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