1 /* 2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.inline.hpp" 27 #include "code/compiledIC.hpp" 28 #include "code/debugInfo.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/oopMap.hpp" 32 #include "memory/allocation.inline.hpp" 33 #include "opto/callnode.hpp" 34 #include "opto/cfgnode.hpp" 35 #include "opto/locknode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/output.hpp" 38 #include "opto/regalloc.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 #include "opto/type.hpp" 42 #include "runtime/handles.inline.hpp" 43 #include "utilities/xmlstream.hpp" 44 45 extern uint size_exception_handler(); 46 extern uint size_deopt_handler(); 47 48 #ifndef PRODUCT 49 #define DEBUG_ARG(x) , x 50 #else 51 #define DEBUG_ARG(x) 52 #endif 53 54 extern int emit_exception_handler(CodeBuffer &cbuf); 55 extern int emit_deopt_handler(CodeBuffer &cbuf); 56 57 //------------------------------Output----------------------------------------- 58 // Convert Nodes to instruction bits and pass off to the VM 59 void Compile::Output() { 60 // RootNode goes 61 assert( _cfg->_broot->_nodes.size() == 0, "" ); 62 63 // The number of new nodes (mostly MachNop) is proportional to 64 // the number of java calls and inner loops which are aligned. 65 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 66 C->inner_loops()*(OptoLoopAlignment-1)), 67 "out of nodes before code generation" ) ) { 68 return; 69 } 70 // Make sure I can find the Start Node 71 Block_Array& bbs = _cfg->_bbs; 72 Block *entry = _cfg->_blocks[1]; 73 Block *broot = _cfg->_broot; 74 75 const StartNode *start = entry->_nodes[0]->as_Start(); 76 77 // Replace StartNode with prolog 78 MachPrologNode *prolog = new (this) MachPrologNode(); 79 entry->_nodes.map( 0, prolog ); 80 bbs.map( prolog->_idx, entry ); 81 bbs.map( start->_idx, NULL ); // start is no longer in any block 82 83 // Virtual methods need an unverified entry point 84 85 if( is_osr_compilation() ) { 86 if( PoisonOSREntry ) { 87 // TODO: Should use a ShouldNotReachHereNode... 88 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 89 } 90 } else { 91 if( _method && !_method->flags().is_static() ) { 92 // Insert unvalidated entry point 93 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 94 } 95 96 } 97 98 99 // Break before main entry point 100 if( (_method && _method->break_at_execute()) 101 #ifndef PRODUCT 102 ||(OptoBreakpoint && is_method_compilation()) 103 ||(OptoBreakpointOSR && is_osr_compilation()) 104 ||(OptoBreakpointC2R && !_method) 105 #endif 106 ) { 107 // checking for _method means that OptoBreakpoint does not apply to 108 // runtime stubs or frame converters 109 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 110 } 111 112 // Insert epilogs before every return 113 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 114 Block *b = _cfg->_blocks[i]; 115 if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point? 116 Node *m = b->end(); 117 if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) { 118 MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 119 b->add_inst( epilog ); 120 bbs.map(epilog->_idx, b); 121 //_regalloc->set_bad(epilog->_idx); // Already initialized this way. 122 } 123 } 124 } 125 126 # ifdef ENABLE_ZAP_DEAD_LOCALS 127 if ( ZapDeadCompiledLocals ) Insert_zap_nodes(); 128 # endif 129 130 uint* blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1); 131 blk_starts[0] = 0; 132 133 // Initialize code buffer and process short branches. 134 CodeBuffer* cb = init_buffer(blk_starts); 135 136 if (cb == NULL || failing()) return; 137 138 ScheduleAndBundle(); 139 140 #ifndef PRODUCT 141 if (trace_opto_output()) { 142 tty->print("\n---- After ScheduleAndBundle ----\n"); 143 for (uint i = 0; i < _cfg->_num_blocks; i++) { 144 tty->print("\nBB#%03d:\n", i); 145 Block *bb = _cfg->_blocks[i]; 146 for (uint j = 0; j < bb->_nodes.size(); j++) { 147 Node *n = bb->_nodes[j]; 148 OptoReg::Name reg = _regalloc->get_reg_first(n); 149 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 150 n->dump(); 151 } 152 } 153 } 154 #endif 155 156 if (failing()) return; 157 158 BuildOopMaps(); 159 160 if (failing()) return; 161 162 fill_buffer(cb, blk_starts); 163 } 164 165 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 166 // Determine if we need to generate a stack overflow check. 167 // Do it if the method is not a stub function and 168 // has java calls or has frame size > vm_page_size/8. 169 return (UseStackBanging && stub_function() == NULL && 170 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); 171 } 172 173 bool Compile::need_register_stack_bang() const { 174 // Determine if we need to generate a register stack overflow check. 175 // This is only used on architectures which have split register 176 // and memory stacks (ie. IA64). 177 // Bang if the method is not a stub function and has java calls 178 return (stub_function() == NULL && has_java_calls()); 179 } 180 181 # ifdef ENABLE_ZAP_DEAD_LOCALS 182 183 184 // In order to catch compiler oop-map bugs, we have implemented 185 // a debugging mode called ZapDeadCompilerLocals. 186 // This mode causes the compiler to insert a call to a runtime routine, 187 // "zap_dead_locals", right before each place in compiled code 188 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 189 // The runtime routine checks that locations mapped as oops are really 190 // oops, that locations mapped as values do not look like oops, 191 // and that locations mapped as dead are not used later 192 // (by zapping them to an invalid address). 193 194 int Compile::_CompiledZap_count = 0; 195 196 void Compile::Insert_zap_nodes() { 197 bool skip = false; 198 199 200 // Dink with static counts because code code without the extra 201 // runtime calls is MUCH faster for debugging purposes 202 203 if ( CompileZapFirst == 0 ) ; // nothing special 204 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 205 else if ( CompileZapFirst == CompiledZap_count() ) 206 warning("starting zap compilation after skipping"); 207 208 if ( CompileZapLast == -1 ) ; // nothing special 209 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 210 else if ( CompileZapLast == CompiledZap_count() ) 211 warning("about to compile last zap"); 212 213 ++_CompiledZap_count; // counts skipped zaps, too 214 215 if ( skip ) return; 216 217 218 if ( _method == NULL ) 219 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 220 221 // Insert call to zap runtime stub before every node with an oop map 222 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 223 Block *b = _cfg->_blocks[i]; 224 for ( uint j = 0; j < b->_nodes.size(); ++j ) { 225 Node *n = b->_nodes[j]; 226 227 // Determining if we should insert a zap-a-lot node in output. 228 // We do that for all nodes that has oopmap info, except for calls 229 // to allocation. Calls to allocation passes in the old top-of-eden pointer 230 // and expect the C code to reset it. Hence, there can be no safepoints between 231 // the inlined-allocation and the call to new_Java, etc. 232 // We also cannot zap monitor calls, as they must hold the microlock 233 // during the call to Zap, which also wants to grab the microlock. 234 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 235 if ( insert ) { // it is MachSafePoint 236 if ( !n->is_MachCall() ) { 237 insert = false; 238 } else if ( n->is_MachCall() ) { 239 MachCallNode* call = n->as_MachCall(); 240 if (call->entry_point() == OptoRuntime::new_instance_Java() || 241 call->entry_point() == OptoRuntime::new_array_Java() || 242 call->entry_point() == OptoRuntime::multianewarray2_Java() || 243 call->entry_point() == OptoRuntime::multianewarray3_Java() || 244 call->entry_point() == OptoRuntime::multianewarray4_Java() || 245 call->entry_point() == OptoRuntime::multianewarray5_Java() || 246 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 247 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 248 ) { 249 insert = false; 250 } 251 } 252 if (insert) { 253 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 254 b->_nodes.insert( j, zap ); 255 _cfg->_bbs.map( zap->_idx, b ); 256 ++j; 257 } 258 } 259 } 260 } 261 } 262 263 264 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 265 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 266 CallStaticJavaNode* ideal_node = 267 new (this) CallStaticJavaNode( tf, 268 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 269 "call zap dead locals stub", 0, TypePtr::BOTTOM); 270 // We need to copy the OopMap from the site we're zapping at. 271 // We have to make a copy, because the zap site might not be 272 // a call site, and zap_dead is a call site. 273 OopMap* clone = node_to_check->oop_map()->deep_copy(); 274 275 // Add the cloned OopMap to the zap node 276 ideal_node->set_oop_map(clone); 277 return _matcher->match_sfpt(ideal_node); 278 } 279 280 //------------------------------is_node_getting_a_safepoint-------------------- 281 bool Compile::is_node_getting_a_safepoint( Node* n) { 282 // This code duplicates the logic prior to the call of add_safepoint 283 // below in this file. 284 if( n->is_MachSafePoint() ) return true; 285 return false; 286 } 287 288 # endif // ENABLE_ZAP_DEAD_LOCALS 289 290 //------------------------------compute_loop_first_inst_sizes------------------ 291 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 292 // of a loop. When aligning a loop we need to provide enough instructions 293 // in cpu's fetch buffer to feed decoders. The loop alignment could be 294 // avoided if we have enough instructions in fetch buffer at the head of a loop. 295 // By default, the size is set to 999999 by Block's constructor so that 296 // a loop will be aligned if the size is not reset here. 297 // 298 // Note: Mach instructions could contain several HW instructions 299 // so the size is estimated only. 300 // 301 void Compile::compute_loop_first_inst_sizes() { 302 // The next condition is used to gate the loop alignment optimization. 303 // Don't aligned a loop if there are enough instructions at the head of a loop 304 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 305 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 306 // equal to 11 bytes which is the largest address NOP instruction. 307 if( MaxLoopPad < OptoLoopAlignment-1 ) { 308 uint last_block = _cfg->_num_blocks-1; 309 for( uint i=1; i <= last_block; i++ ) { 310 Block *b = _cfg->_blocks[i]; 311 // Check the first loop's block which requires an alignment. 312 if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) { 313 uint sum_size = 0; 314 uint inst_cnt = NumberOfLoopInstrToAlign; 315 inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 316 317 // Check subsequent fallthrough blocks if the loop's first 318 // block(s) does not have enough instructions. 319 Block *nb = b; 320 while( inst_cnt > 0 && 321 i < last_block && 322 !_cfg->_blocks[i+1]->has_loop_alignment() && 323 !nb->has_successor(b) ) { 324 i++; 325 nb = _cfg->_blocks[i]; 326 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 327 } // while( inst_cnt > 0 && i < last_block ) 328 329 b->set_first_inst_size(sum_size); 330 } // f( b->head()->is_Loop() ) 331 } // for( i <= last_block ) 332 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 333 } 334 335 //----------------------shorten_branches--------------------------------------- 336 // The architecture description provides short branch variants for some long 337 // branch instructions. Replace eligible long branches with short branches. 338 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 339 340 // ------------------ 341 // Compute size of each block, method size, and relocation information size 342 uint nblocks = _cfg->_num_blocks; 343 344 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 345 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 346 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 347 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 348 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 349 350 bool has_short_branch_candidate = false; 351 352 // Initialize the sizes to 0 353 code_size = 0; // Size in bytes of generated code 354 stub_size = 0; // Size in bytes of all stub entries 355 // Size in bytes of all relocation entries, including those in local stubs. 356 // Start with 2-bytes of reloc info for the unvalidated entry point 357 reloc_size = 1; // Number of relocation entries 358 359 // Make three passes. The first computes pessimistic blk_starts, 360 // relative jmp_offset and reloc_size information. The second performs 361 // short branch substitution using the pessimistic sizing. The 362 // third inserts nops where needed. 363 364 // Step one, perform a pessimistic sizing pass. 365 uint last_call_adr = max_uint; 366 uint last_avoid_back_to_back_adr = max_uint; 367 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 368 for (uint i = 0; i < nblocks; i++) { // For all blocks 369 Block *b = _cfg->_blocks[i]; 370 371 // During short branch replacement, we store the relative (to blk_starts) 372 // offset of jump in jmp_offset, rather than the absolute offset of jump. 373 // This is so that we do not need to recompute sizes of all nodes when 374 // we compute correct blk_starts in our next sizing pass. 375 jmp_offset[i] = 0; 376 jmp_size[i] = 0; 377 jmp_nidx[i] = -1; 378 DEBUG_ONLY( jmp_target[i] = 0; ) 379 DEBUG_ONLY( jmp_rule[i] = 0; ) 380 381 // Sum all instruction sizes to compute block size 382 uint last_inst = b->_nodes.size(); 383 uint blk_size = 0; 384 for (uint j = 0; j < last_inst; j++) { 385 Node* nj = b->_nodes[j]; 386 // Handle machine instruction nodes 387 if (nj->is_Mach()) { 388 MachNode *mach = nj->as_Mach(); 389 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 390 reloc_size += mach->reloc(); 391 if (mach->is_MachCall()) { 392 MachCallNode *mcall = mach->as_MachCall(); 393 // This destination address is NOT PC-relative 394 395 mcall->method_set((intptr_t)mcall->entry_point()); 396 397 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 398 stub_size += CompiledStaticCall::to_interp_stub_size(); 399 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 400 } 401 } else if (mach->is_MachSafePoint()) { 402 // If call/safepoint are adjacent, account for possible 403 // nop to disambiguate the two safepoints. 404 // ScheduleAndBundle() can rearrange nodes in a block, 405 // check for all offsets inside this block. 406 if (last_call_adr >= blk_starts[i]) { 407 blk_size += nop_size; 408 } 409 } 410 if (mach->avoid_back_to_back()) { 411 // Nop is inserted between "avoid back to back" instructions. 412 // ScheduleAndBundle() can rearrange nodes in a block, 413 // check for all offsets inside this block. 414 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 415 blk_size += nop_size; 416 } 417 } 418 if (mach->may_be_short_branch()) { 419 if (!nj->is_MachBranch()) { 420 #ifndef PRODUCT 421 nj->dump(3); 422 #endif 423 Unimplemented(); 424 } 425 assert(jmp_nidx[i] == -1, "block should have only one branch"); 426 jmp_offset[i] = blk_size; 427 jmp_size[i] = nj->size(_regalloc); 428 jmp_nidx[i] = j; 429 has_short_branch_candidate = true; 430 } 431 } 432 blk_size += nj->size(_regalloc); 433 // Remember end of call offset 434 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 435 last_call_adr = blk_starts[i]+blk_size; 436 } 437 // Remember end of avoid_back_to_back offset 438 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back()) { 439 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 440 } 441 } 442 443 // When the next block starts a loop, we may insert pad NOP 444 // instructions. Since we cannot know our future alignment, 445 // assume the worst. 446 if (i< nblocks-1) { 447 Block *nb = _cfg->_blocks[i+1]; 448 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 449 if (max_loop_pad > 0) { 450 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 451 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 452 // If either is the last instruction in this block, bump by 453 // max_loop_pad in lock-step with blk_size, so sizing 454 // calculations in subsequent blocks still can conservatively 455 // detect that it may the last instruction in this block. 456 if (last_call_adr == blk_starts[i]+blk_size) { 457 last_call_adr += max_loop_pad; 458 } 459 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 460 last_avoid_back_to_back_adr += max_loop_pad; 461 } 462 blk_size += max_loop_pad; 463 } 464 } 465 466 // Save block size; update total method size 467 blk_starts[i+1] = blk_starts[i]+blk_size; 468 } 469 470 // Step two, replace eligible long jumps. 471 bool progress = true; 472 uint last_may_be_short_branch_adr = max_uint; 473 while (has_short_branch_candidate && progress) { 474 progress = false; 475 has_short_branch_candidate = false; 476 int adjust_block_start = 0; 477 for (uint i = 0; i < nblocks; i++) { 478 Block *b = _cfg->_blocks[i]; 479 int idx = jmp_nidx[i]; 480 MachNode* mach = (idx == -1) ? NULL: b->_nodes[idx]->as_Mach(); 481 if (mach != NULL && mach->may_be_short_branch()) { 482 #ifdef ASSERT 483 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 484 int j; 485 // Find the branch; ignore trailing NOPs. 486 for (j = b->_nodes.size()-1; j>=0; j--) { 487 Node* n = b->_nodes[j]; 488 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 489 break; 490 } 491 assert(j >= 0 && j == idx && b->_nodes[j] == (Node*)mach, "sanity"); 492 #endif 493 int br_size = jmp_size[i]; 494 int br_offs = blk_starts[i] + jmp_offset[i]; 495 496 // This requires the TRUE branch target be in succs[0] 497 uint bnum = b->non_connector_successor(0)->_pre_order; 498 int offset = blk_starts[bnum] - br_offs; 499 if (bnum > i) { // adjust following block's offset 500 offset -= adjust_block_start; 501 } 502 // In the following code a nop could be inserted before 503 // the branch which will increase the backward distance. 504 bool needs_padding = ((uint)br_offs == last_may_be_short_branch_adr); 505 if (needs_padding && offset <= 0) 506 offset -= nop_size; 507 508 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 509 // We've got a winner. Replace this branch. 510 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 511 512 // Update the jmp_size. 513 int new_size = replacement->size(_regalloc); 514 int diff = br_size - new_size; 515 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 516 // Conservatively take into accound padding between 517 // avoid_back_to_back branches. Previous branch could be 518 // converted into avoid_back_to_back branch during next 519 // rounds. 520 if (needs_padding && replacement->avoid_back_to_back()) { 521 jmp_offset[i] += nop_size; 522 diff -= nop_size; 523 } 524 adjust_block_start += diff; 525 b->_nodes.map(idx, replacement); 526 mach->subsume_by(replacement, C); 527 mach = replacement; 528 progress = true; 529 530 jmp_size[i] = new_size; 531 DEBUG_ONLY( jmp_target[i] = bnum; ); 532 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 533 } else { 534 // The jump distance is not short, try again during next iteration. 535 has_short_branch_candidate = true; 536 } 537 } // (mach->may_be_short_branch()) 538 if (mach != NULL && (mach->may_be_short_branch() || 539 mach->avoid_back_to_back())) { 540 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 541 } 542 blk_starts[i+1] -= adjust_block_start; 543 } 544 } 545 546 #ifdef ASSERT 547 for (uint i = 0; i < nblocks; i++) { // For all blocks 548 if (jmp_target[i] != 0) { 549 int br_size = jmp_size[i]; 550 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 551 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 552 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 553 } 554 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 555 } 556 } 557 #endif 558 559 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 560 // after ScheduleAndBundle(). 561 562 // ------------------ 563 // Compute size for code buffer 564 code_size = blk_starts[nblocks]; 565 566 // Relocation records 567 reloc_size += 1; // Relo entry for exception handler 568 569 // Adjust reloc_size to number of record of relocation info 570 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 571 // a relocation index. 572 // The CodeBuffer will expand the locs array if this estimate is too low. 573 reloc_size *= 10 / sizeof(relocInfo); 574 } 575 576 //------------------------------FillLocArray----------------------------------- 577 // Create a bit of debug info and append it to the array. The mapping is from 578 // Java local or expression stack to constant, register or stack-slot. For 579 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 580 // entry has been taken care of and caller should skip it). 581 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 582 // This should never have accepted Bad before 583 assert(OptoReg::is_valid(regnum), "location must be valid"); 584 return (OptoReg::is_reg(regnum)) 585 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 586 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 587 } 588 589 590 ObjectValue* 591 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 592 for (int i = 0; i < objs->length(); i++) { 593 assert(objs->at(i)->is_object(), "corrupt object cache"); 594 ObjectValue* sv = (ObjectValue*) objs->at(i); 595 if (sv->id() == id) { 596 return sv; 597 } 598 } 599 // Otherwise.. 600 return NULL; 601 } 602 603 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 604 ObjectValue* sv ) { 605 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 606 objs->append(sv); 607 } 608 609 610 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 611 GrowableArray<ScopeValue*> *array, 612 GrowableArray<ScopeValue*> *objs ) { 613 assert( local, "use _top instead of null" ); 614 if (array->length() != idx) { 615 assert(array->length() == idx + 1, "Unexpected array count"); 616 // Old functionality: 617 // return 618 // New functionality: 619 // Assert if the local is not top. In product mode let the new node 620 // override the old entry. 621 assert(local == top(), "LocArray collision"); 622 if (local == top()) { 623 return; 624 } 625 array->pop(); 626 } 627 const Type *t = local->bottom_type(); 628 629 // Is it a safepoint scalar object node? 630 if (local->is_SafePointScalarObject()) { 631 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 632 633 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 634 if (sv == NULL) { 635 ciKlass* cik = t->is_oopptr()->klass(); 636 assert(cik->is_instance_klass() || 637 cik->is_array_klass(), "Not supported allocation."); 638 sv = new ObjectValue(spobj->_idx, 639 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 640 Compile::set_sv_for_object_node(objs, sv); 641 642 uint first_ind = spobj->first_index(); 643 for (uint i = 0; i < spobj->n_fields(); i++) { 644 Node* fld_node = sfpt->in(first_ind+i); 645 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 646 } 647 } 648 array->append(sv); 649 return; 650 } 651 652 // Grab the register number for the local 653 OptoReg::Name regnum = _regalloc->get_reg_first(local); 654 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 655 // Record the double as two float registers. 656 // The register mask for such a value always specifies two adjacent 657 // float registers, with the lower register number even. 658 // Normally, the allocation of high and low words to these registers 659 // is irrelevant, because nearly all operations on register pairs 660 // (e.g., StoreD) treat them as a single unit. 661 // Here, we assume in addition that the words in these two registers 662 // stored "naturally" (by operations like StoreD and double stores 663 // within the interpreter) such that the lower-numbered register 664 // is written to the lower memory address. This may seem like 665 // a machine dependency, but it is not--it is a requirement on 666 // the author of the <arch>.ad file to ensure that, for every 667 // even/odd double-register pair to which a double may be allocated, 668 // the word in the even single-register is stored to the first 669 // memory word. (Note that register numbers are completely 670 // arbitrary, and are not tied to any machine-level encodings.) 671 #ifdef _LP64 672 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 673 array->append(new ConstantIntValue(0)); 674 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 675 } else if ( t->base() == Type::Long ) { 676 array->append(new ConstantIntValue(0)); 677 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 678 } else if ( t->base() == Type::RawPtr ) { 679 // jsr/ret return address which must be restored into a the full 680 // width 64-bit stack slot. 681 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 682 } 683 #else //_LP64 684 #ifdef SPARC 685 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 686 // For SPARC we have to swap high and low words for 687 // long values stored in a single-register (g0-g7). 688 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 689 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 690 } else 691 #endif //SPARC 692 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 693 // Repack the double/long as two jints. 694 // The convention the interpreter uses is that the second local 695 // holds the first raw word of the native double representation. 696 // This is actually reasonable, since locals and stack arrays 697 // grow downwards in all implementations. 698 // (If, on some machine, the interpreter's Java locals or stack 699 // were to grow upwards, the embedded doubles would be word-swapped.) 700 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 701 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 702 } 703 #endif //_LP64 704 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 705 OptoReg::is_reg(regnum) ) { 706 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 707 ? Location::float_in_dbl : Location::normal )); 708 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 709 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 710 ? Location::int_in_long : Location::normal )); 711 } else if( t->base() == Type::NarrowOop ) { 712 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 713 } else { 714 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 715 } 716 return; 717 } 718 719 // No register. It must be constant data. 720 switch (t->base()) { 721 case Type::Half: // Second half of a double 722 ShouldNotReachHere(); // Caller should skip 2nd halves 723 break; 724 case Type::AnyPtr: 725 array->append(new ConstantOopWriteValue(NULL)); 726 break; 727 case Type::AryPtr: 728 case Type::InstPtr: // fall through 729 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 730 break; 731 case Type::NarrowOop: 732 if (t == TypeNarrowOop::NULL_PTR) { 733 array->append(new ConstantOopWriteValue(NULL)); 734 } else { 735 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 736 } 737 break; 738 case Type::Int: 739 array->append(new ConstantIntValue(t->is_int()->get_con())); 740 break; 741 case Type::RawPtr: 742 // A return address (T_ADDRESS). 743 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 744 #ifdef _LP64 745 // Must be restored to the full-width 64-bit stack slot. 746 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 747 #else 748 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 749 #endif 750 break; 751 case Type::FloatCon: { 752 float f = t->is_float_constant()->getf(); 753 array->append(new ConstantIntValue(jint_cast(f))); 754 break; 755 } 756 case Type::DoubleCon: { 757 jdouble d = t->is_double_constant()->getd(); 758 #ifdef _LP64 759 array->append(new ConstantIntValue(0)); 760 array->append(new ConstantDoubleValue(d)); 761 #else 762 // Repack the double as two jints. 763 // The convention the interpreter uses is that the second local 764 // holds the first raw word of the native double representation. 765 // This is actually reasonable, since locals and stack arrays 766 // grow downwards in all implementations. 767 // (If, on some machine, the interpreter's Java locals or stack 768 // were to grow upwards, the embedded doubles would be word-swapped.) 769 jint *dp = (jint*)&d; 770 array->append(new ConstantIntValue(dp[1])); 771 array->append(new ConstantIntValue(dp[0])); 772 #endif 773 break; 774 } 775 case Type::Long: { 776 jlong d = t->is_long()->get_con(); 777 #ifdef _LP64 778 array->append(new ConstantIntValue(0)); 779 array->append(new ConstantLongValue(d)); 780 #else 781 // Repack the long as two jints. 782 // The convention the interpreter uses is that the second local 783 // holds the first raw word of the native double representation. 784 // This is actually reasonable, since locals and stack arrays 785 // grow downwards in all implementations. 786 // (If, on some machine, the interpreter's Java locals or stack 787 // were to grow upwards, the embedded doubles would be word-swapped.) 788 jint *dp = (jint*)&d; 789 array->append(new ConstantIntValue(dp[1])); 790 array->append(new ConstantIntValue(dp[0])); 791 #endif 792 break; 793 } 794 case Type::Top: // Add an illegal value here 795 array->append(new LocationValue(Location())); 796 break; 797 default: 798 ShouldNotReachHere(); 799 break; 800 } 801 } 802 803 // Determine if this node starts a bundle 804 bool Compile::starts_bundle(const Node *n) const { 805 return (_node_bundling_limit > n->_idx && 806 _node_bundling_base[n->_idx].starts_bundle()); 807 } 808 809 //--------------------------Process_OopMap_Node-------------------------------- 810 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 811 812 // Handle special safepoint nodes for synchronization 813 MachSafePointNode *sfn = mach->as_MachSafePoint(); 814 MachCallNode *mcall; 815 816 #ifdef ENABLE_ZAP_DEAD_LOCALS 817 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 818 #endif 819 820 int safepoint_pc_offset = current_offset; 821 bool is_method_handle_invoke = false; 822 bool return_oop = false; 823 824 // Add the safepoint in the DebugInfoRecorder 825 if( !mach->is_MachCall() ) { 826 mcall = NULL; 827 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 828 } else { 829 mcall = mach->as_MachCall(); 830 831 // Is the call a MethodHandle call? 832 if (mcall->is_MachCallJava()) { 833 if (mcall->as_MachCallJava()->_method_handle_invoke) { 834 assert(has_method_handle_invokes(), "must have been set during call generation"); 835 is_method_handle_invoke = true; 836 } 837 } 838 839 // Check if a call returns an object. 840 if (mcall->return_value_is_used() && 841 mcall->tf()->range()->field_at(TypeFunc::Parms)->isa_ptr()) { 842 return_oop = true; 843 } 844 safepoint_pc_offset += mcall->ret_addr_offset(); 845 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 846 } 847 848 // Loop over the JVMState list to add scope information 849 // Do not skip safepoints with a NULL method, they need monitor info 850 JVMState* youngest_jvms = sfn->jvms(); 851 int max_depth = youngest_jvms->depth(); 852 853 // Allocate the object pool for scalar-replaced objects -- the map from 854 // small-integer keys (which can be recorded in the local and ostack 855 // arrays) to descriptions of the object state. 856 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 857 858 // Visit scopes from oldest to youngest. 859 for (int depth = 1; depth <= max_depth; depth++) { 860 JVMState* jvms = youngest_jvms->of_depth(depth); 861 int idx; 862 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 863 // Safepoints that do not have method() set only provide oop-map and monitor info 864 // to support GC; these do not support deoptimization. 865 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 866 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 867 int num_mon = jvms->nof_monitors(); 868 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 869 "JVMS local count must match that of the method"); 870 871 // Add Local and Expression Stack Information 872 873 // Insert locals into the locarray 874 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 875 for( idx = 0; idx < num_locs; idx++ ) { 876 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 877 } 878 879 // Insert expression stack entries into the exparray 880 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 881 for( idx = 0; idx < num_exps; idx++ ) { 882 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 883 } 884 885 // Add in mappings of the monitors 886 assert( !method || 887 !method->is_synchronized() || 888 method->is_native() || 889 num_mon > 0 || 890 !GenerateSynchronizationCode, 891 "monitors must always exist for synchronized methods"); 892 893 // Build the growable array of ScopeValues for exp stack 894 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 895 896 // Loop over monitors and insert into array 897 for(idx = 0; idx < num_mon; idx++) { 898 // Grab the node that defines this monitor 899 Node* box_node = sfn->monitor_box(jvms, idx); 900 Node* obj_node = sfn->monitor_obj(jvms, idx); 901 902 // Create ScopeValue for object 903 ScopeValue *scval = NULL; 904 905 if( obj_node->is_SafePointScalarObject() ) { 906 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 907 scval = Compile::sv_for_node_id(objs, spobj->_idx); 908 if (scval == NULL) { 909 const Type *t = obj_node->bottom_type(); 910 ciKlass* cik = t->is_oopptr()->klass(); 911 assert(cik->is_instance_klass() || 912 cik->is_array_klass(), "Not supported allocation."); 913 ObjectValue* sv = new ObjectValue(spobj->_idx, 914 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 915 Compile::set_sv_for_object_node(objs, sv); 916 917 uint first_ind = spobj->first_index(); 918 for (uint i = 0; i < spobj->n_fields(); i++) { 919 Node* fld_node = sfn->in(first_ind+i); 920 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 921 } 922 scval = sv; 923 } 924 } else if( !obj_node->is_Con() ) { 925 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 926 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 927 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 928 } else { 929 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 930 } 931 } else { 932 const TypePtr *tp = obj_node->bottom_type()->make_ptr(); 933 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 934 } 935 936 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 937 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 938 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 939 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 940 } 941 942 // We dump the object pool first, since deoptimization reads it in first. 943 debug_info()->dump_object_pool(objs); 944 945 // Build first class objects to pass to scope 946 DebugToken *locvals = debug_info()->create_scope_values(locarray); 947 DebugToken *expvals = debug_info()->create_scope_values(exparray); 948 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 949 950 // Make method available for all Safepoints 951 ciMethod* scope_method = method ? method : _method; 952 // Describe the scope here 953 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 954 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 955 // Now we can describe the scope. 956 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 957 } // End jvms loop 958 959 // Mark the end of the scope set. 960 debug_info()->end_safepoint(safepoint_pc_offset); 961 } 962 963 964 965 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 966 class NonSafepointEmitter { 967 Compile* C; 968 JVMState* _pending_jvms; 969 int _pending_offset; 970 971 void emit_non_safepoint(); 972 973 public: 974 NonSafepointEmitter(Compile* compile) { 975 this->C = compile; 976 _pending_jvms = NULL; 977 _pending_offset = 0; 978 } 979 980 void observe_instruction(Node* n, int pc_offset) { 981 if (!C->debug_info()->recording_non_safepoints()) return; 982 983 Node_Notes* nn = C->node_notes_at(n->_idx); 984 if (nn == NULL || nn->jvms() == NULL) return; 985 if (_pending_jvms != NULL && 986 _pending_jvms->same_calls_as(nn->jvms())) { 987 // Repeated JVMS? Stretch it up here. 988 _pending_offset = pc_offset; 989 } else { 990 if (_pending_jvms != NULL && 991 _pending_offset < pc_offset) { 992 emit_non_safepoint(); 993 } 994 _pending_jvms = NULL; 995 if (pc_offset > C->debug_info()->last_pc_offset()) { 996 // This is the only way _pending_jvms can become non-NULL: 997 _pending_jvms = nn->jvms(); 998 _pending_offset = pc_offset; 999 } 1000 } 1001 } 1002 1003 // Stay out of the way of real safepoints: 1004 void observe_safepoint(JVMState* jvms, int pc_offset) { 1005 if (_pending_jvms != NULL && 1006 !_pending_jvms->same_calls_as(jvms) && 1007 _pending_offset < pc_offset) { 1008 emit_non_safepoint(); 1009 } 1010 _pending_jvms = NULL; 1011 } 1012 1013 void flush_at_end() { 1014 if (_pending_jvms != NULL) { 1015 emit_non_safepoint(); 1016 } 1017 _pending_jvms = NULL; 1018 } 1019 }; 1020 1021 void NonSafepointEmitter::emit_non_safepoint() { 1022 JVMState* youngest_jvms = _pending_jvms; 1023 int pc_offset = _pending_offset; 1024 1025 // Clear it now: 1026 _pending_jvms = NULL; 1027 1028 DebugInformationRecorder* debug_info = C->debug_info(); 1029 assert(debug_info->recording_non_safepoints(), "sanity"); 1030 1031 debug_info->add_non_safepoint(pc_offset); 1032 int max_depth = youngest_jvms->depth(); 1033 1034 // Visit scopes from oldest to youngest. 1035 for (int depth = 1; depth <= max_depth; depth++) { 1036 JVMState* jvms = youngest_jvms->of_depth(depth); 1037 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1038 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1039 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1040 } 1041 1042 // Mark the end of the scope set. 1043 debug_info->end_non_safepoint(pc_offset); 1044 } 1045 1046 1047 1048 // helper for fill_buffer bailout logic 1049 static void turn_off_compiler(Compile* C) { 1050 if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) { 1051 // Do not turn off compilation if a single giant method has 1052 // blown the code cache size. 1053 C->record_failure("excessive request to CodeCache"); 1054 } else { 1055 // Let CompilerBroker disable further compilations. 1056 C->record_failure("CodeCache is full"); 1057 } 1058 } 1059 1060 1061 //------------------------------init_buffer------------------------------------ 1062 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 1063 1064 // Set the initially allocated size 1065 int code_req = initial_code_capacity; 1066 int locs_req = initial_locs_capacity; 1067 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1068 int const_req = initial_const_capacity; 1069 1070 int pad_req = NativeCall::instruction_size; 1071 // The extra spacing after the code is necessary on some platforms. 1072 // Sometimes we need to patch in a jump after the last instruction, 1073 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1074 1075 // Compute the byte offset where we can store the deopt pc. 1076 if (fixed_slots() != 0) { 1077 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1078 } 1079 1080 // Compute prolog code size 1081 _method_size = 0; 1082 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1083 #ifdef IA64 1084 if (save_argument_registers()) { 1085 // 4815101: this is a stub with implicit and unknown precision fp args. 1086 // The usual spill mechanism can only generate stfd's in this case, which 1087 // doesn't work if the fp reg to spill contains a single-precision denorm. 1088 // Instead, we hack around the normal spill mechanism using stfspill's and 1089 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1090 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1091 // 1092 // If we ever implement 16-byte 'registers' == stack slots, we can 1093 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1094 // instead of stfd/stfs/ldfd/ldfs. 1095 _frame_slots += 8*(16/BytesPerInt); 1096 } 1097 #endif 1098 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1099 1100 if (has_mach_constant_base_node()) { 1101 // Fill the constant table. 1102 // Note: This must happen before shorten_branches. 1103 for (uint i = 0; i < _cfg->_num_blocks; i++) { 1104 Block* b = _cfg->_blocks[i]; 1105 1106 for (uint j = 0; j < b->_nodes.size(); j++) { 1107 Node* n = b->_nodes[j]; 1108 1109 // If the node is a MachConstantNode evaluate the constant 1110 // value section. 1111 if (n->is_MachConstant()) { 1112 MachConstantNode* machcon = n->as_MachConstant(); 1113 machcon->eval_constant(C); 1114 } 1115 } 1116 } 1117 1118 // Calculate the offsets of the constants and the size of the 1119 // constant table (including the padding to the next section). 1120 constant_table().calculate_offsets_and_size(); 1121 const_req = constant_table().size(); 1122 } 1123 1124 // Initialize the space for the BufferBlob used to find and verify 1125 // instruction size in MachNode::emit_size() 1126 init_scratch_buffer_blob(const_req); 1127 if (failing()) return NULL; // Out of memory 1128 1129 // Pre-compute the length of blocks and replace 1130 // long branches with short if machine supports it. 1131 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1132 1133 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1134 int exception_handler_req = size_exception_handler(); 1135 int deopt_handler_req = size_deopt_handler(); 1136 exception_handler_req += MAX_stubs_size; // add marginal slop for handler 1137 deopt_handler_req += MAX_stubs_size; // add marginal slop for handler 1138 stub_req += MAX_stubs_size; // ensure per-stub margin 1139 code_req += MAX_inst_size; // ensure per-instruction margin 1140 1141 if (StressCodeBuffers) 1142 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1143 1144 int total_req = 1145 const_req + 1146 code_req + 1147 pad_req + 1148 stub_req + 1149 exception_handler_req + 1150 deopt_handler_req; // deopt handler 1151 1152 if (has_method_handle_invokes()) 1153 total_req += deopt_handler_req; // deopt MH handler 1154 1155 CodeBuffer* cb = code_buffer(); 1156 cb->initialize(total_req, locs_req); 1157 1158 // Have we run out of code space? 1159 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1160 turn_off_compiler(this); 1161 return NULL; 1162 } 1163 // Configure the code buffer. 1164 cb->initialize_consts_size(const_req); 1165 cb->initialize_stubs_size(stub_req); 1166 cb->initialize_oop_recorder(env()->oop_recorder()); 1167 1168 // fill in the nop array for bundling computations 1169 MachNode *_nop_list[Bundle::_nop_count]; 1170 Bundle::initialize_nops(_nop_list, this); 1171 1172 return cb; 1173 } 1174 1175 //------------------------------fill_buffer------------------------------------ 1176 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1177 // blk_starts[] contains offsets calculated during short branches processing, 1178 // offsets should not be increased during following steps. 1179 1180 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1181 // of a loop. It is used to determine the padding for loop alignment. 1182 compute_loop_first_inst_sizes(); 1183 1184 // Create oopmap set. 1185 _oop_map_set = new OopMapSet(); 1186 1187 // !!!!! This preserves old handling of oopmaps for now 1188 debug_info()->set_oopmaps(_oop_map_set); 1189 1190 uint nblocks = _cfg->_num_blocks; 1191 // Count and start of implicit null check instructions 1192 uint inct_cnt = 0; 1193 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1194 1195 // Count and start of calls 1196 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1197 1198 uint return_offset = 0; 1199 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1200 1201 int previous_offset = 0; 1202 int current_offset = 0; 1203 int last_call_offset = -1; 1204 int last_avoid_back_to_back_offset = -1; 1205 #ifdef ASSERT 1206 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1207 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1208 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1209 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1210 #endif 1211 1212 // Create an array of unused labels, one for each basic block, if printing is enabled 1213 #ifndef PRODUCT 1214 int *node_offsets = NULL; 1215 uint node_offset_limit = unique(); 1216 1217 if (print_assembly()) 1218 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1219 #endif 1220 1221 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1222 1223 // Emit the constant table. 1224 if (has_mach_constant_base_node()) { 1225 constant_table().emit(*cb); 1226 } 1227 1228 // Create an array of labels, one for each basic block 1229 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1230 for (uint i=0; i <= nblocks; i++) { 1231 blk_labels[i].init(); 1232 } 1233 1234 // ------------------ 1235 // Now fill in the code buffer 1236 Node *delay_slot = NULL; 1237 1238 for (uint i=0; i < nblocks; i++) { 1239 Block *b = _cfg->_blocks[i]; 1240 1241 Node *head = b->head(); 1242 1243 // If this block needs to start aligned (i.e, can be reached other 1244 // than by falling-thru from the previous block), then force the 1245 // start of a new bundle. 1246 if (Pipeline::requires_bundling() && starts_bundle(head)) 1247 cb->flush_bundle(true); 1248 1249 #ifdef ASSERT 1250 if (!b->is_connector()) { 1251 stringStream st; 1252 b->dump_head(&_cfg->_bbs, &st); 1253 MacroAssembler(cb).block_comment(st.as_string()); 1254 } 1255 jmp_target[i] = 0; 1256 jmp_offset[i] = 0; 1257 jmp_size[i] = 0; 1258 jmp_rule[i] = 0; 1259 #endif 1260 int blk_offset = current_offset; 1261 1262 // Define the label at the beginning of the basic block 1263 MacroAssembler(cb).bind(blk_labels[b->_pre_order]); 1264 1265 uint last_inst = b->_nodes.size(); 1266 1267 // Emit block normally, except for last instruction. 1268 // Emit means "dump code bits into code buffer". 1269 for (uint j = 0; j<last_inst; j++) { 1270 1271 // Get the node 1272 Node* n = b->_nodes[j]; 1273 1274 // See if delay slots are supported 1275 if (valid_bundle_info(n) && 1276 node_bundling(n)->used_in_unconditional_delay()) { 1277 assert(delay_slot == NULL, "no use of delay slot node"); 1278 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1279 1280 delay_slot = n; 1281 continue; 1282 } 1283 1284 // If this starts a new instruction group, then flush the current one 1285 // (but allow split bundles) 1286 if (Pipeline::requires_bundling() && starts_bundle(n)) 1287 cb->flush_bundle(false); 1288 1289 // The following logic is duplicated in the code ifdeffed for 1290 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1291 // should be factored out. Or maybe dispersed to the nodes? 1292 1293 // Special handling for SafePoint/Call Nodes 1294 bool is_mcall = false; 1295 if (n->is_Mach()) { 1296 MachNode *mach = n->as_Mach(); 1297 is_mcall = n->is_MachCall(); 1298 bool is_sfn = n->is_MachSafePoint(); 1299 1300 // If this requires all previous instructions be flushed, then do so 1301 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1302 cb->flush_bundle(true); 1303 current_offset = cb->insts_size(); 1304 } 1305 1306 // A padding may be needed again since a previous instruction 1307 // could be moved to delay slot. 1308 1309 // align the instruction if necessary 1310 int padding = mach->compute_padding(current_offset); 1311 // Make sure safepoint node for polling is distinct from a call's 1312 // return by adding a nop if needed. 1313 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1314 padding = nop_size; 1315 } 1316 if (padding == 0 && mach->avoid_back_to_back() && 1317 current_offset == last_avoid_back_to_back_offset) { 1318 // Avoid back to back some instructions. 1319 padding = nop_size; 1320 } 1321 1322 if(padding > 0) { 1323 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1324 int nops_cnt = padding / nop_size; 1325 MachNode *nop = new (this) MachNopNode(nops_cnt); 1326 b->_nodes.insert(j++, nop); 1327 last_inst++; 1328 _cfg->_bbs.map( nop->_idx, b ); 1329 nop->emit(*cb, _regalloc); 1330 cb->flush_bundle(true); 1331 current_offset = cb->insts_size(); 1332 } 1333 1334 // Remember the start of the last call in a basic block 1335 if (is_mcall) { 1336 MachCallNode *mcall = mach->as_MachCall(); 1337 1338 // This destination address is NOT PC-relative 1339 mcall->method_set((intptr_t)mcall->entry_point()); 1340 1341 // Save the return address 1342 call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); 1343 1344 if (mcall->is_MachCallLeaf()) { 1345 is_mcall = false; 1346 is_sfn = false; 1347 } 1348 } 1349 1350 // sfn will be valid whenever mcall is valid now because of inheritance 1351 if (is_sfn || is_mcall) { 1352 1353 // Handle special safepoint nodes for synchronization 1354 if (!is_mcall) { 1355 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1356 // !!!!! Stubs only need an oopmap right now, so bail out 1357 if (sfn->jvms()->method() == NULL) { 1358 // Write the oopmap directly to the code blob??!! 1359 # ifdef ENABLE_ZAP_DEAD_LOCALS 1360 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1361 # endif 1362 continue; 1363 } 1364 } // End synchronization 1365 1366 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1367 current_offset); 1368 Process_OopMap_Node(mach, current_offset); 1369 } // End if safepoint 1370 1371 // If this is a null check, then add the start of the previous instruction to the list 1372 else if( mach->is_MachNullCheck() ) { 1373 inct_starts[inct_cnt++] = previous_offset; 1374 } 1375 1376 // If this is a branch, then fill in the label with the target BB's label 1377 else if (mach->is_MachBranch()) { 1378 // This requires the TRUE branch target be in succs[0] 1379 uint block_num = b->non_connector_successor(0)->_pre_order; 1380 1381 // Try to replace long branch if delay slot is not used, 1382 // it is mostly for back branches since forward branch's 1383 // distance is not updated yet. 1384 bool delay_slot_is_used = valid_bundle_info(n) && 1385 node_bundling(n)->use_unconditional_delay(); 1386 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1387 assert(delay_slot == NULL, "not expecting delay slot node"); 1388 int br_size = n->size(_regalloc); 1389 int offset = blk_starts[block_num] - current_offset; 1390 if (block_num >= i) { 1391 // Current and following block's offset are not 1392 // finilized yet, adjust distance by the difference 1393 // between calculated and final offsets of current block. 1394 offset -= (blk_starts[i] - blk_offset); 1395 } 1396 // In the following code a nop could be inserted before 1397 // the branch which will increase the backward distance. 1398 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1399 if (needs_padding && offset <= 0) 1400 offset -= nop_size; 1401 1402 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1403 // We've got a winner. Replace this branch. 1404 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 1405 1406 // Update the jmp_size. 1407 int new_size = replacement->size(_regalloc); 1408 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1409 // Insert padding between avoid_back_to_back branches. 1410 if (needs_padding && replacement->avoid_back_to_back()) { 1411 MachNode *nop = new (this) MachNopNode(); 1412 b->_nodes.insert(j++, nop); 1413 _cfg->_bbs.map(nop->_idx, b); 1414 last_inst++; 1415 nop->emit(*cb, _regalloc); 1416 cb->flush_bundle(true); 1417 current_offset = cb->insts_size(); 1418 } 1419 #ifdef ASSERT 1420 jmp_target[i] = block_num; 1421 jmp_offset[i] = current_offset - blk_offset; 1422 jmp_size[i] = new_size; 1423 jmp_rule[i] = mach->rule(); 1424 #endif 1425 b->_nodes.map(j, replacement); 1426 mach->subsume_by(replacement, C); 1427 n = replacement; 1428 mach = replacement; 1429 } 1430 } 1431 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1432 } else if (mach->ideal_Opcode() == Op_Jump) { 1433 for (uint h = 0; h < b->_num_succs; h++) { 1434 Block* succs_block = b->_succs[h]; 1435 for (uint j = 1; j < succs_block->num_preds(); j++) { 1436 Node* jpn = succs_block->pred(j); 1437 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1438 uint block_num = succs_block->non_connector()->_pre_order; 1439 Label *blkLabel = &blk_labels[block_num]; 1440 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1441 } 1442 } 1443 } 1444 } 1445 1446 #ifdef ASSERT 1447 // Check that oop-store precedes the card-mark 1448 else if (mach->ideal_Opcode() == Op_StoreCM) { 1449 uint storeCM_idx = j; 1450 int count = 0; 1451 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1452 Node *oop_store = mach->in(prec); // Precedence edge 1453 if (oop_store == NULL) continue; 1454 count++; 1455 uint i4; 1456 for( i4 = 0; i4 < last_inst; ++i4 ) { 1457 if( b->_nodes[i4] == oop_store ) break; 1458 } 1459 // Note: This test can provide a false failure if other precedence 1460 // edges have been added to the storeCMNode. 1461 assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1462 } 1463 assert(count > 0, "storeCM expects at least one precedence edge"); 1464 } 1465 #endif 1466 1467 else if (!n->is_Proj()) { 1468 // Remember the beginning of the previous instruction, in case 1469 // it's followed by a flag-kill and a null-check. Happens on 1470 // Intel all the time, with add-to-memory kind of opcodes. 1471 previous_offset = current_offset; 1472 } 1473 } 1474 1475 // Verify that there is sufficient space remaining 1476 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1477 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1478 turn_off_compiler(this); 1479 return; 1480 } 1481 1482 // Save the offset for the listing 1483 #ifndef PRODUCT 1484 if (node_offsets && n->_idx < node_offset_limit) 1485 node_offsets[n->_idx] = cb->insts_size(); 1486 #endif 1487 1488 // "Normal" instruction case 1489 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1490 n->emit(*cb, _regalloc); 1491 current_offset = cb->insts_size(); 1492 1493 #ifdef ASSERT 1494 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1495 n->dump(); 1496 assert(false, "wrong size of mach node"); 1497 } 1498 #endif 1499 non_safepoints.observe_instruction(n, current_offset); 1500 1501 // mcall is last "call" that can be a safepoint 1502 // record it so we can see if a poll will directly follow it 1503 // in which case we'll need a pad to make the PcDesc sites unique 1504 // see 5010568. This can be slightly inaccurate but conservative 1505 // in the case that return address is not actually at current_offset. 1506 // This is a small price to pay. 1507 1508 if (is_mcall) { 1509 last_call_offset = current_offset; 1510 } 1511 1512 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back()) { 1513 // Avoid back to back some instructions. 1514 last_avoid_back_to_back_offset = current_offset; 1515 } 1516 1517 // See if this instruction has a delay slot 1518 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1519 assert(delay_slot != NULL, "expecting delay slot node"); 1520 1521 // Back up 1 instruction 1522 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1523 1524 // Save the offset for the listing 1525 #ifndef PRODUCT 1526 if (node_offsets && delay_slot->_idx < node_offset_limit) 1527 node_offsets[delay_slot->_idx] = cb->insts_size(); 1528 #endif 1529 1530 // Support a SafePoint in the delay slot 1531 if (delay_slot->is_MachSafePoint()) { 1532 MachNode *mach = delay_slot->as_Mach(); 1533 // !!!!! Stubs only need an oopmap right now, so bail out 1534 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1535 // Write the oopmap directly to the code blob??!! 1536 # ifdef ENABLE_ZAP_DEAD_LOCALS 1537 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1538 # endif 1539 delay_slot = NULL; 1540 continue; 1541 } 1542 1543 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1544 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1545 adjusted_offset); 1546 // Generate an OopMap entry 1547 Process_OopMap_Node(mach, adjusted_offset); 1548 } 1549 1550 // Insert the delay slot instruction 1551 delay_slot->emit(*cb, _regalloc); 1552 1553 // Don't reuse it 1554 delay_slot = NULL; 1555 } 1556 1557 } // End for all instructions in block 1558 1559 // If the next block is the top of a loop, pad this block out to align 1560 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1561 if (i < nblocks-1) { 1562 Block *nb = _cfg->_blocks[i+1]; 1563 int padding = nb->alignment_padding(current_offset); 1564 if( padding > 0 ) { 1565 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1566 b->_nodes.insert( b->_nodes.size(), nop ); 1567 _cfg->_bbs.map( nop->_idx, b ); 1568 nop->emit(*cb, _regalloc); 1569 current_offset = cb->insts_size(); 1570 } 1571 } 1572 // Verify that the distance for generated before forward 1573 // short branches is still valid. 1574 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1575 1576 // Save new block start offset 1577 blk_starts[i] = blk_offset; 1578 } // End of for all blocks 1579 blk_starts[nblocks] = current_offset; 1580 1581 non_safepoints.flush_at_end(); 1582 1583 // Offset too large? 1584 if (failing()) return; 1585 1586 // Define a pseudo-label at the end of the code 1587 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1588 1589 // Compute the size of the first block 1590 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1591 1592 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1593 1594 #ifdef ASSERT 1595 for (uint i = 0; i < nblocks; i++) { // For all blocks 1596 if (jmp_target[i] != 0) { 1597 int br_size = jmp_size[i]; 1598 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1599 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1600 tty->print_cr("target (%d) - jmp_offset(%d) = offset (%d), jump_size(%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_offset[i], offset, br_size, i, jmp_target[i]); 1601 assert(false, "Displacement too large for short jmp"); 1602 } 1603 } 1604 } 1605 #endif 1606 1607 // ------------------ 1608 1609 #ifndef PRODUCT 1610 // Information on the size of the method, without the extraneous code 1611 Scheduling::increment_method_size(cb->insts_size()); 1612 #endif 1613 1614 // ------------------ 1615 // Fill in exception table entries. 1616 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1617 1618 // Only java methods have exception handlers and deopt handlers 1619 if (_method) { 1620 // Emit the exception handler code. 1621 _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); 1622 // Emit the deopt handler code. 1623 _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); 1624 1625 // Emit the MethodHandle deopt handler code (if required). 1626 if (has_method_handle_invokes()) { 1627 // We can use the same code as for the normal deopt handler, we 1628 // just need a different entry point address. 1629 _code_offsets.set_value(CodeOffsets::DeoptMH, emit_deopt_handler(*cb)); 1630 } 1631 } 1632 1633 // One last check for failed CodeBuffer::expand: 1634 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1635 turn_off_compiler(this); 1636 return; 1637 } 1638 1639 #ifndef PRODUCT 1640 // Dump the assembly code, including basic-block numbers 1641 if (print_assembly()) { 1642 ttyLocker ttyl; // keep the following output all in one block 1643 if (!VMThread::should_terminate()) { // test this under the tty lock 1644 // This output goes directly to the tty, not the compiler log. 1645 // To enable tools to match it up with the compilation activity, 1646 // be sure to tag this tty output with the compile ID. 1647 if (xtty != NULL) { 1648 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1649 is_osr_compilation() ? " compile_kind='osr'" : 1650 ""); 1651 } 1652 if (method() != NULL) { 1653 method()->print_metadata(); 1654 } 1655 dump_asm(node_offsets, node_offset_limit); 1656 if (xtty != NULL) { 1657 xtty->tail("opto_assembly"); 1658 } 1659 } 1660 } 1661 #endif 1662 1663 } 1664 1665 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1666 _inc_table.set_size(cnt); 1667 1668 uint inct_cnt = 0; 1669 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1670 Block *b = _cfg->_blocks[i]; 1671 Node *n = NULL; 1672 int j; 1673 1674 // Find the branch; ignore trailing NOPs. 1675 for( j = b->_nodes.size()-1; j>=0; j-- ) { 1676 n = b->_nodes[j]; 1677 if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) 1678 break; 1679 } 1680 1681 // If we didn't find anything, continue 1682 if( j < 0 ) continue; 1683 1684 // Compute ExceptionHandlerTable subtable entry and add it 1685 // (skip empty blocks) 1686 if( n->is_Catch() ) { 1687 1688 // Get the offset of the return from the call 1689 uint call_return = call_returns[b->_pre_order]; 1690 #ifdef ASSERT 1691 assert( call_return > 0, "no call seen for this basic block" ); 1692 while( b->_nodes[--j]->is_MachProj() ) ; 1693 assert( b->_nodes[j]->is_MachCall(), "CatchProj must follow call" ); 1694 #endif 1695 // last instruction is a CatchNode, find it's CatchProjNodes 1696 int nof_succs = b->_num_succs; 1697 // allocate space 1698 GrowableArray<intptr_t> handler_bcis(nof_succs); 1699 GrowableArray<intptr_t> handler_pcos(nof_succs); 1700 // iterate through all successors 1701 for (int j = 0; j < nof_succs; j++) { 1702 Block* s = b->_succs[j]; 1703 bool found_p = false; 1704 for( uint k = 1; k < s->num_preds(); k++ ) { 1705 Node *pk = s->pred(k); 1706 if( pk->is_CatchProj() && pk->in(0) == n ) { 1707 const CatchProjNode* p = pk->as_CatchProj(); 1708 found_p = true; 1709 // add the corresponding handler bci & pco information 1710 if( p->_con != CatchProjNode::fall_through_index ) { 1711 // p leads to an exception handler (and is not fall through) 1712 assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); 1713 // no duplicates, please 1714 if( !handler_bcis.contains(p->handler_bci()) ) { 1715 uint block_num = s->non_connector()->_pre_order; 1716 handler_bcis.append(p->handler_bci()); 1717 handler_pcos.append(blk_labels[block_num].loc_pos()); 1718 } 1719 } 1720 } 1721 } 1722 assert(found_p, "no matching predecessor found"); 1723 // Note: Due to empty block removal, one block may have 1724 // several CatchProj inputs, from the same Catch. 1725 } 1726 1727 // Set the offset of the return from the call 1728 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1729 continue; 1730 } 1731 1732 // Handle implicit null exception table updates 1733 if( n->is_MachNullCheck() ) { 1734 uint block_num = b->non_connector_successor(0)->_pre_order; 1735 _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); 1736 continue; 1737 } 1738 } // End of for all blocks fill in exception table entries 1739 } 1740 1741 // Static Variables 1742 #ifndef PRODUCT 1743 uint Scheduling::_total_nop_size = 0; 1744 uint Scheduling::_total_method_size = 0; 1745 uint Scheduling::_total_branches = 0; 1746 uint Scheduling::_total_unconditional_delays = 0; 1747 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1748 #endif 1749 1750 // Initializer for class Scheduling 1751 1752 Scheduling::Scheduling(Arena *arena, Compile &compile) 1753 : _arena(arena), 1754 _cfg(compile.cfg()), 1755 _bbs(compile.cfg()->_bbs), 1756 _regalloc(compile.regalloc()), 1757 _reg_node(arena), 1758 _bundle_instr_count(0), 1759 _bundle_cycle_number(0), 1760 _scheduled(arena), 1761 _available(arena), 1762 _next_node(NULL), 1763 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1764 _pinch_free_list(arena) 1765 #ifndef PRODUCT 1766 , _branches(0) 1767 , _unconditional_delays(0) 1768 #endif 1769 { 1770 // Create a MachNopNode 1771 _nop = new (&compile) MachNopNode(); 1772 1773 // Now that the nops are in the array, save the count 1774 // (but allow entries for the nops) 1775 _node_bundling_limit = compile.unique(); 1776 uint node_max = _regalloc->node_regs_max_index(); 1777 1778 compile.set_node_bundling_limit(_node_bundling_limit); 1779 1780 // This one is persistent within the Compile class 1781 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1782 1783 // Allocate space for fixed-size arrays 1784 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1785 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1786 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1787 1788 // Clear the arrays 1789 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1790 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1791 memset(_uses, 0, node_max * sizeof(short)); 1792 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1793 1794 // Clear the bundling information 1795 memcpy(_bundle_use_elements, 1796 Pipeline_Use::elaborated_elements, 1797 sizeof(Pipeline_Use::elaborated_elements)); 1798 1799 // Get the last node 1800 Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; 1801 1802 _next_node = bb->_nodes[bb->_nodes.size()-1]; 1803 } 1804 1805 #ifndef PRODUCT 1806 // Scheduling destructor 1807 Scheduling::~Scheduling() { 1808 _total_branches += _branches; 1809 _total_unconditional_delays += _unconditional_delays; 1810 } 1811 #endif 1812 1813 // Step ahead "i" cycles 1814 void Scheduling::step(uint i) { 1815 1816 Bundle *bundle = node_bundling(_next_node); 1817 bundle->set_starts_bundle(); 1818 1819 // Update the bundle record, but leave the flags information alone 1820 if (_bundle_instr_count > 0) { 1821 bundle->set_instr_count(_bundle_instr_count); 1822 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1823 } 1824 1825 // Update the state information 1826 _bundle_instr_count = 0; 1827 _bundle_cycle_number += i; 1828 _bundle_use.step(i); 1829 } 1830 1831 void Scheduling::step_and_clear() { 1832 Bundle *bundle = node_bundling(_next_node); 1833 bundle->set_starts_bundle(); 1834 1835 // Update the bundle record 1836 if (_bundle_instr_count > 0) { 1837 bundle->set_instr_count(_bundle_instr_count); 1838 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1839 1840 _bundle_cycle_number += 1; 1841 } 1842 1843 // Clear the bundling information 1844 _bundle_instr_count = 0; 1845 _bundle_use.reset(); 1846 1847 memcpy(_bundle_use_elements, 1848 Pipeline_Use::elaborated_elements, 1849 sizeof(Pipeline_Use::elaborated_elements)); 1850 } 1851 1852 //------------------------------ScheduleAndBundle------------------------------ 1853 // Perform instruction scheduling and bundling over the sequence of 1854 // instructions in backwards order. 1855 void Compile::ScheduleAndBundle() { 1856 1857 // Don't optimize this if it isn't a method 1858 if (!_method) 1859 return; 1860 1861 // Don't optimize this if scheduling is disabled 1862 if (!do_scheduling()) 1863 return; 1864 1865 // Scheduling code works only with pairs (8 bytes) maximum. 1866 if (max_vector_size() > 8) 1867 return; 1868 1869 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1870 1871 // Create a data structure for all the scheduling information 1872 Scheduling scheduling(Thread::current()->resource_area(), *this); 1873 1874 // Walk backwards over each basic block, computing the needed alignment 1875 // Walk over all the basic blocks 1876 scheduling.DoScheduling(); 1877 } 1878 1879 //------------------------------ComputeLocalLatenciesForward------------------- 1880 // Compute the latency of all the instructions. This is fairly simple, 1881 // because we already have a legal ordering. Walk over the instructions 1882 // from first to last, and compute the latency of the instruction based 1883 // on the latency of the preceding instruction(s). 1884 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1885 #ifndef PRODUCT 1886 if (_cfg->C->trace_opto_output()) 1887 tty->print("# -> ComputeLocalLatenciesForward\n"); 1888 #endif 1889 1890 // Walk over all the schedulable instructions 1891 for( uint j=_bb_start; j < _bb_end; j++ ) { 1892 1893 // This is a kludge, forcing all latency calculations to start at 1. 1894 // Used to allow latency 0 to force an instruction to the beginning 1895 // of the bb 1896 uint latency = 1; 1897 Node *use = bb->_nodes[j]; 1898 uint nlen = use->len(); 1899 1900 // Walk over all the inputs 1901 for ( uint k=0; k < nlen; k++ ) { 1902 Node *def = use->in(k); 1903 if (!def) 1904 continue; 1905 1906 uint l = _node_latency[def->_idx] + use->latency(k); 1907 if (latency < l) 1908 latency = l; 1909 } 1910 1911 _node_latency[use->_idx] = latency; 1912 1913 #ifndef PRODUCT 1914 if (_cfg->C->trace_opto_output()) { 1915 tty->print("# latency %4d: ", latency); 1916 use->dump(); 1917 } 1918 #endif 1919 } 1920 1921 #ifndef PRODUCT 1922 if (_cfg->C->trace_opto_output()) 1923 tty->print("# <- ComputeLocalLatenciesForward\n"); 1924 #endif 1925 1926 } // end ComputeLocalLatenciesForward 1927 1928 // See if this node fits into the present instruction bundle 1929 bool Scheduling::NodeFitsInBundle(Node *n) { 1930 uint n_idx = n->_idx; 1931 1932 // If this is the unconditional delay instruction, then it fits 1933 if (n == _unconditional_delay_slot) { 1934 #ifndef PRODUCT 1935 if (_cfg->C->trace_opto_output()) 1936 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1937 #endif 1938 return (true); 1939 } 1940 1941 // If the node cannot be scheduled this cycle, skip it 1942 if (_current_latency[n_idx] > _bundle_cycle_number) { 1943 #ifndef PRODUCT 1944 if (_cfg->C->trace_opto_output()) 1945 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1946 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1947 #endif 1948 return (false); 1949 } 1950 1951 const Pipeline *node_pipeline = n->pipeline(); 1952 1953 uint instruction_count = node_pipeline->instructionCount(); 1954 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1955 instruction_count = 0; 1956 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1957 instruction_count++; 1958 1959 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1960 #ifndef PRODUCT 1961 if (_cfg->C->trace_opto_output()) 1962 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1963 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1964 #endif 1965 return (false); 1966 } 1967 1968 // Don't allow non-machine nodes to be handled this way 1969 if (!n->is_Mach() && instruction_count == 0) 1970 return (false); 1971 1972 // See if there is any overlap 1973 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1974 1975 if (delay > 0) { 1976 #ifndef PRODUCT 1977 if (_cfg->C->trace_opto_output()) 1978 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1979 #endif 1980 return false; 1981 } 1982 1983 #ifndef PRODUCT 1984 if (_cfg->C->trace_opto_output()) 1985 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1986 #endif 1987 1988 return true; 1989 } 1990 1991 Node * Scheduling::ChooseNodeToBundle() { 1992 uint siz = _available.size(); 1993 1994 if (siz == 0) { 1995 1996 #ifndef PRODUCT 1997 if (_cfg->C->trace_opto_output()) 1998 tty->print("# ChooseNodeToBundle: NULL\n"); 1999 #endif 2000 return (NULL); 2001 } 2002 2003 // Fast path, if only 1 instruction in the bundle 2004 if (siz == 1) { 2005 #ifndef PRODUCT 2006 if (_cfg->C->trace_opto_output()) { 2007 tty->print("# ChooseNodeToBundle (only 1): "); 2008 _available[0]->dump(); 2009 } 2010 #endif 2011 return (_available[0]); 2012 } 2013 2014 // Don't bother, if the bundle is already full 2015 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2016 for ( uint i = 0; i < siz; i++ ) { 2017 Node *n = _available[i]; 2018 2019 // Skip projections, we'll handle them another way 2020 if (n->is_Proj()) 2021 continue; 2022 2023 // This presupposed that instructions are inserted into the 2024 // available list in a legality order; i.e. instructions that 2025 // must be inserted first are at the head of the list 2026 if (NodeFitsInBundle(n)) { 2027 #ifndef PRODUCT 2028 if (_cfg->C->trace_opto_output()) { 2029 tty->print("# ChooseNodeToBundle: "); 2030 n->dump(); 2031 } 2032 #endif 2033 return (n); 2034 } 2035 } 2036 } 2037 2038 // Nothing fits in this bundle, choose the highest priority 2039 #ifndef PRODUCT 2040 if (_cfg->C->trace_opto_output()) { 2041 tty->print("# ChooseNodeToBundle: "); 2042 _available[0]->dump(); 2043 } 2044 #endif 2045 2046 return _available[0]; 2047 } 2048 2049 //------------------------------AddNodeToAvailableList------------------------- 2050 void Scheduling::AddNodeToAvailableList(Node *n) { 2051 assert( !n->is_Proj(), "projections never directly made available" ); 2052 #ifndef PRODUCT 2053 if (_cfg->C->trace_opto_output()) { 2054 tty->print("# AddNodeToAvailableList: "); 2055 n->dump(); 2056 } 2057 #endif 2058 2059 int latency = _current_latency[n->_idx]; 2060 2061 // Insert in latency order (insertion sort) 2062 uint i; 2063 for ( i=0; i < _available.size(); i++ ) 2064 if (_current_latency[_available[i]->_idx] > latency) 2065 break; 2066 2067 // Special Check for compares following branches 2068 if( n->is_Mach() && _scheduled.size() > 0 ) { 2069 int op = n->as_Mach()->ideal_Opcode(); 2070 Node *last = _scheduled[0]; 2071 if( last->is_MachIf() && last->in(1) == n && 2072 ( op == Op_CmpI || 2073 op == Op_CmpU || 2074 op == Op_CmpP || 2075 op == Op_CmpF || 2076 op == Op_CmpD || 2077 op == Op_CmpL ) ) { 2078 2079 // Recalculate position, moving to front of same latency 2080 for ( i=0 ; i < _available.size(); i++ ) 2081 if (_current_latency[_available[i]->_idx] >= latency) 2082 break; 2083 } 2084 } 2085 2086 // Insert the node in the available list 2087 _available.insert(i, n); 2088 2089 #ifndef PRODUCT 2090 if (_cfg->C->trace_opto_output()) 2091 dump_available(); 2092 #endif 2093 } 2094 2095 //------------------------------DecrementUseCounts----------------------------- 2096 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2097 for ( uint i=0; i < n->len(); i++ ) { 2098 Node *def = n->in(i); 2099 if (!def) continue; 2100 if( def->is_Proj() ) // If this is a machine projection, then 2101 def = def->in(0); // propagate usage thru to the base instruction 2102 2103 if( _bbs[def->_idx] != bb ) // Ignore if not block-local 2104 continue; 2105 2106 // Compute the latency 2107 uint l = _bundle_cycle_number + n->latency(i); 2108 if (_current_latency[def->_idx] < l) 2109 _current_latency[def->_idx] = l; 2110 2111 // If this does not have uses then schedule it 2112 if ((--_uses[def->_idx]) == 0) 2113 AddNodeToAvailableList(def); 2114 } 2115 } 2116 2117 //------------------------------AddNodeToBundle-------------------------------- 2118 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2119 #ifndef PRODUCT 2120 if (_cfg->C->trace_opto_output()) { 2121 tty->print("# AddNodeToBundle: "); 2122 n->dump(); 2123 } 2124 #endif 2125 2126 // Remove this from the available list 2127 uint i; 2128 for (i = 0; i < _available.size(); i++) 2129 if (_available[i] == n) 2130 break; 2131 assert(i < _available.size(), "entry in _available list not found"); 2132 _available.remove(i); 2133 2134 // See if this fits in the current bundle 2135 const Pipeline *node_pipeline = n->pipeline(); 2136 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2137 2138 // Check for instructions to be placed in the delay slot. We 2139 // do this before we actually schedule the current instruction, 2140 // because the delay slot follows the current instruction. 2141 if (Pipeline::_branch_has_delay_slot && 2142 node_pipeline->hasBranchDelay() && 2143 !_unconditional_delay_slot) { 2144 2145 uint siz = _available.size(); 2146 2147 // Conditional branches can support an instruction that 2148 // is unconditionally executed and not dependent by the 2149 // branch, OR a conditionally executed instruction if 2150 // the branch is taken. In practice, this means that 2151 // the first instruction at the branch target is 2152 // copied to the delay slot, and the branch goes to 2153 // the instruction after that at the branch target 2154 if ( n->is_MachBranch() ) { 2155 2156 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2157 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2158 2159 #ifndef PRODUCT 2160 _branches++; 2161 #endif 2162 2163 // At least 1 instruction is on the available list 2164 // that is not dependent on the branch 2165 for (uint i = 0; i < siz; i++) { 2166 Node *d = _available[i]; 2167 const Pipeline *avail_pipeline = d->pipeline(); 2168 2169 // Don't allow safepoints in the branch shadow, that will 2170 // cause a number of difficulties 2171 if ( avail_pipeline->instructionCount() == 1 && 2172 !avail_pipeline->hasMultipleBundles() && 2173 !avail_pipeline->hasBranchDelay() && 2174 Pipeline::instr_has_unit_size() && 2175 d->size(_regalloc) == Pipeline::instr_unit_size() && 2176 NodeFitsInBundle(d) && 2177 !node_bundling(d)->used_in_delay()) { 2178 2179 if (d->is_Mach() && !d->is_MachSafePoint()) { 2180 // A node that fits in the delay slot was found, so we need to 2181 // set the appropriate bits in the bundle pipeline information so 2182 // that it correctly indicates resource usage. Later, when we 2183 // attempt to add this instruction to the bundle, we will skip 2184 // setting the resource usage. 2185 _unconditional_delay_slot = d; 2186 node_bundling(n)->set_use_unconditional_delay(); 2187 node_bundling(d)->set_used_in_unconditional_delay(); 2188 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2189 _current_latency[d->_idx] = _bundle_cycle_number; 2190 _next_node = d; 2191 ++_bundle_instr_count; 2192 #ifndef PRODUCT 2193 _unconditional_delays++; 2194 #endif 2195 break; 2196 } 2197 } 2198 } 2199 } 2200 2201 // No delay slot, add a nop to the usage 2202 if (!_unconditional_delay_slot) { 2203 // See if adding an instruction in the delay slot will overflow 2204 // the bundle. 2205 if (!NodeFitsInBundle(_nop)) { 2206 #ifndef PRODUCT 2207 if (_cfg->C->trace_opto_output()) 2208 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2209 #endif 2210 step(1); 2211 } 2212 2213 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2214 _next_node = _nop; 2215 ++_bundle_instr_count; 2216 } 2217 2218 // See if the instruction in the delay slot requires a 2219 // step of the bundles 2220 if (!NodeFitsInBundle(n)) { 2221 #ifndef PRODUCT 2222 if (_cfg->C->trace_opto_output()) 2223 tty->print("# *** STEP(branch won't fit) ***\n"); 2224 #endif 2225 // Update the state information 2226 _bundle_instr_count = 0; 2227 _bundle_cycle_number += 1; 2228 _bundle_use.step(1); 2229 } 2230 } 2231 2232 // Get the number of instructions 2233 uint instruction_count = node_pipeline->instructionCount(); 2234 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2235 instruction_count = 0; 2236 2237 // Compute the latency information 2238 uint delay = 0; 2239 2240 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2241 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2242 if (relative_latency < 0) 2243 relative_latency = 0; 2244 2245 delay = _bundle_use.full_latency(relative_latency, node_usage); 2246 2247 // Does not fit in this bundle, start a new one 2248 if (delay > 0) { 2249 step(delay); 2250 2251 #ifndef PRODUCT 2252 if (_cfg->C->trace_opto_output()) 2253 tty->print("# *** STEP(%d) ***\n", delay); 2254 #endif 2255 } 2256 } 2257 2258 // If this was placed in the delay slot, ignore it 2259 if (n != _unconditional_delay_slot) { 2260 2261 if (delay == 0) { 2262 if (node_pipeline->hasMultipleBundles()) { 2263 #ifndef PRODUCT 2264 if (_cfg->C->trace_opto_output()) 2265 tty->print("# *** STEP(multiple instructions) ***\n"); 2266 #endif 2267 step(1); 2268 } 2269 2270 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2271 #ifndef PRODUCT 2272 if (_cfg->C->trace_opto_output()) 2273 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2274 instruction_count + _bundle_instr_count, 2275 Pipeline::_max_instrs_per_cycle); 2276 #endif 2277 step(1); 2278 } 2279 } 2280 2281 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2282 _bundle_instr_count++; 2283 2284 // Set the node's latency 2285 _current_latency[n->_idx] = _bundle_cycle_number; 2286 2287 // Now merge the functional unit information 2288 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2289 _bundle_use.add_usage(node_usage); 2290 2291 // Increment the number of instructions in this bundle 2292 _bundle_instr_count += instruction_count; 2293 2294 // Remember this node for later 2295 if (n->is_Mach()) 2296 _next_node = n; 2297 } 2298 2299 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2300 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2301 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2302 // into the block. All other scheduled nodes get put in the schedule here. 2303 int op = n->Opcode(); 2304 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2305 (op != Op_Node && // Not an unused antidepedence node and 2306 // not an unallocated boxlock 2307 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2308 2309 // Push any trailing projections 2310 if( bb->_nodes[bb->_nodes.size()-1] != n ) { 2311 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2312 Node *foi = n->fast_out(i); 2313 if( foi->is_Proj() ) 2314 _scheduled.push(foi); 2315 } 2316 } 2317 2318 // Put the instruction in the schedule list 2319 _scheduled.push(n); 2320 } 2321 2322 #ifndef PRODUCT 2323 if (_cfg->C->trace_opto_output()) 2324 dump_available(); 2325 #endif 2326 2327 // Walk all the definitions, decrementing use counts, and 2328 // if a definition has a 0 use count, place it in the available list. 2329 DecrementUseCounts(n,bb); 2330 } 2331 2332 //------------------------------ComputeUseCount-------------------------------- 2333 // This method sets the use count within a basic block. We will ignore all 2334 // uses outside the current basic block. As we are doing a backwards walk, 2335 // any node we reach that has a use count of 0 may be scheduled. This also 2336 // avoids the problem of cyclic references from phi nodes, as long as phi 2337 // nodes are at the front of the basic block. This method also initializes 2338 // the available list to the set of instructions that have no uses within this 2339 // basic block. 2340 void Scheduling::ComputeUseCount(const Block *bb) { 2341 #ifndef PRODUCT 2342 if (_cfg->C->trace_opto_output()) 2343 tty->print("# -> ComputeUseCount\n"); 2344 #endif 2345 2346 // Clear the list of available and scheduled instructions, just in case 2347 _available.clear(); 2348 _scheduled.clear(); 2349 2350 // No delay slot specified 2351 _unconditional_delay_slot = NULL; 2352 2353 #ifdef ASSERT 2354 for( uint i=0; i < bb->_nodes.size(); i++ ) 2355 assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); 2356 #endif 2357 2358 // Force the _uses count to never go to zero for unscheduable pieces 2359 // of the block 2360 for( uint k = 0; k < _bb_start; k++ ) 2361 _uses[bb->_nodes[k]->_idx] = 1; 2362 for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) 2363 _uses[bb->_nodes[l]->_idx] = 1; 2364 2365 // Iterate backwards over the instructions in the block. Don't count the 2366 // branch projections at end or the block header instructions. 2367 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2368 Node *n = bb->_nodes[j]; 2369 if( n->is_Proj() ) continue; // Projections handled another way 2370 2371 // Account for all uses 2372 for ( uint k = 0; k < n->len(); k++ ) { 2373 Node *inp = n->in(k); 2374 if (!inp) continue; 2375 assert(inp != n, "no cycles allowed" ); 2376 if( _bbs[inp->_idx] == bb ) { // Block-local use? 2377 if( inp->is_Proj() ) // Skip through Proj's 2378 inp = inp->in(0); 2379 ++_uses[inp->_idx]; // Count 1 block-local use 2380 } 2381 } 2382 2383 // If this instruction has a 0 use count, then it is available 2384 if (!_uses[n->_idx]) { 2385 _current_latency[n->_idx] = _bundle_cycle_number; 2386 AddNodeToAvailableList(n); 2387 } 2388 2389 #ifndef PRODUCT 2390 if (_cfg->C->trace_opto_output()) { 2391 tty->print("# uses: %3d: ", _uses[n->_idx]); 2392 n->dump(); 2393 } 2394 #endif 2395 } 2396 2397 #ifndef PRODUCT 2398 if (_cfg->C->trace_opto_output()) 2399 tty->print("# <- ComputeUseCount\n"); 2400 #endif 2401 } 2402 2403 // This routine performs scheduling on each basic block in reverse order, 2404 // using instruction latencies and taking into account function unit 2405 // availability. 2406 void Scheduling::DoScheduling() { 2407 #ifndef PRODUCT 2408 if (_cfg->C->trace_opto_output()) 2409 tty->print("# -> DoScheduling\n"); 2410 #endif 2411 2412 Block *succ_bb = NULL; 2413 Block *bb; 2414 2415 // Walk over all the basic blocks in reverse order 2416 for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { 2417 bb = _cfg->_blocks[i]; 2418 2419 #ifndef PRODUCT 2420 if (_cfg->C->trace_opto_output()) { 2421 tty->print("# Schedule BB#%03d (initial)\n", i); 2422 for (uint j = 0; j < bb->_nodes.size(); j++) 2423 bb->_nodes[j]->dump(); 2424 } 2425 #endif 2426 2427 // On the head node, skip processing 2428 if( bb == _cfg->_broot ) 2429 continue; 2430 2431 // Skip empty, connector blocks 2432 if (bb->is_connector()) 2433 continue; 2434 2435 // If the following block is not the sole successor of 2436 // this one, then reset the pipeline information 2437 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2438 #ifndef PRODUCT 2439 if (_cfg->C->trace_opto_output()) { 2440 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2441 _next_node->_idx, _bundle_instr_count); 2442 } 2443 #endif 2444 step_and_clear(); 2445 } 2446 2447 // Leave untouched the starting instruction, any Phis, a CreateEx node 2448 // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. 2449 _bb_end = bb->_nodes.size()-1; 2450 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2451 Node *n = bb->_nodes[_bb_start]; 2452 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2453 // Also, MachIdealNodes do not get scheduled 2454 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2455 MachNode *mach = n->as_Mach(); 2456 int iop = mach->ideal_Opcode(); 2457 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2458 if( iop == Op_Con ) continue; // Do not schedule Top 2459 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2460 mach->pipeline() == MachNode::pipeline_class() && 2461 !n->is_SpillCopy() ) // Breakpoints, Prolog, etc 2462 continue; 2463 break; // Funny loop structure to be sure... 2464 } 2465 // Compute last "interesting" instruction in block - last instruction we 2466 // might schedule. _bb_end points just after last schedulable inst. We 2467 // normally schedule conditional branches (despite them being forced last 2468 // in the block), because they have delay slots we can fill. Calls all 2469 // have their delay slots filled in the template expansions, so we don't 2470 // bother scheduling them. 2471 Node *last = bb->_nodes[_bb_end]; 2472 // Ignore trailing NOPs. 2473 while (_bb_end > 0 && last->is_Mach() && 2474 last->as_Mach()->ideal_Opcode() == Op_Con) { 2475 last = bb->_nodes[--_bb_end]; 2476 } 2477 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2478 if( last->is_Catch() || 2479 // Exclude unreachable path case when Halt node is in a separate block. 2480 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2481 // There must be a prior call. Skip it. 2482 while( !bb->_nodes[--_bb_end]->is_MachCall() ) { 2483 assert( bb->_nodes[_bb_end]->is_MachProj(), "skipping projections after expected call" ); 2484 } 2485 } else if( last->is_MachNullCheck() ) { 2486 // Backup so the last null-checked memory instruction is 2487 // outside the schedulable range. Skip over the nullcheck, 2488 // projection, and the memory nodes. 2489 Node *mem = last->in(1); 2490 do { 2491 _bb_end--; 2492 } while (mem != bb->_nodes[_bb_end]); 2493 } else { 2494 // Set _bb_end to point after last schedulable inst. 2495 _bb_end++; 2496 } 2497 2498 assert( _bb_start <= _bb_end, "inverted block ends" ); 2499 2500 // Compute the register antidependencies for the basic block 2501 ComputeRegisterAntidependencies(bb); 2502 if (_cfg->C->failing()) return; // too many D-U pinch points 2503 2504 // Compute intra-bb latencies for the nodes 2505 ComputeLocalLatenciesForward(bb); 2506 2507 // Compute the usage within the block, and set the list of all nodes 2508 // in the block that have no uses within the block. 2509 ComputeUseCount(bb); 2510 2511 // Schedule the remaining instructions in the block 2512 while ( _available.size() > 0 ) { 2513 Node *n = ChooseNodeToBundle(); 2514 guarantee(n != NULL, "no nodes available"); 2515 AddNodeToBundle(n,bb); 2516 } 2517 2518 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2519 #ifdef ASSERT 2520 for( uint l = _bb_start; l < _bb_end; l++ ) { 2521 Node *n = bb->_nodes[l]; 2522 uint m; 2523 for( m = 0; m < _bb_end-_bb_start; m++ ) 2524 if( _scheduled[m] == n ) 2525 break; 2526 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2527 } 2528 #endif 2529 2530 // Now copy the instructions (in reverse order) back to the block 2531 for ( uint k = _bb_start; k < _bb_end; k++ ) 2532 bb->_nodes.map(k, _scheduled[_bb_end-k-1]); 2533 2534 #ifndef PRODUCT 2535 if (_cfg->C->trace_opto_output()) { 2536 tty->print("# Schedule BB#%03d (final)\n", i); 2537 uint current = 0; 2538 for (uint j = 0; j < bb->_nodes.size(); j++) { 2539 Node *n = bb->_nodes[j]; 2540 if( valid_bundle_info(n) ) { 2541 Bundle *bundle = node_bundling(n); 2542 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2543 tty->print("*** Bundle: "); 2544 bundle->dump(); 2545 } 2546 n->dump(); 2547 } 2548 } 2549 } 2550 #endif 2551 #ifdef ASSERT 2552 verify_good_schedule(bb,"after block local scheduling"); 2553 #endif 2554 } 2555 2556 #ifndef PRODUCT 2557 if (_cfg->C->trace_opto_output()) 2558 tty->print("# <- DoScheduling\n"); 2559 #endif 2560 2561 // Record final node-bundling array location 2562 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2563 2564 } // end DoScheduling 2565 2566 //------------------------------verify_good_schedule--------------------------- 2567 // Verify that no live-range used in the block is killed in the block by a 2568 // wrong DEF. This doesn't verify live-ranges that span blocks. 2569 2570 // Check for edge existence. Used to avoid adding redundant precedence edges. 2571 static bool edge_from_to( Node *from, Node *to ) { 2572 for( uint i=0; i<from->len(); i++ ) 2573 if( from->in(i) == to ) 2574 return true; 2575 return false; 2576 } 2577 2578 #ifdef ASSERT 2579 //------------------------------verify_do_def---------------------------------- 2580 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2581 // Check for bad kills 2582 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2583 Node *prior_use = _reg_node[def]; 2584 if( prior_use && !edge_from_to(prior_use,n) ) { 2585 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2586 n->dump(); 2587 tty->print_cr("..."); 2588 prior_use->dump(); 2589 assert(edge_from_to(prior_use,n),msg); 2590 } 2591 _reg_node.map(def,NULL); // Kill live USEs 2592 } 2593 } 2594 2595 //------------------------------verify_good_schedule--------------------------- 2596 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2597 2598 // Zap to something reasonable for the verify code 2599 _reg_node.clear(); 2600 2601 // Walk over the block backwards. Check to make sure each DEF doesn't 2602 // kill a live value (other than the one it's supposed to). Add each 2603 // USE to the live set. 2604 for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { 2605 Node *n = b->_nodes[i]; 2606 int n_op = n->Opcode(); 2607 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2608 // Fat-proj kills a slew of registers 2609 RegMask rm = n->out_RegMask();// Make local copy 2610 while( rm.is_NotEmpty() ) { 2611 OptoReg::Name kill = rm.find_first_elem(); 2612 rm.Remove(kill); 2613 verify_do_def( n, kill, msg ); 2614 } 2615 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2616 // Get DEF'd registers the normal way 2617 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2618 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2619 } 2620 2621 // Now make all USEs live 2622 for( uint i=1; i<n->req(); i++ ) { 2623 Node *def = n->in(i); 2624 assert(def != 0, "input edge required"); 2625 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2626 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2627 if( OptoReg::is_valid(reg_lo) ) { 2628 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2629 _reg_node.map(reg_lo,n); 2630 } 2631 if( OptoReg::is_valid(reg_hi) ) { 2632 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2633 _reg_node.map(reg_hi,n); 2634 } 2635 } 2636 2637 } 2638 2639 // Zap to something reasonable for the Antidependence code 2640 _reg_node.clear(); 2641 } 2642 #endif 2643 2644 // Conditionally add precedence edges. Avoid putting edges on Projs. 2645 static void add_prec_edge_from_to( Node *from, Node *to ) { 2646 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2647 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2648 from = from->in(0); 2649 } 2650 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2651 !edge_from_to( from, to ) ) // Avoid duplicate edge 2652 from->add_prec(to); 2653 } 2654 2655 //------------------------------anti_do_def------------------------------------ 2656 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2657 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2658 return; 2659 2660 Node *pinch = _reg_node[def_reg]; // Get pinch point 2661 if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? 2662 is_def ) { // Check for a true def (not a kill) 2663 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2664 return; 2665 } 2666 2667 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2668 debug_only( def = (Node*)0xdeadbeef; ) 2669 2670 // After some number of kills there _may_ be a later def 2671 Node *later_def = NULL; 2672 2673 // Finding a kill requires a real pinch-point. 2674 // Check for not already having a pinch-point. 2675 // Pinch points are Op_Node's. 2676 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2677 later_def = pinch; // Must be def/kill as optimistic pinch-point 2678 if ( _pinch_free_list.size() > 0) { 2679 pinch = _pinch_free_list.pop(); 2680 } else { 2681 pinch = new (_cfg->C) Node(1); // Pinch point to-be 2682 } 2683 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2684 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2685 return; 2686 } 2687 _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) 2688 _reg_node.map(def_reg,pinch); // Record pinch-point 2689 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2690 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2691 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2692 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2693 later_def = NULL; // and no later def 2694 } 2695 pinch->set_req(0,later_def); // Hook later def so we can find it 2696 } else { // Else have valid pinch point 2697 if( pinch->in(0) ) // If there is a later-def 2698 later_def = pinch->in(0); // Get it 2699 } 2700 2701 // Add output-dependence edge from later def to kill 2702 if( later_def ) // If there is some original def 2703 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2704 2705 // See if current kill is also a use, and so is forced to be the pinch-point. 2706 if( pinch->Opcode() == Op_Node ) { 2707 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2708 for( uint i=1; i<uses->req(); i++ ) { 2709 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2710 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2711 // Yes, found a use/kill pinch-point 2712 pinch->set_req(0,NULL); // 2713 pinch->replace_by(kill); // Move anti-dep edges up 2714 pinch = kill; 2715 _reg_node.map(def_reg,pinch); 2716 return; 2717 } 2718 } 2719 } 2720 2721 // Add edge from kill to pinch-point 2722 add_prec_edge_from_to(kill,pinch); 2723 } 2724 2725 //------------------------------anti_do_use------------------------------------ 2726 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2727 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2728 return; 2729 Node *pinch = _reg_node[use_reg]; // Get pinch point 2730 // Check for no later def_reg/kill in block 2731 if( pinch && _bbs[pinch->_idx] == b && 2732 // Use has to be block-local as well 2733 _bbs[use->_idx] == b ) { 2734 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2735 pinch->req() == 1 ) { // pinch not yet in block? 2736 pinch->del_req(0); // yank pointer to later-def, also set flag 2737 // Insert the pinch-point in the block just after the last use 2738 b->_nodes.insert(b->find_node(use)+1,pinch); 2739 _bb_end++; // Increase size scheduled region in block 2740 } 2741 2742 add_prec_edge_from_to(pinch,use); 2743 } 2744 } 2745 2746 //------------------------------ComputeRegisterAntidependences----------------- 2747 // We insert antidependences between the reads and following write of 2748 // allocated registers to prevent illegal code motion. Hopefully, the 2749 // number of added references should be fairly small, especially as we 2750 // are only adding references within the current basic block. 2751 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2752 2753 #ifdef ASSERT 2754 verify_good_schedule(b,"before block local scheduling"); 2755 #endif 2756 2757 // A valid schedule, for each register independently, is an endless cycle 2758 // of: a def, then some uses (connected to the def by true dependencies), 2759 // then some kills (defs with no uses), finally the cycle repeats with a new 2760 // def. The uses are allowed to float relative to each other, as are the 2761 // kills. No use is allowed to slide past a kill (or def). This requires 2762 // antidependencies between all uses of a single def and all kills that 2763 // follow, up to the next def. More edges are redundant, because later defs 2764 // & kills are already serialized with true or antidependencies. To keep 2765 // the edge count down, we add a 'pinch point' node if there's more than 2766 // one use or more than one kill/def. 2767 2768 // We add dependencies in one bottom-up pass. 2769 2770 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2771 2772 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2773 // register. If not, we record the DEF/KILL in _reg_node, the 2774 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2775 // "pinch point", a new Node that's in the graph but not in the block. 2776 // We put edges from the prior and current DEF/KILLs to the pinch point. 2777 // We put the pinch point in _reg_node. If there's already a pinch point 2778 // we merely add an edge from the current DEF/KILL to the pinch point. 2779 2780 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2781 // put an edge from the pinch point to the USE. 2782 2783 // To be expedient, the _reg_node array is pre-allocated for the whole 2784 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2785 // or a valid def/kill/pinch-point, or a leftover node from some prior 2786 // block. Leftover node from some prior block is treated like a NULL (no 2787 // prior def, so no anti-dependence needed). Valid def is distinguished by 2788 // it being in the current block. 2789 bool fat_proj_seen = false; 2790 uint last_safept = _bb_end-1; 2791 Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; 2792 Node* last_safept_node = end_node; 2793 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2794 Node *n = b->_nodes[i]; 2795 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2796 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2797 // Fat-proj kills a slew of registers 2798 // This can add edges to 'n' and obscure whether or not it was a def, 2799 // hence the is_def flag. 2800 fat_proj_seen = true; 2801 RegMask rm = n->out_RegMask();// Make local copy 2802 while( rm.is_NotEmpty() ) { 2803 OptoReg::Name kill = rm.find_first_elem(); 2804 rm.Remove(kill); 2805 anti_do_def( b, n, kill, is_def ); 2806 } 2807 } else { 2808 // Get DEF'd registers the normal way 2809 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2810 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2811 } 2812 2813 // Kill projections on a branch should appear to occur on the 2814 // branch, not afterwards, so grab the masks from the projections 2815 // and process them. 2816 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2817 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2818 Node* use = n->fast_out(i); 2819 if (use->is_Proj()) { 2820 RegMask rm = use->out_RegMask();// Make local copy 2821 while( rm.is_NotEmpty() ) { 2822 OptoReg::Name kill = rm.find_first_elem(); 2823 rm.Remove(kill); 2824 anti_do_def( b, n, kill, false ); 2825 } 2826 } 2827 } 2828 } 2829 2830 // Check each register used by this instruction for a following DEF/KILL 2831 // that must occur afterward and requires an anti-dependence edge. 2832 for( uint j=0; j<n->req(); j++ ) { 2833 Node *def = n->in(j); 2834 if( def ) { 2835 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2836 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2837 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2838 } 2839 } 2840 // Do not allow defs of new derived values to float above GC 2841 // points unless the base is definitely available at the GC point. 2842 2843 Node *m = b->_nodes[i]; 2844 2845 // Add precedence edge from following safepoint to use of derived pointer 2846 if( last_safept_node != end_node && 2847 m != last_safept_node) { 2848 for (uint k = 1; k < m->req(); k++) { 2849 const Type *t = m->in(k)->bottom_type(); 2850 if( t->isa_oop_ptr() && 2851 t->is_ptr()->offset() != 0 ) { 2852 last_safept_node->add_prec( m ); 2853 break; 2854 } 2855 } 2856 } 2857 2858 if( n->jvms() ) { // Precedence edge from derived to safept 2859 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2860 if( b->_nodes[last_safept] != last_safept_node ) { 2861 last_safept = b->find_node(last_safept_node); 2862 } 2863 for( uint j=last_safept; j > i; j-- ) { 2864 Node *mach = b->_nodes[j]; 2865 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2866 mach->add_prec( n ); 2867 } 2868 last_safept = i; 2869 last_safept_node = m; 2870 } 2871 } 2872 2873 if (fat_proj_seen) { 2874 // Garbage collect pinch nodes that were not consumed. 2875 // They are usually created by a fat kill MachProj for a call. 2876 garbage_collect_pinch_nodes(); 2877 } 2878 } 2879 2880 //------------------------------garbage_collect_pinch_nodes------------------------------- 2881 2882 // Garbage collect pinch nodes for reuse by other blocks. 2883 // 2884 // The block scheduler's insertion of anti-dependence 2885 // edges creates many pinch nodes when the block contains 2886 // 2 or more Calls. A pinch node is used to prevent a 2887 // combinatorial explosion of edges. If a set of kills for a 2888 // register is anti-dependent on a set of uses (or defs), rather 2889 // than adding an edge in the graph between each pair of kill 2890 // and use (or def), a pinch is inserted between them: 2891 // 2892 // use1 use2 use3 2893 // \ | / 2894 // \ | / 2895 // pinch 2896 // / | \ 2897 // / | \ 2898 // kill1 kill2 kill3 2899 // 2900 // One pinch node is created per register killed when 2901 // the second call is encountered during a backwards pass 2902 // over the block. Most of these pinch nodes are never 2903 // wired into the graph because the register is never 2904 // used or def'ed in the block. 2905 // 2906 void Scheduling::garbage_collect_pinch_nodes() { 2907 #ifndef PRODUCT 2908 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2909 #endif 2910 int trace_cnt = 0; 2911 for (uint k = 0; k < _reg_node.Size(); k++) { 2912 Node* pinch = _reg_node[k]; 2913 if (pinch != NULL && pinch->Opcode() == Op_Node && 2914 // no predecence input edges 2915 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2916 cleanup_pinch(pinch); 2917 _pinch_free_list.push(pinch); 2918 _reg_node.map(k, NULL); 2919 #ifndef PRODUCT 2920 if (_cfg->C->trace_opto_output()) { 2921 trace_cnt++; 2922 if (trace_cnt > 40) { 2923 tty->print("\n"); 2924 trace_cnt = 0; 2925 } 2926 tty->print(" %d", pinch->_idx); 2927 } 2928 #endif 2929 } 2930 } 2931 #ifndef PRODUCT 2932 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2933 #endif 2934 } 2935 2936 // Clean up a pinch node for reuse. 2937 void Scheduling::cleanup_pinch( Node *pinch ) { 2938 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2939 2940 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2941 Node* use = pinch->last_out(i); 2942 uint uses_found = 0; 2943 for (uint j = use->req(); j < use->len(); j++) { 2944 if (use->in(j) == pinch) { 2945 use->rm_prec(j); 2946 uses_found++; 2947 } 2948 } 2949 assert(uses_found > 0, "must be a precedence edge"); 2950 i -= uses_found; // we deleted 1 or more copies of this edge 2951 } 2952 // May have a later_def entry 2953 pinch->set_req(0, NULL); 2954 } 2955 2956 //------------------------------print_statistics------------------------------- 2957 #ifndef PRODUCT 2958 2959 void Scheduling::dump_available() const { 2960 tty->print("#Availist "); 2961 for (uint i = 0; i < _available.size(); i++) 2962 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2963 tty->cr(); 2964 } 2965 2966 // Print Scheduling Statistics 2967 void Scheduling::print_statistics() { 2968 // Print the size added by nops for bundling 2969 tty->print("Nops added %d bytes to total of %d bytes", 2970 _total_nop_size, _total_method_size); 2971 if (_total_method_size > 0) 2972 tty->print(", for %.2f%%", 2973 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2974 tty->print("\n"); 2975 2976 // Print the number of branch shadows filled 2977 if (Pipeline::_branch_has_delay_slot) { 2978 tty->print("Of %d branches, %d had unconditional delay slots filled", 2979 _total_branches, _total_unconditional_delays); 2980 if (_total_branches > 0) 2981 tty->print(", for %.2f%%", 2982 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2983 tty->print("\n"); 2984 } 2985 2986 uint total_instructions = 0, total_bundles = 0; 2987 2988 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2989 uint bundle_count = _total_instructions_per_bundle[i]; 2990 total_instructions += bundle_count * i; 2991 total_bundles += bundle_count; 2992 } 2993 2994 if (total_bundles > 0) 2995 tty->print("Average ILP (excluding nops) is %.2f\n", 2996 ((double)total_instructions) / ((double)total_bundles)); 2997 } 2998 #endif