1 /* 2 * Copyright (c) 1998, 2018, 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 "asm/macroAssembler.inline.hpp" 28 #include "code/compiledIC.hpp" 29 #include "code/debugInfo.hpp" 30 #include "code/debugInfoRec.hpp" 31 #include "compiler/compileBroker.hpp" 32 #include "compiler/compilerDirectives.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "memory/allocation.inline.hpp" 35 #include "opto/ad.hpp" 36 #include "opto/callnode.hpp" 37 #include "opto/cfgnode.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/optoreg.hpp" 41 #include "opto/output.hpp" 42 #include "opto/regalloc.hpp" 43 #include "opto/runtime.hpp" 44 #include "opto/subnode.hpp" 45 #include "opto/type.hpp" 46 #include "runtime/handles.inline.hpp" 47 #include "utilities/xmlstream.hpp" 48 49 #ifndef PRODUCT 50 #define DEBUG_ARG(x) , x 51 #else 52 #define DEBUG_ARG(x) 53 #endif 54 55 // Convert Nodes to instruction bits and pass off to the VM 56 void Compile::Output() { 57 // RootNode goes 58 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 59 60 // The number of new nodes (mostly MachNop) is proportional to 61 // the number of java calls and inner loops which are aligned. 62 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 63 C->inner_loops()*(OptoLoopAlignment-1)), 64 "out of nodes before code generation" ) ) { 65 return; 66 } 67 // Make sure I can find the Start Node 68 Block *entry = _cfg->get_block(1); 69 Block *broot = _cfg->get_root_block(); 70 71 const StartNode *start = entry->head()->as_Start(); 72 73 // Replace StartNode with prolog 74 Label verified_entry; 75 MachPrologNode* prolog = new MachPrologNode(&verified_entry); 76 entry->map_node(prolog, 0); 77 _cfg->map_node_to_block(prolog, entry); 78 _cfg->unmap_node_from_block(start); // start is no longer in any block 79 80 // Virtual methods need an unverified entry point 81 if (is_osr_compilation()) { 82 if (PoisonOSREntry) { 83 // TODO: Should use a ShouldNotReachHereNode... 84 _cfg->insert( broot, 0, new MachBreakpointNode() ); 85 } 86 } else { 87 if (_method && !_method->is_static()) { 88 // Insert unvalidated entry point 89 _cfg->insert(broot, 0, new MachUEPNode()); 90 } 91 if (_method && _method->has_scalarized_args()) { 92 // Add entry point to unpack all value type arguments 93 _cfg->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ false)); 94 if (!_method->is_static()) { 95 // Add verified/unverified entry points to only unpack value type receiver at interface calls 96 _cfg->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ true, /* receiver_only */ true)); 97 _cfg->insert(broot, 0, new MachVEPNode(&verified_entry, /* verified */ false, /* receiver_only */ true)); 98 } 99 } 100 } 101 102 // Break before main entry point 103 if ((_method && C->directive()->BreakAtExecuteOption) || 104 (OptoBreakpoint && is_method_compilation()) || 105 (OptoBreakpointOSR && is_osr_compilation()) || 106 (OptoBreakpointC2R && !_method) ) { 107 // checking for _method means that OptoBreakpoint does not apply to 108 // runtime stubs or frame converters 109 _cfg->insert( entry, 1, new MachBreakpointNode() ); 110 } 111 112 // Insert epilogs before every return 113 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 114 Block* block = _cfg->get_block(i); 115 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 116 Node* m = block->end(); 117 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 118 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 119 block->add_inst(epilog); 120 _cfg->map_node_to_block(epilog, block); 121 } 122 } 123 } 124 125 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 126 blk_starts[0] = 0; 127 128 // Initialize code buffer and process short branches. 129 CodeBuffer* cb = init_buffer(blk_starts); 130 131 if (cb == NULL || failing()) { 132 return; 133 } 134 135 if (_method && _method->has_scalarized_args()) { 136 // Compute the offsets of the entry points required by the value type calling convention 137 if (!_method->is_static()) { 138 uint vep_ro_size = ((MachVEPNode*)broot->get_node(0))->size(_regalloc); 139 uint vvep_ro_size = ((MachVEPNode*)broot->get_node(1))->size(_regalloc); 140 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, vep_ro_size); 141 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, vep_ro_size + vvep_ro_size); 142 } else { 143 _code_offsets.set_value(CodeOffsets::Entry, -1); // will be patched later 144 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, 0); 145 } 146 } 147 148 ScheduleAndBundle(); 149 150 #ifndef PRODUCT 151 if (trace_opto_output()) { 152 tty->print("\n---- After ScheduleAndBundle ----\n"); 153 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 154 tty->print("\nBB#%03d:\n", i); 155 Block* block = _cfg->get_block(i); 156 for (uint j = 0; j < block->number_of_nodes(); j++) { 157 Node* n = block->get_node(j); 158 OptoReg::Name reg = _regalloc->get_reg_first(n); 159 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 160 n->dump(); 161 } 162 } 163 } 164 #endif 165 166 if (failing()) { 167 return; 168 } 169 170 BuildOopMaps(); 171 172 if (failing()) { 173 return; 174 } 175 176 fill_buffer(cb, blk_starts); 177 } 178 179 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 180 // Determine if we need to generate a stack overflow check. 181 // Do it if the method is not a stub function and 182 // has java calls or has frame size > vm_page_size/8. 183 // The debug VM checks that deoptimization doesn't trigger an 184 // unexpected stack overflow (compiled method stack banging should 185 // guarantee it doesn't happen) so we always need the stack bang in 186 // a debug VM. 187 return (UseStackBanging && stub_function() == NULL && 188 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3 189 DEBUG_ONLY(|| true))); 190 } 191 192 bool Compile::need_register_stack_bang() const { 193 // Determine if we need to generate a register stack overflow check. 194 // This is only used on architectures which have split register 195 // and memory stacks (ie. IA64). 196 // Bang if the method is not a stub function and has java calls 197 return (stub_function() == NULL && has_java_calls()); 198 } 199 200 201 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 202 // of a loop. When aligning a loop we need to provide enough instructions 203 // in cpu's fetch buffer to feed decoders. The loop alignment could be 204 // avoided if we have enough instructions in fetch buffer at the head of a loop. 205 // By default, the size is set to 999999 by Block's constructor so that 206 // a loop will be aligned if the size is not reset here. 207 // 208 // Note: Mach instructions could contain several HW instructions 209 // so the size is estimated only. 210 // 211 void Compile::compute_loop_first_inst_sizes() { 212 // The next condition is used to gate the loop alignment optimization. 213 // Don't aligned a loop if there are enough instructions at the head of a loop 214 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 215 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 216 // equal to 11 bytes which is the largest address NOP instruction. 217 if (MaxLoopPad < OptoLoopAlignment - 1) { 218 uint last_block = _cfg->number_of_blocks() - 1; 219 for (uint i = 1; i <= last_block; i++) { 220 Block* block = _cfg->get_block(i); 221 // Check the first loop's block which requires an alignment. 222 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 223 uint sum_size = 0; 224 uint inst_cnt = NumberOfLoopInstrToAlign; 225 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 226 227 // Check subsequent fallthrough blocks if the loop's first 228 // block(s) does not have enough instructions. 229 Block *nb = block; 230 while(inst_cnt > 0 && 231 i < last_block && 232 !_cfg->get_block(i + 1)->has_loop_alignment() && 233 !nb->has_successor(block)) { 234 i++; 235 nb = _cfg->get_block(i); 236 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 237 } // while( inst_cnt > 0 && i < last_block ) 238 239 block->set_first_inst_size(sum_size); 240 } // f( b->head()->is_Loop() ) 241 } // for( i <= last_block ) 242 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 243 } 244 245 // The architecture description provides short branch variants for some long 246 // branch instructions. Replace eligible long branches with short branches. 247 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 248 // Compute size of each block, method size, and relocation information size 249 uint nblocks = _cfg->number_of_blocks(); 250 251 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 252 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 253 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 254 255 // Collect worst case block paddings 256 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 257 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 258 259 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 260 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 261 262 bool has_short_branch_candidate = false; 263 264 // Initialize the sizes to 0 265 code_size = 0; // Size in bytes of generated code 266 stub_size = 0; // Size in bytes of all stub entries 267 // Size in bytes of all relocation entries, including those in local stubs. 268 // Start with 2-bytes of reloc info for the unvalidated entry point 269 reloc_size = 1; // Number of relocation entries 270 271 // Make three passes. The first computes pessimistic blk_starts, 272 // relative jmp_offset and reloc_size information. The second performs 273 // short branch substitution using the pessimistic sizing. The 274 // third inserts nops where needed. 275 276 // Step one, perform a pessimistic sizing pass. 277 uint last_call_adr = max_juint; 278 uint last_avoid_back_to_back_adr = max_juint; 279 uint nop_size = (new MachNopNode())->size(_regalloc); 280 for (uint i = 0; i < nblocks; i++) { // For all blocks 281 Block* block = _cfg->get_block(i); 282 283 // During short branch replacement, we store the relative (to blk_starts) 284 // offset of jump in jmp_offset, rather than the absolute offset of jump. 285 // This is so that we do not need to recompute sizes of all nodes when 286 // we compute correct blk_starts in our next sizing pass. 287 jmp_offset[i] = 0; 288 jmp_size[i] = 0; 289 jmp_nidx[i] = -1; 290 DEBUG_ONLY( jmp_target[i] = 0; ) 291 DEBUG_ONLY( jmp_rule[i] = 0; ) 292 293 // Sum all instruction sizes to compute block size 294 uint last_inst = block->number_of_nodes(); 295 uint blk_size = 0; 296 for (uint j = 0; j < last_inst; j++) { 297 Node* nj = block->get_node(j); 298 // Handle machine instruction nodes 299 if (nj->is_Mach()) { 300 MachNode *mach = nj->as_Mach(); 301 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 302 reloc_size += mach->reloc(); 303 if (mach->is_MachCall()) { 304 // add size information for trampoline stub 305 // class CallStubImpl is platform-specific and defined in the *.ad files. 306 stub_size += CallStubImpl::size_call_trampoline(); 307 reloc_size += CallStubImpl::reloc_call_trampoline(); 308 309 MachCallNode *mcall = mach->as_MachCall(); 310 // This destination address is NOT PC-relative 311 312 if (mcall->entry_point() != NULL) { 313 mcall->method_set((intptr_t)mcall->entry_point()); 314 } 315 316 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 317 stub_size += CompiledStaticCall::to_interp_stub_size(); 318 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 319 #if INCLUDE_AOT 320 stub_size += CompiledStaticCall::to_aot_stub_size(); 321 reloc_size += CompiledStaticCall::reloc_to_aot_stub(); 322 #endif 323 } 324 } else if (mach->is_MachSafePoint()) { 325 // If call/safepoint are adjacent, account for possible 326 // nop to disambiguate the two safepoints. 327 // ScheduleAndBundle() can rearrange nodes in a block, 328 // check for all offsets inside this block. 329 if (last_call_adr >= blk_starts[i]) { 330 blk_size += nop_size; 331 } 332 } 333 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 334 // Nop is inserted between "avoid back to back" instructions. 335 // ScheduleAndBundle() can rearrange nodes in a block, 336 // check for all offsets inside this block. 337 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 338 blk_size += nop_size; 339 } 340 } 341 if (mach->may_be_short_branch()) { 342 if (!nj->is_MachBranch()) { 343 #ifndef PRODUCT 344 nj->dump(3); 345 #endif 346 Unimplemented(); 347 } 348 assert(jmp_nidx[i] == -1, "block should have only one branch"); 349 jmp_offset[i] = blk_size; 350 jmp_size[i] = nj->size(_regalloc); 351 jmp_nidx[i] = j; 352 has_short_branch_candidate = true; 353 } 354 } 355 blk_size += nj->size(_regalloc); 356 // Remember end of call offset 357 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 358 last_call_adr = blk_starts[i]+blk_size; 359 } 360 // Remember end of avoid_back_to_back offset 361 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 362 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 363 } 364 } 365 366 // When the next block starts a loop, we may insert pad NOP 367 // instructions. Since we cannot know our future alignment, 368 // assume the worst. 369 if (i < nblocks - 1) { 370 Block* nb = _cfg->get_block(i + 1); 371 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 372 if (max_loop_pad > 0) { 373 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 374 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 375 // If either is the last instruction in this block, bump by 376 // max_loop_pad in lock-step with blk_size, so sizing 377 // calculations in subsequent blocks still can conservatively 378 // detect that it may the last instruction in this block. 379 if (last_call_adr == blk_starts[i]+blk_size) { 380 last_call_adr += max_loop_pad; 381 } 382 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 383 last_avoid_back_to_back_adr += max_loop_pad; 384 } 385 blk_size += max_loop_pad; 386 block_worst_case_pad[i + 1] = max_loop_pad; 387 } 388 } 389 390 // Save block size; update total method size 391 blk_starts[i+1] = blk_starts[i]+blk_size; 392 } 393 394 // Step two, replace eligible long jumps. 395 bool progress = true; 396 uint last_may_be_short_branch_adr = max_juint; 397 while (has_short_branch_candidate && progress) { 398 progress = false; 399 has_short_branch_candidate = false; 400 int adjust_block_start = 0; 401 for (uint i = 0; i < nblocks; i++) { 402 Block* block = _cfg->get_block(i); 403 int idx = jmp_nidx[i]; 404 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 405 if (mach != NULL && mach->may_be_short_branch()) { 406 #ifdef ASSERT 407 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 408 int j; 409 // Find the branch; ignore trailing NOPs. 410 for (j = block->number_of_nodes()-1; j>=0; j--) { 411 Node* n = block->get_node(j); 412 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 413 break; 414 } 415 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 416 #endif 417 int br_size = jmp_size[i]; 418 int br_offs = blk_starts[i] + jmp_offset[i]; 419 420 // This requires the TRUE branch target be in succs[0] 421 uint bnum = block->non_connector_successor(0)->_pre_order; 422 int offset = blk_starts[bnum] - br_offs; 423 if (bnum > i) { // adjust following block's offset 424 offset -= adjust_block_start; 425 } 426 427 // This block can be a loop header, account for the padding 428 // in the previous block. 429 int block_padding = block_worst_case_pad[i]; 430 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 431 // In the following code a nop could be inserted before 432 // the branch which will increase the backward distance. 433 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 434 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 435 436 if (needs_padding && offset <= 0) 437 offset -= nop_size; 438 439 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 440 // We've got a winner. Replace this branch. 441 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 442 443 // Update the jmp_size. 444 int new_size = replacement->size(_regalloc); 445 int diff = br_size - new_size; 446 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 447 // Conservatively take into account padding between 448 // avoid_back_to_back branches. Previous branch could be 449 // converted into avoid_back_to_back branch during next 450 // rounds. 451 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 452 jmp_offset[i] += nop_size; 453 diff -= nop_size; 454 } 455 adjust_block_start += diff; 456 block->map_node(replacement, idx); 457 mach->subsume_by(replacement, C); 458 mach = replacement; 459 progress = true; 460 461 jmp_size[i] = new_size; 462 DEBUG_ONLY( jmp_target[i] = bnum; ); 463 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 464 } else { 465 // The jump distance is not short, try again during next iteration. 466 has_short_branch_candidate = true; 467 } 468 } // (mach->may_be_short_branch()) 469 if (mach != NULL && (mach->may_be_short_branch() || 470 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 471 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 472 } 473 blk_starts[i+1] -= adjust_block_start; 474 } 475 } 476 477 #ifdef ASSERT 478 for (uint i = 0; i < nblocks; i++) { // For all blocks 479 if (jmp_target[i] != 0) { 480 int br_size = jmp_size[i]; 481 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 482 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 483 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]); 484 } 485 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 486 } 487 } 488 #endif 489 490 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 491 // after ScheduleAndBundle(). 492 493 // ------------------ 494 // Compute size for code buffer 495 code_size = blk_starts[nblocks]; 496 497 // Relocation records 498 reloc_size += 1; // Relo entry for exception handler 499 500 // Adjust reloc_size to number of record of relocation info 501 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 502 // a relocation index. 503 // The CodeBuffer will expand the locs array if this estimate is too low. 504 reloc_size *= 10 / sizeof(relocInfo); 505 } 506 507 //------------------------------FillLocArray----------------------------------- 508 // Create a bit of debug info and append it to the array. The mapping is from 509 // Java local or expression stack to constant, register or stack-slot. For 510 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 511 // entry has been taken care of and caller should skip it). 512 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 513 // This should never have accepted Bad before 514 assert(OptoReg::is_valid(regnum), "location must be valid"); 515 return (OptoReg::is_reg(regnum)) 516 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 517 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 518 } 519 520 521 ObjectValue* 522 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 523 for (int i = 0; i < objs->length(); i++) { 524 assert(objs->at(i)->is_object(), "corrupt object cache"); 525 ObjectValue* sv = (ObjectValue*) objs->at(i); 526 if (sv->id() == id) { 527 return sv; 528 } 529 } 530 // Otherwise.. 531 return NULL; 532 } 533 534 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 535 ObjectValue* sv ) { 536 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 537 objs->append(sv); 538 } 539 540 541 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 542 GrowableArray<ScopeValue*> *array, 543 GrowableArray<ScopeValue*> *objs ) { 544 assert( local, "use _top instead of null" ); 545 if (array->length() != idx) { 546 assert(array->length() == idx + 1, "Unexpected array count"); 547 // Old functionality: 548 // return 549 // New functionality: 550 // Assert if the local is not top. In product mode let the new node 551 // override the old entry. 552 assert(local == top(), "LocArray collision"); 553 if (local == top()) { 554 return; 555 } 556 array->pop(); 557 } 558 const Type *t = local->bottom_type(); 559 560 // Is it a safepoint scalar object node? 561 if (local->is_SafePointScalarObject()) { 562 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 563 564 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 565 if (sv == NULL) { 566 ciKlass* cik = t->is_oopptr()->klass(); 567 assert(cik->is_instance_klass() || 568 cik->is_array_klass(), "Not supported allocation."); 569 sv = new ObjectValue(spobj->_idx, 570 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 571 Compile::set_sv_for_object_node(objs, sv); 572 573 uint first_ind = spobj->first_index(sfpt->jvms()); 574 for (uint i = 0; i < spobj->n_fields(); i++) { 575 Node* fld_node = sfpt->in(first_ind+i); 576 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 577 } 578 } 579 array->append(sv); 580 return; 581 } 582 583 // Grab the register number for the local 584 OptoReg::Name regnum = _regalloc->get_reg_first(local); 585 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 586 // Record the double as two float registers. 587 // The register mask for such a value always specifies two adjacent 588 // float registers, with the lower register number even. 589 // Normally, the allocation of high and low words to these registers 590 // is irrelevant, because nearly all operations on register pairs 591 // (e.g., StoreD) treat them as a single unit. 592 // Here, we assume in addition that the words in these two registers 593 // stored "naturally" (by operations like StoreD and double stores 594 // within the interpreter) such that the lower-numbered register 595 // is written to the lower memory address. This may seem like 596 // a machine dependency, but it is not--it is a requirement on 597 // the author of the <arch>.ad file to ensure that, for every 598 // even/odd double-register pair to which a double may be allocated, 599 // the word in the even single-register is stored to the first 600 // memory word. (Note that register numbers are completely 601 // arbitrary, and are not tied to any machine-level encodings.) 602 #ifdef _LP64 603 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 604 array->append(new ConstantIntValue((jint)0)); 605 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 606 } else if ( t->base() == Type::Long ) { 607 array->append(new ConstantIntValue((jint)0)); 608 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 609 } else if ( t->base() == Type::RawPtr ) { 610 // jsr/ret return address which must be restored into a the full 611 // width 64-bit stack slot. 612 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 613 } 614 #else //_LP64 615 #ifdef SPARC 616 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 617 // For SPARC we have to swap high and low words for 618 // long values stored in a single-register (g0-g7). 619 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 620 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 621 } else 622 #endif //SPARC 623 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 624 // Repack the double/long as two jints. 625 // The convention the interpreter uses is that the second local 626 // holds the first raw word of the native double representation. 627 // This is actually reasonable, since locals and stack arrays 628 // grow downwards in all implementations. 629 // (If, on some machine, the interpreter's Java locals or stack 630 // were to grow upwards, the embedded doubles would be word-swapped.) 631 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 632 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 633 } 634 #endif //_LP64 635 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 636 OptoReg::is_reg(regnum) ) { 637 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 638 ? Location::float_in_dbl : Location::normal )); 639 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 640 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 641 ? Location::int_in_long : Location::normal )); 642 } else if( t->base() == Type::NarrowOop ) { 643 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 644 } else { 645 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 646 } 647 return; 648 } 649 650 // No register. It must be constant data. 651 switch (t->base()) { 652 case Type::Half: // Second half of a double 653 ShouldNotReachHere(); // Caller should skip 2nd halves 654 break; 655 case Type::AnyPtr: 656 array->append(new ConstantOopWriteValue(NULL)); 657 break; 658 case Type::AryPtr: 659 case Type::InstPtr: // fall through 660 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 661 break; 662 case Type::NarrowOop: 663 if (t == TypeNarrowOop::NULL_PTR) { 664 array->append(new ConstantOopWriteValue(NULL)); 665 } else { 666 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 667 } 668 break; 669 case Type::Int: 670 array->append(new ConstantIntValue(t->is_int()->get_con())); 671 break; 672 case Type::RawPtr: 673 // A return address (T_ADDRESS). 674 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 675 #ifdef _LP64 676 // Must be restored to the full-width 64-bit stack slot. 677 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 678 #else 679 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 680 #endif 681 break; 682 case Type::FloatCon: { 683 float f = t->is_float_constant()->getf(); 684 array->append(new ConstantIntValue(jint_cast(f))); 685 break; 686 } 687 case Type::DoubleCon: { 688 jdouble d = t->is_double_constant()->getd(); 689 #ifdef _LP64 690 array->append(new ConstantIntValue((jint)0)); 691 array->append(new ConstantDoubleValue(d)); 692 #else 693 // Repack the double as two jints. 694 // The convention the interpreter uses is that the second local 695 // holds the first raw word of the native double representation. 696 // This is actually reasonable, since locals and stack arrays 697 // grow downwards in all implementations. 698 // (If, on some machine, the interpreter's Java locals or stack 699 // were to grow upwards, the embedded doubles would be word-swapped.) 700 jlong_accessor acc; 701 acc.long_value = jlong_cast(d); 702 array->append(new ConstantIntValue(acc.words[1])); 703 array->append(new ConstantIntValue(acc.words[0])); 704 #endif 705 break; 706 } 707 case Type::Long: { 708 jlong d = t->is_long()->get_con(); 709 #ifdef _LP64 710 array->append(new ConstantIntValue((jint)0)); 711 array->append(new ConstantLongValue(d)); 712 #else 713 // Repack the long as two jints. 714 // The convention the interpreter uses is that the second local 715 // holds the first raw word of the native double representation. 716 // This is actually reasonable, since locals and stack arrays 717 // grow downwards in all implementations. 718 // (If, on some machine, the interpreter's Java locals or stack 719 // were to grow upwards, the embedded doubles would be word-swapped.) 720 jlong_accessor acc; 721 acc.long_value = d; 722 array->append(new ConstantIntValue(acc.words[1])); 723 array->append(new ConstantIntValue(acc.words[0])); 724 #endif 725 break; 726 } 727 case Type::Top: // Add an illegal value here 728 array->append(new LocationValue(Location())); 729 break; 730 default: 731 ShouldNotReachHere(); 732 break; 733 } 734 } 735 736 // Determine if this node starts a bundle 737 bool Compile::starts_bundle(const Node *n) const { 738 return (_node_bundling_limit > n->_idx && 739 _node_bundling_base[n->_idx].starts_bundle()); 740 } 741 742 //--------------------------Process_OopMap_Node-------------------------------- 743 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 744 745 // Handle special safepoint nodes for synchronization 746 MachSafePointNode *sfn = mach->as_MachSafePoint(); 747 MachCallNode *mcall; 748 749 int safepoint_pc_offset = current_offset; 750 bool is_method_handle_invoke = false; 751 bool return_oop = false; 752 bool return_vt = false; 753 754 // Add the safepoint in the DebugInfoRecorder 755 if( !mach->is_MachCall() ) { 756 mcall = NULL; 757 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 758 } else { 759 mcall = mach->as_MachCall(); 760 761 // Is the call a MethodHandle call? 762 if (mcall->is_MachCallJava()) { 763 if (mcall->as_MachCallJava()->_method_handle_invoke) { 764 assert(has_method_handle_invokes(), "must have been set during call generation"); 765 is_method_handle_invoke = true; 766 } 767 } 768 769 // Check if a call returns an object. 770 if (mcall->returns_pointer() || mcall->returns_vt()) { 771 return_oop = true; 772 } 773 if (mcall->returns_vt()) { 774 return_vt = true; 775 } 776 safepoint_pc_offset += mcall->ret_addr_offset(); 777 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 778 } 779 780 // Loop over the JVMState list to add scope information 781 // Do not skip safepoints with a NULL method, they need monitor info 782 JVMState* youngest_jvms = sfn->jvms(); 783 int max_depth = youngest_jvms->depth(); 784 785 // Allocate the object pool for scalar-replaced objects -- the map from 786 // small-integer keys (which can be recorded in the local and ostack 787 // arrays) to descriptions of the object state. 788 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 789 790 // Visit scopes from oldest to youngest. 791 for (int depth = 1; depth <= max_depth; depth++) { 792 JVMState* jvms = youngest_jvms->of_depth(depth); 793 int idx; 794 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 795 // Safepoints that do not have method() set only provide oop-map and monitor info 796 // to support GC; these do not support deoptimization. 797 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 798 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 799 int num_mon = jvms->nof_monitors(); 800 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 801 "JVMS local count must match that of the method"); 802 803 // Add Local and Expression Stack Information 804 805 // Insert locals into the locarray 806 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 807 for( idx = 0; idx < num_locs; idx++ ) { 808 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 809 } 810 811 // Insert expression stack entries into the exparray 812 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 813 for( idx = 0; idx < num_exps; idx++ ) { 814 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 815 } 816 817 // Add in mappings of the monitors 818 assert( !method || 819 !method->is_synchronized() || 820 method->is_native() || 821 num_mon > 0 || 822 !GenerateSynchronizationCode, 823 "monitors must always exist for synchronized methods"); 824 825 // Build the growable array of ScopeValues for exp stack 826 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 827 828 // Loop over monitors and insert into array 829 for (idx = 0; idx < num_mon; idx++) { 830 // Grab the node that defines this monitor 831 Node* box_node = sfn->monitor_box(jvms, idx); 832 Node* obj_node = sfn->monitor_obj(jvms, idx); 833 834 // Create ScopeValue for object 835 ScopeValue *scval = NULL; 836 837 if (obj_node->is_SafePointScalarObject()) { 838 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 839 scval = Compile::sv_for_node_id(objs, spobj->_idx); 840 if (scval == NULL) { 841 const Type *t = spobj->bottom_type(); 842 ciKlass* cik = t->is_oopptr()->klass(); 843 assert(cik->is_instance_klass() || 844 cik->is_array_klass(), "Not supported allocation."); 845 ObjectValue* sv = new ObjectValue(spobj->_idx, 846 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 847 Compile::set_sv_for_object_node(objs, sv); 848 849 uint first_ind = spobj->first_index(youngest_jvms); 850 for (uint i = 0; i < spobj->n_fields(); i++) { 851 Node* fld_node = sfn->in(first_ind+i); 852 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 853 } 854 scval = sv; 855 } 856 } else if (!obj_node->is_Con()) { 857 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 858 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 859 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 860 } else { 861 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 862 } 863 } else { 864 const TypePtr *tp = obj_node->get_ptr_type(); 865 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 866 } 867 868 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 869 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 870 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 871 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 872 } 873 874 // We dump the object pool first, since deoptimization reads it in first. 875 debug_info()->dump_object_pool(objs); 876 877 // Build first class objects to pass to scope 878 DebugToken *locvals = debug_info()->create_scope_values(locarray); 879 DebugToken *expvals = debug_info()->create_scope_values(exparray); 880 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 881 882 // Make method available for all Safepoints 883 ciMethod* scope_method = method ? method : _method; 884 // Describe the scope here 885 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 886 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 887 // Now we can describe the scope. 888 methodHandle null_mh; 889 bool rethrow_exception = false; 890 debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop, return_vt, locvals, expvals, monvals); 891 } // End jvms loop 892 893 // Mark the end of the scope set. 894 debug_info()->end_safepoint(safepoint_pc_offset); 895 } 896 897 898 899 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 900 class NonSafepointEmitter { 901 Compile* C; 902 JVMState* _pending_jvms; 903 int _pending_offset; 904 905 void emit_non_safepoint(); 906 907 public: 908 NonSafepointEmitter(Compile* compile) { 909 this->C = compile; 910 _pending_jvms = NULL; 911 _pending_offset = 0; 912 } 913 914 void observe_instruction(Node* n, int pc_offset) { 915 if (!C->debug_info()->recording_non_safepoints()) return; 916 917 Node_Notes* nn = C->node_notes_at(n->_idx); 918 if (nn == NULL || nn->jvms() == NULL) return; 919 if (_pending_jvms != NULL && 920 _pending_jvms->same_calls_as(nn->jvms())) { 921 // Repeated JVMS? Stretch it up here. 922 _pending_offset = pc_offset; 923 } else { 924 if (_pending_jvms != NULL && 925 _pending_offset < pc_offset) { 926 emit_non_safepoint(); 927 } 928 _pending_jvms = NULL; 929 if (pc_offset > C->debug_info()->last_pc_offset()) { 930 // This is the only way _pending_jvms can become non-NULL: 931 _pending_jvms = nn->jvms(); 932 _pending_offset = pc_offset; 933 } 934 } 935 } 936 937 // Stay out of the way of real safepoints: 938 void observe_safepoint(JVMState* jvms, int pc_offset) { 939 if (_pending_jvms != NULL && 940 !_pending_jvms->same_calls_as(jvms) && 941 _pending_offset < pc_offset) { 942 emit_non_safepoint(); 943 } 944 _pending_jvms = NULL; 945 } 946 947 void flush_at_end() { 948 if (_pending_jvms != NULL) { 949 emit_non_safepoint(); 950 } 951 _pending_jvms = NULL; 952 } 953 }; 954 955 void NonSafepointEmitter::emit_non_safepoint() { 956 JVMState* youngest_jvms = _pending_jvms; 957 int pc_offset = _pending_offset; 958 959 // Clear it now: 960 _pending_jvms = NULL; 961 962 DebugInformationRecorder* debug_info = C->debug_info(); 963 assert(debug_info->recording_non_safepoints(), "sanity"); 964 965 debug_info->add_non_safepoint(pc_offset); 966 int max_depth = youngest_jvms->depth(); 967 968 // Visit scopes from oldest to youngest. 969 for (int depth = 1; depth <= max_depth; depth++) { 970 JVMState* jvms = youngest_jvms->of_depth(depth); 971 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 972 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 973 methodHandle null_mh; 974 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 975 } 976 977 // Mark the end of the scope set. 978 debug_info->end_non_safepoint(pc_offset); 979 } 980 981 //------------------------------init_buffer------------------------------------ 982 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 983 984 // Set the initially allocated size 985 int code_req = initial_code_capacity; 986 int locs_req = initial_locs_capacity; 987 int stub_req = initial_stub_capacity; 988 int const_req = initial_const_capacity; 989 990 int pad_req = NativeCall::instruction_size; 991 // The extra spacing after the code is necessary on some platforms. 992 // Sometimes we need to patch in a jump after the last instruction, 993 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 994 995 // Compute the byte offset where we can store the deopt pc. 996 if (fixed_slots() != 0) { 997 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 998 } 999 if (C->needs_stack_repair()) { 1000 // Compute the byte offset of the stack increment value 1001 _sp_inc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_sp_inc_slot)); 1002 } 1003 1004 // Compute prolog code size 1005 _method_size = 0; 1006 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 1007 #if defined(IA64) && !defined(AIX) 1008 if (save_argument_registers()) { 1009 // 4815101: this is a stub with implicit and unknown precision fp args. 1010 // The usual spill mechanism can only generate stfd's in this case, which 1011 // doesn't work if the fp reg to spill contains a single-precision denorm. 1012 // Instead, we hack around the normal spill mechanism using stfspill's and 1013 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 1014 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 1015 // 1016 // If we ever implement 16-byte 'registers' == stack slots, we can 1017 // get rid of this hack and have SpillCopy generate stfspill/ldffill 1018 // instead of stfd/stfs/ldfd/ldfs. 1019 _frame_slots += 8*(16/BytesPerInt); 1020 } 1021 #endif 1022 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 1023 1024 if (has_mach_constant_base_node()) { 1025 uint add_size = 0; 1026 // Fill the constant table. 1027 // Note: This must happen before shorten_branches. 1028 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1029 Block* b = _cfg->get_block(i); 1030 1031 for (uint j = 0; j < b->number_of_nodes(); j++) { 1032 Node* n = b->get_node(j); 1033 1034 // If the node is a MachConstantNode evaluate the constant 1035 // value section. 1036 if (n->is_MachConstant()) { 1037 MachConstantNode* machcon = n->as_MachConstant(); 1038 machcon->eval_constant(C); 1039 } else if (n->is_Mach()) { 1040 // On Power there are more nodes that issue constants. 1041 add_size += (n->as_Mach()->ins_num_consts() * 8); 1042 } 1043 } 1044 } 1045 1046 // Calculate the offsets of the constants and the size of the 1047 // constant table (including the padding to the next section). 1048 constant_table().calculate_offsets_and_size(); 1049 const_req = constant_table().size() + add_size; 1050 } 1051 1052 // Initialize the space for the BufferBlob used to find and verify 1053 // instruction size in MachNode::emit_size() 1054 init_scratch_buffer_blob(const_req); 1055 if (failing()) return NULL; // Out of memory 1056 1057 // Pre-compute the length of blocks and replace 1058 // long branches with short if machine supports it. 1059 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1060 1061 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1062 // class HandlerImpl is platform-specific and defined in the *.ad files. 1063 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1064 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1065 stub_req += MAX_stubs_size; // ensure per-stub margin 1066 code_req += MAX_inst_size; // ensure per-instruction margin 1067 1068 if (StressCodeBuffers) 1069 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1070 1071 int total_req = 1072 const_req + 1073 code_req + 1074 pad_req + 1075 stub_req + 1076 exception_handler_req + 1077 deopt_handler_req; // deopt handler 1078 1079 if (has_method_handle_invokes()) 1080 total_req += deopt_handler_req; // deopt MH handler 1081 1082 CodeBuffer* cb = code_buffer(); 1083 cb->initialize(total_req, locs_req); 1084 1085 // Have we run out of code space? 1086 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1087 C->record_failure("CodeCache is full"); 1088 return NULL; 1089 } 1090 // Configure the code buffer. 1091 cb->initialize_consts_size(const_req); 1092 cb->initialize_stubs_size(stub_req); 1093 cb->initialize_oop_recorder(env()->oop_recorder()); 1094 1095 // fill in the nop array for bundling computations 1096 MachNode *_nop_list[Bundle::_nop_count]; 1097 Bundle::initialize_nops(_nop_list); 1098 1099 return cb; 1100 } 1101 1102 //------------------------------fill_buffer------------------------------------ 1103 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1104 // blk_starts[] contains offsets calculated during short branches processing, 1105 // offsets should not be increased during following steps. 1106 1107 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1108 // of a loop. It is used to determine the padding for loop alignment. 1109 compute_loop_first_inst_sizes(); 1110 1111 // Create oopmap set. 1112 _oop_map_set = new OopMapSet(); 1113 1114 // !!!!! This preserves old handling of oopmaps for now 1115 debug_info()->set_oopmaps(_oop_map_set); 1116 1117 uint nblocks = _cfg->number_of_blocks(); 1118 // Count and start of implicit null check instructions 1119 uint inct_cnt = 0; 1120 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1121 1122 // Count and start of calls 1123 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1124 1125 uint return_offset = 0; 1126 int nop_size = (new MachNopNode())->size(_regalloc); 1127 1128 int previous_offset = 0; 1129 int current_offset = 0; 1130 int last_call_offset = -1; 1131 int last_avoid_back_to_back_offset = -1; 1132 #ifdef ASSERT 1133 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1134 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1135 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1136 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1137 #endif 1138 1139 // Create an array of unused labels, one for each basic block, if printing is enabled 1140 #ifndef PRODUCT 1141 int *node_offsets = NULL; 1142 uint node_offset_limit = unique(); 1143 1144 if (print_assembly()) 1145 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1146 #endif 1147 1148 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1149 1150 // Emit the constant table. 1151 if (has_mach_constant_base_node()) { 1152 constant_table().emit(*cb); 1153 } 1154 1155 // Create an array of labels, one for each basic block 1156 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1157 for (uint i=0; i <= nblocks; i++) { 1158 blk_labels[i].init(); 1159 } 1160 1161 // ------------------ 1162 // Now fill in the code buffer 1163 Node *delay_slot = NULL; 1164 1165 for (uint i = 0; i < nblocks; i++) { 1166 Block* block = _cfg->get_block(i); 1167 Node* head = block->head(); 1168 1169 // If this block needs to start aligned (i.e, can be reached other 1170 // than by falling-thru from the previous block), then force the 1171 // start of a new bundle. 1172 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1173 cb->flush_bundle(true); 1174 } 1175 1176 #ifdef ASSERT 1177 if (!block->is_connector()) { 1178 stringStream st; 1179 block->dump_head(_cfg, &st); 1180 MacroAssembler(cb).block_comment(st.as_string()); 1181 } 1182 jmp_target[i] = 0; 1183 jmp_offset[i] = 0; 1184 jmp_size[i] = 0; 1185 jmp_rule[i] = 0; 1186 #endif 1187 int blk_offset = current_offset; 1188 1189 // Define the label at the beginning of the basic block 1190 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1191 1192 uint last_inst = block->number_of_nodes(); 1193 1194 // Emit block normally, except for last instruction. 1195 // Emit means "dump code bits into code buffer". 1196 for (uint j = 0; j<last_inst; j++) { 1197 1198 // Get the node 1199 Node* n = block->get_node(j); 1200 1201 // See if delay slots are supported 1202 if (valid_bundle_info(n) && 1203 node_bundling(n)->used_in_unconditional_delay()) { 1204 assert(delay_slot == NULL, "no use of delay slot node"); 1205 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1206 1207 delay_slot = n; 1208 continue; 1209 } 1210 1211 // If this starts a new instruction group, then flush the current one 1212 // (but allow split bundles) 1213 if (Pipeline::requires_bundling() && starts_bundle(n)) 1214 cb->flush_bundle(false); 1215 1216 // Special handling for SafePoint/Call Nodes 1217 bool is_mcall = false; 1218 if (n->is_Mach()) { 1219 MachNode *mach = n->as_Mach(); 1220 is_mcall = n->is_MachCall(); 1221 bool is_sfn = n->is_MachSafePoint(); 1222 1223 // If this requires all previous instructions be flushed, then do so 1224 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1225 cb->flush_bundle(true); 1226 current_offset = cb->insts_size(); 1227 } 1228 1229 // A padding may be needed again since a previous instruction 1230 // could be moved to delay slot. 1231 1232 // align the instruction if necessary 1233 int padding = mach->compute_padding(current_offset); 1234 // Make sure safepoint node for polling is distinct from a call's 1235 // return by adding a nop if needed. 1236 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1237 padding = nop_size; 1238 } 1239 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1240 current_offset == last_avoid_back_to_back_offset) { 1241 // Avoid back to back some instructions. 1242 padding = nop_size; 1243 } 1244 1245 if (padding > 0) { 1246 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1247 int nops_cnt = padding / nop_size; 1248 MachNode *nop = new MachNopNode(nops_cnt); 1249 block->insert_node(nop, j++); 1250 last_inst++; 1251 _cfg->map_node_to_block(nop, block); 1252 // Ensure enough space. 1253 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1254 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1255 C->record_failure("CodeCache is full"); 1256 return; 1257 } 1258 nop->emit(*cb, _regalloc); 1259 cb->flush_bundle(true); 1260 current_offset = cb->insts_size(); 1261 } 1262 1263 // Remember the start of the last call in a basic block 1264 if (is_mcall) { 1265 MachCallNode *mcall = mach->as_MachCall(); 1266 1267 if (mcall->entry_point() != NULL) { 1268 // This destination address is NOT PC-relative 1269 mcall->method_set((intptr_t)mcall->entry_point()); 1270 } 1271 1272 // Save the return address 1273 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1274 1275 if (mcall->is_MachCallLeaf()) { 1276 is_mcall = false; 1277 is_sfn = false; 1278 } 1279 } 1280 1281 // sfn will be valid whenever mcall is valid now because of inheritance 1282 if (is_sfn || is_mcall) { 1283 1284 // Handle special safepoint nodes for synchronization 1285 if (!is_mcall) { 1286 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1287 // !!!!! Stubs only need an oopmap right now, so bail out 1288 if (sfn->jvms()->method() == NULL) { 1289 // Write the oopmap directly to the code blob??!! 1290 continue; 1291 } 1292 } // End synchronization 1293 1294 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1295 current_offset); 1296 Process_OopMap_Node(mach, current_offset); 1297 } // End if safepoint 1298 1299 // If this is a null check, then add the start of the previous instruction to the list 1300 else if( mach->is_MachNullCheck() ) { 1301 inct_starts[inct_cnt++] = previous_offset; 1302 } 1303 1304 // If this is a branch, then fill in the label with the target BB's label 1305 else if (mach->is_MachBranch()) { 1306 // This requires the TRUE branch target be in succs[0] 1307 uint block_num = block->non_connector_successor(0)->_pre_order; 1308 1309 // Try to replace long branch if delay slot is not used, 1310 // it is mostly for back branches since forward branch's 1311 // distance is not updated yet. 1312 bool delay_slot_is_used = valid_bundle_info(n) && 1313 node_bundling(n)->use_unconditional_delay(); 1314 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1315 assert(delay_slot == NULL, "not expecting delay slot node"); 1316 int br_size = n->size(_regalloc); 1317 int offset = blk_starts[block_num] - current_offset; 1318 if (block_num >= i) { 1319 // Current and following block's offset are not 1320 // finalized yet, adjust distance by the difference 1321 // between calculated and final offsets of current block. 1322 offset -= (blk_starts[i] - blk_offset); 1323 } 1324 // In the following code a nop could be inserted before 1325 // the branch which will increase the backward distance. 1326 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1327 if (needs_padding && offset <= 0) 1328 offset -= nop_size; 1329 1330 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1331 // We've got a winner. Replace this branch. 1332 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1333 1334 // Update the jmp_size. 1335 int new_size = replacement->size(_regalloc); 1336 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1337 // Insert padding between avoid_back_to_back branches. 1338 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1339 MachNode *nop = new MachNopNode(); 1340 block->insert_node(nop, j++); 1341 _cfg->map_node_to_block(nop, block); 1342 last_inst++; 1343 nop->emit(*cb, _regalloc); 1344 cb->flush_bundle(true); 1345 current_offset = cb->insts_size(); 1346 } 1347 #ifdef ASSERT 1348 jmp_target[i] = block_num; 1349 jmp_offset[i] = current_offset - blk_offset; 1350 jmp_size[i] = new_size; 1351 jmp_rule[i] = mach->rule(); 1352 #endif 1353 block->map_node(replacement, j); 1354 mach->subsume_by(replacement, C); 1355 n = replacement; 1356 mach = replacement; 1357 } 1358 } 1359 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1360 } else if (mach->ideal_Opcode() == Op_Jump) { 1361 for (uint h = 0; h < block->_num_succs; h++) { 1362 Block* succs_block = block->_succs[h]; 1363 for (uint j = 1; j < succs_block->num_preds(); j++) { 1364 Node* jpn = succs_block->pred(j); 1365 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1366 uint block_num = succs_block->non_connector()->_pre_order; 1367 Label *blkLabel = &blk_labels[block_num]; 1368 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1369 } 1370 } 1371 } 1372 } 1373 #ifdef ASSERT 1374 // Check that oop-store precedes the card-mark 1375 else if (mach->ideal_Opcode() == Op_StoreCM) { 1376 uint storeCM_idx = j; 1377 int count = 0; 1378 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1379 Node *oop_store = mach->in(prec); // Precedence edge 1380 if (oop_store == NULL) continue; 1381 count++; 1382 uint i4; 1383 for (i4 = 0; i4 < last_inst; ++i4) { 1384 if (block->get_node(i4) == oop_store) { 1385 break; 1386 } 1387 } 1388 // Note: This test can provide a false failure if other precedence 1389 // edges have been added to the storeCMNode. 1390 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1391 } 1392 assert(count > 0, "storeCM expects at least one precedence edge"); 1393 } 1394 #endif 1395 else if (!n->is_Proj()) { 1396 // Remember the beginning of the previous instruction, in case 1397 // it's followed by a flag-kill and a null-check. Happens on 1398 // Intel all the time, with add-to-memory kind of opcodes. 1399 previous_offset = current_offset; 1400 } 1401 1402 // Not an else-if! 1403 // If this is a trap based cmp then add its offset to the list. 1404 if (mach->is_TrapBasedCheckNode()) { 1405 inct_starts[inct_cnt++] = current_offset; 1406 } 1407 } 1408 1409 // Verify that there is sufficient space remaining 1410 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1411 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1412 C->record_failure("CodeCache is full"); 1413 return; 1414 } 1415 1416 // Save the offset for the listing 1417 #ifndef PRODUCT 1418 if (node_offsets && n->_idx < node_offset_limit) 1419 node_offsets[n->_idx] = cb->insts_size(); 1420 #endif 1421 1422 // "Normal" instruction case 1423 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1424 n->emit(*cb, _regalloc); 1425 current_offset = cb->insts_size(); 1426 1427 // Above we only verified that there is enough space in the instruction section. 1428 // However, the instruction may emit stubs that cause code buffer expansion. 1429 // Bail out here if expansion failed due to a lack of code cache space. 1430 if (failing()) { 1431 return; 1432 } 1433 1434 #ifdef ASSERT 1435 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1436 n->dump(); 1437 assert(false, "wrong size of mach node"); 1438 } 1439 #endif 1440 non_safepoints.observe_instruction(n, current_offset); 1441 1442 // mcall is last "call" that can be a safepoint 1443 // record it so we can see if a poll will directly follow it 1444 // in which case we'll need a pad to make the PcDesc sites unique 1445 // see 5010568. This can be slightly inaccurate but conservative 1446 // in the case that return address is not actually at current_offset. 1447 // This is a small price to pay. 1448 1449 if (is_mcall) { 1450 last_call_offset = current_offset; 1451 } 1452 1453 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1454 // Avoid back to back some instructions. 1455 last_avoid_back_to_back_offset = current_offset; 1456 } 1457 1458 // See if this instruction has a delay slot 1459 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1460 guarantee(delay_slot != NULL, "expecting delay slot node"); 1461 1462 // Back up 1 instruction 1463 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1464 1465 // Save the offset for the listing 1466 #ifndef PRODUCT 1467 if (node_offsets && delay_slot->_idx < node_offset_limit) 1468 node_offsets[delay_slot->_idx] = cb->insts_size(); 1469 #endif 1470 1471 // Support a SafePoint in the delay slot 1472 if (delay_slot->is_MachSafePoint()) { 1473 MachNode *mach = delay_slot->as_Mach(); 1474 // !!!!! Stubs only need an oopmap right now, so bail out 1475 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1476 // Write the oopmap directly to the code blob??!! 1477 delay_slot = NULL; 1478 continue; 1479 } 1480 1481 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1482 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1483 adjusted_offset); 1484 // Generate an OopMap entry 1485 Process_OopMap_Node(mach, adjusted_offset); 1486 } 1487 1488 // Insert the delay slot instruction 1489 delay_slot->emit(*cb, _regalloc); 1490 1491 // Don't reuse it 1492 delay_slot = NULL; 1493 } 1494 1495 } // End for all instructions in block 1496 1497 // If the next block is the top of a loop, pad this block out to align 1498 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1499 if (i < nblocks-1) { 1500 Block *nb = _cfg->get_block(i + 1); 1501 int padding = nb->alignment_padding(current_offset); 1502 if( padding > 0 ) { 1503 MachNode *nop = new MachNopNode(padding / nop_size); 1504 block->insert_node(nop, block->number_of_nodes()); 1505 _cfg->map_node_to_block(nop, block); 1506 nop->emit(*cb, _regalloc); 1507 current_offset = cb->insts_size(); 1508 } 1509 } 1510 // Verify that the distance for generated before forward 1511 // short branches is still valid. 1512 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1513 1514 // Save new block start offset 1515 blk_starts[i] = blk_offset; 1516 } // End of for all blocks 1517 blk_starts[nblocks] = current_offset; 1518 1519 non_safepoints.flush_at_end(); 1520 1521 // Offset too large? 1522 if (failing()) return; 1523 1524 // Define a pseudo-label at the end of the code 1525 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1526 1527 // Compute the size of the first block 1528 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1529 1530 #ifdef ASSERT 1531 for (uint i = 0; i < nblocks; i++) { // For all blocks 1532 if (jmp_target[i] != 0) { 1533 int br_size = jmp_size[i]; 1534 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1535 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1536 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]); 1537 assert(false, "Displacement too large for short jmp"); 1538 } 1539 } 1540 } 1541 #endif 1542 1543 #ifndef PRODUCT 1544 // Information on the size of the method, without the extraneous code 1545 Scheduling::increment_method_size(cb->insts_size()); 1546 #endif 1547 1548 // ------------------ 1549 // Fill in exception table entries. 1550 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1551 1552 // Only java methods have exception handlers and deopt handlers 1553 // class HandlerImpl is platform-specific and defined in the *.ad files. 1554 if (_method) { 1555 // Emit the exception handler code. 1556 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1557 if (failing()) { 1558 return; // CodeBuffer::expand failed 1559 } 1560 // Emit the deopt handler code. 1561 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1562 1563 // Emit the MethodHandle deopt handler code (if required). 1564 if (has_method_handle_invokes() && !failing()) { 1565 // We can use the same code as for the normal deopt handler, we 1566 // just need a different entry point address. 1567 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1568 } 1569 } 1570 1571 // One last check for failed CodeBuffer::expand: 1572 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1573 C->record_failure("CodeCache is full"); 1574 return; 1575 } 1576 1577 #ifndef PRODUCT 1578 // Dump the assembly code, including basic-block numbers 1579 if (print_assembly()) { 1580 ttyLocker ttyl; // keep the following output all in one block 1581 if (!VMThread::should_terminate()) { // test this under the tty lock 1582 // This output goes directly to the tty, not the compiler log. 1583 // To enable tools to match it up with the compilation activity, 1584 // be sure to tag this tty output with the compile ID. 1585 if (xtty != NULL) { 1586 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1587 is_osr_compilation() ? " compile_kind='osr'" : 1588 ""); 1589 } 1590 if (method() != NULL) { 1591 method()->print_metadata(); 1592 } 1593 dump_asm(node_offsets, node_offset_limit); 1594 if (xtty != NULL) { 1595 // print_metadata and dump_asm above may safepoint which makes us loose the ttylock. 1596 // Retake lock too make sure the end tag is coherent, and that xmlStream->pop_tag is done 1597 // thread safe 1598 ttyLocker ttyl2; 1599 xtty->tail("opto_assembly"); 1600 } 1601 } 1602 } 1603 #endif 1604 1605 } 1606 1607 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1608 _inc_table.set_size(cnt); 1609 1610 uint inct_cnt = 0; 1611 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1612 Block* block = _cfg->get_block(i); 1613 Node *n = NULL; 1614 int j; 1615 1616 // Find the branch; ignore trailing NOPs. 1617 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1618 n = block->get_node(j); 1619 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1620 break; 1621 } 1622 } 1623 1624 // If we didn't find anything, continue 1625 if (j < 0) { 1626 continue; 1627 } 1628 1629 // Compute ExceptionHandlerTable subtable entry and add it 1630 // (skip empty blocks) 1631 if (n->is_Catch()) { 1632 1633 // Get the offset of the return from the call 1634 uint call_return = call_returns[block->_pre_order]; 1635 #ifdef ASSERT 1636 assert( call_return > 0, "no call seen for this basic block" ); 1637 while (block->get_node(--j)->is_MachProj()) ; 1638 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1639 #endif 1640 // last instruction is a CatchNode, find it's CatchProjNodes 1641 int nof_succs = block->_num_succs; 1642 // allocate space 1643 GrowableArray<intptr_t> handler_bcis(nof_succs); 1644 GrowableArray<intptr_t> handler_pcos(nof_succs); 1645 // iterate through all successors 1646 for (int j = 0; j < nof_succs; j++) { 1647 Block* s = block->_succs[j]; 1648 bool found_p = false; 1649 for (uint k = 1; k < s->num_preds(); k++) { 1650 Node* pk = s->pred(k); 1651 if (pk->is_CatchProj() && pk->in(0) == n) { 1652 const CatchProjNode* p = pk->as_CatchProj(); 1653 found_p = true; 1654 // add the corresponding handler bci & pco information 1655 if (p->_con != CatchProjNode::fall_through_index) { 1656 // p leads to an exception handler (and is not fall through) 1657 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1658 // no duplicates, please 1659 if (!handler_bcis.contains(p->handler_bci())) { 1660 uint block_num = s->non_connector()->_pre_order; 1661 handler_bcis.append(p->handler_bci()); 1662 handler_pcos.append(blk_labels[block_num].loc_pos()); 1663 } 1664 } 1665 } 1666 } 1667 assert(found_p, "no matching predecessor found"); 1668 // Note: Due to empty block removal, one block may have 1669 // several CatchProj inputs, from the same Catch. 1670 } 1671 1672 // Set the offset of the return from the call 1673 assert(handler_bcis.find(-1) != -1, "must have default handler"); 1674 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1675 continue; 1676 } 1677 1678 // Handle implicit null exception table updates 1679 if (n->is_MachNullCheck()) { 1680 uint block_num = block->non_connector_successor(0)->_pre_order; 1681 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1682 continue; 1683 } 1684 // Handle implicit exception table updates: trap instructions. 1685 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1686 uint block_num = block->non_connector_successor(0)->_pre_order; 1687 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1688 continue; 1689 } 1690 } // End of for all blocks fill in exception table entries 1691 } 1692 1693 // Static Variables 1694 #ifndef PRODUCT 1695 uint Scheduling::_total_nop_size = 0; 1696 uint Scheduling::_total_method_size = 0; 1697 uint Scheduling::_total_branches = 0; 1698 uint Scheduling::_total_unconditional_delays = 0; 1699 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1700 #endif 1701 1702 // Initializer for class Scheduling 1703 1704 Scheduling::Scheduling(Arena *arena, Compile &compile) 1705 : _arena(arena), 1706 _cfg(compile.cfg()), 1707 _regalloc(compile.regalloc()), 1708 _scheduled(arena), 1709 _available(arena), 1710 _reg_node(arena), 1711 _pinch_free_list(arena), 1712 _next_node(NULL), 1713 _bundle_instr_count(0), 1714 _bundle_cycle_number(0), 1715 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]) 1716 #ifndef PRODUCT 1717 , _branches(0) 1718 , _unconditional_delays(0) 1719 #endif 1720 { 1721 // Create a MachNopNode 1722 _nop = new MachNopNode(); 1723 1724 // Now that the nops are in the array, save the count 1725 // (but allow entries for the nops) 1726 _node_bundling_limit = compile.unique(); 1727 uint node_max = _regalloc->node_regs_max_index(); 1728 1729 compile.set_node_bundling_limit(_node_bundling_limit); 1730 1731 // This one is persistent within the Compile class 1732 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1733 1734 // Allocate space for fixed-size arrays 1735 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1736 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1737 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1738 1739 // Clear the arrays 1740 for (uint i = 0; i < node_max; i++) { 1741 ::new (&_node_bundling_base[i]) Bundle(); 1742 } 1743 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1744 memset(_uses, 0, node_max * sizeof(short)); 1745 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1746 1747 // Clear the bundling information 1748 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1749 1750 // Get the last node 1751 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1752 1753 _next_node = block->get_node(block->number_of_nodes() - 1); 1754 } 1755 1756 #ifndef PRODUCT 1757 // Scheduling destructor 1758 Scheduling::~Scheduling() { 1759 _total_branches += _branches; 1760 _total_unconditional_delays += _unconditional_delays; 1761 } 1762 #endif 1763 1764 // Step ahead "i" cycles 1765 void Scheduling::step(uint i) { 1766 1767 Bundle *bundle = node_bundling(_next_node); 1768 bundle->set_starts_bundle(); 1769 1770 // Update the bundle record, but leave the flags information alone 1771 if (_bundle_instr_count > 0) { 1772 bundle->set_instr_count(_bundle_instr_count); 1773 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1774 } 1775 1776 // Update the state information 1777 _bundle_instr_count = 0; 1778 _bundle_cycle_number += i; 1779 _bundle_use.step(i); 1780 } 1781 1782 void Scheduling::step_and_clear() { 1783 Bundle *bundle = node_bundling(_next_node); 1784 bundle->set_starts_bundle(); 1785 1786 // Update the bundle record 1787 if (_bundle_instr_count > 0) { 1788 bundle->set_instr_count(_bundle_instr_count); 1789 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1790 1791 _bundle_cycle_number += 1; 1792 } 1793 1794 // Clear the bundling information 1795 _bundle_instr_count = 0; 1796 _bundle_use.reset(); 1797 1798 memcpy(_bundle_use_elements, 1799 Pipeline_Use::elaborated_elements, 1800 sizeof(Pipeline_Use::elaborated_elements)); 1801 } 1802 1803 // Perform instruction scheduling and bundling over the sequence of 1804 // instructions in backwards order. 1805 void Compile::ScheduleAndBundle() { 1806 1807 // Don't optimize this if it isn't a method 1808 if (!_method) 1809 return; 1810 1811 // Don't optimize this if scheduling is disabled 1812 if (!do_scheduling()) 1813 return; 1814 1815 // Scheduling code works only with pairs (16 bytes) maximum. 1816 if (max_vector_size() > 16) 1817 return; 1818 1819 TracePhase tp("isched", &timers[_t_instrSched]); 1820 1821 // Create a data structure for all the scheduling information 1822 Scheduling scheduling(Thread::current()->resource_area(), *this); 1823 1824 // Walk backwards over each basic block, computing the needed alignment 1825 // Walk over all the basic blocks 1826 scheduling.DoScheduling(); 1827 } 1828 1829 // Compute the latency of all the instructions. This is fairly simple, 1830 // because we already have a legal ordering. Walk over the instructions 1831 // from first to last, and compute the latency of the instruction based 1832 // on the latency of the preceding instruction(s). 1833 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1834 #ifndef PRODUCT 1835 if (_cfg->C->trace_opto_output()) 1836 tty->print("# -> ComputeLocalLatenciesForward\n"); 1837 #endif 1838 1839 // Walk over all the schedulable instructions 1840 for( uint j=_bb_start; j < _bb_end; j++ ) { 1841 1842 // This is a kludge, forcing all latency calculations to start at 1. 1843 // Used to allow latency 0 to force an instruction to the beginning 1844 // of the bb 1845 uint latency = 1; 1846 Node *use = bb->get_node(j); 1847 uint nlen = use->len(); 1848 1849 // Walk over all the inputs 1850 for ( uint k=0; k < nlen; k++ ) { 1851 Node *def = use->in(k); 1852 if (!def) 1853 continue; 1854 1855 uint l = _node_latency[def->_idx] + use->latency(k); 1856 if (latency < l) 1857 latency = l; 1858 } 1859 1860 _node_latency[use->_idx] = latency; 1861 1862 #ifndef PRODUCT 1863 if (_cfg->C->trace_opto_output()) { 1864 tty->print("# latency %4d: ", latency); 1865 use->dump(); 1866 } 1867 #endif 1868 } 1869 1870 #ifndef PRODUCT 1871 if (_cfg->C->trace_opto_output()) 1872 tty->print("# <- ComputeLocalLatenciesForward\n"); 1873 #endif 1874 1875 } // end ComputeLocalLatenciesForward 1876 1877 // See if this node fits into the present instruction bundle 1878 bool Scheduling::NodeFitsInBundle(Node *n) { 1879 uint n_idx = n->_idx; 1880 1881 // If this is the unconditional delay instruction, then it fits 1882 if (n == _unconditional_delay_slot) { 1883 #ifndef PRODUCT 1884 if (_cfg->C->trace_opto_output()) 1885 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1886 #endif 1887 return (true); 1888 } 1889 1890 // If the node cannot be scheduled this cycle, skip it 1891 if (_current_latency[n_idx] > _bundle_cycle_number) { 1892 #ifndef PRODUCT 1893 if (_cfg->C->trace_opto_output()) 1894 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1895 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1896 #endif 1897 return (false); 1898 } 1899 1900 const Pipeline *node_pipeline = n->pipeline(); 1901 1902 uint instruction_count = node_pipeline->instructionCount(); 1903 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1904 instruction_count = 0; 1905 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1906 instruction_count++; 1907 1908 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1909 #ifndef PRODUCT 1910 if (_cfg->C->trace_opto_output()) 1911 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1912 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1913 #endif 1914 return (false); 1915 } 1916 1917 // Don't allow non-machine nodes to be handled this way 1918 if (!n->is_Mach() && instruction_count == 0) 1919 return (false); 1920 1921 // See if there is any overlap 1922 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1923 1924 if (delay > 0) { 1925 #ifndef PRODUCT 1926 if (_cfg->C->trace_opto_output()) 1927 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1928 #endif 1929 return false; 1930 } 1931 1932 #ifndef PRODUCT 1933 if (_cfg->C->trace_opto_output()) 1934 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1935 #endif 1936 1937 return true; 1938 } 1939 1940 Node * Scheduling::ChooseNodeToBundle() { 1941 uint siz = _available.size(); 1942 1943 if (siz == 0) { 1944 1945 #ifndef PRODUCT 1946 if (_cfg->C->trace_opto_output()) 1947 tty->print("# ChooseNodeToBundle: NULL\n"); 1948 #endif 1949 return (NULL); 1950 } 1951 1952 // Fast path, if only 1 instruction in the bundle 1953 if (siz == 1) { 1954 #ifndef PRODUCT 1955 if (_cfg->C->trace_opto_output()) { 1956 tty->print("# ChooseNodeToBundle (only 1): "); 1957 _available[0]->dump(); 1958 } 1959 #endif 1960 return (_available[0]); 1961 } 1962 1963 // Don't bother, if the bundle is already full 1964 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1965 for ( uint i = 0; i < siz; i++ ) { 1966 Node *n = _available[i]; 1967 1968 // Skip projections, we'll handle them another way 1969 if (n->is_Proj()) 1970 continue; 1971 1972 // This presupposed that instructions are inserted into the 1973 // available list in a legality order; i.e. instructions that 1974 // must be inserted first are at the head of the list 1975 if (NodeFitsInBundle(n)) { 1976 #ifndef PRODUCT 1977 if (_cfg->C->trace_opto_output()) { 1978 tty->print("# ChooseNodeToBundle: "); 1979 n->dump(); 1980 } 1981 #endif 1982 return (n); 1983 } 1984 } 1985 } 1986 1987 // Nothing fits in this bundle, choose the highest priority 1988 #ifndef PRODUCT 1989 if (_cfg->C->trace_opto_output()) { 1990 tty->print("# ChooseNodeToBundle: "); 1991 _available[0]->dump(); 1992 } 1993 #endif 1994 1995 return _available[0]; 1996 } 1997 1998 void Scheduling::AddNodeToAvailableList(Node *n) { 1999 assert( !n->is_Proj(), "projections never directly made available" ); 2000 #ifndef PRODUCT 2001 if (_cfg->C->trace_opto_output()) { 2002 tty->print("# AddNodeToAvailableList: "); 2003 n->dump(); 2004 } 2005 #endif 2006 2007 int latency = _current_latency[n->_idx]; 2008 2009 // Insert in latency order (insertion sort) 2010 uint i; 2011 for ( i=0; i < _available.size(); i++ ) 2012 if (_current_latency[_available[i]->_idx] > latency) 2013 break; 2014 2015 // Special Check for compares following branches 2016 if( n->is_Mach() && _scheduled.size() > 0 ) { 2017 int op = n->as_Mach()->ideal_Opcode(); 2018 Node *last = _scheduled[0]; 2019 if( last->is_MachIf() && last->in(1) == n && 2020 ( op == Op_CmpI || 2021 op == Op_CmpU || 2022 op == Op_CmpUL || 2023 op == Op_CmpP || 2024 op == Op_CmpF || 2025 op == Op_CmpD || 2026 op == Op_CmpL ) ) { 2027 2028 // Recalculate position, moving to front of same latency 2029 for ( i=0 ; i < _available.size(); i++ ) 2030 if (_current_latency[_available[i]->_idx] >= latency) 2031 break; 2032 } 2033 } 2034 2035 // Insert the node in the available list 2036 _available.insert(i, n); 2037 2038 #ifndef PRODUCT 2039 if (_cfg->C->trace_opto_output()) 2040 dump_available(); 2041 #endif 2042 } 2043 2044 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 2045 for ( uint i=0; i < n->len(); i++ ) { 2046 Node *def = n->in(i); 2047 if (!def) continue; 2048 if( def->is_Proj() ) // If this is a machine projection, then 2049 def = def->in(0); // propagate usage thru to the base instruction 2050 2051 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2052 continue; 2053 } 2054 2055 // Compute the latency 2056 uint l = _bundle_cycle_number + n->latency(i); 2057 if (_current_latency[def->_idx] < l) 2058 _current_latency[def->_idx] = l; 2059 2060 // If this does not have uses then schedule it 2061 if ((--_uses[def->_idx]) == 0) 2062 AddNodeToAvailableList(def); 2063 } 2064 } 2065 2066 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2067 #ifndef PRODUCT 2068 if (_cfg->C->trace_opto_output()) { 2069 tty->print("# AddNodeToBundle: "); 2070 n->dump(); 2071 } 2072 #endif 2073 2074 // Remove this from the available list 2075 uint i; 2076 for (i = 0; i < _available.size(); i++) 2077 if (_available[i] == n) 2078 break; 2079 assert(i < _available.size(), "entry in _available list not found"); 2080 _available.remove(i); 2081 2082 // See if this fits in the current bundle 2083 const Pipeline *node_pipeline = n->pipeline(); 2084 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2085 2086 // Check for instructions to be placed in the delay slot. We 2087 // do this before we actually schedule the current instruction, 2088 // because the delay slot follows the current instruction. 2089 if (Pipeline::_branch_has_delay_slot && 2090 node_pipeline->hasBranchDelay() && 2091 !_unconditional_delay_slot) { 2092 2093 uint siz = _available.size(); 2094 2095 // Conditional branches can support an instruction that 2096 // is unconditionally executed and not dependent by the 2097 // branch, OR a conditionally executed instruction if 2098 // the branch is taken. In practice, this means that 2099 // the first instruction at the branch target is 2100 // copied to the delay slot, and the branch goes to 2101 // the instruction after that at the branch target 2102 if ( n->is_MachBranch() ) { 2103 2104 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2105 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2106 2107 #ifndef PRODUCT 2108 _branches++; 2109 #endif 2110 2111 // At least 1 instruction is on the available list 2112 // that is not dependent on the branch 2113 for (uint i = 0; i < siz; i++) { 2114 Node *d = _available[i]; 2115 const Pipeline *avail_pipeline = d->pipeline(); 2116 2117 // Don't allow safepoints in the branch shadow, that will 2118 // cause a number of difficulties 2119 if ( avail_pipeline->instructionCount() == 1 && 2120 !avail_pipeline->hasMultipleBundles() && 2121 !avail_pipeline->hasBranchDelay() && 2122 Pipeline::instr_has_unit_size() && 2123 d->size(_regalloc) == Pipeline::instr_unit_size() && 2124 NodeFitsInBundle(d) && 2125 !node_bundling(d)->used_in_delay()) { 2126 2127 if (d->is_Mach() && !d->is_MachSafePoint()) { 2128 // A node that fits in the delay slot was found, so we need to 2129 // set the appropriate bits in the bundle pipeline information so 2130 // that it correctly indicates resource usage. Later, when we 2131 // attempt to add this instruction to the bundle, we will skip 2132 // setting the resource usage. 2133 _unconditional_delay_slot = d; 2134 node_bundling(n)->set_use_unconditional_delay(); 2135 node_bundling(d)->set_used_in_unconditional_delay(); 2136 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2137 _current_latency[d->_idx] = _bundle_cycle_number; 2138 _next_node = d; 2139 ++_bundle_instr_count; 2140 #ifndef PRODUCT 2141 _unconditional_delays++; 2142 #endif 2143 break; 2144 } 2145 } 2146 } 2147 } 2148 2149 // No delay slot, add a nop to the usage 2150 if (!_unconditional_delay_slot) { 2151 // See if adding an instruction in the delay slot will overflow 2152 // the bundle. 2153 if (!NodeFitsInBundle(_nop)) { 2154 #ifndef PRODUCT 2155 if (_cfg->C->trace_opto_output()) 2156 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2157 #endif 2158 step(1); 2159 } 2160 2161 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2162 _next_node = _nop; 2163 ++_bundle_instr_count; 2164 } 2165 2166 // See if the instruction in the delay slot requires a 2167 // step of the bundles 2168 if (!NodeFitsInBundle(n)) { 2169 #ifndef PRODUCT 2170 if (_cfg->C->trace_opto_output()) 2171 tty->print("# *** STEP(branch won't fit) ***\n"); 2172 #endif 2173 // Update the state information 2174 _bundle_instr_count = 0; 2175 _bundle_cycle_number += 1; 2176 _bundle_use.step(1); 2177 } 2178 } 2179 2180 // Get the number of instructions 2181 uint instruction_count = node_pipeline->instructionCount(); 2182 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2183 instruction_count = 0; 2184 2185 // Compute the latency information 2186 uint delay = 0; 2187 2188 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2189 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2190 if (relative_latency < 0) 2191 relative_latency = 0; 2192 2193 delay = _bundle_use.full_latency(relative_latency, node_usage); 2194 2195 // Does not fit in this bundle, start a new one 2196 if (delay > 0) { 2197 step(delay); 2198 2199 #ifndef PRODUCT 2200 if (_cfg->C->trace_opto_output()) 2201 tty->print("# *** STEP(%d) ***\n", delay); 2202 #endif 2203 } 2204 } 2205 2206 // If this was placed in the delay slot, ignore it 2207 if (n != _unconditional_delay_slot) { 2208 2209 if (delay == 0) { 2210 if (node_pipeline->hasMultipleBundles()) { 2211 #ifndef PRODUCT 2212 if (_cfg->C->trace_opto_output()) 2213 tty->print("# *** STEP(multiple instructions) ***\n"); 2214 #endif 2215 step(1); 2216 } 2217 2218 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2219 #ifndef PRODUCT 2220 if (_cfg->C->trace_opto_output()) 2221 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2222 instruction_count + _bundle_instr_count, 2223 Pipeline::_max_instrs_per_cycle); 2224 #endif 2225 step(1); 2226 } 2227 } 2228 2229 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2230 _bundle_instr_count++; 2231 2232 // Set the node's latency 2233 _current_latency[n->_idx] = _bundle_cycle_number; 2234 2235 // Now merge the functional unit information 2236 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2237 _bundle_use.add_usage(node_usage); 2238 2239 // Increment the number of instructions in this bundle 2240 _bundle_instr_count += instruction_count; 2241 2242 // Remember this node for later 2243 if (n->is_Mach()) 2244 _next_node = n; 2245 } 2246 2247 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2248 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2249 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2250 // into the block. All other scheduled nodes get put in the schedule here. 2251 int op = n->Opcode(); 2252 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2253 (op != Op_Node && // Not an unused antidepedence node and 2254 // not an unallocated boxlock 2255 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2256 2257 // Push any trailing projections 2258 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2259 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2260 Node *foi = n->fast_out(i); 2261 if( foi->is_Proj() ) 2262 _scheduled.push(foi); 2263 } 2264 } 2265 2266 // Put the instruction in the schedule list 2267 _scheduled.push(n); 2268 } 2269 2270 #ifndef PRODUCT 2271 if (_cfg->C->trace_opto_output()) 2272 dump_available(); 2273 #endif 2274 2275 // Walk all the definitions, decrementing use counts, and 2276 // if a definition has a 0 use count, place it in the available list. 2277 DecrementUseCounts(n,bb); 2278 } 2279 2280 // This method sets the use count within a basic block. We will ignore all 2281 // uses outside the current basic block. As we are doing a backwards walk, 2282 // any node we reach that has a use count of 0 may be scheduled. This also 2283 // avoids the problem of cyclic references from phi nodes, as long as phi 2284 // nodes are at the front of the basic block. This method also initializes 2285 // the available list to the set of instructions that have no uses within this 2286 // basic block. 2287 void Scheduling::ComputeUseCount(const Block *bb) { 2288 #ifndef PRODUCT 2289 if (_cfg->C->trace_opto_output()) 2290 tty->print("# -> ComputeUseCount\n"); 2291 #endif 2292 2293 // Clear the list of available and scheduled instructions, just in case 2294 _available.clear(); 2295 _scheduled.clear(); 2296 2297 // No delay slot specified 2298 _unconditional_delay_slot = NULL; 2299 2300 #ifdef ASSERT 2301 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2302 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2303 #endif 2304 2305 // Force the _uses count to never go to zero for unscheduable pieces 2306 // of the block 2307 for( uint k = 0; k < _bb_start; k++ ) 2308 _uses[bb->get_node(k)->_idx] = 1; 2309 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2310 _uses[bb->get_node(l)->_idx] = 1; 2311 2312 // Iterate backwards over the instructions in the block. Don't count the 2313 // branch projections at end or the block header instructions. 2314 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2315 Node *n = bb->get_node(j); 2316 if( n->is_Proj() ) continue; // Projections handled another way 2317 2318 // Account for all uses 2319 for ( uint k = 0; k < n->len(); k++ ) { 2320 Node *inp = n->in(k); 2321 if (!inp) continue; 2322 assert(inp != n, "no cycles allowed" ); 2323 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2324 if (inp->is_Proj()) { // Skip through Proj's 2325 inp = inp->in(0); 2326 } 2327 ++_uses[inp->_idx]; // Count 1 block-local use 2328 } 2329 } 2330 2331 // If this instruction has a 0 use count, then it is available 2332 if (!_uses[n->_idx]) { 2333 _current_latency[n->_idx] = _bundle_cycle_number; 2334 AddNodeToAvailableList(n); 2335 } 2336 2337 #ifndef PRODUCT 2338 if (_cfg->C->trace_opto_output()) { 2339 tty->print("# uses: %3d: ", _uses[n->_idx]); 2340 n->dump(); 2341 } 2342 #endif 2343 } 2344 2345 #ifndef PRODUCT 2346 if (_cfg->C->trace_opto_output()) 2347 tty->print("# <- ComputeUseCount\n"); 2348 #endif 2349 } 2350 2351 // This routine performs scheduling on each basic block in reverse order, 2352 // using instruction latencies and taking into account function unit 2353 // availability. 2354 void Scheduling::DoScheduling() { 2355 #ifndef PRODUCT 2356 if (_cfg->C->trace_opto_output()) 2357 tty->print("# -> DoScheduling\n"); 2358 #endif 2359 2360 Block *succ_bb = NULL; 2361 Block *bb; 2362 2363 // Walk over all the basic blocks in reverse order 2364 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2365 bb = _cfg->get_block(i); 2366 2367 #ifndef PRODUCT 2368 if (_cfg->C->trace_opto_output()) { 2369 tty->print("# Schedule BB#%03d (initial)\n", i); 2370 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2371 bb->get_node(j)->dump(); 2372 } 2373 } 2374 #endif 2375 2376 // On the head node, skip processing 2377 if (bb == _cfg->get_root_block()) { 2378 continue; 2379 } 2380 2381 // Skip empty, connector blocks 2382 if (bb->is_connector()) 2383 continue; 2384 2385 // If the following block is not the sole successor of 2386 // this one, then reset the pipeline information 2387 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2388 #ifndef PRODUCT 2389 if (_cfg->C->trace_opto_output()) { 2390 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2391 _next_node->_idx, _bundle_instr_count); 2392 } 2393 #endif 2394 step_and_clear(); 2395 } 2396 2397 // Leave untouched the starting instruction, any Phis, a CreateEx node 2398 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2399 _bb_end = bb->number_of_nodes()-1; 2400 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2401 Node *n = bb->get_node(_bb_start); 2402 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2403 // Also, MachIdealNodes do not get scheduled 2404 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2405 MachNode *mach = n->as_Mach(); 2406 int iop = mach->ideal_Opcode(); 2407 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2408 if( iop == Op_Con ) continue; // Do not schedule Top 2409 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2410 mach->pipeline() == MachNode::pipeline_class() && 2411 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2412 continue; 2413 break; // Funny loop structure to be sure... 2414 } 2415 // Compute last "interesting" instruction in block - last instruction we 2416 // might schedule. _bb_end points just after last schedulable inst. We 2417 // normally schedule conditional branches (despite them being forced last 2418 // in the block), because they have delay slots we can fill. Calls all 2419 // have their delay slots filled in the template expansions, so we don't 2420 // bother scheduling them. 2421 Node *last = bb->get_node(_bb_end); 2422 // Ignore trailing NOPs. 2423 while (_bb_end > 0 && last->is_Mach() && 2424 last->as_Mach()->ideal_Opcode() == Op_Con) { 2425 last = bb->get_node(--_bb_end); 2426 } 2427 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2428 if( last->is_Catch() || 2429 // Exclude unreachable path case when Halt node is in a separate block. 2430 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2431 // There must be a prior call. Skip it. 2432 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2433 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2434 } 2435 } else if( last->is_MachNullCheck() ) { 2436 // Backup so the last null-checked memory instruction is 2437 // outside the schedulable range. Skip over the nullcheck, 2438 // projection, and the memory nodes. 2439 Node *mem = last->in(1); 2440 do { 2441 _bb_end--; 2442 } while (mem != bb->get_node(_bb_end)); 2443 } else { 2444 // Set _bb_end to point after last schedulable inst. 2445 _bb_end++; 2446 } 2447 2448 assert( _bb_start <= _bb_end, "inverted block ends" ); 2449 2450 // Compute the register antidependencies for the basic block 2451 ComputeRegisterAntidependencies(bb); 2452 if (_cfg->C->failing()) return; // too many D-U pinch points 2453 2454 // Compute intra-bb latencies for the nodes 2455 ComputeLocalLatenciesForward(bb); 2456 2457 // Compute the usage within the block, and set the list of all nodes 2458 // in the block that have no uses within the block. 2459 ComputeUseCount(bb); 2460 2461 // Schedule the remaining instructions in the block 2462 while ( _available.size() > 0 ) { 2463 Node *n = ChooseNodeToBundle(); 2464 guarantee(n != NULL, "no nodes available"); 2465 AddNodeToBundle(n,bb); 2466 } 2467 2468 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2469 #ifdef ASSERT 2470 for( uint l = _bb_start; l < _bb_end; l++ ) { 2471 Node *n = bb->get_node(l); 2472 uint m; 2473 for( m = 0; m < _bb_end-_bb_start; m++ ) 2474 if( _scheduled[m] == n ) 2475 break; 2476 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2477 } 2478 #endif 2479 2480 // Now copy the instructions (in reverse order) back to the block 2481 for ( uint k = _bb_start; k < _bb_end; k++ ) 2482 bb->map_node(_scheduled[_bb_end-k-1], k); 2483 2484 #ifndef PRODUCT 2485 if (_cfg->C->trace_opto_output()) { 2486 tty->print("# Schedule BB#%03d (final)\n", i); 2487 uint current = 0; 2488 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2489 Node *n = bb->get_node(j); 2490 if( valid_bundle_info(n) ) { 2491 Bundle *bundle = node_bundling(n); 2492 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2493 tty->print("*** Bundle: "); 2494 bundle->dump(); 2495 } 2496 n->dump(); 2497 } 2498 } 2499 } 2500 #endif 2501 #ifdef ASSERT 2502 verify_good_schedule(bb,"after block local scheduling"); 2503 #endif 2504 } 2505 2506 #ifndef PRODUCT 2507 if (_cfg->C->trace_opto_output()) 2508 tty->print("# <- DoScheduling\n"); 2509 #endif 2510 2511 // Record final node-bundling array location 2512 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2513 2514 } // end DoScheduling 2515 2516 // Verify that no live-range used in the block is killed in the block by a 2517 // wrong DEF. This doesn't verify live-ranges that span blocks. 2518 2519 // Check for edge existence. Used to avoid adding redundant precedence edges. 2520 static bool edge_from_to( Node *from, Node *to ) { 2521 for( uint i=0; i<from->len(); i++ ) 2522 if( from->in(i) == to ) 2523 return true; 2524 return false; 2525 } 2526 2527 #ifdef ASSERT 2528 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2529 // Check for bad kills 2530 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2531 Node *prior_use = _reg_node[def]; 2532 if( prior_use && !edge_from_to(prior_use,n) ) { 2533 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2534 n->dump(); 2535 tty->print_cr("..."); 2536 prior_use->dump(); 2537 assert(edge_from_to(prior_use,n), "%s", msg); 2538 } 2539 _reg_node.map(def,NULL); // Kill live USEs 2540 } 2541 } 2542 2543 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2544 2545 // Zap to something reasonable for the verify code 2546 _reg_node.clear(); 2547 2548 // Walk over the block backwards. Check to make sure each DEF doesn't 2549 // kill a live value (other than the one it's supposed to). Add each 2550 // USE to the live set. 2551 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2552 Node *n = b->get_node(i); 2553 int n_op = n->Opcode(); 2554 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2555 // Fat-proj kills a slew of registers 2556 RegMask rm = n->out_RegMask();// Make local copy 2557 while( rm.is_NotEmpty() ) { 2558 OptoReg::Name kill = rm.find_first_elem(); 2559 rm.Remove(kill); 2560 verify_do_def( n, kill, msg ); 2561 } 2562 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2563 // Get DEF'd registers the normal way 2564 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2565 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2566 } 2567 2568 // Now make all USEs live 2569 for( uint i=1; i<n->req(); i++ ) { 2570 Node *def = n->in(i); 2571 assert(def != 0, "input edge required"); 2572 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2573 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2574 if( OptoReg::is_valid(reg_lo) ) { 2575 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), "%s", msg); 2576 _reg_node.map(reg_lo,n); 2577 } 2578 if( OptoReg::is_valid(reg_hi) ) { 2579 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), "%s", msg); 2580 _reg_node.map(reg_hi,n); 2581 } 2582 } 2583 2584 } 2585 2586 // Zap to something reasonable for the Antidependence code 2587 _reg_node.clear(); 2588 } 2589 #endif 2590 2591 // Conditionally add precedence edges. Avoid putting edges on Projs. 2592 static void add_prec_edge_from_to( Node *from, Node *to ) { 2593 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2594 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2595 from = from->in(0); 2596 } 2597 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2598 !edge_from_to( from, to ) ) // Avoid duplicate edge 2599 from->add_prec(to); 2600 } 2601 2602 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2603 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2604 return; 2605 2606 Node *pinch = _reg_node[def_reg]; // Get pinch point 2607 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2608 is_def ) { // Check for a true def (not a kill) 2609 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2610 return; 2611 } 2612 2613 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2614 debug_only( def = (Node*)((intptr_t)0xdeadbeef); ) 2615 2616 // After some number of kills there _may_ be a later def 2617 Node *later_def = NULL; 2618 2619 // Finding a kill requires a real pinch-point. 2620 // Check for not already having a pinch-point. 2621 // Pinch points are Op_Node's. 2622 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2623 later_def = pinch; // Must be def/kill as optimistic pinch-point 2624 if ( _pinch_free_list.size() > 0) { 2625 pinch = _pinch_free_list.pop(); 2626 } else { 2627 pinch = new Node(1); // Pinch point to-be 2628 } 2629 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2630 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2631 return; 2632 } 2633 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2634 _reg_node.map(def_reg,pinch); // Record pinch-point 2635 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2636 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2637 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2638 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2639 later_def = NULL; // and no later def 2640 } 2641 pinch->set_req(0,later_def); // Hook later def so we can find it 2642 } else { // Else have valid pinch point 2643 if( pinch->in(0) ) // If there is a later-def 2644 later_def = pinch->in(0); // Get it 2645 } 2646 2647 // Add output-dependence edge from later def to kill 2648 if( later_def ) // If there is some original def 2649 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2650 2651 // See if current kill is also a use, and so is forced to be the pinch-point. 2652 if( pinch->Opcode() == Op_Node ) { 2653 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2654 for( uint i=1; i<uses->req(); i++ ) { 2655 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2656 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2657 // Yes, found a use/kill pinch-point 2658 pinch->set_req(0,NULL); // 2659 pinch->replace_by(kill); // Move anti-dep edges up 2660 pinch = kill; 2661 _reg_node.map(def_reg,pinch); 2662 return; 2663 } 2664 } 2665 } 2666 2667 // Add edge from kill to pinch-point 2668 add_prec_edge_from_to(kill,pinch); 2669 } 2670 2671 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2672 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2673 return; 2674 Node *pinch = _reg_node[use_reg]; // Get pinch point 2675 // Check for no later def_reg/kill in block 2676 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2677 // Use has to be block-local as well 2678 _cfg->get_block_for_node(use) == b) { 2679 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2680 pinch->req() == 1 ) { // pinch not yet in block? 2681 pinch->del_req(0); // yank pointer to later-def, also set flag 2682 // Insert the pinch-point in the block just after the last use 2683 b->insert_node(pinch, b->find_node(use) + 1); 2684 _bb_end++; // Increase size scheduled region in block 2685 } 2686 2687 add_prec_edge_from_to(pinch,use); 2688 } 2689 } 2690 2691 // We insert antidependences between the reads and following write of 2692 // allocated registers to prevent illegal code motion. Hopefully, the 2693 // number of added references should be fairly small, especially as we 2694 // are only adding references within the current basic block. 2695 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2696 2697 #ifdef ASSERT 2698 verify_good_schedule(b,"before block local scheduling"); 2699 #endif 2700 2701 // A valid schedule, for each register independently, is an endless cycle 2702 // of: a def, then some uses (connected to the def by true dependencies), 2703 // then some kills (defs with no uses), finally the cycle repeats with a new 2704 // def. The uses are allowed to float relative to each other, as are the 2705 // kills. No use is allowed to slide past a kill (or def). This requires 2706 // antidependencies between all uses of a single def and all kills that 2707 // follow, up to the next def. More edges are redundant, because later defs 2708 // & kills are already serialized with true or antidependencies. To keep 2709 // the edge count down, we add a 'pinch point' node if there's more than 2710 // one use or more than one kill/def. 2711 2712 // We add dependencies in one bottom-up pass. 2713 2714 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2715 2716 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2717 // register. If not, we record the DEF/KILL in _reg_node, the 2718 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2719 // "pinch point", a new Node that's in the graph but not in the block. 2720 // We put edges from the prior and current DEF/KILLs to the pinch point. 2721 // We put the pinch point in _reg_node. If there's already a pinch point 2722 // we merely add an edge from the current DEF/KILL to the pinch point. 2723 2724 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2725 // put an edge from the pinch point to the USE. 2726 2727 // To be expedient, the _reg_node array is pre-allocated for the whole 2728 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2729 // or a valid def/kill/pinch-point, or a leftover node from some prior 2730 // block. Leftover node from some prior block is treated like a NULL (no 2731 // prior def, so no anti-dependence needed). Valid def is distinguished by 2732 // it being in the current block. 2733 bool fat_proj_seen = false; 2734 uint last_safept = _bb_end-1; 2735 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2736 Node* last_safept_node = end_node; 2737 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2738 Node *n = b->get_node(i); 2739 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2740 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2741 // Fat-proj kills a slew of registers 2742 // This can add edges to 'n' and obscure whether or not it was a def, 2743 // hence the is_def flag. 2744 fat_proj_seen = true; 2745 RegMask rm = n->out_RegMask();// Make local copy 2746 while( rm.is_NotEmpty() ) { 2747 OptoReg::Name kill = rm.find_first_elem(); 2748 rm.Remove(kill); 2749 anti_do_def( b, n, kill, is_def ); 2750 } 2751 } else { 2752 // Get DEF'd registers the normal way 2753 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2754 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2755 } 2756 2757 // Kill projections on a branch should appear to occur on the 2758 // branch, not afterwards, so grab the masks from the projections 2759 // and process them. 2760 if (n->is_MachBranch() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump)) { 2761 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2762 Node* use = n->fast_out(i); 2763 if (use->is_Proj()) { 2764 RegMask rm = use->out_RegMask();// Make local copy 2765 while( rm.is_NotEmpty() ) { 2766 OptoReg::Name kill = rm.find_first_elem(); 2767 rm.Remove(kill); 2768 anti_do_def( b, n, kill, false ); 2769 } 2770 } 2771 } 2772 } 2773 2774 // Check each register used by this instruction for a following DEF/KILL 2775 // that must occur afterward and requires an anti-dependence edge. 2776 for( uint j=0; j<n->req(); j++ ) { 2777 Node *def = n->in(j); 2778 if( def ) { 2779 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2780 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2781 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2782 } 2783 } 2784 // Do not allow defs of new derived values to float above GC 2785 // points unless the base is definitely available at the GC point. 2786 2787 Node *m = b->get_node(i); 2788 2789 // Add precedence edge from following safepoint to use of derived pointer 2790 if( last_safept_node != end_node && 2791 m != last_safept_node) { 2792 for (uint k = 1; k < m->req(); k++) { 2793 const Type *t = m->in(k)->bottom_type(); 2794 if( t->isa_oop_ptr() && 2795 t->is_ptr()->offset() != 0 ) { 2796 last_safept_node->add_prec( m ); 2797 break; 2798 } 2799 } 2800 } 2801 2802 if( n->jvms() ) { // Precedence edge from derived to safept 2803 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2804 if( b->get_node(last_safept) != last_safept_node ) { 2805 last_safept = b->find_node(last_safept_node); 2806 } 2807 for( uint j=last_safept; j > i; j-- ) { 2808 Node *mach = b->get_node(j); 2809 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2810 mach->add_prec( n ); 2811 } 2812 last_safept = i; 2813 last_safept_node = m; 2814 } 2815 } 2816 2817 if (fat_proj_seen) { 2818 // Garbage collect pinch nodes that were not consumed. 2819 // They are usually created by a fat kill MachProj for a call. 2820 garbage_collect_pinch_nodes(); 2821 } 2822 } 2823 2824 // Garbage collect pinch nodes for reuse by other blocks. 2825 // 2826 // The block scheduler's insertion of anti-dependence 2827 // edges creates many pinch nodes when the block contains 2828 // 2 or more Calls. A pinch node is used to prevent a 2829 // combinatorial explosion of edges. If a set of kills for a 2830 // register is anti-dependent on a set of uses (or defs), rather 2831 // than adding an edge in the graph between each pair of kill 2832 // and use (or def), a pinch is inserted between them: 2833 // 2834 // use1 use2 use3 2835 // \ | / 2836 // \ | / 2837 // pinch 2838 // / | \ 2839 // / | \ 2840 // kill1 kill2 kill3 2841 // 2842 // One pinch node is created per register killed when 2843 // the second call is encountered during a backwards pass 2844 // over the block. Most of these pinch nodes are never 2845 // wired into the graph because the register is never 2846 // used or def'ed in the block. 2847 // 2848 void Scheduling::garbage_collect_pinch_nodes() { 2849 #ifndef PRODUCT 2850 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2851 #endif 2852 int trace_cnt = 0; 2853 for (uint k = 0; k < _reg_node.Size(); k++) { 2854 Node* pinch = _reg_node[k]; 2855 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2856 // no predecence input edges 2857 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2858 cleanup_pinch(pinch); 2859 _pinch_free_list.push(pinch); 2860 _reg_node.map(k, NULL); 2861 #ifndef PRODUCT 2862 if (_cfg->C->trace_opto_output()) { 2863 trace_cnt++; 2864 if (trace_cnt > 40) { 2865 tty->print("\n"); 2866 trace_cnt = 0; 2867 } 2868 tty->print(" %d", pinch->_idx); 2869 } 2870 #endif 2871 } 2872 } 2873 #ifndef PRODUCT 2874 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2875 #endif 2876 } 2877 2878 // Clean up a pinch node for reuse. 2879 void Scheduling::cleanup_pinch( Node *pinch ) { 2880 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2881 2882 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2883 Node* use = pinch->last_out(i); 2884 uint uses_found = 0; 2885 for (uint j = use->req(); j < use->len(); j++) { 2886 if (use->in(j) == pinch) { 2887 use->rm_prec(j); 2888 uses_found++; 2889 } 2890 } 2891 assert(uses_found > 0, "must be a precedence edge"); 2892 i -= uses_found; // we deleted 1 or more copies of this edge 2893 } 2894 // May have a later_def entry 2895 pinch->set_req(0, NULL); 2896 } 2897 2898 #ifndef PRODUCT 2899 2900 void Scheduling::dump_available() const { 2901 tty->print("#Availist "); 2902 for (uint i = 0; i < _available.size(); i++) 2903 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2904 tty->cr(); 2905 } 2906 2907 // Print Scheduling Statistics 2908 void Scheduling::print_statistics() { 2909 // Print the size added by nops for bundling 2910 tty->print("Nops added %d bytes to total of %d bytes", 2911 _total_nop_size, _total_method_size); 2912 if (_total_method_size > 0) 2913 tty->print(", for %.2f%%", 2914 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2915 tty->print("\n"); 2916 2917 // Print the number of branch shadows filled 2918 if (Pipeline::_branch_has_delay_slot) { 2919 tty->print("Of %d branches, %d had unconditional delay slots filled", 2920 _total_branches, _total_unconditional_delays); 2921 if (_total_branches > 0) 2922 tty->print(", for %.2f%%", 2923 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2924 tty->print("\n"); 2925 } 2926 2927 uint total_instructions = 0, total_bundles = 0; 2928 2929 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2930 uint bundle_count = _total_instructions_per_bundle[i]; 2931 total_instructions += bundle_count * i; 2932 total_bundles += bundle_count; 2933 } 2934 2935 if (total_bundles > 0) 2936 tty->print("Average ILP (excluding nops) is %.2f\n", 2937 ((double)total_instructions) / ((double)total_bundles)); 2938 } 2939 #endif