); 1661} 1662 1663void Scheduling::step_and_clear() { 1664 Bundle *bundle = node_bundling(_next_node); 1665 bundle->set_starts_bundle(); 1666 1667 // Update the bundle record 1668 if (_bundle_instr_count > 0) { 1669 bundle->set_instr_count(_bundle_instr_count); 1670 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1671 1672 _bundle_cycle_number += 1; 1673 } 1674 1675 // Clear the bundling information 1676 _bundle_instr_count = 0; 1677 _bundle_use.reset(); 1678 1679 memcpy(_bundle_use_elements, 1680 Pipeline_Use::elaborated_elements, 1681 sizeof(Pipeline_Use::elaborated_elements)); 1682} 1683 1684//------------------------------ScheduleAndBundle------------------------------ 1685// Perform instruction scheduling and bundling over the sequence of 1686// instructions in backwards order. 1687void Compile::ScheduleAndBundle() { 1688 1689 // Don't optimize this if it isn't a method 1690 if (!_method) 1691 return; 1692 1693 // Don't optimize this if scheduling is disabled 1694 if (!do_scheduling()) 1695 return; 1696 1697 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1698 1699 // Create a data structure for all the scheduling information 1700 Scheduling scheduling(Thread::current()->resource_area(), *this); 1701 1702 // Walk backwards over each basic block, computing the needed alignment 1703 // Walk over all the basic blocks 1704 scheduling.DoScheduling(); 1705} 1706 1707//------------------------------ComputeLocalLatenciesForward------------------- 1708// Compute the latency of all the instructions. This is fairly simple, 1709// because we already have a legal ordering. Walk over the instructions 1710// from first to last, and compute the latency of the instruction based 1711// on the latency of the preceding instruction(s). 1712void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1713#ifndef PRODUCT 1714 if (_cfg->C->trace_opto_output()) 1715 tty->print("# -> ComputeLocalLatenciesForward\n"); 1716#endif 1717 1718 // Walk over all the schedulable instructions 1719 for( uint j=_bb_start; j < _bb_end; j++ ) { 1720 1721 // This is a kludge, forcing all latency calculations to start at 1. 1722 // Used to allow latency 0 to force an instruction to the beginning 1723 // of the bb 1724 uint latency = 1; 1725 Node *use = bb->_nodes[j]; 1726 uint nlen = use->len(); 1727 1728 // Walk over all the inputs 1729 for ( uint k=0; k < nlen; k++ ) { 1730 Node *def = use->in(k); 1731 if (!def) 1732 continue; 1733 1734 uint l = _node_latency[def->_idx] + use->latency(k); 1735 if (latency < l) 1736 latency = l; 1737 } 1738 1739 _node_latency[use->_idx] = latency; 1740 1741#ifndef PRODUCT 1742 if (_cfg->C->trace_opto_output()) { 1743 tty->print("# latency %4d: ", latency); 1744 use->dump(); 1745 } 1746#endif 1747 } 1748 1749#ifndef PRODUCT 1750 if (_cfg->C->trace_opto_output()) 1751 tty->print("# <- ComputeLocalLatenciesForward\n"); 1752#endif 1753 1754} // end ComputeLocalLatenciesForward 1755 1756// See if this node fits into the present instruction bundle 1757bool Scheduling::NodeFitsInBundle(Node *n) { 1758 uint n_idx = n->_idx; 1759 1760 // If this is the unconditional delay instruction, then it fits 1761 if (n == _unconditional_delay_slot) { 1762#ifndef PRODUCT 1763 if (_cfg->C->trace_opto_output()) 1764 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1765#endif 1766 return (true); 1767 } 1768 1769 // If the node cannot be scheduled this cycle, skip it 1770 if (_current_latency[n_idx] > _bundle_cycle_number) { 1771#ifndef PRODUCT 1772 if (_cfg->C->trace_opto_output()) 1773 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1774 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1775#endif 1776 return (false); 1777 } 1778 1779 const Pipeline *node_pipeline = n->pipeline(); 1780 1781 uint instruction_count = node_pipeline->instructionCount(); 1782 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1783 instruction_count = 0; 1784 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1785 instruction_count++; 1786 1787 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1788#ifndef PRODUCT 1789 if (_cfg->C->trace_opto_output()) 1790 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1791 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1792#endif 1793 return (false); 1794 } 1795 1796 // Don't allow non-machine nodes to be handled this way 1797 if (!n->is_Mach() && instruction_count == 0) 1798 return (false); 1799 1800 // See if there is any overlap 1801 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1802 1803 if (delay > 0) { 1804#ifndef PRODUCT 1805 if (_cfg->C->trace_opto_output()) 1806 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1807#endif 1808 return false; 1809 } 1810 1811#ifndef PRODUCT 1812 if (_cfg->C->trace_opto_output()) 1813 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1814#endif 1815 1816 return true; 1817} 1818 1819Node * Scheduling::ChooseNodeToBundle() { 1820 uint siz = _available.size(); 1821 1822 if (siz == 0) { 1823 1824#ifndef PRODUCT 1825 if (_cfg->C->trace_opto_output()) 1826 tty->print("# ChooseNodeToBundle: NULL\n"); 1827#endif 1828 return (NULL); 1829 } 1830 1831 // Fast path, if only 1 instruction in the bundle 1832 if (siz == 1) { 1833#ifndef PRODUCT 1834 if (_cfg->C->trace_opto_output()) { 1835 tty->print("# ChooseNodeToBundle (only 1): "); 1836 _available[0]->dump(); 1837 } 1838#endif 1839 return (_available[0]); 1840 } 1841 1842 // Don't bother, if the bundle is already full 1843 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1844 for ( uint i = 0; i < siz; i++ ) { 1845 Node *n = _available[i]; 1846 1847 // Skip projections, we'll handle them another way 1848 if (n->is_Proj()) 1849 continue; 1850 1851 // This presupposed that instructions are inserted into the 1852 // available list in a legality order; i.e. instructions that 1853 // must be inserted first are at the head of the list 1854 if (NodeFitsInBundle(n)) { 1855#ifndef PRODUCT 1856 if (_cfg->C->trace_opto_output()) { 1857 tty->print("# ChooseNodeToBundle: "); 1858 n->dump(); 1859 } 1860#endif 1861 return (n); 1862 } 1863 } 1864 } 1865 1866 // Nothing fits in this bundle, choose the highest priority 1867#ifndef PRODUCT 1868 if (_cfg->C->trace_opto_output()) { 1869 tty->print("# ChooseNodeToBundle: "); 1870 _available[0]->dump(); 1871 } 1872#endif 1873 1874 return _available[0]; 1875} 1876 1877//------------------------------AddNodeToAvailableList------------------------- 1878void Scheduling::AddNodeToAvailableList(Node *n) { 1879 assert( !n->is_Proj(), "projections never directly made available" ); 1880#ifndef PRODUCT 1881 if (_cfg->C->trace_opto_output()) { 1882 tty->print("# AddNodeToAvailableList: "); 1883 n->dump(); 1884 } 1885#endif 1886 1887 int latency = _current_latency[n->_idx]; 1888 1889 // Insert in latency order (insertion sort) 1890 uint i; 1891 for ( i=0; i < _available.size(); i++ ) 1892 if (_current_latency[_available[i]->_idx] > latency) 1893 break; 1894 1895 // Special Check for compares following branches 1896 if( n->is_Mach() && _scheduled.size() > 0 ) { 1897 int op = n->as_Mach()->ideal_Opcode(); 1898 Node *last = _scheduled[0]; 1899 if( last->is_MachIf() && last->in(1) == n && 1900 ( op == Op_CmpI || 1901 op == Op_CmpU || 1902 op == Op_CmpP || 1903 op == Op_CmpF || 1904 op == Op_CmpD || 1905 op == Op_CmpL ) ) { 1906 1907 // Recalculate position, moving to front of same latency 1908 for ( i=0 ; i < _available.size(); i++ ) 1909 if (_current_latency[_available[i]->_idx] >= latency) 1910 break; 1911 } 1912 } 1913 1914 // Insert the node in the available list 1915 _available.insert(i, n); 1916 1917#ifndef PRODUCT 1918 if (_cfg->C->trace_opto_output()) 1919 dump_available(); 1920#endif 1921} 1922 1923//------------------------------DecrementUseCounts----------------------------- 1924void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 1925 for ( uint i=0; i < n->len(); i++ ) { 1926 Node *def = n->in(i); 1927 if (!def) continue; 1928 if( def->is_Proj() ) // If this is a machine projection, then 1929 def = def->in(0); // propagate usage thru to the base instruction 1930 1931 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 1932 continue; 1933 1934 // Compute the latency 1935 uint l = _bundle_cycle_number + n->latency(i); 1936 if (_current_latency[def->_idx] < l) 1937 _current_latency[def->_idx] = l; 1938 1939 // If this does not have uses then schedule it 1940 if ((--_uses[def->_idx]) == 0) 1941 AddNodeToAvailableList(def); 1942 } 1943} 1944 1945//------------------------------AddNodeToBundle-------------------------------- 1946void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 1947#ifndef PRODUCT 1948 if (_cfg->C->trace_opto_output()) { 1949 tty->print("# AddNodeToBundle: "); 1950 n->dump(); 1951 } 1952#endif 1953 1954 // Remove this from the available list 1955 uint i; 1956 for (i = 0; i < _available.size(); i++) 1957 if (_available[i] == n) 1958 break; 1959 assert(i < _available.size(), "entry in _available list not found"); 1960 _available.remove(i); 1961 1962 // See if this fits in the current bundle 1963 const Pipeline *node_pipeline = n->pipeline(); 1964 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 1965 1966 // Check for instructions to be placed in the delay slot. We 1967 // do this before we actually schedule the current instruction, 1968 // because the delay slot follows the current instruction. 1969 if (Pipeline::_branch_has_delay_slot && 1970 node_pipeline->hasBranchDelay() && 1971 !_unconditional_delay_slot) { 1972 1973 uint siz = _available.size(); 1974 1975 // Conditional branches can support an instruction that 1976 // is unconditionally executed and not dependent by the 1977 // branch, OR a conditionally executed instruction if 1978 // the branch is taken. In practice, this means that 1979 // the first instruction at the branch target is 1980 // copied to the delay slot, and the branch goes to 1981 // the instruction after that at the branch target 1982 if ( n->is_Mach() && n->is_Branch() ) { 1983 1984 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 1985 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 1986 1987#ifndef PRODUCT 1988 _branches++; 1989#endif 1990 1991 // At least 1 instruction is on the available list 1992 // that is not dependent on the branch 1993 for (uint i = 0; i < siz; i++) { 1994 Node *d = _available[i]; 1995 const Pipeline *avail_pipeline = d->pipeline(); 1996 1997 // Don't allow safepoints in the branch shadow, that will 1998 // cause a number of difficulties 1999 if ( avail_pipeline->instructionCount() == 1 && 2000 !avail_pipeline->hasMultipleBundles() && 2001 !avail_pipeline->hasBranchDelay() && 2002 Pipeline::instr_has_unit_size() && 2003 d->size(_regalloc) == Pipeline::instr_unit_size() && 2004 NodeFitsInBundle(d) && 2005 !node_bundling(d)->used_in_delay()) { 2006 2007 if (d->is_Mach() && !d->is_MachSafePoint()) { 2008 // A node that fits in the delay slot was found, so we need to 2009 // set the appropriate bits in the bundle pipeline information so 2010 // that it correctly indicates resource usage. Later, when we 2011 // attempt to add this instruction to the bundle, we will skip 2012 // setting the resource usage. 2013 _unconditional_delay_slot = d; 2014 node_bundling(n)->set_use_unconditional_delay(); 2015 node_bundling(d)->set_used_in_unconditional_delay(); 2016 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2017 _current_latency[d->_idx] = _bundle_cycle_number; 2018 _next_node = d; 2019 ++_bundle_instr_count; 2020#ifndef PRODUCT 2021 _unconditional_delays++; 2022#endif 2023 break; 2024 } 2025 } 2026 } 2027 } 2028 2029 // No delay slot, add a nop to the usage 2030 if (!_unconditional_delay_slot) { 2031 // See if adding an instruction in the delay slot will overflow 2032 // the bundle. 2033 if (!NodeFitsInBundle(_nop)) { 2034#ifndef PRODUCT 2035 if (_cfg->C->trace_opto_output()) 2036 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2037#endif 2038 step(1); 2039 } 2040 2041 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2042 _next_node = _nop; 2043 ++_bundle_instr_count; 2044 } 2045 2046 // See if the instruction in the delay slot requires a 2047 // step of the bundles 2048 if (!NodeFitsInBundle(n)) { 2049#ifndef PRODUCT 2050 if (_cfg->C->trace_opto_output()) 2051 tty->print("# *** STEP(branch won't fit) ***\n"); 2052#endif 2053 // Update the state information 2054 _bundle_instr_count = 0; 2055 _bundle_cycle_number += 1; 2056 _bundle_use.step(1); 2057 } 2058 } 2059 2060 // Get the number of instructions 2061 uint instruction_count = node_pipeline->instructionCount(); 2062 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2063 instruction_count = 0; 2064 2065 // Compute the latency information 2066 uint delay = 0; 2067 2068 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2069 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2070 if (relative_latency < 0) 2071 relative_latency = 0; 2072 2073 delay = _bundle_use.full_latency(relative_latency, node_usage); 2074 2075 // Does not fit in this bundle, start a new one 2076 if (delay > 0) { 2077 step(delay); 2078 2079#ifndef PRODUCT 2080 if (_cfg->C->trace_opto_output()) 2081 tty->print("# *** STEP(%d) ***\n", delay); 2082#endif 2083 } 2084 } 2085 2086 // If this was placed in the delay slot, ignore it 2087 if (n != _unconditional_delay_slot) { 2088 2089 if (delay == 0) { 2090 if (node_pipeline->hasMultipleBundles()) { 2091#ifndef PRODUCT 2092 if (_cfg->C->trace_opto_output()) 2093 tty->print("# *** STEP(multiple instructions) ***\n"); 2094#endif 2095 step(1); 2096 } 2097 2098 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2099#ifndef PRODUCT 2100 if (_cfg->C->trace_opto_output()) 2101 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2102 instruction_count + _bundle_instr_count, 2103 Pipeline::_max_instrs_per_cycle); 2104#endif 2105 step(1); 2106 } 2107 } 2108 2109 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2110 _bundle_instr_count++; 2111 2112 // Set the node's latency 2113 _current_latency[n->_idx] = _bundle_cycle_number; 2114 2115 // Now merge the functional unit information 2116 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2117 _bundle_use.add_usage(node_usage); 2118 2119 // Increment the number of instructions in this bundle 2120 _bundle_instr_count += instruction_count; 2121 2122 // Remember this node for later 2123 if (n->is_Mach()) 2124 _next_node = n; 2125 } 2126 2127 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2128 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2129 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2130 // into the block. All other scheduled nodes get put in the schedule here. 2131 int op = n->Opcode(); 2132 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2133 (op != Op_Node && // Not an unused antidepedence node and 2134 // not an unallocated boxlock 2135 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2136 2137 // Push any trailing projections 2138 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2139 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2140 Node *foi = n->fast_out(i); 2141 if( foi->is_Proj() ) 2142 _scheduled.push(foi); 2143 } 2144 } 2145 2146 // Put the instruction in the schedule list 2147 _scheduled.push(n); 2148 } 2149 2150#ifndef PRODUCT 2151 if (_cfg->C->trace_opto_output()) 2152 dump_available(); 2153#endif 2154 2155 // Walk all the definitions, decrementing use counts, and 2156 // if a definition has a 0 use count, place it in the available list. 2157 DecrementUseCounts(n,bb); 2158} 2159 2160//------------------------------ComputeUseCount-------------------------------- 2161// This method sets the use count within a basic block. We will ignore all 2162// uses outside the current basic block. As we are doing a backwards walk, 2163// any node we reach that has a use count of 0 may be scheduled. This also 2164// avoids the problem of cyclic references from phi nodes, as long as phi 2165// nodes are at the front of the basic block. This method also initializes 2166// the available list to the set of instructions that have no uses within this 2167// basic block. 2168void Scheduling::ComputeUseCount(const Block *bb) { 2169#ifndef PRODUCT 2170 if (_cfg->C->trace_opto_output()) 2171 tty->print("# -> ComputeUseCount\n"); 2172#endif 2173 2174 // Clear the list of available and scheduled instructions, just in case 2175 _available.clear(); 2176 _scheduled.clear(); 2177 2178 // No delay slot specified 2179 _unconditional_delay_slot = NULL; 2180 2181#ifdef ASSERT 2182 for( uint i=0; i < bb->_nodes.size(); i++ ) 2183 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2184#endif 2185 2186 // Force the _uses count to never go to zero for unscheduable pieces 2187 // of the block 2188 for( uint k = 0; k < _bb_start; k++ ) 2189 _uses[bb->_nodes[k]->_idx] = 1; 2190 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2191 _uses[bb->_nodes[l]->_idx] = 1; 2192 2193 // Iterate backwards over the instructions in the block. Don't count the 2194 // branch projections at end or the block header instructions. 2195 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2196 Node *n = bb->_nodes[j]; 2197 if( n->is_Proj() ) continue; // Projections handled another way 2198 2199 // Account for all uses 2200 for ( uint k = 0; k < n->len(); k++ ) { 2201 Node *inp = n->in(k); 2202 if (!inp) continue; 2203 assert(inp != n, "no cycles allowed" ); 2204 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2205 if( inp->is_Proj() ) // Skip through Proj's 2206 inp = inp->in(0); 2207 ++_uses[inp->_idx]; // Count 1 block-local use 2208 } 2209 } 2210 2211 // If this instruction has a 0 use count, then it is available 2212 if (!_uses[n->_idx]) { 2213 _current_latency[n->_idx] = _bundle_cycle_number; 2214 AddNodeToAvailableList(n); 2215 } 2216 2217#ifndef PRODUCT 2218 if (_cfg->C->trace_opto_output()) { 2219 tty->print("# uses: %3d: ", _uses[n->_idx]); 2220 n->dump(); 2221 } 2222#endif 2223 } 2224 2225#ifndef PRODUCT 2226 if (_cfg->C->trace_opto_output()) 2227 tty->print("# <- ComputeUseCount\n"); 2228#endif 2229} 2230 2231// This routine performs scheduling on each basic block in reverse order, 2232// using instruction latencies and taking into account function unit 2233// availability. 2234void Scheduling::DoScheduling() { 2235#ifndef PRODUCT 2236 if (_cfg->C->trace_opto_output()) 2237 tty->print("# -> DoScheduling\n"); 2238#endif 2239 2240 Block *succ_bb = NULL; 2241 Block *bb; 2242 2243 // Walk over all the basic blocks in reverse order 2244 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2245 bb = _cfg->_blocks[i]; 2246 2247#ifndef PRODUCT 2248 if (_cfg->C->trace_opto_output()) { 2249 tty->print("# Schedule BB#%03d (initial)\n", i); 2250 for (uint j = 0; j < bb->_nodes.size(); j++) 2251 bb->_nodes[j]->dump(); 2252 } 2253#endif 2254 2255 // On the head node, skip processing 2256 if( bb == _cfg->_broot ) 2257 continue; 2258 2259 // Skip empty, connector blocks 2260 if (bb->is_connector()) 2261 continue; 2262 2263 // If the following block is not the sole successor of 2264 // this one, then reset the pipeline information 2265 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2266#ifndef PRODUCT 2267 if (_cfg->C->trace_opto_output()) { 2268 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2269 _next_node->_idx, _bundle_instr_count); 2270 } 2271#endif 2272 step_and_clear(); 2273 } 2274 2275 // Leave untouched the starting instruction, any Phis, a CreateEx node 2276 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2277 _bb_end = bb->_nodes.size()-1; 2278 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2279 Node *n = bb->_nodes[_bb_start]; 2280 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2281 // Also, MachIdealNodes do not get scheduled 2282 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2283 MachNode *mach = n->as_Mach(); 2284 int iop = mach->ideal_Opcode(); 2285 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2286 if( iop == Op_Con ) continue; // Do not schedule Top 2287 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2288 mach->pipeline() == MachNode::pipeline_class() && 2289 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2290 continue; 2291 break; // Funny loop structure to be sure... 2292 } 2293 // Compute last "interesting" instruction in block - last instruction we 2294 // might schedule. _bb_end points just after last schedulable inst. We 2295 // normally schedule conditional branches (despite them being forced last 2296 // in the block), because they have delay slots we can fill. Calls all 2297 // have their delay slots filled in the template expansions, so we don't 2298 // bother scheduling them. 2299 Node *last = bb->_nodes[_bb_end]; 2300 if( last->is_Catch() || 2301 // Exclude unreachable path case when Halt node is in a separate block. 2302 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2303 // There must be a prior call. Skip it. 2304 while( !bb->_nodes[--_bb_end]->is_Call() ) { 2305 assert( bb->_nodes[_bb_end]->is_Proj(), "skipping projections after expected call" ); 2306 } 2307 } else if( last->is_MachNullCheck() ) { 2308 // Backup so the last null-checked memory instruction is 2309 // outside the schedulable range. Skip over the nullcheck, 2310 // projection, and the memory nodes. 2311 Node *mem = last->in(1); 2312 do { 2313 _bb_end--; 2314 } while (mem != bb->_nodes[_bb_end]); 2315 } else { 2316 // Set _bb_end to point after last schedulable inst. 2317 _bb_end++; 2318 } 2319 2320 assert( _bb_start <= _bb_end, "inverted block ends" ); 2321 2322 // Compute the register antidependencies for the basic block 2323 ComputeRegisterAntidependencies(bb); 2324 if (_cfg->C->failing()) return; // too many D-U pinch points 2325 2326 // Compute intra-bb latencies for the nodes 2327 ComputeLocalLatenciesForward(bb); 2328 2329 // Compute the usage within the block, and set the list of all nodes 2330 // in the block that have no uses within the block. 2331 ComputeUseCount(bb); 2332 2333 // Schedule the remaining instructions in the block 2334 while ( _available.size() > 0 ) { 2335 Node *n = ChooseNodeToBundle(); 2336 AddNodeToBundle(n,bb); 2337 } 2338 2339 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2340#ifdef ASSERT 2341 for( uint l = _bb_start; l < _bb_end; l++ ) { 2342 Node *n = bb->_nodes[l]; 2343 uint m; 2344 for( m = 0; m < _bb_end-_bb_start; m++ ) 2345 if( _scheduled[m] == n ) 2346 break; 2347 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2348 } 2349#endif 2350 2351 // Now copy the instructions (in reverse order) back to the block 2352 for ( uint k = _bb_start; k < _bb_end; k++ ) 2353 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2354 2355#ifndef PRODUCT 2356 if (_cfg->C->trace_opto_output()) { 2357 tty->print("# Schedule BB#%03d (final)\n", i); 2358 uint current = 0; 2359 for (uint j = 0; j < bb->_nodes.size(); j++) { 2360 Node *n = bb->_nodes[j]; 2361 if( valid_bundle_info(n) ) { 2362 Bundle *bundle = node_bundling(n); 2363 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2364 tty->print("*** Bundle: "); 2365 bundle->dump(); 2366 } 2367 n->dump(); 2368 } 2369 } 2370 } 2371#endif 2372#ifdef ASSERT 2373 verify_good_schedule(bb,"after block local scheduling"); 2374#endif 2375 } 2376 2377#ifndef PRODUCT 2378 if (_cfg->C->trace_opto_output()) 2379 tty->print("# <- DoScheduling\n"); 2380#endif 2381 2382 // Record final node-bundling array location 2383 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2384 2385} // end DoScheduling 2386 2387//------------------------------verify_good_schedule--------------------------- 2388// Verify that no live-range used in the block is killed in the block by a 2389// wrong DEF. This doesn't verify live-ranges that span blocks. 2390 2391// Check for edge existence. Used to avoid adding redundant precedence edges. 2392static bool edge_from_to( Node *from, Node *to ) { 2393 for( uint i=0; i<from->len(); i++ ) 2394 if( from->in(i) == to ) 2395 return true; 2396 return false; 2397} 2398 2399#ifdef ASSERT 2400//------------------------------verify_do_def---------------------------------- 2401void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2402 // Check for bad kills 2403 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2404 Node *prior_use = _reg_node[def]; 2405 if( prior_use && !edge_from_to(prior_use,n) ) { 2406 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2407 n->dump(); 2408 tty->print_cr("..."); 2409 prior_use->dump(); 2410 assert(edge_from_to(prior_use,n),msg); 2411 } 2412 _reg_node.map(def,NULL); // Kill live USEs 2413 } 2414} 2415 2416//------------------------------verify_good_schedule--------------------------- 2417void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2418 2419 // Zap to something reasonable for the verify code 2420 _reg_node.clear(); 2421 2422 // Walk over the block backwards. Check to make sure each DEF doesn't 2423 // kill a live value (other than the one it's supposed to). Add each 2424 // USE to the live set. 2425 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2426 Node *n = b->_nodes[i]; 2427 int n_op = n->Opcode(); 2428 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2429 // Fat-proj kills a slew of registers 2430 RegMask rm = n->out_RegMask();// Make local copy 2431 while( rm.is_NotEmpty() ) { 2432 OptoReg::Name kill = rm.find_first_elem(); 2433 rm.Remove(kill); 2434 verify_do_def( n, kill, msg ); 2435 } 2436 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2437 // Get DEF'd registers the normal way 2438 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2439 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2440 } 2441 2442 // Now make all USEs live 2443 for( uint i=1; i<n->req(); i++ ) { 2444 Node *def = n->in(i); 2445 assert(def != 0, "input edge required"); 2446 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2447 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2448 if( OptoReg::is_valid(reg_lo) ) { 2449 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2450 _reg_node.map(reg_lo,n); 2451 } 2452 if( OptoReg::is_valid(reg_hi) ) { 2453 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2454 _reg_node.map(reg_hi,n); 2455 } 2456 } 2457 2458 } 2459 2460 // Zap to something reasonable for the Antidependence code 2461 _reg_node.clear(); 2462} 2463#endif 2464 2465// Conditionally add precedence edges. Avoid putting edges on Projs. 2466static void add_prec_edge_from_to( Node *from, Node *to ) { 2467 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2468 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2469 from = from->in(0); 2470 } 2471 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2472 !edge_from_to( from, to ) ) // Avoid duplicate edge 2473 from->add_prec(to); 2474} 2475 2476//------------------------------anti_do_def------------------------------------ 2477void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2478 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2479 return; 2480 2481 Node *pinch = _reg_node[def_reg]; // Get pinch point 2482 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2483 is_def ) { // Check for a true def (not a kill) 2484 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2485 return; 2486 } 2487 2488 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2489 debug_only( def = (Node*)0xdeadbeef; ) 2490 2491 // After some number of kills there _may_ be a later def 2492 Node *later_def = NULL; 2493 2494 // Finding a kill requires a real pinch-point. 2495 // Check for not already having a pinch-point. 2496 // Pinch points are Op_Node's. 2497 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2498 later_def = pinch; // Must be def/kill as optimistic pinch-point 2499 if ( _pinch_free_list.size() > 0) { 2500 pinch = _pinch_free_list.pop(); 2501 } else { 2502 pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be 2503 } 2504 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2505 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2506 return; 2507 } 2508 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2509 _reg_node.map(def_reg,pinch); // Record pinch-point 2510 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2511 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2512 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2513 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2514 later_def = NULL; // and no later def 2515 } 2516 pinch->set_req(0,later_def); // Hook later def so we can find it 2517 } else { // Else have valid pinch point 2518 if( pinch->in(0) ) // If there is a later-def 2519 later_def = pinch->in(0); // Get it 2520 } 2521 2522 // Add output-dependence edge from later def to kill 2523 if( later_def ) // If there is some original def 2524 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2525 2526 // See if current kill is also a use, and so is forced to be the pinch-point. 2527 if( pinch->Opcode() == Op_Node ) { 2528 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2529 for( uint i=1; i<uses->req(); i++ ) { 2530 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2531 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2532 // Yes, found a use/kill pinch-point 2533 pinch->set_req(0,NULL); // 2534 pinch->replace_by(kill); // Move anti-dep edges up 2535 pinch = kill; 2536 _reg_node.map(def_reg,pinch); 2537 return; 2538 } 2539 } 2540 } 2541 2542 // Add edge from kill to pinch-point 2543 add_prec_edge_from_to(kill,pinch); 2544} 2545 2546//------------------------------anti_do_use------------------------------------ 2547void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2548 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2549 return; 2550 Node *pinch = _reg_node[use_reg]; // Get pinch point 2551 // Check for no later def_reg/kill in block 2552 if( pinch && _bbs[pinch->_idx] == b && 2553 // Use has to be block-local as well 2554 _bbs[use->_idx] == b ) { 2555 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2556 pinch->req() == 1 ) { // pinch not yet in block? 2557 pinch->del_req(0); // yank pointer to later-def, also set flag 2558 // Insert the pinch-point in the block just after the last use 2559 b->_nodes.insert(b->find_node(use)+1,pinch); 2560 _bb_end++; // Increase size scheduled region in block 2561 } 2562 2563 add_prec_edge_from_to(pinch,use); 2564 } 2565} 2566 2567//------------------------------ComputeRegisterAntidependences----------------- 2568// We insert antidependences between the reads and following write of 2569// allocated registers to prevent illegal code motion. Hopefully, the 2570// number of added references should be fairly small, especially as we 2571// are only adding references within the current basic block. 2572void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2573 2574#ifdef ASSERT 2575 verify_good_schedule(b,"before block local scheduling"); 2576#endif 2577 2578 // A valid schedule, for each register independently, is an endless cycle 2579 // of: a def, then some uses (connected to the def by true dependencies), 2580 // then some kills (defs with no uses), finally the cycle repeats with a new 2581 // def. The uses are allowed to float relative to each other, as are the 2582 // kills. No use is allowed to slide past a kill (or def). This requires 2583 // antidependencies between all uses of a single def and all kills that 2584 // follow, up to the next def. More edges are redundant, because later defs 2585 // & kills are already serialized with true or antidependencies. To keep 2586 // the edge count down, we add a 'pinch point' node if there's more than 2587 // one use or more than one kill/def. 2588 2589 // We add dependencies in one bottom-up pass. 2590 2591 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2592 2593 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2594 // register. If not, we record the DEF/KILL in _reg_node, the 2595 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2596 // "pinch point", a new Node that's in the graph but not in the block. 2597 // We put edges from the prior and current DEF/KILLs to the pinch point. 2598 // We put the pinch point in _reg_node. If there's already a pinch point 2599 // we merely add an edge from the current DEF/KILL to the pinch point. 2600 2601 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2602 // put an edge from the pinch point to the USE. 2603 2604 // To be expedient, the _reg_node array is pre-allocated for the whole 2605 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2606 // or a valid def/kill/pinch-point, or a leftover node from some prior 2607 // block. Leftover node from some prior block is treated like a NULL (no 2608 // prior def, so no anti-dependence needed). Valid def is distinguished by 2609 // it being in the current block. 2610 bool fat_proj_seen = false; 2611 uint last_safept = _bb_end-1; 2612 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2613 Node* last_safept_node = end_node; 2614 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2615 Node *n = b->_nodes[i]; 2616 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2617 if( n->Opcode() == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2618 // Fat-proj kills a slew of registers 2619 // This can add edges to 'n' and obscure whether or not it was a def, 2620 // hence the is_def flag. 2621 fat_proj_seen = true; 2622 RegMask rm = n->out_RegMask();// Make local copy 2623 while( rm.is_NotEmpty() ) { 2624 OptoReg::Name kill = rm.find_first_elem(); 2625 rm.Remove(kill); 2626 anti_do_def( b, n, kill, is_def ); 2627 } 2628 } else { 2629 // Get DEF'd registers the normal way 2630 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2631 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2632 } 2633 2634 // Check each register used by this instruction for a following DEF/KILL 2635 // that must occur afterward and requires an anti-dependence edge. 2636 for( uint j=0; j<n->req(); j++ ) { 2637 Node *def = n->in(j); 2638 if( def ) { 2639 assert( def->Opcode() != Op_MachProj || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2640 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2641 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2642 } 2643 } 2644 // Do not allow defs of new derived values to float above GC 2645 // points unless the base is definitely available at the GC point. 2646 2647 Node *m = b->_nodes[i]; 2648 2649 // Add precedence edge from following safepoint to use of derived pointer 2650 if( last_safept_node != end_node && 2651 m != last_safept_node) { 2652 for (uint k = 1; k < m->req(); k++) { 2653 const Type *t = m->in(k)->bottom_type(); 2654 if( t->isa_oop_ptr() && 2655 t->is_ptr()->offset() != 0 ) { 2656 last_safept_node->add_prec( m ); 2657 break; 2658 } 2659 } 2660 } 2661 2662 if( n->jvms() ) { // Precedence edge from derived to safept 2663 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2664 if( b->_nodes[last_safept] != last_safept_node ) { 2665 last_safept = b->find_node(last_safept_node); 2666 } 2667 for( uint j=last_safept; j > i; j-- ) { 2668 Node *mach = b->_nodes[j]; 2669 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2670 mach->add_prec( n ); 2671 } 2672 last_safept = i; 2673 last_safept_node = m; 2674 } 2675 } 2676 2677 if (fat_proj_seen) { 2678 // Garbage collect pinch nodes that were not consumed. 2679 // They are usually created by a fat kill MachProj for a call. 2680 garbage_collect_pinch_nodes(); 2681 } 2682} 2683 2684//------------------------------garbage_collect_pinch_nodes------------------------------- 2685 2686// Garbage collect pinch nodes for reuse by other blocks. 2687// 2688// The block scheduler's insertion of anti-dependence 2689// edges creates many pinch nodes when the block contains 2690// 2 or more Calls. A pinch node is used to prevent a 2691// combinatorial explosion of edges. If a set of kills for a 2692// register is anti-dependent on a set of uses (or defs), rather 2693// than adding an edge in the graph between each pair of kill 2694// and use (or def), a pinch is inserted between them: 2695// 2696// use1 use2 use3 2697// \ | / 2698// \ | / 2699// pinch 2700// / | \ 2701// / | \ 2702// kill1 kill2 kill3 2703// 2704// One pinch node is created per register killed when 2705// the second call is encountered during a backwards pass 2706// over the block. Most of these pinch nodes are never 2707// wired into the graph because the register is never 2708// used or def'ed in the block. 2709// 2710void Scheduling::garbage_collect_pinch_nodes() { 2711#ifndef PRODUCT 2712 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2713#endif 2714 int trace_cnt = 0; 2715 for (uint k = 0; k < _reg_node.Size(); k++) { 2716 Node* pinch = _reg_node[k]; 2717 if (pinch != NULL && pinch->Opcode() == Op_Node && 2718 // no predecence input edges 2719 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2720 cleanup_pinch(pinch); 2721 _pinch_free_list.push(pinch); 2722 _reg_node.map(k, NULL); 2723#ifndef PRODUCT 2724 if (_cfg->C->trace_opto_output()) { 2725 trace_cnt++; 2726 if (trace_cnt > 40) { 2727 tty->print("\n"); 2728 trace_cnt = 0; 2729 } 2730 tty->print(" %d", pinch->_idx); 2731 } 2732#endif 2733 } 2734 } 2735#ifndef PRODUCT 2736 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2737#endif 2738} 2739 2740// Clean up a pinch node for reuse. 2741void Scheduling::cleanup_pinch( Node *pinch ) { 2742 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2743 2744 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2745 Node* use = pinch->last_out(i); 2746 uint uses_found = 0; 2747 for (uint j = use->req(); j < use->len(); j++) { 2748 if (use->in(j) == pinch) { 2749 use->rm_prec(j); 2750 uses_found++; 2751 } 2752 } 2753 assert(uses_found > 0, "must be a precedence edge"); 2754 i -= uses_found; // we deleted 1 or more copies of this edge 2755 } 2756 // May have a later_def entry 2757 pinch->set_req(0, NULL); 2758} 2759 2760//------------------------------print_statistics------------------------------- 2761#ifndef PRODUCT 2762 2763void Scheduling::dump_available() const { 2764 tty->print("#Availist "); 2765 for (uint i = 0; i < _available.size(); i++) 2766 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2767 tty->cr(); 2768} 2769 2770// Print Scheduling Statistics 2771void Scheduling::print_statistics() { 2772 // Print the size added by nops for bundling 2773 tty->print("Nops added %d bytes to total of %d bytes", 2774 _total_nop_size, _total_method_size); 2775 if (_total_method_size > 0) 2776 tty->print(", for %.2f%%", 2777 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2778 tty->print("\n"); 2779 2780 // Print the number of branch shadows filled 2781 if (Pipeline::_branch_has_delay_slot) { 2782 tty->print("Of %d branches, %d had unconditional delay slots filled", 2783 _total_branches, _total_unconditional_delays); 2784 if (_total_branches > 0) 2785 tty->print(", for %.2f%%", 2786 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2787 tty->print("\n"); 2788 } 2789 2790 uint total_instructions = 0, total_bundles = 0; 2791 2792 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2793 uint bundle_count = _total_instructions_per_bundle[i]; 2794 total_instructions += bundle_count * i; 2795 total_bundles += bundle_count; 2796 } 2797 2798 if (total_bundles > 0) 2799 tty->print("Average ILP (excluding nops) is %.2f\n", 2800 ((double)total_instructions) / ((double)total_bundles)); 2801} 2802#endif 2803