1 /* 2 * Copyright (c) 1998, 2014, 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 #ifndef PRODUCT 46 #define DEBUG_ARG(x) , x 47 #else 48 #define DEBUG_ARG(x) 49 #endif 50 51 // Convert Nodes to instruction bits and pass off to the VM 52 void Compile::Output() { 53 // RootNode goes 54 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 55 56 // The number of new nodes (mostly MachNop) is proportional to 57 // the number of java calls and inner loops which are aligned. 58 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 59 C->inner_loops()*(OptoLoopAlignment-1)), 60 "out of nodes before code generation" ) ) { 61 return; 62 } 63 // Make sure I can find the Start Node 64 Block *entry = _cfg->get_block(1); 65 Block *broot = _cfg->get_root_block(); 66 67 const StartNode *start = entry->head()->as_Start(); 68 69 // Replace StartNode with prolog 70 MachPrologNode *prolog = new (this) MachPrologNode(); 71 entry->map_node(prolog, 0); 72 _cfg->map_node_to_block(prolog, entry); 73 _cfg->unmap_node_from_block(start); // start is no longer in any block 74 75 // Virtual methods need an unverified entry point 76 77 if( is_osr_compilation() ) { 78 if( PoisonOSREntry ) { 79 // TODO: Should use a ShouldNotReachHereNode... 80 _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); 81 } 82 } else { 83 if( _method && !_method->flags().is_static() ) { 84 // Insert unvalidated entry point 85 _cfg->insert( broot, 0, new (this) MachUEPNode() ); 86 } 87 88 } 89 90 91 // Break before main entry point 92 if( (_method && _method->break_at_execute()) 93 #ifndef PRODUCT 94 ||(OptoBreakpoint && is_method_compilation()) 95 ||(OptoBreakpointOSR && is_osr_compilation()) 96 ||(OptoBreakpointC2R && !_method) 97 #endif 98 ) { 99 // checking for _method means that OptoBreakpoint does not apply to 100 // runtime stubs or frame converters 101 _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); 102 } 103 104 // Insert epilogs before every return 105 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 106 Block* block = _cfg->get_block(i); 107 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 108 Node* m = block->end(); 109 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 110 MachEpilogNode* epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 111 block->add_inst(epilog); 112 _cfg->map_node_to_block(epilog, block); 113 } 114 } 115 } 116 117 # ifdef ENABLE_ZAP_DEAD_LOCALS 118 if (ZapDeadCompiledLocals) { 119 Insert_zap_nodes(); 120 } 121 # endif 122 123 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 124 blk_starts[0] = 0; 125 126 // Initialize code buffer and process short branches. 127 CodeBuffer* cb = init_buffer(blk_starts); 128 129 if (cb == NULL || failing()) { 130 return; 131 } 132 133 ScheduleAndBundle(); 134 135 #ifndef PRODUCT 136 if (trace_opto_output()) { 137 tty->print("\n---- After ScheduleAndBundle ----\n"); 138 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 139 tty->print("\nBB#%03d:\n", i); 140 Block* block = _cfg->get_block(i); 141 for (uint j = 0; j < block->number_of_nodes(); j++) { 142 Node* n = block->get_node(j); 143 OptoReg::Name reg = _regalloc->get_reg_first(n); 144 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 145 n->dump(); 146 } 147 } 148 } 149 #endif 150 151 if (failing()) { 152 return; 153 } 154 155 BuildOopMaps(); 156 157 if (failing()) { 158 return; 159 } 160 161 fill_buffer(cb, blk_starts); 162 } 163 164 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 165 // Determine if we need to generate a stack overflow check. 166 // Do it if the method is not a stub function and 167 // has java calls or has frame size > vm_page_size/8. 168 // The debug VM checks that deoptimization doesn't trigger an 169 // unexpected stack overflow (compiled method stack banging should 170 // guarantee it doesn't happen) so we always need the stack bang in 171 // a debug VM. 172 return (UseStackBanging && stub_function() == NULL && 173 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3 174 DEBUG_ONLY(|| true))); 175 } 176 177 bool Compile::need_register_stack_bang() const { 178 // Determine if we need to generate a register stack overflow check. 179 // This is only used on architectures which have split register 180 // and memory stacks (ie. IA64). 181 // Bang if the method is not a stub function and has java calls 182 return (stub_function() == NULL && has_java_calls()); 183 } 184 185 # ifdef ENABLE_ZAP_DEAD_LOCALS 186 187 188 // In order to catch compiler oop-map bugs, we have implemented 189 // a debugging mode called ZapDeadCompilerLocals. 190 // This mode causes the compiler to insert a call to a runtime routine, 191 // "zap_dead_locals", right before each place in compiled code 192 // that could potentially be a gc-point (i.e., a safepoint or oop map point). 193 // The runtime routine checks that locations mapped as oops are really 194 // oops, that locations mapped as values do not look like oops, 195 // and that locations mapped as dead are not used later 196 // (by zapping them to an invalid address). 197 198 int Compile::_CompiledZap_count = 0; 199 200 void Compile::Insert_zap_nodes() { 201 bool skip = false; 202 203 204 // Dink with static counts because code code without the extra 205 // runtime calls is MUCH faster for debugging purposes 206 207 if ( CompileZapFirst == 0 ) ; // nothing special 208 else if ( CompileZapFirst > CompiledZap_count() ) skip = true; 209 else if ( CompileZapFirst == CompiledZap_count() ) 210 warning("starting zap compilation after skipping"); 211 212 if ( CompileZapLast == -1 ) ; // nothing special 213 else if ( CompileZapLast < CompiledZap_count() ) skip = true; 214 else if ( CompileZapLast == CompiledZap_count() ) 215 warning("about to compile last zap"); 216 217 ++_CompiledZap_count; // counts skipped zaps, too 218 219 if ( skip ) return; 220 221 222 if ( _method == NULL ) 223 return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care 224 225 // Insert call to zap runtime stub before every node with an oop map 226 for( uint i=0; i<_cfg->number_of_blocks(); i++ ) { 227 Block *b = _cfg->get_block(i); 228 for ( uint j = 0; j < b->number_of_nodes(); ++j ) { 229 Node *n = b->get_node(j); 230 231 // Determining if we should insert a zap-a-lot node in output. 232 // We do that for all nodes that has oopmap info, except for calls 233 // to allocation. Calls to allocation passes in the old top-of-eden pointer 234 // and expect the C code to reset it. Hence, there can be no safepoints between 235 // the inlined-allocation and the call to new_Java, etc. 236 // We also cannot zap monitor calls, as they must hold the microlock 237 // during the call to Zap, which also wants to grab the microlock. 238 bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); 239 if ( insert ) { // it is MachSafePoint 240 if ( !n->is_MachCall() ) { 241 insert = false; 242 } else if ( n->is_MachCall() ) { 243 MachCallNode* call = n->as_MachCall(); 244 if (call->entry_point() == OptoRuntime::new_instance_Java() || 245 call->entry_point() == OptoRuntime::new_array_Java() || 246 call->entry_point() == OptoRuntime::multianewarray2_Java() || 247 call->entry_point() == OptoRuntime::multianewarray3_Java() || 248 call->entry_point() == OptoRuntime::multianewarray4_Java() || 249 call->entry_point() == OptoRuntime::multianewarray5_Java() || 250 call->entry_point() == OptoRuntime::slow_arraycopy_Java() || 251 call->entry_point() == OptoRuntime::complete_monitor_locking_Java() 252 ) { 253 insert = false; 254 } 255 } 256 if (insert) { 257 Node *zap = call_zap_node(n->as_MachSafePoint(), i); 258 b->insert_node(zap, j); 259 _cfg->map_node_to_block(zap, b); 260 ++j; 261 } 262 } 263 } 264 } 265 } 266 267 268 Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { 269 const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); 270 CallStaticJavaNode* ideal_node = 271 new (this) CallStaticJavaNode( tf, 272 OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), 273 "call zap dead locals stub", 0, TypePtr::BOTTOM); 274 // We need to copy the OopMap from the site we're zapping at. 275 // We have to make a copy, because the zap site might not be 276 // a call site, and zap_dead is a call site. 277 OopMap* clone = node_to_check->oop_map()->deep_copy(); 278 279 // Add the cloned OopMap to the zap node 280 ideal_node->set_oop_map(clone); 281 return _matcher->match_sfpt(ideal_node); 282 } 283 284 bool Compile::is_node_getting_a_safepoint( Node* n) { 285 // This code duplicates the logic prior to the call of add_safepoint 286 // below in this file. 287 if( n->is_MachSafePoint() ) return true; 288 return false; 289 } 290 291 # endif // ENABLE_ZAP_DEAD_LOCALS 292 293 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 294 // of a loop. When aligning a loop we need to provide enough instructions 295 // in cpu's fetch buffer to feed decoders. The loop alignment could be 296 // avoided if we have enough instructions in fetch buffer at the head of a loop. 297 // By default, the size is set to 999999 by Block's constructor so that 298 // a loop will be aligned if the size is not reset here. 299 // 300 // Note: Mach instructions could contain several HW instructions 301 // so the size is estimated only. 302 // 303 void Compile::compute_loop_first_inst_sizes() { 304 // The next condition is used to gate the loop alignment optimization. 305 // Don't aligned a loop if there are enough instructions at the head of a loop 306 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 307 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 308 // equal to 11 bytes which is the largest address NOP instruction. 309 if (MaxLoopPad < OptoLoopAlignment - 1) { 310 uint last_block = _cfg->number_of_blocks() - 1; 311 for (uint i = 1; i <= last_block; i++) { 312 Block* block = _cfg->get_block(i); 313 // Check the first loop's block which requires an alignment. 314 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 315 uint sum_size = 0; 316 uint inst_cnt = NumberOfLoopInstrToAlign; 317 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 318 319 // Check subsequent fallthrough blocks if the loop's first 320 // block(s) does not have enough instructions. 321 Block *nb = block; 322 while(inst_cnt > 0 && 323 i < last_block && 324 !_cfg->get_block(i + 1)->has_loop_alignment() && 325 !nb->has_successor(block)) { 326 i++; 327 nb = _cfg->get_block(i); 328 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 329 } // while( inst_cnt > 0 && i < last_block ) 330 331 block->set_first_inst_size(sum_size); 332 } // f( b->head()->is_Loop() ) 333 } // for( i <= last_block ) 334 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 335 } 336 337 // The architecture description provides short branch variants for some long 338 // branch instructions. Replace eligible long branches with short branches. 339 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 340 // Compute size of each block, method size, and relocation information size 341 uint nblocks = _cfg->number_of_blocks(); 342 343 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 344 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 345 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 346 347 // Collect worst case block paddings 348 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 349 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 350 351 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 352 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 353 354 bool has_short_branch_candidate = false; 355 356 // Initialize the sizes to 0 357 code_size = 0; // Size in bytes of generated code 358 stub_size = 0; // Size in bytes of all stub entries 359 // Size in bytes of all relocation entries, including those in local stubs. 360 // Start with 2-bytes of reloc info for the unvalidated entry point 361 reloc_size = 1; // Number of relocation entries 362 363 // Make three passes. The first computes pessimistic blk_starts, 364 // relative jmp_offset and reloc_size information. The second performs 365 // short branch substitution using the pessimistic sizing. The 366 // third inserts nops where needed. 367 368 // Step one, perform a pessimistic sizing pass. 369 uint last_call_adr = max_uint; 370 uint last_avoid_back_to_back_adr = max_uint; 371 uint nop_size = (new (this) MachNopNode())->size(_regalloc); 372 for (uint i = 0; i < nblocks; i++) { // For all blocks 373 Block* block = _cfg->get_block(i); 374 375 // During short branch replacement, we store the relative (to blk_starts) 376 // offset of jump in jmp_offset, rather than the absolute offset of jump. 377 // This is so that we do not need to recompute sizes of all nodes when 378 // we compute correct blk_starts in our next sizing pass. 379 jmp_offset[i] = 0; 380 jmp_size[i] = 0; 381 jmp_nidx[i] = -1; 382 DEBUG_ONLY( jmp_target[i] = 0; ) 383 DEBUG_ONLY( jmp_rule[i] = 0; ) 384 385 // Sum all instruction sizes to compute block size 386 uint last_inst = block->number_of_nodes(); 387 uint blk_size = 0; 388 for (uint j = 0; j < last_inst; j++) { 389 Node* nj = block->get_node(j); 390 // Handle machine instruction nodes 391 if (nj->is_Mach()) { 392 MachNode *mach = nj->as_Mach(); 393 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 394 reloc_size += mach->reloc(); 395 if (mach->is_MachCall()) { 396 // add size information for trampoline stub 397 // class CallStubImpl is platform-specific and defined in the *.ad files. 398 stub_size += CallStubImpl::size_call_trampoline(); 399 reloc_size += CallStubImpl::reloc_call_trampoline(); 400 401 MachCallNode *mcall = mach->as_MachCall(); 402 // This destination address is NOT PC-relative 403 404 mcall->method_set((intptr_t)mcall->entry_point()); 405 406 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 407 stub_size += CompiledStaticCall::to_interp_stub_size(); 408 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 409 } 410 } else if (mach->is_MachSafePoint()) { 411 // If call/safepoint are adjacent, account for possible 412 // nop to disambiguate the two safepoints. 413 // ScheduleAndBundle() can rearrange nodes in a block, 414 // check for all offsets inside this block. 415 if (last_call_adr >= blk_starts[i]) { 416 blk_size += nop_size; 417 } 418 } 419 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 420 // Nop is inserted between "avoid back to back" instructions. 421 // ScheduleAndBundle() can rearrange nodes in a block, 422 // check for all offsets inside this block. 423 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 424 blk_size += nop_size; 425 } 426 } 427 if (mach->may_be_short_branch()) { 428 if (!nj->is_MachBranch()) { 429 #ifndef PRODUCT 430 nj->dump(3); 431 #endif 432 Unimplemented(); 433 } 434 assert(jmp_nidx[i] == -1, "block should have only one branch"); 435 jmp_offset[i] = blk_size; 436 jmp_size[i] = nj->size(_regalloc); 437 jmp_nidx[i] = j; 438 has_short_branch_candidate = true; 439 } 440 } 441 blk_size += nj->size(_regalloc); 442 // Remember end of call offset 443 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 444 last_call_adr = blk_starts[i]+blk_size; 445 } 446 // Remember end of avoid_back_to_back offset 447 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 448 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 449 } 450 } 451 452 // When the next block starts a loop, we may insert pad NOP 453 // instructions. Since we cannot know our future alignment, 454 // assume the worst. 455 if (i < nblocks - 1) { 456 Block* nb = _cfg->get_block(i + 1); 457 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 458 if (max_loop_pad > 0) { 459 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 460 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 461 // If either is the last instruction in this block, bump by 462 // max_loop_pad in lock-step with blk_size, so sizing 463 // calculations in subsequent blocks still can conservatively 464 // detect that it may the last instruction in this block. 465 if (last_call_adr == blk_starts[i]+blk_size) { 466 last_call_adr += max_loop_pad; 467 } 468 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 469 last_avoid_back_to_back_adr += max_loop_pad; 470 } 471 blk_size += max_loop_pad; 472 block_worst_case_pad[i + 1] = max_loop_pad; 473 } 474 } 475 476 // Save block size; update total method size 477 blk_starts[i+1] = blk_starts[i]+blk_size; 478 } 479 480 // Step two, replace eligible long jumps. 481 bool progress = true; 482 uint last_may_be_short_branch_adr = max_uint; 483 while (has_short_branch_candidate && progress) { 484 progress = false; 485 has_short_branch_candidate = false; 486 int adjust_block_start = 0; 487 for (uint i = 0; i < nblocks; i++) { 488 Block* block = _cfg->get_block(i); 489 int idx = jmp_nidx[i]; 490 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 491 if (mach != NULL && mach->may_be_short_branch()) { 492 #ifdef ASSERT 493 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 494 int j; 495 // Find the branch; ignore trailing NOPs. 496 for (j = block->number_of_nodes()-1; j>=0; j--) { 497 Node* n = block->get_node(j); 498 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 499 break; 500 } 501 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 502 #endif 503 int br_size = jmp_size[i]; 504 int br_offs = blk_starts[i] + jmp_offset[i]; 505 506 // This requires the TRUE branch target be in succs[0] 507 uint bnum = block->non_connector_successor(0)->_pre_order; 508 int offset = blk_starts[bnum] - br_offs; 509 if (bnum > i) { // adjust following block's offset 510 offset -= adjust_block_start; 511 } 512 513 // This block can be a loop header, account for the padding 514 // in the previous block. 515 int block_padding = block_worst_case_pad[i]; 516 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 517 // In the following code a nop could be inserted before 518 // the branch which will increase the backward distance. 519 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 520 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 521 522 if (needs_padding && offset <= 0) 523 offset -= nop_size; 524 525 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 526 // We've got a winner. Replace this branch. 527 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 528 529 // Update the jmp_size. 530 int new_size = replacement->size(_regalloc); 531 int diff = br_size - new_size; 532 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 533 // Conservatively take into account padding between 534 // avoid_back_to_back branches. Previous branch could be 535 // converted into avoid_back_to_back branch during next 536 // rounds. 537 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 538 jmp_offset[i] += nop_size; 539 diff -= nop_size; 540 } 541 adjust_block_start += diff; 542 block->map_node(replacement, idx); 543 mach->subsume_by(replacement, C); 544 mach = replacement; 545 progress = true; 546 547 jmp_size[i] = new_size; 548 DEBUG_ONLY( jmp_target[i] = bnum; ); 549 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 550 } else { 551 // The jump distance is not short, try again during next iteration. 552 has_short_branch_candidate = true; 553 } 554 } // (mach->may_be_short_branch()) 555 if (mach != NULL && (mach->may_be_short_branch() || 556 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 557 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 558 } 559 blk_starts[i+1] -= adjust_block_start; 560 } 561 } 562 563 #ifdef ASSERT 564 for (uint i = 0; i < nblocks; i++) { // For all blocks 565 if (jmp_target[i] != 0) { 566 int br_size = jmp_size[i]; 567 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 568 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 569 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]); 570 } 571 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 572 } 573 } 574 #endif 575 576 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 577 // after ScheduleAndBundle(). 578 579 // ------------------ 580 // Compute size for code buffer 581 code_size = blk_starts[nblocks]; 582 583 // Relocation records 584 reloc_size += 1; // Relo entry for exception handler 585 586 // Adjust reloc_size to number of record of relocation info 587 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 588 // a relocation index. 589 // The CodeBuffer will expand the locs array if this estimate is too low. 590 reloc_size *= 10 / sizeof(relocInfo); 591 } 592 593 //------------------------------FillLocArray----------------------------------- 594 // Create a bit of debug info and append it to the array. The mapping is from 595 // Java local or expression stack to constant, register or stack-slot. For 596 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 597 // entry has been taken care of and caller should skip it). 598 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 599 // This should never have accepted Bad before 600 assert(OptoReg::is_valid(regnum), "location must be valid"); 601 return (OptoReg::is_reg(regnum)) 602 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 603 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 604 } 605 606 607 ObjectValue* 608 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 609 for (int i = 0; i < objs->length(); i++) { 610 assert(objs->at(i)->is_object(), "corrupt object cache"); 611 ObjectValue* sv = (ObjectValue*) objs->at(i); 612 if (sv->id() == id) { 613 return sv; 614 } 615 } 616 // Otherwise.. 617 return NULL; 618 } 619 620 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 621 ObjectValue* sv ) { 622 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 623 objs->append(sv); 624 } 625 626 627 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 628 GrowableArray<ScopeValue*> *array, 629 GrowableArray<ScopeValue*> *objs ) { 630 assert( local, "use _top instead of null" ); 631 if (array->length() != idx) { 632 assert(array->length() == idx + 1, "Unexpected array count"); 633 // Old functionality: 634 // return 635 // New functionality: 636 // Assert if the local is not top. In product mode let the new node 637 // override the old entry. 638 assert(local == top(), "LocArray collision"); 639 if (local == top()) { 640 return; 641 } 642 array->pop(); 643 } 644 const Type *t = local->bottom_type(); 645 646 // Is it a safepoint scalar object node? 647 if (local->is_SafePointScalarObject()) { 648 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 649 650 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 651 if (sv == NULL) { 652 ciKlass* cik = t->is_oopptr()->klass(); 653 assert(cik->is_instance_klass() || 654 cik->is_array_klass(), "Not supported allocation."); 655 sv = new ObjectValue(spobj->_idx, 656 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 657 Compile::set_sv_for_object_node(objs, sv); 658 659 uint first_ind = spobj->first_index(sfpt->jvms()); 660 for (uint i = 0; i < spobj->n_fields(); i++) { 661 Node* fld_node = sfpt->in(first_ind+i); 662 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 663 } 664 } 665 array->append(sv); 666 return; 667 } 668 669 // Grab the register number for the local 670 OptoReg::Name regnum = _regalloc->get_reg_first(local); 671 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 672 // Record the double as two float registers. 673 // The register mask for such a value always specifies two adjacent 674 // float registers, with the lower register number even. 675 // Normally, the allocation of high and low words to these registers 676 // is irrelevant, because nearly all operations on register pairs 677 // (e.g., StoreD) treat them as a single unit. 678 // Here, we assume in addition that the words in these two registers 679 // stored "naturally" (by operations like StoreD and double stores 680 // within the interpreter) such that the lower-numbered register 681 // is written to the lower memory address. This may seem like 682 // a machine dependency, but it is not--it is a requirement on 683 // the author of the <arch>.ad file to ensure that, for every 684 // even/odd double-register pair to which a double may be allocated, 685 // the word in the even single-register is stored to the first 686 // memory word. (Note that register numbers are completely 687 // arbitrary, and are not tied to any machine-level encodings.) 688 #ifdef _LP64 689 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 690 array->append(new ConstantIntValue(0)); 691 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 692 } else if ( t->base() == Type::Long ) { 693 array->append(new ConstantIntValue(0)); 694 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 695 } else if ( t->base() == Type::RawPtr ) { 696 // jsr/ret return address which must be restored into a the full 697 // width 64-bit stack slot. 698 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 699 } 700 #else //_LP64 701 #ifdef SPARC 702 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 703 // For SPARC we have to swap high and low words for 704 // long values stored in a single-register (g0-g7). 705 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 706 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 707 } else 708 #endif //SPARC 709 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 710 // Repack the double/long as two jints. 711 // The convention the interpreter uses is that the second local 712 // holds the first raw word of the native double representation. 713 // This is actually reasonable, since locals and stack arrays 714 // grow downwards in all implementations. 715 // (If, on some machine, the interpreter's Java locals or stack 716 // were to grow upwards, the embedded doubles would be word-swapped.) 717 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 718 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 719 } 720 #endif //_LP64 721 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 722 OptoReg::is_reg(regnum) ) { 723 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 724 ? Location::float_in_dbl : Location::normal )); 725 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 726 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 727 ? Location::int_in_long : Location::normal )); 728 } else if( t->base() == Type::NarrowOop ) { 729 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 730 } else { 731 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 732 } 733 return; 734 } 735 736 // No register. It must be constant data. 737 switch (t->base()) { 738 case Type::Half: // Second half of a double 739 ShouldNotReachHere(); // Caller should skip 2nd halves 740 break; 741 case Type::AnyPtr: 742 array->append(new ConstantOopWriteValue(NULL)); 743 break; 744 case Type::AryPtr: 745 case Type::InstPtr: // fall through 746 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 747 break; 748 case Type::NarrowOop: 749 if (t == TypeNarrowOop::NULL_PTR) { 750 array->append(new ConstantOopWriteValue(NULL)); 751 } else { 752 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 753 } 754 break; 755 case Type::Int: 756 array->append(new ConstantIntValue(t->is_int()->get_con())); 757 break; 758 case Type::RawPtr: 759 // A return address (T_ADDRESS). 760 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 761 #ifdef _LP64 762 // Must be restored to the full-width 64-bit stack slot. 763 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 764 #else 765 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 766 #endif 767 break; 768 case Type::FloatCon: { 769 float f = t->is_float_constant()->getf(); 770 array->append(new ConstantIntValue(jint_cast(f))); 771 break; 772 } 773 case Type::DoubleCon: { 774 jdouble d = t->is_double_constant()->getd(); 775 #ifdef _LP64 776 array->append(new ConstantIntValue(0)); 777 array->append(new ConstantDoubleValue(d)); 778 #else 779 // Repack the double as two jints. 780 // The convention the interpreter uses is that the second local 781 // holds the first raw word of the native double representation. 782 // This is actually reasonable, since locals and stack arrays 783 // grow downwards in all implementations. 784 // (If, on some machine, the interpreter's Java locals or stack 785 // were to grow upwards, the embedded doubles would be word-swapped.) 786 jlong_accessor acc; 787 acc.long_value = jlong_cast(d); 788 array->append(new ConstantIntValue(acc.words[1])); 789 array->append(new ConstantIntValue(acc.words[0])); 790 #endif 791 break; 792 } 793 case Type::Long: { 794 jlong d = t->is_long()->get_con(); 795 #ifdef _LP64 796 array->append(new ConstantIntValue(0)); 797 array->append(new ConstantLongValue(d)); 798 #else 799 // Repack the long as two jints. 800 // The convention the interpreter uses is that the second local 801 // holds the first raw word of the native double representation. 802 // This is actually reasonable, since locals and stack arrays 803 // grow downwards in all implementations. 804 // (If, on some machine, the interpreter's Java locals or stack 805 // were to grow upwards, the embedded doubles would be word-swapped.) 806 jlong_accessor acc; 807 acc.long_value = d; 808 array->append(new ConstantIntValue(acc.words[1])); 809 array->append(new ConstantIntValue(acc.words[0])); 810 #endif 811 break; 812 } 813 case Type::Top: // Add an illegal value here 814 array->append(new LocationValue(Location())); 815 break; 816 default: 817 ShouldNotReachHere(); 818 break; 819 } 820 } 821 822 // Determine if this node starts a bundle 823 bool Compile::starts_bundle(const Node *n) const { 824 return (_node_bundling_limit > n->_idx && 825 _node_bundling_base[n->_idx].starts_bundle()); 826 } 827 828 //--------------------------Process_OopMap_Node-------------------------------- 829 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 830 831 // Handle special safepoint nodes for synchronization 832 MachSafePointNode *sfn = mach->as_MachSafePoint(); 833 MachCallNode *mcall; 834 835 #ifdef ENABLE_ZAP_DEAD_LOCALS 836 assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); 837 #endif 838 839 int safepoint_pc_offset = current_offset; 840 bool is_method_handle_invoke = false; 841 bool return_oop = false; 842 843 // Add the safepoint in the DebugInfoRecorder 844 if( !mach->is_MachCall() ) { 845 mcall = NULL; 846 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 847 } else { 848 mcall = mach->as_MachCall(); 849 850 // Is the call a MethodHandle call? 851 if (mcall->is_MachCallJava()) { 852 if (mcall->as_MachCallJava()->_method_handle_invoke) { 853 assert(has_method_handle_invokes(), "must have been set during call generation"); 854 is_method_handle_invoke = true; 855 } 856 } 857 858 // Check if a call returns an object. 859 if (mcall->returns_pointer()) { 860 return_oop = true; 861 } 862 safepoint_pc_offset += mcall->ret_addr_offset(); 863 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 864 } 865 866 // Loop over the JVMState list to add scope information 867 // Do not skip safepoints with a NULL method, they need monitor info 868 JVMState* youngest_jvms = sfn->jvms(); 869 int max_depth = youngest_jvms->depth(); 870 871 // Allocate the object pool for scalar-replaced objects -- the map from 872 // small-integer keys (which can be recorded in the local and ostack 873 // arrays) to descriptions of the object state. 874 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 875 876 // Visit scopes from oldest to youngest. 877 for (int depth = 1; depth <= max_depth; depth++) { 878 JVMState* jvms = youngest_jvms->of_depth(depth); 879 int idx; 880 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 881 // Safepoints that do not have method() set only provide oop-map and monitor info 882 // to support GC; these do not support deoptimization. 883 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 884 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 885 int num_mon = jvms->nof_monitors(); 886 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 887 "JVMS local count must match that of the method"); 888 889 // Add Local and Expression Stack Information 890 891 // Insert locals into the locarray 892 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 893 for( idx = 0; idx < num_locs; idx++ ) { 894 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 895 } 896 897 // Insert expression stack entries into the exparray 898 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 899 for( idx = 0; idx < num_exps; idx++ ) { 900 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 901 } 902 903 // Add in mappings of the monitors 904 assert( !method || 905 !method->is_synchronized() || 906 method->is_native() || 907 num_mon > 0 || 908 !GenerateSynchronizationCode, 909 "monitors must always exist for synchronized methods"); 910 911 // Build the growable array of ScopeValues for exp stack 912 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 913 914 // Loop over monitors and insert into array 915 for (idx = 0; idx < num_mon; idx++) { 916 // Grab the node that defines this monitor 917 Node* box_node = sfn->monitor_box(jvms, idx); 918 Node* obj_node = sfn->monitor_obj(jvms, idx); 919 920 // Create ScopeValue for object 921 ScopeValue *scval = NULL; 922 923 if (obj_node->is_SafePointScalarObject()) { 924 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 925 scval = Compile::sv_for_node_id(objs, spobj->_idx); 926 if (scval == NULL) { 927 const Type *t = spobj->bottom_type(); 928 ciKlass* cik = t->is_oopptr()->klass(); 929 assert(cik->is_instance_klass() || 930 cik->is_array_klass(), "Not supported allocation."); 931 ObjectValue* sv = new ObjectValue(spobj->_idx, 932 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 933 Compile::set_sv_for_object_node(objs, sv); 934 935 uint first_ind = spobj->first_index(youngest_jvms); 936 for (uint i = 0; i < spobj->n_fields(); i++) { 937 Node* fld_node = sfn->in(first_ind+i); 938 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 939 } 940 scval = sv; 941 } 942 } else if (!obj_node->is_Con()) { 943 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 944 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 945 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 946 } else { 947 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 948 } 949 } else { 950 const TypePtr *tp = obj_node->get_ptr_type(); 951 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 952 } 953 954 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 955 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 956 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 957 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 958 } 959 960 // We dump the object pool first, since deoptimization reads it in first. 961 debug_info()->dump_object_pool(objs); 962 963 // Build first class objects to pass to scope 964 DebugToken *locvals = debug_info()->create_scope_values(locarray); 965 DebugToken *expvals = debug_info()->create_scope_values(exparray); 966 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 967 968 // Make method available for all Safepoints 969 ciMethod* scope_method = method ? method : _method; 970 // Describe the scope here 971 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 972 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 973 // Now we can describe the scope. 974 debug_info()->describe_scope(safepoint_pc_offset, scope_method, jvms->bci(), jvms->should_reexecute(), is_method_handle_invoke, return_oop, locvals, expvals, monvals); 975 } // End jvms loop 976 977 // Mark the end of the scope set. 978 debug_info()->end_safepoint(safepoint_pc_offset); 979 } 980 981 982 983 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 984 class NonSafepointEmitter { 985 Compile* C; 986 JVMState* _pending_jvms; 987 int _pending_offset; 988 989 void emit_non_safepoint(); 990 991 public: 992 NonSafepointEmitter(Compile* compile) { 993 this->C = compile; 994 _pending_jvms = NULL; 995 _pending_offset = 0; 996 } 997 998 void observe_instruction(Node* n, int pc_offset) { 999 if (!C->debug_info()->recording_non_safepoints()) return; 1000 1001 Node_Notes* nn = C->node_notes_at(n->_idx); 1002 if (nn == NULL || nn->jvms() == NULL) return; 1003 if (_pending_jvms != NULL && 1004 _pending_jvms->same_calls_as(nn->jvms())) { 1005 // Repeated JVMS? Stretch it up here. 1006 _pending_offset = pc_offset; 1007 } else { 1008 if (_pending_jvms != NULL && 1009 _pending_offset < pc_offset) { 1010 emit_non_safepoint(); 1011 } 1012 _pending_jvms = NULL; 1013 if (pc_offset > C->debug_info()->last_pc_offset()) { 1014 // This is the only way _pending_jvms can become non-NULL: 1015 _pending_jvms = nn->jvms(); 1016 _pending_offset = pc_offset; 1017 } 1018 } 1019 } 1020 1021 // Stay out of the way of real safepoints: 1022 void observe_safepoint(JVMState* jvms, int pc_offset) { 1023 if (_pending_jvms != NULL && 1024 !_pending_jvms->same_calls_as(jvms) && 1025 _pending_offset < pc_offset) { 1026 emit_non_safepoint(); 1027 } 1028 _pending_jvms = NULL; 1029 } 1030 1031 void flush_at_end() { 1032 if (_pending_jvms != NULL) { 1033 emit_non_safepoint(); 1034 } 1035 _pending_jvms = NULL; 1036 } 1037 }; 1038 1039 void NonSafepointEmitter::emit_non_safepoint() { 1040 JVMState* youngest_jvms = _pending_jvms; 1041 int pc_offset = _pending_offset; 1042 1043 // Clear it now: 1044 _pending_jvms = NULL; 1045 1046 DebugInformationRecorder* debug_info = C->debug_info(); 1047 assert(debug_info->recording_non_safepoints(), "sanity"); 1048 1049 debug_info->add_non_safepoint(pc_offset); 1050 int max_depth = youngest_jvms->depth(); 1051 1052 // Visit scopes from oldest to youngest. 1053 for (int depth = 1; depth <= max_depth; depth++) { 1054 JVMState* jvms = youngest_jvms->of_depth(depth); 1055 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 1056 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 1057 debug_info->describe_scope(pc_offset, method, jvms->bci(), jvms->should_reexecute()); 1058 } 1059 1060 // Mark the end of the scope set. 1061 debug_info->end_non_safepoint(pc_offset); 1062 } 1063 1064 //------------------------------init_buffer------------------------------------ 1065 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 1066 1067 // Set the initially allocated size 1068 int code_req = initial_code_capacity; 1069 int locs_req = initial_locs_capacity; 1070 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 1071 int const_req = initial_const_capacity; 1072 1073 int pad_req = NativeCall::instruction_size; 1074 // The extra spacing after the code is necessary on some platforms. 1075 // Sometimes we need to patch in a jump after the last instruction, 1076 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 1077 1078 // Compute the byte offset where we can store the deopt pc. 1079 if (fixed_slots() != 0) { 1080 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 1081 } 1082 1083 // Compute prolog code size 1084 _method_size = 0; 1085 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1086 #if defined(IA64) && !defined(AIX) 1087 if (save_argument_registers()) { 1088 // 4815101: this is a stub with implicit and unknown precision fp args. 1089 // The usual spill mechanism can only generate stfd's in this case, which 1090 // doesn't work if the fp reg to spill contains a single-precision denorm. 1091 // Instead, we hack around the normal spill mechanism using stfspill's and 1092 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1093 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1094 // 1095 // If we ever implement 16-byte 'registers' == stack slots, we can 1096 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1097 // instead of stfd/stfs/ldfd/ldfs. 1098 _frame_slots += 8*(16/BytesPerInt); 1099 } 1100 #endif 1101 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1102 1103 if (has_mach_constant_base_node()) { 1104 uint add_size = 0; 1105 // Fill the constant table. 1106 // Note: This must happen before shorten_branches. 1107 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1108 Block* b = _cfg->get_block(i); 1109 1110 for (uint j = 0; j < b->number_of_nodes(); j++) { 1111 Node* n = b->get_node(j); 1112 1113 // If the node is a MachConstantNode evaluate the constant 1114 // value section. 1115 if (n->is_MachConstant()) { 1116 MachConstantNode* machcon = n->as_MachConstant(); 1117 machcon->eval_constant(C); 1118 } else if (n->is_Mach()) { 1119 // On Power there are more nodes that issue constants. 1120 add_size += (n->as_Mach()->ins_num_consts() * 8); 1121 } 1122 } 1123 } 1124 1125 // Calculate the offsets of the constants and the size of the 1126 // constant table (including the padding to the next section). 1127 constant_table().calculate_offsets_and_size(); 1128 const_req = constant_table().size() + add_size; 1129 } 1130 1131 // Initialize the space for the BufferBlob used to find and verify 1132 // instruction size in MachNode::emit_size() 1133 init_scratch_buffer_blob(const_req); 1134 if (failing()) return NULL; // Out of memory 1135 1136 // Pre-compute the length of blocks and replace 1137 // long branches with short if machine supports it. 1138 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1139 1140 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1141 // class HandlerImpl is platform-specific and defined in the *.ad files. 1142 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1143 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1144 stub_req += MAX_stubs_size; // ensure per-stub margin 1145 code_req += MAX_inst_size; // ensure per-instruction margin 1146 1147 if (StressCodeBuffers) 1148 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1149 1150 int total_req = 1151 const_req + 1152 code_req + 1153 pad_req + 1154 stub_req + 1155 exception_handler_req + 1156 deopt_handler_req; // deopt handler 1157 1158 if (has_method_handle_invokes()) 1159 total_req += deopt_handler_req; // deopt MH handler 1160 1161 CodeBuffer* cb = code_buffer(); 1162 cb->initialize(total_req, locs_req); 1163 1164 // Have we run out of code space? 1165 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1166 C->record_failure("CodeCache is full"); 1167 return NULL; 1168 } 1169 // Configure the code buffer. 1170 cb->initialize_consts_size(const_req); 1171 cb->initialize_stubs_size(stub_req); 1172 cb->initialize_oop_recorder(env()->oop_recorder()); 1173 1174 // fill in the nop array for bundling computations 1175 MachNode *_nop_list[Bundle::_nop_count]; 1176 Bundle::initialize_nops(_nop_list, this); 1177 1178 return cb; 1179 } 1180 1181 //------------------------------fill_buffer------------------------------------ 1182 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1183 // blk_starts[] contains offsets calculated during short branches processing, 1184 // offsets should not be increased during following steps. 1185 1186 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1187 // of a loop. It is used to determine the padding for loop alignment. 1188 compute_loop_first_inst_sizes(); 1189 1190 // Create oopmap set. 1191 _oop_map_set = new OopMapSet(); 1192 1193 // !!!!! This preserves old handling of oopmaps for now 1194 debug_info()->set_oopmaps(_oop_map_set); 1195 1196 uint nblocks = _cfg->number_of_blocks(); 1197 // Count and start of implicit null check instructions 1198 uint inct_cnt = 0; 1199 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1200 1201 // Count and start of calls 1202 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1203 1204 uint return_offset = 0; 1205 int nop_size = (new (this) MachNopNode())->size(_regalloc); 1206 1207 int previous_offset = 0; 1208 int current_offset = 0; 1209 int last_call_offset = -1; 1210 int last_avoid_back_to_back_offset = -1; 1211 #ifdef ASSERT 1212 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1213 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1214 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1215 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1216 #endif 1217 1218 // Create an array of unused labels, one for each basic block, if printing is enabled 1219 #ifndef PRODUCT 1220 int *node_offsets = NULL; 1221 uint node_offset_limit = unique(); 1222 1223 if (print_assembly()) 1224 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1225 #endif 1226 1227 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1228 1229 // Emit the constant table. 1230 if (has_mach_constant_base_node()) { 1231 constant_table().emit(*cb); 1232 } 1233 1234 // Create an array of labels, one for each basic block 1235 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1236 for (uint i=0; i <= nblocks; i++) { 1237 blk_labels[i].init(); 1238 } 1239 1240 // ------------------ 1241 // Now fill in the code buffer 1242 Node *delay_slot = NULL; 1243 1244 for (uint i = 0; i < nblocks; i++) { 1245 Block* block = _cfg->get_block(i); 1246 Node* head = block->head(); 1247 1248 // If this block needs to start aligned (i.e, can be reached other 1249 // than by falling-thru from the previous block), then force the 1250 // start of a new bundle. 1251 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1252 cb->flush_bundle(true); 1253 } 1254 1255 #ifdef ASSERT 1256 if (!block->is_connector()) { 1257 stringStream st; 1258 block->dump_head(_cfg, &st); 1259 MacroAssembler(cb).block_comment(st.as_string()); 1260 } 1261 jmp_target[i] = 0; 1262 jmp_offset[i] = 0; 1263 jmp_size[i] = 0; 1264 jmp_rule[i] = 0; 1265 #endif 1266 int blk_offset = current_offset; 1267 1268 // Define the label at the beginning of the basic block 1269 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1270 1271 uint last_inst = block->number_of_nodes(); 1272 1273 // Emit block normally, except for last instruction. 1274 // Emit means "dump code bits into code buffer". 1275 for (uint j = 0; j<last_inst; j++) { 1276 1277 // Get the node 1278 Node* n = block->get_node(j); 1279 1280 // See if delay slots are supported 1281 if (valid_bundle_info(n) && 1282 node_bundling(n)->used_in_unconditional_delay()) { 1283 assert(delay_slot == NULL, "no use of delay slot node"); 1284 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1285 1286 delay_slot = n; 1287 continue; 1288 } 1289 1290 // If this starts a new instruction group, then flush the current one 1291 // (but allow split bundles) 1292 if (Pipeline::requires_bundling() && starts_bundle(n)) 1293 cb->flush_bundle(false); 1294 1295 // The following logic is duplicated in the code ifdeffed for 1296 // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It 1297 // should be factored out. Or maybe dispersed to the nodes? 1298 1299 // Special handling for SafePoint/Call Nodes 1300 bool is_mcall = false; 1301 if (n->is_Mach()) { 1302 MachNode *mach = n->as_Mach(); 1303 is_mcall = n->is_MachCall(); 1304 bool is_sfn = n->is_MachSafePoint(); 1305 1306 // If this requires all previous instructions be flushed, then do so 1307 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1308 cb->flush_bundle(true); 1309 current_offset = cb->insts_size(); 1310 } 1311 1312 // A padding may be needed again since a previous instruction 1313 // could be moved to delay slot. 1314 1315 // align the instruction if necessary 1316 int padding = mach->compute_padding(current_offset); 1317 // Make sure safepoint node for polling is distinct from a call's 1318 // return by adding a nop if needed. 1319 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1320 padding = nop_size; 1321 } 1322 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1323 current_offset == last_avoid_back_to_back_offset) { 1324 // Avoid back to back some instructions. 1325 padding = nop_size; 1326 } 1327 1328 if(padding > 0) { 1329 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1330 int nops_cnt = padding / nop_size; 1331 MachNode *nop = new (this) MachNopNode(nops_cnt); 1332 block->insert_node(nop, j++); 1333 last_inst++; 1334 _cfg->map_node_to_block(nop, block); 1335 nop->emit(*cb, _regalloc); 1336 cb->flush_bundle(true); 1337 current_offset = cb->insts_size(); 1338 } 1339 1340 // Remember the start of the last call in a basic block 1341 if (is_mcall) { 1342 MachCallNode *mcall = mach->as_MachCall(); 1343 1344 // This destination address is NOT PC-relative 1345 mcall->method_set((intptr_t)mcall->entry_point()); 1346 1347 // Save the return address 1348 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1349 1350 if (mcall->is_MachCallLeaf()) { 1351 is_mcall = false; 1352 is_sfn = false; 1353 } 1354 } 1355 1356 // sfn will be valid whenever mcall is valid now because of inheritance 1357 if (is_sfn || is_mcall) { 1358 1359 // Handle special safepoint nodes for synchronization 1360 if (!is_mcall) { 1361 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1362 // !!!!! Stubs only need an oopmap right now, so bail out 1363 if (sfn->jvms()->method() == NULL) { 1364 // Write the oopmap directly to the code blob??!! 1365 # ifdef ENABLE_ZAP_DEAD_LOCALS 1366 assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); 1367 # endif 1368 continue; 1369 } 1370 } // End synchronization 1371 1372 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1373 current_offset); 1374 Process_OopMap_Node(mach, current_offset); 1375 } // End if safepoint 1376 1377 // If this is a null check, then add the start of the previous instruction to the list 1378 else if( mach->is_MachNullCheck() ) { 1379 inct_starts[inct_cnt++] = previous_offset; 1380 } 1381 1382 // If this is a branch, then fill in the label with the target BB's label 1383 else if (mach->is_MachBranch()) { 1384 // This requires the TRUE branch target be in succs[0] 1385 uint block_num = block->non_connector_successor(0)->_pre_order; 1386 1387 // Try to replace long branch if delay slot is not used, 1388 // it is mostly for back branches since forward branch's 1389 // distance is not updated yet. 1390 bool delay_slot_is_used = valid_bundle_info(n) && 1391 node_bundling(n)->use_unconditional_delay(); 1392 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1393 assert(delay_slot == NULL, "not expecting delay slot node"); 1394 int br_size = n->size(_regalloc); 1395 int offset = blk_starts[block_num] - current_offset; 1396 if (block_num >= i) { 1397 // Current and following block's offset are not 1398 // finalized yet, adjust distance by the difference 1399 // between calculated and final offsets of current block. 1400 offset -= (blk_starts[i] - blk_offset); 1401 } 1402 // In the following code a nop could be inserted before 1403 // the branch which will increase the backward distance. 1404 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1405 if (needs_padding && offset <= 0) 1406 offset -= nop_size; 1407 1408 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1409 // We've got a winner. Replace this branch. 1410 MachNode* replacement = mach->as_MachBranch()->short_branch_version(this); 1411 1412 // Update the jmp_size. 1413 int new_size = replacement->size(_regalloc); 1414 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1415 // Insert padding between avoid_back_to_back branches. 1416 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1417 MachNode *nop = new (this) MachNopNode(); 1418 block->insert_node(nop, j++); 1419 _cfg->map_node_to_block(nop, block); 1420 last_inst++; 1421 nop->emit(*cb, _regalloc); 1422 cb->flush_bundle(true); 1423 current_offset = cb->insts_size(); 1424 } 1425 #ifdef ASSERT 1426 jmp_target[i] = block_num; 1427 jmp_offset[i] = current_offset - blk_offset; 1428 jmp_size[i] = new_size; 1429 jmp_rule[i] = mach->rule(); 1430 #endif 1431 block->map_node(replacement, j); 1432 mach->subsume_by(replacement, C); 1433 n = replacement; 1434 mach = replacement; 1435 } 1436 } 1437 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1438 } else if (mach->ideal_Opcode() == Op_Jump) { 1439 for (uint h = 0; h < block->_num_succs; h++) { 1440 Block* succs_block = block->_succs[h]; 1441 for (uint j = 1; j < succs_block->num_preds(); j++) { 1442 Node* jpn = succs_block->pred(j); 1443 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1444 uint block_num = succs_block->non_connector()->_pre_order; 1445 Label *blkLabel = &blk_labels[block_num]; 1446 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1447 } 1448 } 1449 } 1450 } 1451 #ifdef ASSERT 1452 // Check that oop-store precedes the card-mark 1453 else if (mach->ideal_Opcode() == Op_StoreCM) { 1454 uint storeCM_idx = j; 1455 int count = 0; 1456 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1457 Node *oop_store = mach->in(prec); // Precedence edge 1458 if (oop_store == NULL) continue; 1459 count++; 1460 uint i4; 1461 for (i4 = 0; i4 < last_inst; ++i4) { 1462 if (block->get_node(i4) == oop_store) { 1463 break; 1464 } 1465 } 1466 // Note: This test can provide a false failure if other precedence 1467 // edges have been added to the storeCMNode. 1468 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1469 } 1470 assert(count > 0, "storeCM expects at least one precedence edge"); 1471 } 1472 #endif 1473 else if (!n->is_Proj()) { 1474 // Remember the beginning of the previous instruction, in case 1475 // it's followed by a flag-kill and a null-check. Happens on 1476 // Intel all the time, with add-to-memory kind of opcodes. 1477 previous_offset = current_offset; 1478 } 1479 1480 // Not an else-if! 1481 // If this is a trap based cmp then add its offset to the list. 1482 if (mach->is_TrapBasedCheckNode()) { 1483 inct_starts[inct_cnt++] = current_offset; 1484 } 1485 } 1486 1487 // Verify that there is sufficient space remaining 1488 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1489 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1490 C->record_failure("CodeCache is full"); 1491 return; 1492 } 1493 1494 // Save the offset for the listing 1495 #ifndef PRODUCT 1496 if (node_offsets && n->_idx < node_offset_limit) 1497 node_offsets[n->_idx] = cb->insts_size(); 1498 #endif 1499 1500 // "Normal" instruction case 1501 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1502 n->emit(*cb, _regalloc); 1503 current_offset = cb->insts_size(); 1504 1505 // Above we only verified that there is enough space in the instruction section. 1506 // However, the instruction may emit stubs that cause code buffer expansion. 1507 // Bail out here if expansion failed due to a lack of code cache space. 1508 if (failing()) { 1509 return; 1510 } 1511 1512 #ifdef ASSERT 1513 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1514 n->dump(); 1515 assert(false, "wrong size of mach node"); 1516 } 1517 #endif 1518 non_safepoints.observe_instruction(n, current_offset); 1519 1520 // mcall is last "call" that can be a safepoint 1521 // record it so we can see if a poll will directly follow it 1522 // in which case we'll need a pad to make the PcDesc sites unique 1523 // see 5010568. This can be slightly inaccurate but conservative 1524 // in the case that return address is not actually at current_offset. 1525 // This is a small price to pay. 1526 1527 if (is_mcall) { 1528 last_call_offset = current_offset; 1529 } 1530 1531 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1532 // Avoid back to back some instructions. 1533 last_avoid_back_to_back_offset = current_offset; 1534 } 1535 1536 // See if this instruction has a delay slot 1537 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1538 assert(delay_slot != NULL, "expecting delay slot node"); 1539 1540 // Back up 1 instruction 1541 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1542 1543 // Save the offset for the listing 1544 #ifndef PRODUCT 1545 if (node_offsets && delay_slot->_idx < node_offset_limit) 1546 node_offsets[delay_slot->_idx] = cb->insts_size(); 1547 #endif 1548 1549 // Support a SafePoint in the delay slot 1550 if (delay_slot->is_MachSafePoint()) { 1551 MachNode *mach = delay_slot->as_Mach(); 1552 // !!!!! Stubs only need an oopmap right now, so bail out 1553 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1554 // Write the oopmap directly to the code blob??!! 1555 # ifdef ENABLE_ZAP_DEAD_LOCALS 1556 assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); 1557 # endif 1558 delay_slot = NULL; 1559 continue; 1560 } 1561 1562 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1563 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1564 adjusted_offset); 1565 // Generate an OopMap entry 1566 Process_OopMap_Node(mach, adjusted_offset); 1567 } 1568 1569 // Insert the delay slot instruction 1570 delay_slot->emit(*cb, _regalloc); 1571 1572 // Don't reuse it 1573 delay_slot = NULL; 1574 } 1575 1576 } // End for all instructions in block 1577 1578 // If the next block is the top of a loop, pad this block out to align 1579 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1580 if (i < nblocks-1) { 1581 Block *nb = _cfg->get_block(i + 1); 1582 int padding = nb->alignment_padding(current_offset); 1583 if( padding > 0 ) { 1584 MachNode *nop = new (this) MachNopNode(padding / nop_size); 1585 block->insert_node(nop, block->number_of_nodes()); 1586 _cfg->map_node_to_block(nop, block); 1587 nop->emit(*cb, _regalloc); 1588 current_offset = cb->insts_size(); 1589 } 1590 } 1591 // Verify that the distance for generated before forward 1592 // short branches is still valid. 1593 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1594 1595 // Save new block start offset 1596 blk_starts[i] = blk_offset; 1597 } // End of for all blocks 1598 blk_starts[nblocks] = current_offset; 1599 1600 non_safepoints.flush_at_end(); 1601 1602 // Offset too large? 1603 if (failing()) return; 1604 1605 // Define a pseudo-label at the end of the code 1606 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1607 1608 // Compute the size of the first block 1609 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1610 1611 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1612 1613 #ifdef ASSERT 1614 for (uint i = 0; i < nblocks; i++) { // For all blocks 1615 if (jmp_target[i] != 0) { 1616 int br_size = jmp_size[i]; 1617 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1618 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1619 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]); 1620 assert(false, "Displacement too large for short jmp"); 1621 } 1622 } 1623 } 1624 #endif 1625 1626 #ifndef PRODUCT 1627 // Information on the size of the method, without the extraneous code 1628 Scheduling::increment_method_size(cb->insts_size()); 1629 #endif 1630 1631 // ------------------ 1632 // Fill in exception table entries. 1633 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1634 1635 // Only java methods have exception handlers and deopt handlers 1636 // class HandlerImpl is platform-specific and defined in the *.ad files. 1637 if (_method) { 1638 // Emit the exception handler code. 1639 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1640 if (failing()) { 1641 return; // CodeBuffer::expand failed 1642 } 1643 // Emit the deopt handler code. 1644 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1645 1646 // Emit the MethodHandle deopt handler code (if required). 1647 if (has_method_handle_invokes() && !failing()) { 1648 // We can use the same code as for the normal deopt handler, we 1649 // just need a different entry point address. 1650 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1651 } 1652 } 1653 1654 // One last check for failed CodeBuffer::expand: 1655 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1656 C->record_failure("CodeCache is full"); 1657 return; 1658 } 1659 1660 #ifndef PRODUCT 1661 // Dump the assembly code, including basic-block numbers 1662 if (print_assembly()) { 1663 ttyLocker ttyl; // keep the following output all in one block 1664 if (!VMThread::should_terminate()) { // test this under the tty lock 1665 // This output goes directly to the tty, not the compiler log. 1666 // To enable tools to match it up with the compilation activity, 1667 // be sure to tag this tty output with the compile ID. 1668 if (xtty != NULL) { 1669 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1670 is_osr_compilation() ? " compile_kind='osr'" : 1671 ""); 1672 } 1673 if (method() != NULL) { 1674 method()->print_metadata(); 1675 } 1676 dump_asm(node_offsets, node_offset_limit); 1677 if (xtty != NULL) { 1678 xtty->tail("opto_assembly"); 1679 } 1680 } 1681 } 1682 #endif 1683 1684 } 1685 1686 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1687 _inc_table.set_size(cnt); 1688 1689 uint inct_cnt = 0; 1690 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1691 Block* block = _cfg->get_block(i); 1692 Node *n = NULL; 1693 int j; 1694 1695 // Find the branch; ignore trailing NOPs. 1696 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1697 n = block->get_node(j); 1698 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1699 break; 1700 } 1701 } 1702 1703 // If we didn't find anything, continue 1704 if (j < 0) { 1705 continue; 1706 } 1707 1708 // Compute ExceptionHandlerTable subtable entry and add it 1709 // (skip empty blocks) 1710 if (n->is_Catch()) { 1711 1712 // Get the offset of the return from the call 1713 uint call_return = call_returns[block->_pre_order]; 1714 #ifdef ASSERT 1715 assert( call_return > 0, "no call seen for this basic block" ); 1716 while (block->get_node(--j)->is_MachProj()) ; 1717 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1718 #endif 1719 // last instruction is a CatchNode, find it's CatchProjNodes 1720 int nof_succs = block->_num_succs; 1721 // allocate space 1722 GrowableArray<intptr_t> handler_bcis(nof_succs); 1723 GrowableArray<intptr_t> handler_pcos(nof_succs); 1724 // iterate through all successors 1725 for (int j = 0; j < nof_succs; j++) { 1726 Block* s = block->_succs[j]; 1727 bool found_p = false; 1728 for (uint k = 1; k < s->num_preds(); k++) { 1729 Node* pk = s->pred(k); 1730 if (pk->is_CatchProj() && pk->in(0) == n) { 1731 const CatchProjNode* p = pk->as_CatchProj(); 1732 found_p = true; 1733 // add the corresponding handler bci & pco information 1734 if (p->_con != CatchProjNode::fall_through_index) { 1735 // p leads to an exception handler (and is not fall through) 1736 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1737 // no duplicates, please 1738 if (!handler_bcis.contains(p->handler_bci())) { 1739 uint block_num = s->non_connector()->_pre_order; 1740 handler_bcis.append(p->handler_bci()); 1741 handler_pcos.append(blk_labels[block_num].loc_pos()); 1742 } 1743 } 1744 } 1745 } 1746 assert(found_p, "no matching predecessor found"); 1747 // Note: Due to empty block removal, one block may have 1748 // several CatchProj inputs, from the same Catch. 1749 } 1750 1751 // Set the offset of the return from the call 1752 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1753 continue; 1754 } 1755 1756 // Handle implicit null exception table updates 1757 if (n->is_MachNullCheck()) { 1758 uint block_num = block->non_connector_successor(0)->_pre_order; 1759 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1760 continue; 1761 } 1762 // Handle implicit exception table updates: trap instructions. 1763 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1764 uint block_num = block->non_connector_successor(0)->_pre_order; 1765 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1766 continue; 1767 } 1768 } // End of for all blocks fill in exception table entries 1769 } 1770 1771 // Static Variables 1772 #ifndef PRODUCT 1773 uint Scheduling::_total_nop_size = 0; 1774 uint Scheduling::_total_method_size = 0; 1775 uint Scheduling::_total_branches = 0; 1776 uint Scheduling::_total_unconditional_delays = 0; 1777 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1778 #endif 1779 1780 // Initializer for class Scheduling 1781 1782 Scheduling::Scheduling(Arena *arena, Compile &compile) 1783 : _arena(arena), 1784 _cfg(compile.cfg()), 1785 _regalloc(compile.regalloc()), 1786 _reg_node(arena), 1787 _bundle_instr_count(0), 1788 _bundle_cycle_number(0), 1789 _scheduled(arena), 1790 _available(arena), 1791 _next_node(NULL), 1792 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1793 _pinch_free_list(arena) 1794 #ifndef PRODUCT 1795 , _branches(0) 1796 , _unconditional_delays(0) 1797 #endif 1798 { 1799 // Create a MachNopNode 1800 _nop = new (&compile) MachNopNode(); 1801 1802 // Now that the nops are in the array, save the count 1803 // (but allow entries for the nops) 1804 _node_bundling_limit = compile.unique(); 1805 uint node_max = _regalloc->node_regs_max_index(); 1806 1807 compile.set_node_bundling_limit(_node_bundling_limit); 1808 1809 // This one is persistent within the Compile class 1810 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1811 1812 // Allocate space for fixed-size arrays 1813 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1814 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1815 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1816 1817 // Clear the arrays 1818 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1819 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1820 memset(_uses, 0, node_max * sizeof(short)); 1821 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1822 1823 // Clear the bundling information 1824 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1825 1826 // Get the last node 1827 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1828 1829 _next_node = block->get_node(block->number_of_nodes() - 1); 1830 } 1831 1832 #ifndef PRODUCT 1833 // Scheduling destructor 1834 Scheduling::~Scheduling() { 1835 _total_branches += _branches; 1836 _total_unconditional_delays += _unconditional_delays; 1837 } 1838 #endif 1839 1840 // Step ahead "i" cycles 1841 void Scheduling::step(uint i) { 1842 1843 Bundle *bundle = node_bundling(_next_node); 1844 bundle->set_starts_bundle(); 1845 1846 // Update the bundle record, but leave the flags information alone 1847 if (_bundle_instr_count > 0) { 1848 bundle->set_instr_count(_bundle_instr_count); 1849 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1850 } 1851 1852 // Update the state information 1853 _bundle_instr_count = 0; 1854 _bundle_cycle_number += i; 1855 _bundle_use.step(i); 1856 } 1857 1858 void Scheduling::step_and_clear() { 1859 Bundle *bundle = node_bundling(_next_node); 1860 bundle->set_starts_bundle(); 1861 1862 // Update the bundle record 1863 if (_bundle_instr_count > 0) { 1864 bundle->set_instr_count(_bundle_instr_count); 1865 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1866 1867 _bundle_cycle_number += 1; 1868 } 1869 1870 // Clear the bundling information 1871 _bundle_instr_count = 0; 1872 _bundle_use.reset(); 1873 1874 memcpy(_bundle_use_elements, 1875 Pipeline_Use::elaborated_elements, 1876 sizeof(Pipeline_Use::elaborated_elements)); 1877 } 1878 1879 // Perform instruction scheduling and bundling over the sequence of 1880 // instructions in backwards order. 1881 void Compile::ScheduleAndBundle() { 1882 1883 // Don't optimize this if it isn't a method 1884 if (!_method) 1885 return; 1886 1887 // Don't optimize this if scheduling is disabled 1888 if (!do_scheduling()) 1889 return; 1890 1891 // Scheduling code works only with pairs (8 bytes) maximum. 1892 if (max_vector_size() > 8) 1893 return; 1894 1895 NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) 1896 1897 // Create a data structure for all the scheduling information 1898 Scheduling scheduling(Thread::current()->resource_area(), *this); 1899 1900 // Walk backwards over each basic block, computing the needed alignment 1901 // Walk over all the basic blocks 1902 scheduling.DoScheduling(); 1903 } 1904 1905 // Compute the latency of all the instructions. This is fairly simple, 1906 // because we already have a legal ordering. Walk over the instructions 1907 // from first to last, and compute the latency of the instruction based 1908 // on the latency of the preceding instruction(s). 1909 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1910 #ifndef PRODUCT 1911 if (_cfg->C->trace_opto_output()) 1912 tty->print("# -> ComputeLocalLatenciesForward\n"); 1913 #endif 1914 1915 // Walk over all the schedulable instructions 1916 for( uint j=_bb_start; j < _bb_end; j++ ) { 1917 1918 // This is a kludge, forcing all latency calculations to start at 1. 1919 // Used to allow latency 0 to force an instruction to the beginning 1920 // of the bb 1921 uint latency = 1; 1922 Node *use = bb->get_node(j); 1923 uint nlen = use->len(); 1924 1925 // Walk over all the inputs 1926 for ( uint k=0; k < nlen; k++ ) { 1927 Node *def = use->in(k); 1928 if (!def) 1929 continue; 1930 1931 uint l = _node_latency[def->_idx] + use->latency(k); 1932 if (latency < l) 1933 latency = l; 1934 } 1935 1936 _node_latency[use->_idx] = latency; 1937 1938 #ifndef PRODUCT 1939 if (_cfg->C->trace_opto_output()) { 1940 tty->print("# latency %4d: ", latency); 1941 use->dump(); 1942 } 1943 #endif 1944 } 1945 1946 #ifndef PRODUCT 1947 if (_cfg->C->trace_opto_output()) 1948 tty->print("# <- ComputeLocalLatenciesForward\n"); 1949 #endif 1950 1951 } // end ComputeLocalLatenciesForward 1952 1953 // See if this node fits into the present instruction bundle 1954 bool Scheduling::NodeFitsInBundle(Node *n) { 1955 uint n_idx = n->_idx; 1956 1957 // If this is the unconditional delay instruction, then it fits 1958 if (n == _unconditional_delay_slot) { 1959 #ifndef PRODUCT 1960 if (_cfg->C->trace_opto_output()) 1961 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1962 #endif 1963 return (true); 1964 } 1965 1966 // If the node cannot be scheduled this cycle, skip it 1967 if (_current_latency[n_idx] > _bundle_cycle_number) { 1968 #ifndef PRODUCT 1969 if (_cfg->C->trace_opto_output()) 1970 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1971 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1972 #endif 1973 return (false); 1974 } 1975 1976 const Pipeline *node_pipeline = n->pipeline(); 1977 1978 uint instruction_count = node_pipeline->instructionCount(); 1979 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1980 instruction_count = 0; 1981 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1982 instruction_count++; 1983 1984 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1985 #ifndef PRODUCT 1986 if (_cfg->C->trace_opto_output()) 1987 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1988 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1989 #endif 1990 return (false); 1991 } 1992 1993 // Don't allow non-machine nodes to be handled this way 1994 if (!n->is_Mach() && instruction_count == 0) 1995 return (false); 1996 1997 // See if there is any overlap 1998 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1999 2000 if (delay > 0) { 2001 #ifndef PRODUCT 2002 if (_cfg->C->trace_opto_output()) 2003 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 2004 #endif 2005 return false; 2006 } 2007 2008 #ifndef PRODUCT 2009 if (_cfg->C->trace_opto_output()) 2010 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 2011 #endif 2012 2013 return true; 2014 } 2015 2016 Node * Scheduling::ChooseNodeToBundle() { 2017 uint siz = _available.size(); 2018 2019 if (siz == 0) { 2020 2021 #ifndef PRODUCT 2022 if (_cfg->C->trace_opto_output()) 2023 tty->print("# ChooseNodeToBundle: NULL\n"); 2024 #endif 2025 return (NULL); 2026 } 2027 2028 // Fast path, if only 1 instruction in the bundle 2029 if (siz == 1) { 2030 #ifndef PRODUCT 2031 if (_cfg->C->trace_opto_output()) { 2032 tty->print("# ChooseNodeToBundle (only 1): "); 2033 _available[0]->dump(); 2034 } 2035 #endif 2036 return (_available[0]); 2037 } 2038 2039 // Don't bother, if the bundle is already full 2040 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 2041 for ( uint i = 0; i < siz; i++ ) { 2042 Node *n = _available[i]; 2043 2044 // Skip projections, we'll handle them another way 2045 if (n->is_Proj()) 2046 continue; 2047 2048 // This presupposed that instructions are inserted into the 2049 // available list in a legality order; i.e. instructions that 2050 // must be inserted first are at the head of the list 2051 if (NodeFitsInBundle(n)) { 2052 #ifndef PRODUCT 2053 if (_cfg->C->trace_opto_output()) { 2054 tty->print("# ChooseNodeToBundle: "); 2055 n->dump(); 2056 } 2057 #endif 2058 return (n); 2059 } 2060 } 2061 } 2062 2063 // Nothing fits in this bundle, choose the highest priority 2064 #ifndef PRODUCT 2065 if (_cfg->C->trace_opto_output()) { 2066 tty->print("# ChooseNodeToBundle: "); 2067 _available[0]->dump(); 2068 } 2069 #endif 2070 2071 return _available[0]; 2072 } 2073 2074 void Scheduling::AddNodeToAvailableList(Node *n) { 2075 assert( !n->is_Proj(), "projections never directly made available" ); 2076 #ifndef PRODUCT 2077 if (_cfg->C->trace_opto_output()) { 2078 tty->print("# AddNodeToAvailableList: "); 2079 n->dump(); 2080 } 2081 #endif 2082 2083 int latency = _current_latency[n->_idx]; 2084 2085 // Insert in latency order (insertion sort) 2086 uint i; 2087 for ( i=0; i < _available.size(); i++ ) 2088 if (_current_latency[_available[i]->_idx] > latency) 2089 break; 2090 2091 // Special Check for compares following branches 2092 if( n->is_Mach() && _scheduled.size() > 0 ) { 2093 int op = n->as_Mach()->ideal_Opcode(); 2094 Node *last = _scheduled[0]; 2095 if( last->is_MachIf() && last->in(1) == n && 2096 ( op == Op_CmpI || 2097 op == Op_CmpU || 2098 op == Op_CmpP || 2099 op == Op_CmpF || 2100 op == Op_CmpD || 2101 op == Op_CmpL ) ) { 2102 2103 // Recalculate position, moving to front of same latency 2104 for ( i=0 ; i < _available.size(); i++ ) 2105 if (_current_latency[_available[i]->_idx] >= latency) 2106 break; 2107 } 2108 } 2109 2110 // Insert the node in the available list 2111 _available.insert(i, n); 2112 2113 #ifndef PRODUCT 2114 if (_cfg->C->trace_opto_output()) 2115 dump_available(); 2116 #endif 2117 } 2118 2119 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2120 for ( uint i=0; i < n->len(); i++ ) { 2121 Node *def = n->in(i); 2122 if (!def) continue; 2123 if( def->is_Proj() ) // If this is a machine projection, then 2124 def = def->in(0); // propagate usage thru to the base instruction 2125 2126 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2127 continue; 2128 } 2129 2130 // Compute the latency 2131 uint l = _bundle_cycle_number + n->latency(i); 2132 if (_current_latency[def->_idx] < l) 2133 _current_latency[def->_idx] = l; 2134 2135 // If this does not have uses then schedule it 2136 if ((--_uses[def->_idx]) == 0) 2137 AddNodeToAvailableList(def); 2138 } 2139 } 2140 2141 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2142 #ifndef PRODUCT 2143 if (_cfg->C->trace_opto_output()) { 2144 tty->print("# AddNodeToBundle: "); 2145 n->dump(); 2146 } 2147 #endif 2148 2149 // Remove this from the available list 2150 uint i; 2151 for (i = 0; i < _available.size(); i++) 2152 if (_available[i] == n) 2153 break; 2154 assert(i < _available.size(), "entry in _available list not found"); 2155 _available.remove(i); 2156 2157 // See if this fits in the current bundle 2158 const Pipeline *node_pipeline = n->pipeline(); 2159 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2160 2161 // Check for instructions to be placed in the delay slot. We 2162 // do this before we actually schedule the current instruction, 2163 // because the delay slot follows the current instruction. 2164 if (Pipeline::_branch_has_delay_slot && 2165 node_pipeline->hasBranchDelay() && 2166 !_unconditional_delay_slot) { 2167 2168 uint siz = _available.size(); 2169 2170 // Conditional branches can support an instruction that 2171 // is unconditionally executed and not dependent by the 2172 // branch, OR a conditionally executed instruction if 2173 // the branch is taken. In practice, this means that 2174 // the first instruction at the branch target is 2175 // copied to the delay slot, and the branch goes to 2176 // the instruction after that at the branch target 2177 if ( n->is_MachBranch() ) { 2178 2179 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2180 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2181 2182 #ifndef PRODUCT 2183 _branches++; 2184 #endif 2185 2186 // At least 1 instruction is on the available list 2187 // that is not dependent on the branch 2188 for (uint i = 0; i < siz; i++) { 2189 Node *d = _available[i]; 2190 const Pipeline *avail_pipeline = d->pipeline(); 2191 2192 // Don't allow safepoints in the branch shadow, that will 2193 // cause a number of difficulties 2194 if ( avail_pipeline->instructionCount() == 1 && 2195 !avail_pipeline->hasMultipleBundles() && 2196 !avail_pipeline->hasBranchDelay() && 2197 Pipeline::instr_has_unit_size() && 2198 d->size(_regalloc) == Pipeline::instr_unit_size() && 2199 NodeFitsInBundle(d) && 2200 !node_bundling(d)->used_in_delay()) { 2201 2202 if (d->is_Mach() && !d->is_MachSafePoint()) { 2203 // A node that fits in the delay slot was found, so we need to 2204 // set the appropriate bits in the bundle pipeline information so 2205 // that it correctly indicates resource usage. Later, when we 2206 // attempt to add this instruction to the bundle, we will skip 2207 // setting the resource usage. 2208 _unconditional_delay_slot = d; 2209 node_bundling(n)->set_use_unconditional_delay(); 2210 node_bundling(d)->set_used_in_unconditional_delay(); 2211 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2212 _current_latency[d->_idx] = _bundle_cycle_number; 2213 _next_node = d; 2214 ++_bundle_instr_count; 2215 #ifndef PRODUCT 2216 _unconditional_delays++; 2217 #endif 2218 break; 2219 } 2220 } 2221 } 2222 } 2223 2224 // No delay slot, add a nop to the usage 2225 if (!_unconditional_delay_slot) { 2226 // See if adding an instruction in the delay slot will overflow 2227 // the bundle. 2228 if (!NodeFitsInBundle(_nop)) { 2229 #ifndef PRODUCT 2230 if (_cfg->C->trace_opto_output()) 2231 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2232 #endif 2233 step(1); 2234 } 2235 2236 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2237 _next_node = _nop; 2238 ++_bundle_instr_count; 2239 } 2240 2241 // See if the instruction in the delay slot requires a 2242 // step of the bundles 2243 if (!NodeFitsInBundle(n)) { 2244 #ifndef PRODUCT 2245 if (_cfg->C->trace_opto_output()) 2246 tty->print("# *** STEP(branch won't fit) ***\n"); 2247 #endif 2248 // Update the state information 2249 _bundle_instr_count = 0; 2250 _bundle_cycle_number += 1; 2251 _bundle_use.step(1); 2252 } 2253 } 2254 2255 // Get the number of instructions 2256 uint instruction_count = node_pipeline->instructionCount(); 2257 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2258 instruction_count = 0; 2259 2260 // Compute the latency information 2261 uint delay = 0; 2262 2263 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2264 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2265 if (relative_latency < 0) 2266 relative_latency = 0; 2267 2268 delay = _bundle_use.full_latency(relative_latency, node_usage); 2269 2270 // Does not fit in this bundle, start a new one 2271 if (delay > 0) { 2272 step(delay); 2273 2274 #ifndef PRODUCT 2275 if (_cfg->C->trace_opto_output()) 2276 tty->print("# *** STEP(%d) ***\n", delay); 2277 #endif 2278 } 2279 } 2280 2281 // If this was placed in the delay slot, ignore it 2282 if (n != _unconditional_delay_slot) { 2283 2284 if (delay == 0) { 2285 if (node_pipeline->hasMultipleBundles()) { 2286 #ifndef PRODUCT 2287 if (_cfg->C->trace_opto_output()) 2288 tty->print("# *** STEP(multiple instructions) ***\n"); 2289 #endif 2290 step(1); 2291 } 2292 2293 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2294 #ifndef PRODUCT 2295 if (_cfg->C->trace_opto_output()) 2296 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2297 instruction_count + _bundle_instr_count, 2298 Pipeline::_max_instrs_per_cycle); 2299 #endif 2300 step(1); 2301 } 2302 } 2303 2304 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2305 _bundle_instr_count++; 2306 2307 // Set the node's latency 2308 _current_latency[n->_idx] = _bundle_cycle_number; 2309 2310 // Now merge the functional unit information 2311 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2312 _bundle_use.add_usage(node_usage); 2313 2314 // Increment the number of instructions in this bundle 2315 _bundle_instr_count += instruction_count; 2316 2317 // Remember this node for later 2318 if (n->is_Mach()) 2319 _next_node = n; 2320 } 2321 2322 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2323 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2324 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2325 // into the block. All other scheduled nodes get put in the schedule here. 2326 int op = n->Opcode(); 2327 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2328 (op != Op_Node && // Not an unused antidepedence node and 2329 // not an unallocated boxlock 2330 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2331 2332 // Push any trailing projections 2333 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2334 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2335 Node *foi = n->fast_out(i); 2336 if( foi->is_Proj() ) 2337 _scheduled.push(foi); 2338 } 2339 } 2340 2341 // Put the instruction in the schedule list 2342 _scheduled.push(n); 2343 } 2344 2345 #ifndef PRODUCT 2346 if (_cfg->C->trace_opto_output()) 2347 dump_available(); 2348 #endif 2349 2350 // Walk all the definitions, decrementing use counts, and 2351 // if a definition has a 0 use count, place it in the available list. 2352 DecrementUseCounts(n,bb); 2353 } 2354 2355 // This method sets the use count within a basic block. We will ignore all 2356 // uses outside the current basic block. As we are doing a backwards walk, 2357 // any node we reach that has a use count of 0 may be scheduled. This also 2358 // avoids the problem of cyclic references from phi nodes, as long as phi 2359 // nodes are at the front of the basic block. This method also initializes 2360 // the available list to the set of instructions that have no uses within this 2361 // basic block. 2362 void Scheduling::ComputeUseCount(const Block *bb) { 2363 #ifndef PRODUCT 2364 if (_cfg->C->trace_opto_output()) 2365 tty->print("# -> ComputeUseCount\n"); 2366 #endif 2367 2368 // Clear the list of available and scheduled instructions, just in case 2369 _available.clear(); 2370 _scheduled.clear(); 2371 2372 // No delay slot specified 2373 _unconditional_delay_slot = NULL; 2374 2375 #ifdef ASSERT 2376 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2377 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2378 #endif 2379 2380 // Force the _uses count to never go to zero for unscheduable pieces 2381 // of the block 2382 for( uint k = 0; k < _bb_start; k++ ) 2383 _uses[bb->get_node(k)->_idx] = 1; 2384 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2385 _uses[bb->get_node(l)->_idx] = 1; 2386 2387 // Iterate backwards over the instructions in the block. Don't count the 2388 // branch projections at end or the block header instructions. 2389 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2390 Node *n = bb->get_node(j); 2391 if( n->is_Proj() ) continue; // Projections handled another way 2392 2393 // Account for all uses 2394 for ( uint k = 0; k < n->len(); k++ ) { 2395 Node *inp = n->in(k); 2396 if (!inp) continue; 2397 assert(inp != n, "no cycles allowed" ); 2398 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2399 if (inp->is_Proj()) { // Skip through Proj's 2400 inp = inp->in(0); 2401 } 2402 ++_uses[inp->_idx]; // Count 1 block-local use 2403 } 2404 } 2405 2406 // If this instruction has a 0 use count, then it is available 2407 if (!_uses[n->_idx]) { 2408 _current_latency[n->_idx] = _bundle_cycle_number; 2409 AddNodeToAvailableList(n); 2410 } 2411 2412 #ifndef PRODUCT 2413 if (_cfg->C->trace_opto_output()) { 2414 tty->print("# uses: %3d: ", _uses[n->_idx]); 2415 n->dump(); 2416 } 2417 #endif 2418 } 2419 2420 #ifndef PRODUCT 2421 if (_cfg->C->trace_opto_output()) 2422 tty->print("# <- ComputeUseCount\n"); 2423 #endif 2424 } 2425 2426 // This routine performs scheduling on each basic block in reverse order, 2427 // using instruction latencies and taking into account function unit 2428 // availability. 2429 void Scheduling::DoScheduling() { 2430 #ifndef PRODUCT 2431 if (_cfg->C->trace_opto_output()) 2432 tty->print("# -> DoScheduling\n"); 2433 #endif 2434 2435 Block *succ_bb = NULL; 2436 Block *bb; 2437 2438 // Walk over all the basic blocks in reverse order 2439 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2440 bb = _cfg->get_block(i); 2441 2442 #ifndef PRODUCT 2443 if (_cfg->C->trace_opto_output()) { 2444 tty->print("# Schedule BB#%03d (initial)\n", i); 2445 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2446 bb->get_node(j)->dump(); 2447 } 2448 } 2449 #endif 2450 2451 // On the head node, skip processing 2452 if (bb == _cfg->get_root_block()) { 2453 continue; 2454 } 2455 2456 // Skip empty, connector blocks 2457 if (bb->is_connector()) 2458 continue; 2459 2460 // If the following block is not the sole successor of 2461 // this one, then reset the pipeline information 2462 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2463 #ifndef PRODUCT 2464 if (_cfg->C->trace_opto_output()) { 2465 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2466 _next_node->_idx, _bundle_instr_count); 2467 } 2468 #endif 2469 step_and_clear(); 2470 } 2471 2472 // Leave untouched the starting instruction, any Phis, a CreateEx node 2473 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2474 _bb_end = bb->number_of_nodes()-1; 2475 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2476 Node *n = bb->get_node(_bb_start); 2477 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2478 // Also, MachIdealNodes do not get scheduled 2479 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2480 MachNode *mach = n->as_Mach(); 2481 int iop = mach->ideal_Opcode(); 2482 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2483 if( iop == Op_Con ) continue; // Do not schedule Top 2484 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2485 mach->pipeline() == MachNode::pipeline_class() && 2486 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2487 continue; 2488 break; // Funny loop structure to be sure... 2489 } 2490 // Compute last "interesting" instruction in block - last instruction we 2491 // might schedule. _bb_end points just after last schedulable inst. We 2492 // normally schedule conditional branches (despite them being forced last 2493 // in the block), because they have delay slots we can fill. Calls all 2494 // have their delay slots filled in the template expansions, so we don't 2495 // bother scheduling them. 2496 Node *last = bb->get_node(_bb_end); 2497 // Ignore trailing NOPs. 2498 while (_bb_end > 0 && last->is_Mach() && 2499 last->as_Mach()->ideal_Opcode() == Op_Con) { 2500 last = bb->get_node(--_bb_end); 2501 } 2502 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2503 if( last->is_Catch() || 2504 // Exclude unreachable path case when Halt node is in a separate block. 2505 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2506 // There must be a prior call. Skip it. 2507 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2508 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2509 } 2510 } else if( last->is_MachNullCheck() ) { 2511 // Backup so the last null-checked memory instruction is 2512 // outside the schedulable range. Skip over the nullcheck, 2513 // projection, and the memory nodes. 2514 Node *mem = last->in(1); 2515 do { 2516 _bb_end--; 2517 } while (mem != bb->get_node(_bb_end)); 2518 } else { 2519 // Set _bb_end to point after last schedulable inst. 2520 _bb_end++; 2521 } 2522 2523 assert( _bb_start <= _bb_end, "inverted block ends" ); 2524 2525 // Compute the register antidependencies for the basic block 2526 ComputeRegisterAntidependencies(bb); 2527 if (_cfg->C->failing()) return; // too many D-U pinch points 2528 2529 // Compute intra-bb latencies for the nodes 2530 ComputeLocalLatenciesForward(bb); 2531 2532 // Compute the usage within the block, and set the list of all nodes 2533 // in the block that have no uses within the block. 2534 ComputeUseCount(bb); 2535 2536 // Schedule the remaining instructions in the block 2537 while ( _available.size() > 0 ) { 2538 Node *n = ChooseNodeToBundle(); 2539 guarantee(n != NULL, "no nodes available"); 2540 AddNodeToBundle(n,bb); 2541 } 2542 2543 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2544 #ifdef ASSERT 2545 for( uint l = _bb_start; l < _bb_end; l++ ) { 2546 Node *n = bb->get_node(l); 2547 uint m; 2548 for( m = 0; m < _bb_end-_bb_start; m++ ) 2549 if( _scheduled[m] == n ) 2550 break; 2551 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2552 } 2553 #endif 2554 2555 // Now copy the instructions (in reverse order) back to the block 2556 for ( uint k = _bb_start; k < _bb_end; k++ ) 2557 bb->map_node(_scheduled[_bb_end-k-1], k); 2558 2559 #ifndef PRODUCT 2560 if (_cfg->C->trace_opto_output()) { 2561 tty->print("# Schedule BB#%03d (final)\n", i); 2562 uint current = 0; 2563 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2564 Node *n = bb->get_node(j); 2565 if( valid_bundle_info(n) ) { 2566 Bundle *bundle = node_bundling(n); 2567 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2568 tty->print("*** Bundle: "); 2569 bundle->dump(); 2570 } 2571 n->dump(); 2572 } 2573 } 2574 } 2575 #endif 2576 #ifdef ASSERT 2577 verify_good_schedule(bb,"after block local scheduling"); 2578 #endif 2579 } 2580 2581 #ifndef PRODUCT 2582 if (_cfg->C->trace_opto_output()) 2583 tty->print("# <- DoScheduling\n"); 2584 #endif 2585 2586 // Record final node-bundling array location 2587 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2588 2589 } // end DoScheduling 2590 2591 // Verify that no live-range used in the block is killed in the block by a 2592 // wrong DEF. This doesn't verify live-ranges that span blocks. 2593 2594 // Check for edge existence. Used to avoid adding redundant precedence edges. 2595 static bool edge_from_to( Node *from, Node *to ) { 2596 for( uint i=0; i<from->len(); i++ ) 2597 if( from->in(i) == to ) 2598 return true; 2599 return false; 2600 } 2601 2602 #ifdef ASSERT 2603 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2604 // Check for bad kills 2605 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2606 Node *prior_use = _reg_node[def]; 2607 if( prior_use && !edge_from_to(prior_use,n) ) { 2608 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2609 n->dump(); 2610 tty->print_cr("..."); 2611 prior_use->dump(); 2612 assert(edge_from_to(prior_use,n),msg); 2613 } 2614 _reg_node.map(def,NULL); // Kill live USEs 2615 } 2616 } 2617 2618 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2619 2620 // Zap to something reasonable for the verify code 2621 _reg_node.clear(); 2622 2623 // Walk over the block backwards. Check to make sure each DEF doesn't 2624 // kill a live value (other than the one it's supposed to). Add each 2625 // USE to the live set. 2626 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2627 Node *n = b->get_node(i); 2628 int n_op = n->Opcode(); 2629 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2630 // Fat-proj kills a slew of registers 2631 RegMask rm = n->out_RegMask();// Make local copy 2632 while( rm.is_NotEmpty() ) { 2633 OptoReg::Name kill = rm.find_first_elem(); 2634 rm.Remove(kill); 2635 verify_do_def( n, kill, msg ); 2636 } 2637 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2638 // Get DEF'd registers the normal way 2639 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2640 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2641 } 2642 2643 // Now make all USEs live 2644 for( uint i=1; i<n->req(); i++ ) { 2645 Node *def = n->in(i); 2646 assert(def != 0, "input edge required"); 2647 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2648 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2649 if( OptoReg::is_valid(reg_lo) ) { 2650 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2651 _reg_node.map(reg_lo,n); 2652 } 2653 if( OptoReg::is_valid(reg_hi) ) { 2654 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2655 _reg_node.map(reg_hi,n); 2656 } 2657 } 2658 2659 } 2660 2661 // Zap to something reasonable for the Antidependence code 2662 _reg_node.clear(); 2663 } 2664 #endif 2665 2666 // Conditionally add precedence edges. Avoid putting edges on Projs. 2667 static void add_prec_edge_from_to( Node *from, Node *to ) { 2668 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2669 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2670 from = from->in(0); 2671 } 2672 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2673 !edge_from_to( from, to ) ) // Avoid duplicate edge 2674 from->add_prec(to); 2675 } 2676 2677 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2678 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2679 return; 2680 2681 Node *pinch = _reg_node[def_reg]; // Get pinch point 2682 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2683 is_def ) { // Check for a true def (not a kill) 2684 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2685 return; 2686 } 2687 2688 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2689 debug_only( def = (Node*)0xdeadbeef; ) 2690 2691 // After some number of kills there _may_ be a later def 2692 Node *later_def = NULL; 2693 2694 // Finding a kill requires a real pinch-point. 2695 // Check for not already having a pinch-point. 2696 // Pinch points are Op_Node's. 2697 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2698 later_def = pinch; // Must be def/kill as optimistic pinch-point 2699 if ( _pinch_free_list.size() > 0) { 2700 pinch = _pinch_free_list.pop(); 2701 } else { 2702 pinch = new (_cfg->C) Node(1); // Pinch point to-be 2703 } 2704 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2705 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2706 return; 2707 } 2708 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2709 _reg_node.map(def_reg,pinch); // Record pinch-point 2710 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2711 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2712 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2713 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2714 later_def = NULL; // and no later def 2715 } 2716 pinch->set_req(0,later_def); // Hook later def so we can find it 2717 } else { // Else have valid pinch point 2718 if( pinch->in(0) ) // If there is a later-def 2719 later_def = pinch->in(0); // Get it 2720 } 2721 2722 // Add output-dependence edge from later def to kill 2723 if( later_def ) // If there is some original def 2724 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2725 2726 // See if current kill is also a use, and so is forced to be the pinch-point. 2727 if( pinch->Opcode() == Op_Node ) { 2728 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2729 for( uint i=1; i<uses->req(); i++ ) { 2730 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2731 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2732 // Yes, found a use/kill pinch-point 2733 pinch->set_req(0,NULL); // 2734 pinch->replace_by(kill); // Move anti-dep edges up 2735 pinch = kill; 2736 _reg_node.map(def_reg,pinch); 2737 return; 2738 } 2739 } 2740 } 2741 2742 // Add edge from kill to pinch-point 2743 add_prec_edge_from_to(kill,pinch); 2744 } 2745 2746 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2747 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2748 return; 2749 Node *pinch = _reg_node[use_reg]; // Get pinch point 2750 // Check for no later def_reg/kill in block 2751 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2752 // Use has to be block-local as well 2753 _cfg->get_block_for_node(use) == b) { 2754 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2755 pinch->req() == 1 ) { // pinch not yet in block? 2756 pinch->del_req(0); // yank pointer to later-def, also set flag 2757 // Insert the pinch-point in the block just after the last use 2758 b->insert_node(pinch, b->find_node(use) + 1); 2759 _bb_end++; // Increase size scheduled region in block 2760 } 2761 2762 add_prec_edge_from_to(pinch,use); 2763 } 2764 } 2765 2766 // We insert antidependences between the reads and following write of 2767 // allocated registers to prevent illegal code motion. Hopefully, the 2768 // number of added references should be fairly small, especially as we 2769 // are only adding references within the current basic block. 2770 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2771 2772 #ifdef ASSERT 2773 verify_good_schedule(b,"before block local scheduling"); 2774 #endif 2775 2776 // A valid schedule, for each register independently, is an endless cycle 2777 // of: a def, then some uses (connected to the def by true dependencies), 2778 // then some kills (defs with no uses), finally the cycle repeats with a new 2779 // def. The uses are allowed to float relative to each other, as are the 2780 // kills. No use is allowed to slide past a kill (or def). This requires 2781 // antidependencies between all uses of a single def and all kills that 2782 // follow, up to the next def. More edges are redundant, because later defs 2783 // & kills are already serialized with true or antidependencies. To keep 2784 // the edge count down, we add a 'pinch point' node if there's more than 2785 // one use or more than one kill/def. 2786 2787 // We add dependencies in one bottom-up pass. 2788 2789 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2790 2791 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2792 // register. If not, we record the DEF/KILL in _reg_node, the 2793 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2794 // "pinch point", a new Node that's in the graph but not in the block. 2795 // We put edges from the prior and current DEF/KILLs to the pinch point. 2796 // We put the pinch point in _reg_node. If there's already a pinch point 2797 // we merely add an edge from the current DEF/KILL to the pinch point. 2798 2799 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2800 // put an edge from the pinch point to the USE. 2801 2802 // To be expedient, the _reg_node array is pre-allocated for the whole 2803 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2804 // or a valid def/kill/pinch-point, or a leftover node from some prior 2805 // block. Leftover node from some prior block is treated like a NULL (no 2806 // prior def, so no anti-dependence needed). Valid def is distinguished by 2807 // it being in the current block. 2808 bool fat_proj_seen = false; 2809 uint last_safept = _bb_end-1; 2810 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2811 Node* last_safept_node = end_node; 2812 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2813 Node *n = b->get_node(i); 2814 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2815 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2816 // Fat-proj kills a slew of registers 2817 // This can add edges to 'n' and obscure whether or not it was a def, 2818 // hence the is_def flag. 2819 fat_proj_seen = true; 2820 RegMask rm = n->out_RegMask();// Make local copy 2821 while( rm.is_NotEmpty() ) { 2822 OptoReg::Name kill = rm.find_first_elem(); 2823 rm.Remove(kill); 2824 anti_do_def( b, n, kill, is_def ); 2825 } 2826 } else { 2827 // Get DEF'd registers the normal way 2828 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2829 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2830 } 2831 2832 // Kill projections on a branch should appear to occur on the 2833 // branch, not afterwards, so grab the masks from the projections 2834 // and process them. 2835 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2836 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2837 Node* use = n->fast_out(i); 2838 if (use->is_Proj()) { 2839 RegMask rm = use->out_RegMask();// Make local copy 2840 while( rm.is_NotEmpty() ) { 2841 OptoReg::Name kill = rm.find_first_elem(); 2842 rm.Remove(kill); 2843 anti_do_def( b, n, kill, false ); 2844 } 2845 } 2846 } 2847 } 2848 2849 // Check each register used by this instruction for a following DEF/KILL 2850 // that must occur afterward and requires an anti-dependence edge. 2851 for( uint j=0; j<n->req(); j++ ) { 2852 Node *def = n->in(j); 2853 if( def ) { 2854 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2855 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2856 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2857 } 2858 } 2859 // Do not allow defs of new derived values to float above GC 2860 // points unless the base is definitely available at the GC point. 2861 2862 Node *m = b->get_node(i); 2863 2864 // Add precedence edge from following safepoint to use of derived pointer 2865 if( last_safept_node != end_node && 2866 m != last_safept_node) { 2867 for (uint k = 1; k < m->req(); k++) { 2868 const Type *t = m->in(k)->bottom_type(); 2869 if( t->isa_oop_ptr() && 2870 t->is_ptr()->offset() != 0 ) { 2871 last_safept_node->add_prec( m ); 2872 break; 2873 } 2874 } 2875 } 2876 2877 if( n->jvms() ) { // Precedence edge from derived to safept 2878 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2879 if( b->get_node(last_safept) != last_safept_node ) { 2880 last_safept = b->find_node(last_safept_node); 2881 } 2882 for( uint j=last_safept; j > i; j-- ) { 2883 Node *mach = b->get_node(j); 2884 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2885 mach->add_prec( n ); 2886 } 2887 last_safept = i; 2888 last_safept_node = m; 2889 } 2890 } 2891 2892 if (fat_proj_seen) { 2893 // Garbage collect pinch nodes that were not consumed. 2894 // They are usually created by a fat kill MachProj for a call. 2895 garbage_collect_pinch_nodes(); 2896 } 2897 } 2898 2899 // Garbage collect pinch nodes for reuse by other blocks. 2900 // 2901 // The block scheduler's insertion of anti-dependence 2902 // edges creates many pinch nodes when the block contains 2903 // 2 or more Calls. A pinch node is used to prevent a 2904 // combinatorial explosion of edges. If a set of kills for a 2905 // register is anti-dependent on a set of uses (or defs), rather 2906 // than adding an edge in the graph between each pair of kill 2907 // and use (or def), a pinch is inserted between them: 2908 // 2909 // use1 use2 use3 2910 // \ | / 2911 // \ | / 2912 // pinch 2913 // / | \ 2914 // / | \ 2915 // kill1 kill2 kill3 2916 // 2917 // One pinch node is created per register killed when 2918 // the second call is encountered during a backwards pass 2919 // over the block. Most of these pinch nodes are never 2920 // wired into the graph because the register is never 2921 // used or def'ed in the block. 2922 // 2923 void Scheduling::garbage_collect_pinch_nodes() { 2924 #ifndef PRODUCT 2925 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2926 #endif 2927 int trace_cnt = 0; 2928 for (uint k = 0; k < _reg_node.Size(); k++) { 2929 Node* pinch = _reg_node[k]; 2930 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2931 // no predecence input edges 2932 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2933 cleanup_pinch(pinch); 2934 _pinch_free_list.push(pinch); 2935 _reg_node.map(k, NULL); 2936 #ifndef PRODUCT 2937 if (_cfg->C->trace_opto_output()) { 2938 trace_cnt++; 2939 if (trace_cnt > 40) { 2940 tty->print("\n"); 2941 trace_cnt = 0; 2942 } 2943 tty->print(" %d", pinch->_idx); 2944 } 2945 #endif 2946 } 2947 } 2948 #ifndef PRODUCT 2949 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2950 #endif 2951 } 2952 2953 // Clean up a pinch node for reuse. 2954 void Scheduling::cleanup_pinch( Node *pinch ) { 2955 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2956 2957 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2958 Node* use = pinch->last_out(i); 2959 uint uses_found = 0; 2960 for (uint j = use->req(); j < use->len(); j++) { 2961 if (use->in(j) == pinch) { 2962 use->rm_prec(j); 2963 uses_found++; 2964 } 2965 } 2966 assert(uses_found > 0, "must be a precedence edge"); 2967 i -= uses_found; // we deleted 1 or more copies of this edge 2968 } 2969 // May have a later_def entry 2970 pinch->set_req(0, NULL); 2971 } 2972 2973 #ifndef PRODUCT 2974 2975 void Scheduling::dump_available() const { 2976 tty->print("#Availist "); 2977 for (uint i = 0; i < _available.size(); i++) 2978 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2979 tty->cr(); 2980 } 2981 2982 // Print Scheduling Statistics 2983 void Scheduling::print_statistics() { 2984 // Print the size added by nops for bundling 2985 tty->print("Nops added %d bytes to total of %d bytes", 2986 _total_nop_size, _total_method_size); 2987 if (_total_method_size > 0) 2988 tty->print(", for %.2f%%", 2989 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2990 tty->print("\n"); 2991 2992 // Print the number of branch shadows filled 2993 if (Pipeline::_branch_has_delay_slot) { 2994 tty->print("Of %d branches, %d had unconditional delay slots filled", 2995 _total_branches, _total_unconditional_delays); 2996 if (_total_branches > 0) 2997 tty->print(", for %.2f%%", 2998 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2999 tty->print("\n"); 3000 } 3001 3002 uint total_instructions = 0, total_bundles = 0; 3003 3004 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 3005 uint bundle_count = _total_instructions_per_bundle[i]; 3006 total_instructions += bundle_count * i; 3007 total_bundles += bundle_count; 3008 } 3009 3010 if (total_bundles > 0) 3011 tty->print("Average ILP (excluding nops) is %.2f\n", 3012 ((double)total_instructions) / ((double)total_bundles)); 3013 } 3014 #endif