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