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