1 /* 2 * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.inline.hpp" 27 #include "code/compiledIC.hpp" 28 #include "code/debugInfo.hpp" 29 #include "code/debugInfoRec.hpp" 30 #include "compiler/compileBroker.hpp" 31 #include "compiler/oopMap.hpp" 32 #include "memory/allocation.inline.hpp" 33 #include "opto/ad.hpp" 34 #include "opto/callnode.hpp" 35 #include "opto/cfgnode.hpp" 36 #include "opto/locknode.hpp" 37 #include "opto/machnode.hpp" 38 #include "opto/optoreg.hpp" 39 #include "opto/output.hpp" 40 #include "opto/regalloc.hpp" 41 #include "opto/runtime.hpp" 42 #include "opto/subnode.hpp" 43 #include "opto/type.hpp" 44 #include "runtime/handles.inline.hpp" 45 #include "utilities/xmlstream.hpp" 46 47 #ifndef PRODUCT 48 #define DEBUG_ARG(x) , x 49 #else 50 #define DEBUG_ARG(x) 51 #endif 52 53 // Convert Nodes to instruction bits and pass off to the VM 54 void Compile::Output() { 55 // RootNode goes 56 assert( _cfg->get_root_block()->number_of_nodes() == 0, "" ); 57 58 // The number of new nodes (mostly MachNop) is proportional to 59 // the number of java calls and inner loops which are aligned. 60 if ( C->check_node_count((NodeLimitFudgeFactor + C->java_calls()*3 + 61 C->inner_loops()*(OptoLoopAlignment-1)), 62 "out of nodes before code generation" ) ) { 63 return; 64 } 65 // Make sure I can find the Start Node 66 Block *entry = _cfg->get_block(1); 67 Block *broot = _cfg->get_root_block(); 68 69 const StartNode *start = entry->head()->as_Start(); 70 71 // Replace StartNode with prolog 72 MachPrologNode *prolog = new MachPrologNode(); 73 entry->map_node(prolog, 0); 74 _cfg->map_node_to_block(prolog, entry); 75 _cfg->unmap_node_from_block(start); // start is no longer in any block 76 77 // Virtual methods need an unverified entry point 78 79 if( is_osr_compilation() ) { 80 if( PoisonOSREntry ) { 81 // TODO: Should use a ShouldNotReachHereNode... 82 _cfg->insert( broot, 0, new MachBreakpointNode() ); 83 } 84 } else { 85 if( _method && !_method->flags().is_static() ) { 86 // Insert unvalidated entry point 87 _cfg->insert( broot, 0, new MachUEPNode() ); 88 } 89 90 } 91 92 93 // Break before main entry point 94 if( (_method && _method->break_at_execute()) 95 #ifndef PRODUCT 96 ||(OptoBreakpoint && is_method_compilation()) 97 ||(OptoBreakpointOSR && is_osr_compilation()) 98 ||(OptoBreakpointC2R && !_method) 99 #endif 100 ) { 101 // checking for _method means that OptoBreakpoint does not apply to 102 // runtime stubs or frame converters 103 _cfg->insert( entry, 1, new MachBreakpointNode() ); 104 } 105 106 // Insert epilogs before every return 107 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 108 Block* block = _cfg->get_block(i); 109 if (!block->is_connector() && block->non_connector_successor(0) == _cfg->get_root_block()) { // Found a program exit point? 110 Node* m = block->end(); 111 if (m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt) { 112 MachEpilogNode* epilog = new MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); 113 block->add_inst(epilog); 114 _cfg->map_node_to_block(epilog, block); 115 } 116 } 117 } 118 119 uint* blk_starts = NEW_RESOURCE_ARRAY(uint, _cfg->number_of_blocks() + 1); 120 blk_starts[0] = 0; 121 122 // Initialize code buffer and process short branches. 123 CodeBuffer* cb = init_buffer(blk_starts); 124 125 if (cb == NULL || failing()) { 126 return; 127 } 128 129 ScheduleAndBundle(); 130 131 #ifndef PRODUCT 132 if (trace_opto_output()) { 133 tty->print("\n---- After ScheduleAndBundle ----\n"); 134 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 135 tty->print("\nBB#%03d:\n", i); 136 Block* block = _cfg->get_block(i); 137 for (uint j = 0; j < block->number_of_nodes(); j++) { 138 Node* n = block->get_node(j); 139 OptoReg::Name reg = _regalloc->get_reg_first(n); 140 tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); 141 n->dump(); 142 } 143 } 144 } 145 #endif 146 147 if (failing()) { 148 return; 149 } 150 151 BuildOopMaps(); 152 153 if (failing()) { 154 return; 155 } 156 157 fill_buffer(cb, blk_starts); 158 } 159 160 bool Compile::need_stack_bang(int frame_size_in_bytes) const { 161 // Determine if we need to generate a stack overflow check. 162 // Do it if the method is not a stub function and 163 // has java calls or has frame size > vm_page_size/8. 164 // The debug VM checks that deoptimization doesn't trigger an 165 // unexpected stack overflow (compiled method stack banging should 166 // guarantee it doesn't happen) so we always need the stack bang in 167 // a debug VM. 168 return (UseStackBanging && stub_function() == NULL && 169 (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3 170 DEBUG_ONLY(|| true))); 171 } 172 173 bool Compile::need_register_stack_bang() const { 174 // Determine if we need to generate a register stack overflow check. 175 // This is only used on architectures which have split register 176 // and memory stacks (ie. IA64). 177 // Bang if the method is not a stub function and has java calls 178 return (stub_function() == NULL && has_java_calls()); 179 } 180 181 182 // Compute the size of first NumberOfLoopInstrToAlign instructions at the top 183 // of a loop. When aligning a loop we need to provide enough instructions 184 // in cpu's fetch buffer to feed decoders. The loop alignment could be 185 // avoided if we have enough instructions in fetch buffer at the head of a loop. 186 // By default, the size is set to 999999 by Block's constructor so that 187 // a loop will be aligned if the size is not reset here. 188 // 189 // Note: Mach instructions could contain several HW instructions 190 // so the size is estimated only. 191 // 192 void Compile::compute_loop_first_inst_sizes() { 193 // The next condition is used to gate the loop alignment optimization. 194 // Don't aligned a loop if there are enough instructions at the head of a loop 195 // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad 196 // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is 197 // equal to 11 bytes which is the largest address NOP instruction. 198 if (MaxLoopPad < OptoLoopAlignment - 1) { 199 uint last_block = _cfg->number_of_blocks() - 1; 200 for (uint i = 1; i <= last_block; i++) { 201 Block* block = _cfg->get_block(i); 202 // Check the first loop's block which requires an alignment. 203 if (block->loop_alignment() > (uint)relocInfo::addr_unit()) { 204 uint sum_size = 0; 205 uint inst_cnt = NumberOfLoopInstrToAlign; 206 inst_cnt = block->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 207 208 // Check subsequent fallthrough blocks if the loop's first 209 // block(s) does not have enough instructions. 210 Block *nb = block; 211 while(inst_cnt > 0 && 212 i < last_block && 213 !_cfg->get_block(i + 1)->has_loop_alignment() && 214 !nb->has_successor(block)) { 215 i++; 216 nb = _cfg->get_block(i); 217 inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); 218 } // while( inst_cnt > 0 && i < last_block ) 219 220 block->set_first_inst_size(sum_size); 221 } // f( b->head()->is_Loop() ) 222 } // for( i <= last_block ) 223 } // if( MaxLoopPad < OptoLoopAlignment-1 ) 224 } 225 226 // The architecture description provides short branch variants for some long 227 // branch instructions. Replace eligible long branches with short branches. 228 void Compile::shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size) { 229 // Compute size of each block, method size, and relocation information size 230 uint nblocks = _cfg->number_of_blocks(); 231 232 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 233 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 234 int* jmp_nidx = NEW_RESOURCE_ARRAY(int ,nblocks); 235 236 // Collect worst case block paddings 237 int* block_worst_case_pad = NEW_RESOURCE_ARRAY(int, nblocks); 238 memset(block_worst_case_pad, 0, nblocks * sizeof(int)); 239 240 DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); ) 241 DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); ) 242 243 bool has_short_branch_candidate = false; 244 245 // Initialize the sizes to 0 246 code_size = 0; // Size in bytes of generated code 247 stub_size = 0; // Size in bytes of all stub entries 248 // Size in bytes of all relocation entries, including those in local stubs. 249 // Start with 2-bytes of reloc info for the unvalidated entry point 250 reloc_size = 1; // Number of relocation entries 251 252 // Make three passes. The first computes pessimistic blk_starts, 253 // relative jmp_offset and reloc_size information. The second performs 254 // short branch substitution using the pessimistic sizing. The 255 // third inserts nops where needed. 256 257 // Step one, perform a pessimistic sizing pass. 258 uint last_call_adr = max_juint; 259 uint last_avoid_back_to_back_adr = max_juint; 260 uint nop_size = (new MachNopNode())->size(_regalloc); 261 for (uint i = 0; i < nblocks; i++) { // For all blocks 262 Block* block = _cfg->get_block(i); 263 264 // During short branch replacement, we store the relative (to blk_starts) 265 // offset of jump in jmp_offset, rather than the absolute offset of jump. 266 // This is so that we do not need to recompute sizes of all nodes when 267 // we compute correct blk_starts in our next sizing pass. 268 jmp_offset[i] = 0; 269 jmp_size[i] = 0; 270 jmp_nidx[i] = -1; 271 DEBUG_ONLY( jmp_target[i] = 0; ) 272 DEBUG_ONLY( jmp_rule[i] = 0; ) 273 274 // Sum all instruction sizes to compute block size 275 uint last_inst = block->number_of_nodes(); 276 uint blk_size = 0; 277 for (uint j = 0; j < last_inst; j++) { 278 Node* nj = block->get_node(j); 279 // Handle machine instruction nodes 280 if (nj->is_Mach()) { 281 MachNode *mach = nj->as_Mach(); 282 blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding 283 reloc_size += mach->reloc(); 284 if (mach->is_MachCall()) { 285 // add size information for trampoline stub 286 // class CallStubImpl is platform-specific and defined in the *.ad files. 287 stub_size += CallStubImpl::size_call_trampoline(); 288 reloc_size += CallStubImpl::reloc_call_trampoline(); 289 290 MachCallNode *mcall = mach->as_MachCall(); 291 // This destination address is NOT PC-relative 292 293 mcall->method_set((intptr_t)mcall->entry_point()); 294 295 if (mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method) { 296 stub_size += CompiledStaticCall::to_interp_stub_size(); 297 reloc_size += CompiledStaticCall::reloc_to_interp_stub(); 298 } 299 } else if (mach->is_MachSafePoint()) { 300 // If call/safepoint are adjacent, account for possible 301 // nop to disambiguate the two safepoints. 302 // ScheduleAndBundle() can rearrange nodes in a block, 303 // check for all offsets inside this block. 304 if (last_call_adr >= blk_starts[i]) { 305 blk_size += nop_size; 306 } 307 } 308 if (mach->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 309 // Nop is inserted between "avoid back to back" instructions. 310 // ScheduleAndBundle() can rearrange nodes in a block, 311 // check for all offsets inside this block. 312 if (last_avoid_back_to_back_adr >= blk_starts[i]) { 313 blk_size += nop_size; 314 } 315 } 316 if (mach->may_be_short_branch()) { 317 if (!nj->is_MachBranch()) { 318 #ifndef PRODUCT 319 nj->dump(3); 320 #endif 321 Unimplemented(); 322 } 323 assert(jmp_nidx[i] == -1, "block should have only one branch"); 324 jmp_offset[i] = blk_size; 325 jmp_size[i] = nj->size(_regalloc); 326 jmp_nidx[i] = j; 327 has_short_branch_candidate = true; 328 } 329 } 330 blk_size += nj->size(_regalloc); 331 // Remember end of call offset 332 if (nj->is_MachCall() && !nj->is_MachCallLeaf()) { 333 last_call_adr = blk_starts[i]+blk_size; 334 } 335 // Remember end of avoid_back_to_back offset 336 if (nj->is_Mach() && nj->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 337 last_avoid_back_to_back_adr = blk_starts[i]+blk_size; 338 } 339 } 340 341 // When the next block starts a loop, we may insert pad NOP 342 // instructions. Since we cannot know our future alignment, 343 // assume the worst. 344 if (i < nblocks - 1) { 345 Block* nb = _cfg->get_block(i + 1); 346 int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); 347 if (max_loop_pad > 0) { 348 assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); 349 // Adjust last_call_adr and/or last_avoid_back_to_back_adr. 350 // If either is the last instruction in this block, bump by 351 // max_loop_pad in lock-step with blk_size, so sizing 352 // calculations in subsequent blocks still can conservatively 353 // detect that it may the last instruction in this block. 354 if (last_call_adr == blk_starts[i]+blk_size) { 355 last_call_adr += max_loop_pad; 356 } 357 if (last_avoid_back_to_back_adr == blk_starts[i]+blk_size) { 358 last_avoid_back_to_back_adr += max_loop_pad; 359 } 360 blk_size += max_loop_pad; 361 block_worst_case_pad[i + 1] = max_loop_pad; 362 } 363 } 364 365 // Save block size; update total method size 366 blk_starts[i+1] = blk_starts[i]+blk_size; 367 } 368 369 // Step two, replace eligible long jumps. 370 bool progress = true; 371 uint last_may_be_short_branch_adr = max_juint; 372 while (has_short_branch_candidate && progress) { 373 progress = false; 374 has_short_branch_candidate = false; 375 int adjust_block_start = 0; 376 for (uint i = 0; i < nblocks; i++) { 377 Block* block = _cfg->get_block(i); 378 int idx = jmp_nidx[i]; 379 MachNode* mach = (idx == -1) ? NULL: block->get_node(idx)->as_Mach(); 380 if (mach != NULL && mach->may_be_short_branch()) { 381 #ifdef ASSERT 382 assert(jmp_size[i] > 0 && mach->is_MachBranch(), "sanity"); 383 int j; 384 // Find the branch; ignore trailing NOPs. 385 for (j = block->number_of_nodes()-1; j>=0; j--) { 386 Node* n = block->get_node(j); 387 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) 388 break; 389 } 390 assert(j >= 0 && j == idx && block->get_node(j) == (Node*)mach, "sanity"); 391 #endif 392 int br_size = jmp_size[i]; 393 int br_offs = blk_starts[i] + jmp_offset[i]; 394 395 // This requires the TRUE branch target be in succs[0] 396 uint bnum = block->non_connector_successor(0)->_pre_order; 397 int offset = blk_starts[bnum] - br_offs; 398 if (bnum > i) { // adjust following block's offset 399 offset -= adjust_block_start; 400 } 401 402 // This block can be a loop header, account for the padding 403 // in the previous block. 404 int block_padding = block_worst_case_pad[i]; 405 assert(i == 0 || block_padding == 0 || br_offs >= block_padding, "Should have at least a padding on top"); 406 // In the following code a nop could be inserted before 407 // the branch which will increase the backward distance. 408 bool needs_padding = ((uint)(br_offs - block_padding) == last_may_be_short_branch_adr); 409 assert(!needs_padding || jmp_offset[i] == 0, "padding only branches at the beginning of block"); 410 411 if (needs_padding && offset <= 0) 412 offset -= nop_size; 413 414 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 415 // We've got a winner. Replace this branch. 416 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 417 418 // Update the jmp_size. 419 int new_size = replacement->size(_regalloc); 420 int diff = br_size - new_size; 421 assert(diff >= (int)nop_size, "short_branch size should be smaller"); 422 // Conservatively take into account padding between 423 // avoid_back_to_back branches. Previous branch could be 424 // converted into avoid_back_to_back branch during next 425 // rounds. 426 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 427 jmp_offset[i] += nop_size; 428 diff -= nop_size; 429 } 430 adjust_block_start += diff; 431 block->map_node(replacement, idx); 432 mach->subsume_by(replacement, C); 433 mach = replacement; 434 progress = true; 435 436 jmp_size[i] = new_size; 437 DEBUG_ONLY( jmp_target[i] = bnum; ); 438 DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); 439 } else { 440 // The jump distance is not short, try again during next iteration. 441 has_short_branch_candidate = true; 442 } 443 } // (mach->may_be_short_branch()) 444 if (mach != NULL && (mach->may_be_short_branch() || 445 mach->avoid_back_to_back(MachNode::AVOID_AFTER))) { 446 last_may_be_short_branch_adr = blk_starts[i] + jmp_offset[i] + jmp_size[i]; 447 } 448 blk_starts[i+1] -= adjust_block_start; 449 } 450 } 451 452 #ifdef ASSERT 453 for (uint i = 0; i < nblocks; i++) { // For all blocks 454 if (jmp_target[i] != 0) { 455 int br_size = jmp_size[i]; 456 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 457 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 458 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]); 459 } 460 assert(_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset), "Displacement too large for short jmp"); 461 } 462 } 463 #endif 464 465 // Step 3, compute the offsets of all blocks, will be done in fill_buffer() 466 // after ScheduleAndBundle(). 467 468 // ------------------ 469 // Compute size for code buffer 470 code_size = blk_starts[nblocks]; 471 472 // Relocation records 473 reloc_size += 1; // Relo entry for exception handler 474 475 // Adjust reloc_size to number of record of relocation info 476 // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for 477 // a relocation index. 478 // The CodeBuffer will expand the locs array if this estimate is too low. 479 reloc_size *= 10 / sizeof(relocInfo); 480 } 481 482 //------------------------------FillLocArray----------------------------------- 483 // Create a bit of debug info and append it to the array. The mapping is from 484 // Java local or expression stack to constant, register or stack-slot. For 485 // doubles, insert 2 mappings and return 1 (to tell the caller that the next 486 // entry has been taken care of and caller should skip it). 487 static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { 488 // This should never have accepted Bad before 489 assert(OptoReg::is_valid(regnum), "location must be valid"); 490 return (OptoReg::is_reg(regnum)) 491 ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) 492 : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); 493 } 494 495 496 ObjectValue* 497 Compile::sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id) { 498 for (int i = 0; i < objs->length(); i++) { 499 assert(objs->at(i)->is_object(), "corrupt object cache"); 500 ObjectValue* sv = (ObjectValue*) objs->at(i); 501 if (sv->id() == id) { 502 return sv; 503 } 504 } 505 // Otherwise.. 506 return NULL; 507 } 508 509 void Compile::set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, 510 ObjectValue* sv ) { 511 assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); 512 objs->append(sv); 513 } 514 515 516 void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, 517 GrowableArray<ScopeValue*> *array, 518 GrowableArray<ScopeValue*> *objs ) { 519 assert( local, "use _top instead of null" ); 520 if (array->length() != idx) { 521 assert(array->length() == idx + 1, "Unexpected array count"); 522 // Old functionality: 523 // return 524 // New functionality: 525 // Assert if the local is not top. In product mode let the new node 526 // override the old entry. 527 assert(local == top(), "LocArray collision"); 528 if (local == top()) { 529 return; 530 } 531 array->pop(); 532 } 533 const Type *t = local->bottom_type(); 534 535 // Is it a safepoint scalar object node? 536 if (local->is_SafePointScalarObject()) { 537 SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); 538 539 ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); 540 if (sv == NULL) { 541 ciKlass* cik = t->is_oopptr()->klass(); 542 assert(cik->is_instance_klass() || 543 cik->is_array_klass(), "Not supported allocation."); 544 sv = new ObjectValue(spobj->_idx, 545 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 546 Compile::set_sv_for_object_node(objs, sv); 547 548 uint first_ind = spobj->first_index(sfpt->jvms()); 549 for (uint i = 0; i < spobj->n_fields(); i++) { 550 Node* fld_node = sfpt->in(first_ind+i); 551 (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); 552 } 553 } 554 array->append(sv); 555 return; 556 } 557 558 // Grab the register number for the local 559 OptoReg::Name regnum = _regalloc->get_reg_first(local); 560 if( OptoReg::is_valid(regnum) ) {// Got a register/stack? 561 // Record the double as two float registers. 562 // The register mask for such a value always specifies two adjacent 563 // float registers, with the lower register number even. 564 // Normally, the allocation of high and low words to these registers 565 // is irrelevant, because nearly all operations on register pairs 566 // (e.g., StoreD) treat them as a single unit. 567 // Here, we assume in addition that the words in these two registers 568 // stored "naturally" (by operations like StoreD and double stores 569 // within the interpreter) such that the lower-numbered register 570 // is written to the lower memory address. This may seem like 571 // a machine dependency, but it is not--it is a requirement on 572 // the author of the <arch>.ad file to ensure that, for every 573 // even/odd double-register pair to which a double may be allocated, 574 // the word in the even single-register is stored to the first 575 // memory word. (Note that register numbers are completely 576 // arbitrary, and are not tied to any machine-level encodings.) 577 #ifdef _LP64 578 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { 579 array->append(new ConstantIntValue(0)); 580 array->append(new_loc_value( _regalloc, regnum, Location::dbl )); 581 } else if ( t->base() == Type::Long ) { 582 array->append(new ConstantIntValue(0)); 583 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 584 } else if ( t->base() == Type::RawPtr ) { 585 // jsr/ret return address which must be restored into a the full 586 // width 64-bit stack slot. 587 array->append(new_loc_value( _regalloc, regnum, Location::lng )); 588 } 589 #else //_LP64 590 #ifdef SPARC 591 if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { 592 // For SPARC we have to swap high and low words for 593 // long values stored in a single-register (g0-g7). 594 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 595 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 596 } else 597 #endif //SPARC 598 if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { 599 // Repack the double/long as two jints. 600 // The convention the interpreter uses is that the second local 601 // holds the first raw word of the native double representation. 602 // This is actually reasonable, since locals and stack arrays 603 // grow downwards in all implementations. 604 // (If, on some machine, the interpreter's Java locals or stack 605 // were to grow upwards, the embedded doubles would be word-swapped.) 606 array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); 607 array->append(new_loc_value( _regalloc, regnum , Location::normal )); 608 } 609 #endif //_LP64 610 else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && 611 OptoReg::is_reg(regnum) ) { 612 array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double() 613 ? Location::float_in_dbl : Location::normal )); 614 } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { 615 array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long 616 ? Location::int_in_long : Location::normal )); 617 } else if( t->base() == Type::NarrowOop ) { 618 array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); 619 } else { 620 array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); 621 } 622 return; 623 } 624 625 // No register. It must be constant data. 626 switch (t->base()) { 627 case Type::Half: // Second half of a double 628 ShouldNotReachHere(); // Caller should skip 2nd halves 629 break; 630 case Type::AnyPtr: 631 array->append(new ConstantOopWriteValue(NULL)); 632 break; 633 case Type::AryPtr: 634 case Type::InstPtr: // fall through 635 array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->constant_encoding())); 636 break; 637 case Type::NarrowOop: 638 if (t == TypeNarrowOop::NULL_PTR) { 639 array->append(new ConstantOopWriteValue(NULL)); 640 } else { 641 array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->constant_encoding())); 642 } 643 break; 644 case Type::Int: 645 array->append(new ConstantIntValue(t->is_int()->get_con())); 646 break; 647 case Type::RawPtr: 648 // A return address (T_ADDRESS). 649 assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); 650 #ifdef _LP64 651 // Must be restored to the full-width 64-bit stack slot. 652 array->append(new ConstantLongValue(t->is_ptr()->get_con())); 653 #else 654 array->append(new ConstantIntValue(t->is_ptr()->get_con())); 655 #endif 656 break; 657 case Type::FloatCon: { 658 float f = t->is_float_constant()->getf(); 659 array->append(new ConstantIntValue(jint_cast(f))); 660 break; 661 } 662 case Type::DoubleCon: { 663 jdouble d = t->is_double_constant()->getd(); 664 #ifdef _LP64 665 array->append(new ConstantIntValue(0)); 666 array->append(new ConstantDoubleValue(d)); 667 #else 668 // Repack the double as two jints. 669 // The convention the interpreter uses is that the second local 670 // holds the first raw word of the native double representation. 671 // This is actually reasonable, since locals and stack arrays 672 // grow downwards in all implementations. 673 // (If, on some machine, the interpreter's Java locals or stack 674 // were to grow upwards, the embedded doubles would be word-swapped.) 675 jlong_accessor acc; 676 acc.long_value = jlong_cast(d); 677 array->append(new ConstantIntValue(acc.words[1])); 678 array->append(new ConstantIntValue(acc.words[0])); 679 #endif 680 break; 681 } 682 case Type::Long: { 683 jlong d = t->is_long()->get_con(); 684 #ifdef _LP64 685 array->append(new ConstantIntValue(0)); 686 array->append(new ConstantLongValue(d)); 687 #else 688 // Repack the long as two jints. 689 // The convention the interpreter uses is that the second local 690 // holds the first raw word of the native double representation. 691 // This is actually reasonable, since locals and stack arrays 692 // grow downwards in all implementations. 693 // (If, on some machine, the interpreter's Java locals or stack 694 // were to grow upwards, the embedded doubles would be word-swapped.) 695 jlong_accessor acc; 696 acc.long_value = d; 697 array->append(new ConstantIntValue(acc.words[1])); 698 array->append(new ConstantIntValue(acc.words[0])); 699 #endif 700 break; 701 } 702 case Type::Top: // Add an illegal value here 703 array->append(new LocationValue(Location())); 704 break; 705 default: 706 ShouldNotReachHere(); 707 break; 708 } 709 } 710 711 // Determine if this node starts a bundle 712 bool Compile::starts_bundle(const Node *n) const { 713 return (_node_bundling_limit > n->_idx && 714 _node_bundling_base[n->_idx].starts_bundle()); 715 } 716 717 //--------------------------Process_OopMap_Node-------------------------------- 718 void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { 719 720 // Handle special safepoint nodes for synchronization 721 MachSafePointNode *sfn = mach->as_MachSafePoint(); 722 MachCallNode *mcall; 723 724 int safepoint_pc_offset = current_offset; 725 bool is_method_handle_invoke = false; 726 bool return_oop = false; 727 728 // Add the safepoint in the DebugInfoRecorder 729 if( !mach->is_MachCall() ) { 730 mcall = NULL; 731 debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); 732 } else { 733 mcall = mach->as_MachCall(); 734 735 // Is the call a MethodHandle call? 736 if (mcall->is_MachCallJava()) { 737 if (mcall->as_MachCallJava()->_method_handle_invoke) { 738 assert(has_method_handle_invokes(), "must have been set during call generation"); 739 is_method_handle_invoke = true; 740 } 741 } 742 743 // Check if a call returns an object. 744 if (mcall->returns_pointer()) { 745 return_oop = true; 746 } 747 safepoint_pc_offset += mcall->ret_addr_offset(); 748 debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); 749 } 750 751 // Loop over the JVMState list to add scope information 752 // Do not skip safepoints with a NULL method, they need monitor info 753 JVMState* youngest_jvms = sfn->jvms(); 754 int max_depth = youngest_jvms->depth(); 755 756 // Allocate the object pool for scalar-replaced objects -- the map from 757 // small-integer keys (which can be recorded in the local and ostack 758 // arrays) to descriptions of the object state. 759 GrowableArray<ScopeValue*> *objs = new GrowableArray<ScopeValue*>(); 760 761 // Visit scopes from oldest to youngest. 762 for (int depth = 1; depth <= max_depth; depth++) { 763 JVMState* jvms = youngest_jvms->of_depth(depth); 764 int idx; 765 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 766 // Safepoints that do not have method() set only provide oop-map and monitor info 767 // to support GC; these do not support deoptimization. 768 int num_locs = (method == NULL) ? 0 : jvms->loc_size(); 769 int num_exps = (method == NULL) ? 0 : jvms->stk_size(); 770 int num_mon = jvms->nof_monitors(); 771 assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), 772 "JVMS local count must match that of the method"); 773 774 // Add Local and Expression Stack Information 775 776 // Insert locals into the locarray 777 GrowableArray<ScopeValue*> *locarray = new GrowableArray<ScopeValue*>(num_locs); 778 for( idx = 0; idx < num_locs; idx++ ) { 779 FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); 780 } 781 782 // Insert expression stack entries into the exparray 783 GrowableArray<ScopeValue*> *exparray = new GrowableArray<ScopeValue*>(num_exps); 784 for( idx = 0; idx < num_exps; idx++ ) { 785 FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); 786 } 787 788 // Add in mappings of the monitors 789 assert( !method || 790 !method->is_synchronized() || 791 method->is_native() || 792 num_mon > 0 || 793 !GenerateSynchronizationCode, 794 "monitors must always exist for synchronized methods"); 795 796 // Build the growable array of ScopeValues for exp stack 797 GrowableArray<MonitorValue*> *monarray = new GrowableArray<MonitorValue*>(num_mon); 798 799 // Loop over monitors and insert into array 800 for (idx = 0; idx < num_mon; idx++) { 801 // Grab the node that defines this monitor 802 Node* box_node = sfn->monitor_box(jvms, idx); 803 Node* obj_node = sfn->monitor_obj(jvms, idx); 804 805 // Create ScopeValue for object 806 ScopeValue *scval = NULL; 807 808 if (obj_node->is_SafePointScalarObject()) { 809 SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); 810 scval = Compile::sv_for_node_id(objs, spobj->_idx); 811 if (scval == NULL) { 812 const Type *t = spobj->bottom_type(); 813 ciKlass* cik = t->is_oopptr()->klass(); 814 assert(cik->is_instance_klass() || 815 cik->is_array_klass(), "Not supported allocation."); 816 ObjectValue* sv = new ObjectValue(spobj->_idx, 817 new ConstantOopWriteValue(cik->java_mirror()->constant_encoding())); 818 Compile::set_sv_for_object_node(objs, sv); 819 820 uint first_ind = spobj->first_index(youngest_jvms); 821 for (uint i = 0; i < spobj->n_fields(); i++) { 822 Node* fld_node = sfn->in(first_ind+i); 823 (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); 824 } 825 scval = sv; 826 } 827 } else if (!obj_node->is_Con()) { 828 OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); 829 if( obj_node->bottom_type()->base() == Type::NarrowOop ) { 830 scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); 831 } else { 832 scval = new_loc_value( _regalloc, obj_reg, Location::oop ); 833 } 834 } else { 835 const TypePtr *tp = obj_node->get_ptr_type(); 836 scval = new ConstantOopWriteValue(tp->is_oopptr()->const_oop()->constant_encoding()); 837 } 838 839 OptoReg::Name box_reg = BoxLockNode::reg(box_node); 840 Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); 841 bool eliminated = (box_node->is_BoxLock() && box_node->as_BoxLock()->is_eliminated()); 842 monarray->append(new MonitorValue(scval, basic_lock, eliminated)); 843 } 844 845 // We dump the object pool first, since deoptimization reads it in first. 846 debug_info()->dump_object_pool(objs); 847 848 // Build first class objects to pass to scope 849 DebugToken *locvals = debug_info()->create_scope_values(locarray); 850 DebugToken *expvals = debug_info()->create_scope_values(exparray); 851 DebugToken *monvals = debug_info()->create_monitor_values(monarray); 852 853 // Make method available for all Safepoints 854 ciMethod* scope_method = method ? method : _method; 855 // Describe the scope here 856 assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); 857 assert(!jvms->should_reexecute() || depth == max_depth, "reexecute allowed only for the youngest"); 858 // Now we can describe the scope. 859 methodHandle null_mh; 860 bool rethrow_exception = false; 861 debug_info()->describe_scope(safepoint_pc_offset, null_mh, scope_method, jvms->bci(), jvms->should_reexecute(), rethrow_exception, is_method_handle_invoke, return_oop, locvals, expvals, monvals); 862 } // End jvms loop 863 864 // Mark the end of the scope set. 865 debug_info()->end_safepoint(safepoint_pc_offset); 866 } 867 868 869 870 // A simplified version of Process_OopMap_Node, to handle non-safepoints. 871 class NonSafepointEmitter { 872 Compile* C; 873 JVMState* _pending_jvms; 874 int _pending_offset; 875 876 void emit_non_safepoint(); 877 878 public: 879 NonSafepointEmitter(Compile* compile) { 880 this->C = compile; 881 _pending_jvms = NULL; 882 _pending_offset = 0; 883 } 884 885 void observe_instruction(Node* n, int pc_offset) { 886 if (!C->debug_info()->recording_non_safepoints()) return; 887 888 Node_Notes* nn = C->node_notes_at(n->_idx); 889 if (nn == NULL || nn->jvms() == NULL) return; 890 if (_pending_jvms != NULL && 891 _pending_jvms->same_calls_as(nn->jvms())) { 892 // Repeated JVMS? Stretch it up here. 893 _pending_offset = pc_offset; 894 } else { 895 if (_pending_jvms != NULL && 896 _pending_offset < pc_offset) { 897 emit_non_safepoint(); 898 } 899 _pending_jvms = NULL; 900 if (pc_offset > C->debug_info()->last_pc_offset()) { 901 // This is the only way _pending_jvms can become non-NULL: 902 _pending_jvms = nn->jvms(); 903 _pending_offset = pc_offset; 904 } 905 } 906 } 907 908 // Stay out of the way of real safepoints: 909 void observe_safepoint(JVMState* jvms, int pc_offset) { 910 if (_pending_jvms != NULL && 911 !_pending_jvms->same_calls_as(jvms) && 912 _pending_offset < pc_offset) { 913 emit_non_safepoint(); 914 } 915 _pending_jvms = NULL; 916 } 917 918 void flush_at_end() { 919 if (_pending_jvms != NULL) { 920 emit_non_safepoint(); 921 } 922 _pending_jvms = NULL; 923 } 924 }; 925 926 void NonSafepointEmitter::emit_non_safepoint() { 927 JVMState* youngest_jvms = _pending_jvms; 928 int pc_offset = _pending_offset; 929 930 // Clear it now: 931 _pending_jvms = NULL; 932 933 DebugInformationRecorder* debug_info = C->debug_info(); 934 assert(debug_info->recording_non_safepoints(), "sanity"); 935 936 debug_info->add_non_safepoint(pc_offset); 937 int max_depth = youngest_jvms->depth(); 938 939 // Visit scopes from oldest to youngest. 940 for (int depth = 1; depth <= max_depth; depth++) { 941 JVMState* jvms = youngest_jvms->of_depth(depth); 942 ciMethod* method = jvms->has_method() ? jvms->method() : NULL; 943 assert(!jvms->should_reexecute() || depth==max_depth, "reexecute allowed only for the youngest"); 944 methodHandle null_mh; 945 debug_info->describe_scope(pc_offset, null_mh, method, jvms->bci(), jvms->should_reexecute()); 946 } 947 948 // Mark the end of the scope set. 949 debug_info->end_non_safepoint(pc_offset); 950 } 951 952 //------------------------------init_buffer------------------------------------ 953 CodeBuffer* Compile::init_buffer(uint* blk_starts) { 954 955 // Set the initially allocated size 956 int code_req = initial_code_capacity; 957 int locs_req = initial_locs_capacity; 958 int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; 959 int const_req = initial_const_capacity; 960 961 int pad_req = NativeCall::instruction_size; 962 // The extra spacing after the code is necessary on some platforms. 963 // Sometimes we need to patch in a jump after the last instruction, 964 // if the nmethod has been deoptimized. (See 4932387, 4894843.) 965 966 // Compute the byte offset where we can store the deopt pc. 967 if (fixed_slots() != 0) { 968 _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); 969 } 970 971 // Compute prolog code size 972 _method_size = 0; 973 _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; 974 #if defined(IA64) && !defined(AIX) 975 if (save_argument_registers()) { 976 // 4815101: this is a stub with implicit and unknown precision fp args. 977 // The usual spill mechanism can only generate stfd's in this case, which 978 // doesn't work if the fp reg to spill contains a single-precision denorm. 979 // Instead, we hack around the normal spill mechanism using stfspill's and 980 // ldffill's in the MachProlog and MachEpilog emit methods. We allocate 981 // space here for the fp arg regs (f8-f15) we're going to thusly spill. 982 // 983 // If we ever implement 16-byte 'registers' == stack slots, we can 984 // get rid of this hack and have SpillCopy generate stfspill/ldffill 985 // instead of stfd/stfs/ldfd/ldfs. 986 _frame_slots += 8*(16/BytesPerInt); 987 } 988 #endif 989 assert(_frame_slots >= 0 && _frame_slots < 1000000, "sanity check"); 990 991 if (has_mach_constant_base_node()) { 992 uint add_size = 0; 993 // Fill the constant table. 994 // Note: This must happen before shorten_branches. 995 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 996 Block* b = _cfg->get_block(i); 997 998 for (uint j = 0; j < b->number_of_nodes(); j++) { 999 Node* n = b->get_node(j); 1000 1001 // If the node is a MachConstantNode evaluate the constant 1002 // value section. 1003 if (n->is_MachConstant()) { 1004 MachConstantNode* machcon = n->as_MachConstant(); 1005 machcon->eval_constant(C); 1006 } else if (n->is_Mach()) { 1007 // On Power there are more nodes that issue constants. 1008 add_size += (n->as_Mach()->ins_num_consts() * 8); 1009 } 1010 } 1011 } 1012 1013 // Calculate the offsets of the constants and the size of the 1014 // constant table (including the padding to the next section). 1015 constant_table().calculate_offsets_and_size(); 1016 const_req = constant_table().size() + add_size; 1017 } 1018 1019 // Initialize the space for the BufferBlob used to find and verify 1020 // instruction size in MachNode::emit_size() 1021 init_scratch_buffer_blob(const_req); 1022 if (failing()) return NULL; // Out of memory 1023 1024 // Pre-compute the length of blocks and replace 1025 // long branches with short if machine supports it. 1026 shorten_branches(blk_starts, code_req, locs_req, stub_req); 1027 1028 // nmethod and CodeBuffer count stubs & constants as part of method's code. 1029 // class HandlerImpl is platform-specific and defined in the *.ad files. 1030 int exception_handler_req = HandlerImpl::size_exception_handler() + MAX_stubs_size; // add marginal slop for handler 1031 int deopt_handler_req = HandlerImpl::size_deopt_handler() + MAX_stubs_size; // add marginal slop for handler 1032 stub_req += MAX_stubs_size; // ensure per-stub margin 1033 code_req += MAX_inst_size; // ensure per-instruction margin 1034 1035 if (StressCodeBuffers) 1036 code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion 1037 1038 int total_req = 1039 const_req + 1040 code_req + 1041 pad_req + 1042 stub_req + 1043 exception_handler_req + 1044 deopt_handler_req; // deopt handler 1045 1046 if (has_method_handle_invokes()) 1047 total_req += deopt_handler_req; // deopt MH handler 1048 1049 CodeBuffer* cb = code_buffer(); 1050 cb->initialize(total_req, locs_req); 1051 1052 // Have we run out of code space? 1053 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1054 C->record_failure("CodeCache is full"); 1055 return NULL; 1056 } 1057 // Configure the code buffer. 1058 cb->initialize_consts_size(const_req); 1059 cb->initialize_stubs_size(stub_req); 1060 cb->initialize_oop_recorder(env()->oop_recorder()); 1061 1062 // fill in the nop array for bundling computations 1063 MachNode *_nop_list[Bundle::_nop_count]; 1064 Bundle::initialize_nops(_nop_list); 1065 1066 return cb; 1067 } 1068 1069 //------------------------------fill_buffer------------------------------------ 1070 void Compile::fill_buffer(CodeBuffer* cb, uint* blk_starts) { 1071 // blk_starts[] contains offsets calculated during short branches processing, 1072 // offsets should not be increased during following steps. 1073 1074 // Compute the size of first NumberOfLoopInstrToAlign instructions at head 1075 // of a loop. It is used to determine the padding for loop alignment. 1076 compute_loop_first_inst_sizes(); 1077 1078 // Create oopmap set. 1079 _oop_map_set = new OopMapSet(); 1080 1081 // !!!!! This preserves old handling of oopmaps for now 1082 debug_info()->set_oopmaps(_oop_map_set); 1083 1084 uint nblocks = _cfg->number_of_blocks(); 1085 // Count and start of implicit null check instructions 1086 uint inct_cnt = 0; 1087 uint *inct_starts = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1088 1089 // Count and start of calls 1090 uint *call_returns = NEW_RESOURCE_ARRAY(uint, nblocks+1); 1091 1092 uint return_offset = 0; 1093 int nop_size = (new MachNopNode())->size(_regalloc); 1094 1095 int previous_offset = 0; 1096 int current_offset = 0; 1097 int last_call_offset = -1; 1098 int last_avoid_back_to_back_offset = -1; 1099 #ifdef ASSERT 1100 uint* jmp_target = NEW_RESOURCE_ARRAY(uint,nblocks); 1101 uint* jmp_offset = NEW_RESOURCE_ARRAY(uint,nblocks); 1102 uint* jmp_size = NEW_RESOURCE_ARRAY(uint,nblocks); 1103 uint* jmp_rule = NEW_RESOURCE_ARRAY(uint,nblocks); 1104 #endif 1105 1106 // Create an array of unused labels, one for each basic block, if printing is enabled 1107 #ifndef PRODUCT 1108 int *node_offsets = NULL; 1109 uint node_offset_limit = unique(); 1110 1111 if (print_assembly()) 1112 node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); 1113 #endif 1114 1115 NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily 1116 1117 // Emit the constant table. 1118 if (has_mach_constant_base_node()) { 1119 constant_table().emit(*cb); 1120 } 1121 1122 // Create an array of labels, one for each basic block 1123 Label *blk_labels = NEW_RESOURCE_ARRAY(Label, nblocks+1); 1124 for (uint i=0; i <= nblocks; i++) { 1125 blk_labels[i].init(); 1126 } 1127 1128 // ------------------ 1129 // Now fill in the code buffer 1130 Node *delay_slot = NULL; 1131 1132 for (uint i = 0; i < nblocks; i++) { 1133 Block* block = _cfg->get_block(i); 1134 Node* head = block->head(); 1135 1136 // If this block needs to start aligned (i.e, can be reached other 1137 // than by falling-thru from the previous block), then force the 1138 // start of a new bundle. 1139 if (Pipeline::requires_bundling() && starts_bundle(head)) { 1140 cb->flush_bundle(true); 1141 } 1142 1143 #ifdef ASSERT 1144 if (!block->is_connector()) { 1145 stringStream st; 1146 block->dump_head(_cfg, &st); 1147 MacroAssembler(cb).block_comment(st.as_string()); 1148 } 1149 jmp_target[i] = 0; 1150 jmp_offset[i] = 0; 1151 jmp_size[i] = 0; 1152 jmp_rule[i] = 0; 1153 #endif 1154 int blk_offset = current_offset; 1155 1156 // Define the label at the beginning of the basic block 1157 MacroAssembler(cb).bind(blk_labels[block->_pre_order]); 1158 1159 uint last_inst = block->number_of_nodes(); 1160 1161 // Emit block normally, except for last instruction. 1162 // Emit means "dump code bits into code buffer". 1163 for (uint j = 0; j<last_inst; j++) { 1164 1165 // Get the node 1166 Node* n = block->get_node(j); 1167 1168 // See if delay slots are supported 1169 if (valid_bundle_info(n) && 1170 node_bundling(n)->used_in_unconditional_delay()) { 1171 assert(delay_slot == NULL, "no use of delay slot node"); 1172 assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); 1173 1174 delay_slot = n; 1175 continue; 1176 } 1177 1178 // If this starts a new instruction group, then flush the current one 1179 // (but allow split bundles) 1180 if (Pipeline::requires_bundling() && starts_bundle(n)) 1181 cb->flush_bundle(false); 1182 1183 // Special handling for SafePoint/Call Nodes 1184 bool is_mcall = false; 1185 if (n->is_Mach()) { 1186 MachNode *mach = n->as_Mach(); 1187 is_mcall = n->is_MachCall(); 1188 bool is_sfn = n->is_MachSafePoint(); 1189 1190 // If this requires all previous instructions be flushed, then do so 1191 if (is_sfn || is_mcall || mach->alignment_required() != 1) { 1192 cb->flush_bundle(true); 1193 current_offset = cb->insts_size(); 1194 } 1195 1196 // A padding may be needed again since a previous instruction 1197 // could be moved to delay slot. 1198 1199 // align the instruction if necessary 1200 int padding = mach->compute_padding(current_offset); 1201 // Make sure safepoint node for polling is distinct from a call's 1202 // return by adding a nop if needed. 1203 if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset) { 1204 padding = nop_size; 1205 } 1206 if (padding == 0 && mach->avoid_back_to_back(MachNode::AVOID_BEFORE) && 1207 current_offset == last_avoid_back_to_back_offset) { 1208 // Avoid back to back some instructions. 1209 padding = nop_size; 1210 } 1211 1212 if(padding > 0) { 1213 assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); 1214 int nops_cnt = padding / nop_size; 1215 MachNode *nop = new MachNopNode(nops_cnt); 1216 block->insert_node(nop, j++); 1217 last_inst++; 1218 _cfg->map_node_to_block(nop, block); 1219 nop->emit(*cb, _regalloc); 1220 cb->flush_bundle(true); 1221 current_offset = cb->insts_size(); 1222 } 1223 1224 // Remember the start of the last call in a basic block 1225 if (is_mcall) { 1226 MachCallNode *mcall = mach->as_MachCall(); 1227 1228 // This destination address is NOT PC-relative 1229 mcall->method_set((intptr_t)mcall->entry_point()); 1230 1231 // Save the return address 1232 call_returns[block->_pre_order] = current_offset + mcall->ret_addr_offset(); 1233 1234 if (mcall->is_MachCallLeaf()) { 1235 is_mcall = false; 1236 is_sfn = false; 1237 } 1238 } 1239 1240 // sfn will be valid whenever mcall is valid now because of inheritance 1241 if (is_sfn || is_mcall) { 1242 1243 // Handle special safepoint nodes for synchronization 1244 if (!is_mcall) { 1245 MachSafePointNode *sfn = mach->as_MachSafePoint(); 1246 // !!!!! Stubs only need an oopmap right now, so bail out 1247 if (sfn->jvms()->method() == NULL) { 1248 // Write the oopmap directly to the code blob??!! 1249 continue; 1250 } 1251 } // End synchronization 1252 1253 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1254 current_offset); 1255 Process_OopMap_Node(mach, current_offset); 1256 } // End if safepoint 1257 1258 // If this is a null check, then add the start of the previous instruction to the list 1259 else if( mach->is_MachNullCheck() ) { 1260 inct_starts[inct_cnt++] = previous_offset; 1261 } 1262 1263 // If this is a branch, then fill in the label with the target BB's label 1264 else if (mach->is_MachBranch()) { 1265 // This requires the TRUE branch target be in succs[0] 1266 uint block_num = block->non_connector_successor(0)->_pre_order; 1267 1268 // Try to replace long branch if delay slot is not used, 1269 // it is mostly for back branches since forward branch's 1270 // distance is not updated yet. 1271 bool delay_slot_is_used = valid_bundle_info(n) && 1272 node_bundling(n)->use_unconditional_delay(); 1273 if (!delay_slot_is_used && mach->may_be_short_branch()) { 1274 assert(delay_slot == NULL, "not expecting delay slot node"); 1275 int br_size = n->size(_regalloc); 1276 int offset = blk_starts[block_num] - current_offset; 1277 if (block_num >= i) { 1278 // Current and following block's offset are not 1279 // finalized yet, adjust distance by the difference 1280 // between calculated and final offsets of current block. 1281 offset -= (blk_starts[i] - blk_offset); 1282 } 1283 // In the following code a nop could be inserted before 1284 // the branch which will increase the backward distance. 1285 bool needs_padding = (current_offset == last_avoid_back_to_back_offset); 1286 if (needs_padding && offset <= 0) 1287 offset -= nop_size; 1288 1289 if (_matcher->is_short_branch_offset(mach->rule(), br_size, offset)) { 1290 // We've got a winner. Replace this branch. 1291 MachNode* replacement = mach->as_MachBranch()->short_branch_version(); 1292 1293 // Update the jmp_size. 1294 int new_size = replacement->size(_regalloc); 1295 assert((br_size - new_size) >= (int)nop_size, "short_branch size should be smaller"); 1296 // Insert padding between avoid_back_to_back branches. 1297 if (needs_padding && replacement->avoid_back_to_back(MachNode::AVOID_BEFORE)) { 1298 MachNode *nop = new MachNopNode(); 1299 block->insert_node(nop, j++); 1300 _cfg->map_node_to_block(nop, block); 1301 last_inst++; 1302 nop->emit(*cb, _regalloc); 1303 cb->flush_bundle(true); 1304 current_offset = cb->insts_size(); 1305 } 1306 #ifdef ASSERT 1307 jmp_target[i] = block_num; 1308 jmp_offset[i] = current_offset - blk_offset; 1309 jmp_size[i] = new_size; 1310 jmp_rule[i] = mach->rule(); 1311 #endif 1312 block->map_node(replacement, j); 1313 mach->subsume_by(replacement, C); 1314 n = replacement; 1315 mach = replacement; 1316 } 1317 } 1318 mach->as_MachBranch()->label_set( &blk_labels[block_num], block_num ); 1319 } else if (mach->ideal_Opcode() == Op_Jump) { 1320 for (uint h = 0; h < block->_num_succs; h++) { 1321 Block* succs_block = block->_succs[h]; 1322 for (uint j = 1; j < succs_block->num_preds(); j++) { 1323 Node* jpn = succs_block->pred(j); 1324 if (jpn->is_JumpProj() && jpn->in(0) == mach) { 1325 uint block_num = succs_block->non_connector()->_pre_order; 1326 Label *blkLabel = &blk_labels[block_num]; 1327 mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); 1328 } 1329 } 1330 } 1331 } 1332 #ifdef ASSERT 1333 // Check that oop-store precedes the card-mark 1334 else if (mach->ideal_Opcode() == Op_StoreCM) { 1335 uint storeCM_idx = j; 1336 int count = 0; 1337 for (uint prec = mach->req(); prec < mach->len(); prec++) { 1338 Node *oop_store = mach->in(prec); // Precedence edge 1339 if (oop_store == NULL) continue; 1340 count++; 1341 uint i4; 1342 for (i4 = 0; i4 < last_inst; ++i4) { 1343 if (block->get_node(i4) == oop_store) { 1344 break; 1345 } 1346 } 1347 // Note: This test can provide a false failure if other precedence 1348 // edges have been added to the storeCMNode. 1349 assert(i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); 1350 } 1351 assert(count > 0, "storeCM expects at least one precedence edge"); 1352 } 1353 #endif 1354 else if (!n->is_Proj()) { 1355 // Remember the beginning of the previous instruction, in case 1356 // it's followed by a flag-kill and a null-check. Happens on 1357 // Intel all the time, with add-to-memory kind of opcodes. 1358 previous_offset = current_offset; 1359 } 1360 1361 // Not an else-if! 1362 // If this is a trap based cmp then add its offset to the list. 1363 if (mach->is_TrapBasedCheckNode()) { 1364 inct_starts[inct_cnt++] = current_offset; 1365 } 1366 } 1367 1368 // Verify that there is sufficient space remaining 1369 cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); 1370 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1371 C->record_failure("CodeCache is full"); 1372 return; 1373 } 1374 1375 // Save the offset for the listing 1376 #ifndef PRODUCT 1377 if (node_offsets && n->_idx < node_offset_limit) 1378 node_offsets[n->_idx] = cb->insts_size(); 1379 #endif 1380 1381 // "Normal" instruction case 1382 DEBUG_ONLY( uint instr_offset = cb->insts_size(); ) 1383 n->emit(*cb, _regalloc); 1384 current_offset = cb->insts_size(); 1385 1386 // Above we only verified that there is enough space in the instruction section. 1387 // However, the instruction may emit stubs that cause code buffer expansion. 1388 // Bail out here if expansion failed due to a lack of code cache space. 1389 if (failing()) { 1390 return; 1391 } 1392 1393 #ifdef ASSERT 1394 if (n->size(_regalloc) < (current_offset-instr_offset)) { 1395 n->dump(); 1396 assert(false, "wrong size of mach node"); 1397 } 1398 #endif 1399 non_safepoints.observe_instruction(n, current_offset); 1400 1401 // mcall is last "call" that can be a safepoint 1402 // record it so we can see if a poll will directly follow it 1403 // in which case we'll need a pad to make the PcDesc sites unique 1404 // see 5010568. This can be slightly inaccurate but conservative 1405 // in the case that return address is not actually at current_offset. 1406 // This is a small price to pay. 1407 1408 if (is_mcall) { 1409 last_call_offset = current_offset; 1410 } 1411 1412 if (n->is_Mach() && n->as_Mach()->avoid_back_to_back(MachNode::AVOID_AFTER)) { 1413 // Avoid back to back some instructions. 1414 last_avoid_back_to_back_offset = current_offset; 1415 } 1416 1417 // See if this instruction has a delay slot 1418 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1419 assert(delay_slot != NULL, "expecting delay slot node"); 1420 1421 // Back up 1 instruction 1422 cb->set_insts_end(cb->insts_end() - Pipeline::instr_unit_size()); 1423 1424 // Save the offset for the listing 1425 #ifndef PRODUCT 1426 if (node_offsets && delay_slot->_idx < node_offset_limit) 1427 node_offsets[delay_slot->_idx] = cb->insts_size(); 1428 #endif 1429 1430 // Support a SafePoint in the delay slot 1431 if (delay_slot->is_MachSafePoint()) { 1432 MachNode *mach = delay_slot->as_Mach(); 1433 // !!!!! Stubs only need an oopmap right now, so bail out 1434 if (!mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL) { 1435 // Write the oopmap directly to the code blob??!! 1436 delay_slot = NULL; 1437 continue; 1438 } 1439 1440 int adjusted_offset = current_offset - Pipeline::instr_unit_size(); 1441 non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), 1442 adjusted_offset); 1443 // Generate an OopMap entry 1444 Process_OopMap_Node(mach, adjusted_offset); 1445 } 1446 1447 // Insert the delay slot instruction 1448 delay_slot->emit(*cb, _regalloc); 1449 1450 // Don't reuse it 1451 delay_slot = NULL; 1452 } 1453 1454 } // End for all instructions in block 1455 1456 // If the next block is the top of a loop, pad this block out to align 1457 // the loop top a little. Helps prevent pipe stalls at loop back branches. 1458 if (i < nblocks-1) { 1459 Block *nb = _cfg->get_block(i + 1); 1460 int padding = nb->alignment_padding(current_offset); 1461 if( padding > 0 ) { 1462 MachNode *nop = new MachNopNode(padding / nop_size); 1463 block->insert_node(nop, block->number_of_nodes()); 1464 _cfg->map_node_to_block(nop, block); 1465 nop->emit(*cb, _regalloc); 1466 current_offset = cb->insts_size(); 1467 } 1468 } 1469 // Verify that the distance for generated before forward 1470 // short branches is still valid. 1471 guarantee((int)(blk_starts[i+1] - blk_starts[i]) >= (current_offset - blk_offset), "shouldn't increase block size"); 1472 1473 // Save new block start offset 1474 blk_starts[i] = blk_offset; 1475 } // End of for all blocks 1476 blk_starts[nblocks] = current_offset; 1477 1478 non_safepoints.flush_at_end(); 1479 1480 // Offset too large? 1481 if (failing()) return; 1482 1483 // Define a pseudo-label at the end of the code 1484 MacroAssembler(cb).bind( blk_labels[nblocks] ); 1485 1486 // Compute the size of the first block 1487 _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); 1488 1489 assert(cb->insts_size() < 500000, "method is unreasonably large"); 1490 1491 #ifdef ASSERT 1492 for (uint i = 0; i < nblocks; i++) { // For all blocks 1493 if (jmp_target[i] != 0) { 1494 int br_size = jmp_size[i]; 1495 int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_offset[i]); 1496 if (!_matcher->is_short_branch_offset(jmp_rule[i], br_size, offset)) { 1497 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]); 1498 assert(false, "Displacement too large for short jmp"); 1499 } 1500 } 1501 } 1502 #endif 1503 1504 #ifndef PRODUCT 1505 // Information on the size of the method, without the extraneous code 1506 Scheduling::increment_method_size(cb->insts_size()); 1507 #endif 1508 1509 // ------------------ 1510 // Fill in exception table entries. 1511 FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); 1512 1513 // Only java methods have exception handlers and deopt handlers 1514 // class HandlerImpl is platform-specific and defined in the *.ad files. 1515 if (_method) { 1516 // Emit the exception handler code. 1517 _code_offsets.set_value(CodeOffsets::Exceptions, HandlerImpl::emit_exception_handler(*cb)); 1518 if (failing()) { 1519 return; // CodeBuffer::expand failed 1520 } 1521 // Emit the deopt handler code. 1522 _code_offsets.set_value(CodeOffsets::Deopt, HandlerImpl::emit_deopt_handler(*cb)); 1523 1524 // Emit the MethodHandle deopt handler code (if required). 1525 if (has_method_handle_invokes() && !failing()) { 1526 // We can use the same code as for the normal deopt handler, we 1527 // just need a different entry point address. 1528 _code_offsets.set_value(CodeOffsets::DeoptMH, HandlerImpl::emit_deopt_handler(*cb)); 1529 } 1530 } 1531 1532 // One last check for failed CodeBuffer::expand: 1533 if ((cb->blob() == NULL) || (!CompileBroker::should_compile_new_jobs())) { 1534 C->record_failure("CodeCache is full"); 1535 return; 1536 } 1537 1538 #ifndef PRODUCT 1539 // Dump the assembly code, including basic-block numbers 1540 if (print_assembly()) { 1541 ttyLocker ttyl; // keep the following output all in one block 1542 if (!VMThread::should_terminate()) { // test this under the tty lock 1543 // This output goes directly to the tty, not the compiler log. 1544 // To enable tools to match it up with the compilation activity, 1545 // be sure to tag this tty output with the compile ID. 1546 if (xtty != NULL) { 1547 xtty->head("opto_assembly compile_id='%d'%s", compile_id(), 1548 is_osr_compilation() ? " compile_kind='osr'" : 1549 ""); 1550 } 1551 if (method() != NULL) { 1552 method()->print_metadata(); 1553 } 1554 dump_asm(node_offsets, node_offset_limit); 1555 if (xtty != NULL) { 1556 xtty->tail("opto_assembly"); 1557 } 1558 } 1559 } 1560 #endif 1561 1562 } 1563 1564 void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { 1565 _inc_table.set_size(cnt); 1566 1567 uint inct_cnt = 0; 1568 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 1569 Block* block = _cfg->get_block(i); 1570 Node *n = NULL; 1571 int j; 1572 1573 // Find the branch; ignore trailing NOPs. 1574 for (j = block->number_of_nodes() - 1; j >= 0; j--) { 1575 n = block->get_node(j); 1576 if (!n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con) { 1577 break; 1578 } 1579 } 1580 1581 // If we didn't find anything, continue 1582 if (j < 0) { 1583 continue; 1584 } 1585 1586 // Compute ExceptionHandlerTable subtable entry and add it 1587 // (skip empty blocks) 1588 if (n->is_Catch()) { 1589 1590 // Get the offset of the return from the call 1591 uint call_return = call_returns[block->_pre_order]; 1592 #ifdef ASSERT 1593 assert( call_return > 0, "no call seen for this basic block" ); 1594 while (block->get_node(--j)->is_MachProj()) ; 1595 assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); 1596 #endif 1597 // last instruction is a CatchNode, find it's CatchProjNodes 1598 int nof_succs = block->_num_succs; 1599 // allocate space 1600 GrowableArray<intptr_t> handler_bcis(nof_succs); 1601 GrowableArray<intptr_t> handler_pcos(nof_succs); 1602 // iterate through all successors 1603 for (int j = 0; j < nof_succs; j++) { 1604 Block* s = block->_succs[j]; 1605 bool found_p = false; 1606 for (uint k = 1; k < s->num_preds(); k++) { 1607 Node* pk = s->pred(k); 1608 if (pk->is_CatchProj() && pk->in(0) == n) { 1609 const CatchProjNode* p = pk->as_CatchProj(); 1610 found_p = true; 1611 // add the corresponding handler bci & pco information 1612 if (p->_con != CatchProjNode::fall_through_index) { 1613 // p leads to an exception handler (and is not fall through) 1614 assert(s == _cfg->get_block(s->_pre_order), "bad numbering"); 1615 // no duplicates, please 1616 if (!handler_bcis.contains(p->handler_bci())) { 1617 uint block_num = s->non_connector()->_pre_order; 1618 handler_bcis.append(p->handler_bci()); 1619 handler_pcos.append(blk_labels[block_num].loc_pos()); 1620 } 1621 } 1622 } 1623 } 1624 assert(found_p, "no matching predecessor found"); 1625 // Note: Due to empty block removal, one block may have 1626 // several CatchProj inputs, from the same Catch. 1627 } 1628 1629 // Set the offset of the return from the call 1630 _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); 1631 continue; 1632 } 1633 1634 // Handle implicit null exception table updates 1635 if (n->is_MachNullCheck()) { 1636 uint block_num = block->non_connector_successor(0)->_pre_order; 1637 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1638 continue; 1639 } 1640 // Handle implicit exception table updates: trap instructions. 1641 if (n->is_Mach() && n->as_Mach()->is_TrapBasedCheckNode()) { 1642 uint block_num = block->non_connector_successor(0)->_pre_order; 1643 _inc_table.append(inct_starts[inct_cnt++], blk_labels[block_num].loc_pos()); 1644 continue; 1645 } 1646 } // End of for all blocks fill in exception table entries 1647 } 1648 1649 // Static Variables 1650 #ifndef PRODUCT 1651 uint Scheduling::_total_nop_size = 0; 1652 uint Scheduling::_total_method_size = 0; 1653 uint Scheduling::_total_branches = 0; 1654 uint Scheduling::_total_unconditional_delays = 0; 1655 uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; 1656 #endif 1657 1658 // Initializer for class Scheduling 1659 1660 Scheduling::Scheduling(Arena *arena, Compile &compile) 1661 : _arena(arena), 1662 _cfg(compile.cfg()), 1663 _regalloc(compile.regalloc()), 1664 _reg_node(arena), 1665 _bundle_instr_count(0), 1666 _bundle_cycle_number(0), 1667 _scheduled(arena), 1668 _available(arena), 1669 _next_node(NULL), 1670 _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), 1671 _pinch_free_list(arena) 1672 #ifndef PRODUCT 1673 , _branches(0) 1674 , _unconditional_delays(0) 1675 #endif 1676 { 1677 // Create a MachNopNode 1678 _nop = new MachNopNode(); 1679 1680 // Now that the nops are in the array, save the count 1681 // (but allow entries for the nops) 1682 _node_bundling_limit = compile.unique(); 1683 uint node_max = _regalloc->node_regs_max_index(); 1684 1685 compile.set_node_bundling_limit(_node_bundling_limit); 1686 1687 // This one is persistent within the Compile class 1688 _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); 1689 1690 // Allocate space for fixed-size arrays 1691 _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1692 _uses = NEW_ARENA_ARRAY(arena, short, node_max); 1693 _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); 1694 1695 // Clear the arrays 1696 memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); 1697 memset(_node_latency, 0, node_max * sizeof(unsigned short)); 1698 memset(_uses, 0, node_max * sizeof(short)); 1699 memset(_current_latency, 0, node_max * sizeof(unsigned short)); 1700 1701 // Clear the bundling information 1702 memcpy(_bundle_use_elements, Pipeline_Use::elaborated_elements, sizeof(Pipeline_Use::elaborated_elements)); 1703 1704 // Get the last node 1705 Block* block = _cfg->get_block(_cfg->number_of_blocks() - 1); 1706 1707 _next_node = block->get_node(block->number_of_nodes() - 1); 1708 } 1709 1710 #ifndef PRODUCT 1711 // Scheduling destructor 1712 Scheduling::~Scheduling() { 1713 _total_branches += _branches; 1714 _total_unconditional_delays += _unconditional_delays; 1715 } 1716 #endif 1717 1718 // Step ahead "i" cycles 1719 void Scheduling::step(uint i) { 1720 1721 Bundle *bundle = node_bundling(_next_node); 1722 bundle->set_starts_bundle(); 1723 1724 // Update the bundle record, but leave the flags information alone 1725 if (_bundle_instr_count > 0) { 1726 bundle->set_instr_count(_bundle_instr_count); 1727 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1728 } 1729 1730 // Update the state information 1731 _bundle_instr_count = 0; 1732 _bundle_cycle_number += i; 1733 _bundle_use.step(i); 1734 } 1735 1736 void Scheduling::step_and_clear() { 1737 Bundle *bundle = node_bundling(_next_node); 1738 bundle->set_starts_bundle(); 1739 1740 // Update the bundle record 1741 if (_bundle_instr_count > 0) { 1742 bundle->set_instr_count(_bundle_instr_count); 1743 bundle->set_resources_used(_bundle_use.resourcesUsed()); 1744 1745 _bundle_cycle_number += 1; 1746 } 1747 1748 // Clear the bundling information 1749 _bundle_instr_count = 0; 1750 _bundle_use.reset(); 1751 1752 memcpy(_bundle_use_elements, 1753 Pipeline_Use::elaborated_elements, 1754 sizeof(Pipeline_Use::elaborated_elements)); 1755 } 1756 1757 // Perform instruction scheduling and bundling over the sequence of 1758 // instructions in backwards order. 1759 void Compile::ScheduleAndBundle() { 1760 1761 // Don't optimize this if it isn't a method 1762 if (!_method) 1763 return; 1764 1765 // Don't optimize this if scheduling is disabled 1766 if (!do_scheduling()) 1767 return; 1768 1769 // Scheduling code works only with pairs (16 bytes) maximum. 1770 if (max_vector_size() > 16) 1771 return; 1772 1773 TracePhase tp("isched", &timers[_t_instrSched]); 1774 1775 // Create a data structure for all the scheduling information 1776 Scheduling scheduling(Thread::current()->resource_area(), *this); 1777 1778 // Walk backwards over each basic block, computing the needed alignment 1779 // Walk over all the basic blocks 1780 scheduling.DoScheduling(); 1781 } 1782 1783 // Compute the latency of all the instructions. This is fairly simple, 1784 // because we already have a legal ordering. Walk over the instructions 1785 // from first to last, and compute the latency of the instruction based 1786 // on the latency of the preceding instruction(s). 1787 void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { 1788 #ifndef PRODUCT 1789 if (_cfg->C->trace_opto_output()) 1790 tty->print("# -> ComputeLocalLatenciesForward\n"); 1791 #endif 1792 1793 // Walk over all the schedulable instructions 1794 for( uint j=_bb_start; j < _bb_end; j++ ) { 1795 1796 // This is a kludge, forcing all latency calculations to start at 1. 1797 // Used to allow latency 0 to force an instruction to the beginning 1798 // of the bb 1799 uint latency = 1; 1800 Node *use = bb->get_node(j); 1801 uint nlen = use->len(); 1802 1803 // Walk over all the inputs 1804 for ( uint k=0; k < nlen; k++ ) { 1805 Node *def = use->in(k); 1806 if (!def) 1807 continue; 1808 1809 uint l = _node_latency[def->_idx] + use->latency(k); 1810 if (latency < l) 1811 latency = l; 1812 } 1813 1814 _node_latency[use->_idx] = latency; 1815 1816 #ifndef PRODUCT 1817 if (_cfg->C->trace_opto_output()) { 1818 tty->print("# latency %4d: ", latency); 1819 use->dump(); 1820 } 1821 #endif 1822 } 1823 1824 #ifndef PRODUCT 1825 if (_cfg->C->trace_opto_output()) 1826 tty->print("# <- ComputeLocalLatenciesForward\n"); 1827 #endif 1828 1829 } // end ComputeLocalLatenciesForward 1830 1831 // See if this node fits into the present instruction bundle 1832 bool Scheduling::NodeFitsInBundle(Node *n) { 1833 uint n_idx = n->_idx; 1834 1835 // If this is the unconditional delay instruction, then it fits 1836 if (n == _unconditional_delay_slot) { 1837 #ifndef PRODUCT 1838 if (_cfg->C->trace_opto_output()) 1839 tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); 1840 #endif 1841 return (true); 1842 } 1843 1844 // If the node cannot be scheduled this cycle, skip it 1845 if (_current_latency[n_idx] > _bundle_cycle_number) { 1846 #ifndef PRODUCT 1847 if (_cfg->C->trace_opto_output()) 1848 tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", 1849 n->_idx, _current_latency[n_idx], _bundle_cycle_number); 1850 #endif 1851 return (false); 1852 } 1853 1854 const Pipeline *node_pipeline = n->pipeline(); 1855 1856 uint instruction_count = node_pipeline->instructionCount(); 1857 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 1858 instruction_count = 0; 1859 else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 1860 instruction_count++; 1861 1862 if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { 1863 #ifndef PRODUCT 1864 if (_cfg->C->trace_opto_output()) 1865 tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", 1866 n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); 1867 #endif 1868 return (false); 1869 } 1870 1871 // Don't allow non-machine nodes to be handled this way 1872 if (!n->is_Mach() && instruction_count == 0) 1873 return (false); 1874 1875 // See if there is any overlap 1876 uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); 1877 1878 if (delay > 0) { 1879 #ifndef PRODUCT 1880 if (_cfg->C->trace_opto_output()) 1881 tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); 1882 #endif 1883 return false; 1884 } 1885 1886 #ifndef PRODUCT 1887 if (_cfg->C->trace_opto_output()) 1888 tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); 1889 #endif 1890 1891 return true; 1892 } 1893 1894 Node * Scheduling::ChooseNodeToBundle() { 1895 uint siz = _available.size(); 1896 1897 if (siz == 0) { 1898 1899 #ifndef PRODUCT 1900 if (_cfg->C->trace_opto_output()) 1901 tty->print("# ChooseNodeToBundle: NULL\n"); 1902 #endif 1903 return (NULL); 1904 } 1905 1906 // Fast path, if only 1 instruction in the bundle 1907 if (siz == 1) { 1908 #ifndef PRODUCT 1909 if (_cfg->C->trace_opto_output()) { 1910 tty->print("# ChooseNodeToBundle (only 1): "); 1911 _available[0]->dump(); 1912 } 1913 #endif 1914 return (_available[0]); 1915 } 1916 1917 // Don't bother, if the bundle is already full 1918 if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { 1919 for ( uint i = 0; i < siz; i++ ) { 1920 Node *n = _available[i]; 1921 1922 // Skip projections, we'll handle them another way 1923 if (n->is_Proj()) 1924 continue; 1925 1926 // This presupposed that instructions are inserted into the 1927 // available list in a legality order; i.e. instructions that 1928 // must be inserted first are at the head of the list 1929 if (NodeFitsInBundle(n)) { 1930 #ifndef PRODUCT 1931 if (_cfg->C->trace_opto_output()) { 1932 tty->print("# ChooseNodeToBundle: "); 1933 n->dump(); 1934 } 1935 #endif 1936 return (n); 1937 } 1938 } 1939 } 1940 1941 // Nothing fits in this bundle, choose the highest priority 1942 #ifndef PRODUCT 1943 if (_cfg->C->trace_opto_output()) { 1944 tty->print("# ChooseNodeToBundle: "); 1945 _available[0]->dump(); 1946 } 1947 #endif 1948 1949 return _available[0]; 1950 } 1951 1952 void Scheduling::AddNodeToAvailableList(Node *n) { 1953 assert( !n->is_Proj(), "projections never directly made available" ); 1954 #ifndef PRODUCT 1955 if (_cfg->C->trace_opto_output()) { 1956 tty->print("# AddNodeToAvailableList: "); 1957 n->dump(); 1958 } 1959 #endif 1960 1961 int latency = _current_latency[n->_idx]; 1962 1963 // Insert in latency order (insertion sort) 1964 uint i; 1965 for ( i=0; i < _available.size(); i++ ) 1966 if (_current_latency[_available[i]->_idx] > latency) 1967 break; 1968 1969 // Special Check for compares following branches 1970 if( n->is_Mach() && _scheduled.size() > 0 ) { 1971 int op = n->as_Mach()->ideal_Opcode(); 1972 Node *last = _scheduled[0]; 1973 if( last->is_MachIf() && last->in(1) == n && 1974 ( op == Op_CmpI || 1975 op == Op_CmpU || 1976 op == Op_CmpP || 1977 op == Op_CmpF || 1978 op == Op_CmpD || 1979 op == Op_CmpL ) ) { 1980 1981 // Recalculate position, moving to front of same latency 1982 for ( i=0 ; i < _available.size(); i++ ) 1983 if (_current_latency[_available[i]->_idx] >= latency) 1984 break; 1985 } 1986 } 1987 1988 // Insert the node in the available list 1989 _available.insert(i, n); 1990 1991 #ifndef PRODUCT 1992 if (_cfg->C->trace_opto_output()) 1993 dump_available(); 1994 #endif 1995 } 1996 1997 void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { 1998 for ( uint i=0; i < n->len(); i++ ) { 1999 Node *def = n->in(i); 2000 if (!def) continue; 2001 if( def->is_Proj() ) // If this is a machine projection, then 2002 def = def->in(0); // propagate usage thru to the base instruction 2003 2004 if(_cfg->get_block_for_node(def) != bb) { // Ignore if not block-local 2005 continue; 2006 } 2007 2008 // Compute the latency 2009 uint l = _bundle_cycle_number + n->latency(i); 2010 if (_current_latency[def->_idx] < l) 2011 _current_latency[def->_idx] = l; 2012 2013 // If this does not have uses then schedule it 2014 if ((--_uses[def->_idx]) == 0) 2015 AddNodeToAvailableList(def); 2016 } 2017 } 2018 2019 void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { 2020 #ifndef PRODUCT 2021 if (_cfg->C->trace_opto_output()) { 2022 tty->print("# AddNodeToBundle: "); 2023 n->dump(); 2024 } 2025 #endif 2026 2027 // Remove this from the available list 2028 uint i; 2029 for (i = 0; i < _available.size(); i++) 2030 if (_available[i] == n) 2031 break; 2032 assert(i < _available.size(), "entry in _available list not found"); 2033 _available.remove(i); 2034 2035 // See if this fits in the current bundle 2036 const Pipeline *node_pipeline = n->pipeline(); 2037 const Pipeline_Use& node_usage = node_pipeline->resourceUse(); 2038 2039 // Check for instructions to be placed in the delay slot. We 2040 // do this before we actually schedule the current instruction, 2041 // because the delay slot follows the current instruction. 2042 if (Pipeline::_branch_has_delay_slot && 2043 node_pipeline->hasBranchDelay() && 2044 !_unconditional_delay_slot) { 2045 2046 uint siz = _available.size(); 2047 2048 // Conditional branches can support an instruction that 2049 // is unconditionally executed and not dependent by the 2050 // branch, OR a conditionally executed instruction if 2051 // the branch is taken. In practice, this means that 2052 // the first instruction at the branch target is 2053 // copied to the delay slot, and the branch goes to 2054 // the instruction after that at the branch target 2055 if ( n->is_MachBranch() ) { 2056 2057 assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); 2058 assert( !n->is_Catch(), "should not look for delay slot for Catch" ); 2059 2060 #ifndef PRODUCT 2061 _branches++; 2062 #endif 2063 2064 // At least 1 instruction is on the available list 2065 // that is not dependent on the branch 2066 for (uint i = 0; i < siz; i++) { 2067 Node *d = _available[i]; 2068 const Pipeline *avail_pipeline = d->pipeline(); 2069 2070 // Don't allow safepoints in the branch shadow, that will 2071 // cause a number of difficulties 2072 if ( avail_pipeline->instructionCount() == 1 && 2073 !avail_pipeline->hasMultipleBundles() && 2074 !avail_pipeline->hasBranchDelay() && 2075 Pipeline::instr_has_unit_size() && 2076 d->size(_regalloc) == Pipeline::instr_unit_size() && 2077 NodeFitsInBundle(d) && 2078 !node_bundling(d)->used_in_delay()) { 2079 2080 if (d->is_Mach() && !d->is_MachSafePoint()) { 2081 // A node that fits in the delay slot was found, so we need to 2082 // set the appropriate bits in the bundle pipeline information so 2083 // that it correctly indicates resource usage. Later, when we 2084 // attempt to add this instruction to the bundle, we will skip 2085 // setting the resource usage. 2086 _unconditional_delay_slot = d; 2087 node_bundling(n)->set_use_unconditional_delay(); 2088 node_bundling(d)->set_used_in_unconditional_delay(); 2089 _bundle_use.add_usage(avail_pipeline->resourceUse()); 2090 _current_latency[d->_idx] = _bundle_cycle_number; 2091 _next_node = d; 2092 ++_bundle_instr_count; 2093 #ifndef PRODUCT 2094 _unconditional_delays++; 2095 #endif 2096 break; 2097 } 2098 } 2099 } 2100 } 2101 2102 // No delay slot, add a nop to the usage 2103 if (!_unconditional_delay_slot) { 2104 // See if adding an instruction in the delay slot will overflow 2105 // the bundle. 2106 if (!NodeFitsInBundle(_nop)) { 2107 #ifndef PRODUCT 2108 if (_cfg->C->trace_opto_output()) 2109 tty->print("# *** STEP(1 instruction for delay slot) ***\n"); 2110 #endif 2111 step(1); 2112 } 2113 2114 _bundle_use.add_usage(_nop->pipeline()->resourceUse()); 2115 _next_node = _nop; 2116 ++_bundle_instr_count; 2117 } 2118 2119 // See if the instruction in the delay slot requires a 2120 // step of the bundles 2121 if (!NodeFitsInBundle(n)) { 2122 #ifndef PRODUCT 2123 if (_cfg->C->trace_opto_output()) 2124 tty->print("# *** STEP(branch won't fit) ***\n"); 2125 #endif 2126 // Update the state information 2127 _bundle_instr_count = 0; 2128 _bundle_cycle_number += 1; 2129 _bundle_use.step(1); 2130 } 2131 } 2132 2133 // Get the number of instructions 2134 uint instruction_count = node_pipeline->instructionCount(); 2135 if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) 2136 instruction_count = 0; 2137 2138 // Compute the latency information 2139 uint delay = 0; 2140 2141 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { 2142 int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; 2143 if (relative_latency < 0) 2144 relative_latency = 0; 2145 2146 delay = _bundle_use.full_latency(relative_latency, node_usage); 2147 2148 // Does not fit in this bundle, start a new one 2149 if (delay > 0) { 2150 step(delay); 2151 2152 #ifndef PRODUCT 2153 if (_cfg->C->trace_opto_output()) 2154 tty->print("# *** STEP(%d) ***\n", delay); 2155 #endif 2156 } 2157 } 2158 2159 // If this was placed in the delay slot, ignore it 2160 if (n != _unconditional_delay_slot) { 2161 2162 if (delay == 0) { 2163 if (node_pipeline->hasMultipleBundles()) { 2164 #ifndef PRODUCT 2165 if (_cfg->C->trace_opto_output()) 2166 tty->print("# *** STEP(multiple instructions) ***\n"); 2167 #endif 2168 step(1); 2169 } 2170 2171 else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { 2172 #ifndef PRODUCT 2173 if (_cfg->C->trace_opto_output()) 2174 tty->print("# *** STEP(%d >= %d instructions) ***\n", 2175 instruction_count + _bundle_instr_count, 2176 Pipeline::_max_instrs_per_cycle); 2177 #endif 2178 step(1); 2179 } 2180 } 2181 2182 if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) 2183 _bundle_instr_count++; 2184 2185 // Set the node's latency 2186 _current_latency[n->_idx] = _bundle_cycle_number; 2187 2188 // Now merge the functional unit information 2189 if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) 2190 _bundle_use.add_usage(node_usage); 2191 2192 // Increment the number of instructions in this bundle 2193 _bundle_instr_count += instruction_count; 2194 2195 // Remember this node for later 2196 if (n->is_Mach()) 2197 _next_node = n; 2198 } 2199 2200 // It's possible to have a BoxLock in the graph and in the _bbs mapping but 2201 // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. 2202 // 'Schedule' them (basically ignore in the schedule) but do not insert them 2203 // into the block. All other scheduled nodes get put in the schedule here. 2204 int op = n->Opcode(); 2205 if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR 2206 (op != Op_Node && // Not an unused antidepedence node and 2207 // not an unallocated boxlock 2208 (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { 2209 2210 // Push any trailing projections 2211 if( bb->get_node(bb->number_of_nodes()-1) != n ) { 2212 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2213 Node *foi = n->fast_out(i); 2214 if( foi->is_Proj() ) 2215 _scheduled.push(foi); 2216 } 2217 } 2218 2219 // Put the instruction in the schedule list 2220 _scheduled.push(n); 2221 } 2222 2223 #ifndef PRODUCT 2224 if (_cfg->C->trace_opto_output()) 2225 dump_available(); 2226 #endif 2227 2228 // Walk all the definitions, decrementing use counts, and 2229 // if a definition has a 0 use count, place it in the available list. 2230 DecrementUseCounts(n,bb); 2231 } 2232 2233 // This method sets the use count within a basic block. We will ignore all 2234 // uses outside the current basic block. As we are doing a backwards walk, 2235 // any node we reach that has a use count of 0 may be scheduled. This also 2236 // avoids the problem of cyclic references from phi nodes, as long as phi 2237 // nodes are at the front of the basic block. This method also initializes 2238 // the available list to the set of instructions that have no uses within this 2239 // basic block. 2240 void Scheduling::ComputeUseCount(const Block *bb) { 2241 #ifndef PRODUCT 2242 if (_cfg->C->trace_opto_output()) 2243 tty->print("# -> ComputeUseCount\n"); 2244 #endif 2245 2246 // Clear the list of available and scheduled instructions, just in case 2247 _available.clear(); 2248 _scheduled.clear(); 2249 2250 // No delay slot specified 2251 _unconditional_delay_slot = NULL; 2252 2253 #ifdef ASSERT 2254 for( uint i=0; i < bb->number_of_nodes(); i++ ) 2255 assert( _uses[bb->get_node(i)->_idx] == 0, "_use array not clean" ); 2256 #endif 2257 2258 // Force the _uses count to never go to zero for unscheduable pieces 2259 // of the block 2260 for( uint k = 0; k < _bb_start; k++ ) 2261 _uses[bb->get_node(k)->_idx] = 1; 2262 for( uint l = _bb_end; l < bb->number_of_nodes(); l++ ) 2263 _uses[bb->get_node(l)->_idx] = 1; 2264 2265 // Iterate backwards over the instructions in the block. Don't count the 2266 // branch projections at end or the block header instructions. 2267 for( uint j = _bb_end-1; j >= _bb_start; j-- ) { 2268 Node *n = bb->get_node(j); 2269 if( n->is_Proj() ) continue; // Projections handled another way 2270 2271 // Account for all uses 2272 for ( uint k = 0; k < n->len(); k++ ) { 2273 Node *inp = n->in(k); 2274 if (!inp) continue; 2275 assert(inp != n, "no cycles allowed" ); 2276 if (_cfg->get_block_for_node(inp) == bb) { // Block-local use? 2277 if (inp->is_Proj()) { // Skip through Proj's 2278 inp = inp->in(0); 2279 } 2280 ++_uses[inp->_idx]; // Count 1 block-local use 2281 } 2282 } 2283 2284 // If this instruction has a 0 use count, then it is available 2285 if (!_uses[n->_idx]) { 2286 _current_latency[n->_idx] = _bundle_cycle_number; 2287 AddNodeToAvailableList(n); 2288 } 2289 2290 #ifndef PRODUCT 2291 if (_cfg->C->trace_opto_output()) { 2292 tty->print("# uses: %3d: ", _uses[n->_idx]); 2293 n->dump(); 2294 } 2295 #endif 2296 } 2297 2298 #ifndef PRODUCT 2299 if (_cfg->C->trace_opto_output()) 2300 tty->print("# <- ComputeUseCount\n"); 2301 #endif 2302 } 2303 2304 // This routine performs scheduling on each basic block in reverse order, 2305 // using instruction latencies and taking into account function unit 2306 // availability. 2307 void Scheduling::DoScheduling() { 2308 #ifndef PRODUCT 2309 if (_cfg->C->trace_opto_output()) 2310 tty->print("# -> DoScheduling\n"); 2311 #endif 2312 2313 Block *succ_bb = NULL; 2314 Block *bb; 2315 2316 // Walk over all the basic blocks in reverse order 2317 for (int i = _cfg->number_of_blocks() - 1; i >= 0; succ_bb = bb, i--) { 2318 bb = _cfg->get_block(i); 2319 2320 #ifndef PRODUCT 2321 if (_cfg->C->trace_opto_output()) { 2322 tty->print("# Schedule BB#%03d (initial)\n", i); 2323 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2324 bb->get_node(j)->dump(); 2325 } 2326 } 2327 #endif 2328 2329 // On the head node, skip processing 2330 if (bb == _cfg->get_root_block()) { 2331 continue; 2332 } 2333 2334 // Skip empty, connector blocks 2335 if (bb->is_connector()) 2336 continue; 2337 2338 // If the following block is not the sole successor of 2339 // this one, then reset the pipeline information 2340 if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { 2341 #ifndef PRODUCT 2342 if (_cfg->C->trace_opto_output()) { 2343 tty->print("*** bundle start of next BB, node %d, for %d instructions\n", 2344 _next_node->_idx, _bundle_instr_count); 2345 } 2346 #endif 2347 step_and_clear(); 2348 } 2349 2350 // Leave untouched the starting instruction, any Phis, a CreateEx node 2351 // or Top. bb->get_node(_bb_start) is the first schedulable instruction. 2352 _bb_end = bb->number_of_nodes()-1; 2353 for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { 2354 Node *n = bb->get_node(_bb_start); 2355 // Things not matched, like Phinodes and ProjNodes don't get scheduled. 2356 // Also, MachIdealNodes do not get scheduled 2357 if( !n->is_Mach() ) continue; // Skip non-machine nodes 2358 MachNode *mach = n->as_Mach(); 2359 int iop = mach->ideal_Opcode(); 2360 if( iop == Op_CreateEx ) continue; // CreateEx is pinned 2361 if( iop == Op_Con ) continue; // Do not schedule Top 2362 if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes 2363 mach->pipeline() == MachNode::pipeline_class() && 2364 !n->is_SpillCopy() && !n->is_MachMerge() ) // Breakpoints, Prolog, etc 2365 continue; 2366 break; // Funny loop structure to be sure... 2367 } 2368 // Compute last "interesting" instruction in block - last instruction we 2369 // might schedule. _bb_end points just after last schedulable inst. We 2370 // normally schedule conditional branches (despite them being forced last 2371 // in the block), because they have delay slots we can fill. Calls all 2372 // have their delay slots filled in the template expansions, so we don't 2373 // bother scheduling them. 2374 Node *last = bb->get_node(_bb_end); 2375 // Ignore trailing NOPs. 2376 while (_bb_end > 0 && last->is_Mach() && 2377 last->as_Mach()->ideal_Opcode() == Op_Con) { 2378 last = bb->get_node(--_bb_end); 2379 } 2380 assert(!last->is_Mach() || last->as_Mach()->ideal_Opcode() != Op_Con, ""); 2381 if( last->is_Catch() || 2382 // Exclude unreachable path case when Halt node is in a separate block. 2383 (_bb_end > 1 && last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { 2384 // There must be a prior call. Skip it. 2385 while( !bb->get_node(--_bb_end)->is_MachCall() ) { 2386 assert( bb->get_node(_bb_end)->is_MachProj(), "skipping projections after expected call" ); 2387 } 2388 } else if( last->is_MachNullCheck() ) { 2389 // Backup so the last null-checked memory instruction is 2390 // outside the schedulable range. Skip over the nullcheck, 2391 // projection, and the memory nodes. 2392 Node *mem = last->in(1); 2393 do { 2394 _bb_end--; 2395 } while (mem != bb->get_node(_bb_end)); 2396 } else { 2397 // Set _bb_end to point after last schedulable inst. 2398 _bb_end++; 2399 } 2400 2401 assert( _bb_start <= _bb_end, "inverted block ends" ); 2402 2403 // Compute the register antidependencies for the basic block 2404 ComputeRegisterAntidependencies(bb); 2405 if (_cfg->C->failing()) return; // too many D-U pinch points 2406 2407 // Compute intra-bb latencies for the nodes 2408 ComputeLocalLatenciesForward(bb); 2409 2410 // Compute the usage within the block, and set the list of all nodes 2411 // in the block that have no uses within the block. 2412 ComputeUseCount(bb); 2413 2414 // Schedule the remaining instructions in the block 2415 while ( _available.size() > 0 ) { 2416 Node *n = ChooseNodeToBundle(); 2417 guarantee(n != NULL, "no nodes available"); 2418 AddNodeToBundle(n,bb); 2419 } 2420 2421 assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); 2422 #ifdef ASSERT 2423 for( uint l = _bb_start; l < _bb_end; l++ ) { 2424 Node *n = bb->get_node(l); 2425 uint m; 2426 for( m = 0; m < _bb_end-_bb_start; m++ ) 2427 if( _scheduled[m] == n ) 2428 break; 2429 assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); 2430 } 2431 #endif 2432 2433 // Now copy the instructions (in reverse order) back to the block 2434 for ( uint k = _bb_start; k < _bb_end; k++ ) 2435 bb->map_node(_scheduled[_bb_end-k-1], k); 2436 2437 #ifndef PRODUCT 2438 if (_cfg->C->trace_opto_output()) { 2439 tty->print("# Schedule BB#%03d (final)\n", i); 2440 uint current = 0; 2441 for (uint j = 0; j < bb->number_of_nodes(); j++) { 2442 Node *n = bb->get_node(j); 2443 if( valid_bundle_info(n) ) { 2444 Bundle *bundle = node_bundling(n); 2445 if (bundle->instr_count() > 0 || bundle->flags() > 0) { 2446 tty->print("*** Bundle: "); 2447 bundle->dump(); 2448 } 2449 n->dump(); 2450 } 2451 } 2452 } 2453 #endif 2454 #ifdef ASSERT 2455 verify_good_schedule(bb,"after block local scheduling"); 2456 #endif 2457 } 2458 2459 #ifndef PRODUCT 2460 if (_cfg->C->trace_opto_output()) 2461 tty->print("# <- DoScheduling\n"); 2462 #endif 2463 2464 // Record final node-bundling array location 2465 _regalloc->C->set_node_bundling_base(_node_bundling_base); 2466 2467 } // end DoScheduling 2468 2469 // Verify that no live-range used in the block is killed in the block by a 2470 // wrong DEF. This doesn't verify live-ranges that span blocks. 2471 2472 // Check for edge existence. Used to avoid adding redundant precedence edges. 2473 static bool edge_from_to( Node *from, Node *to ) { 2474 for( uint i=0; i<from->len(); i++ ) 2475 if( from->in(i) == to ) 2476 return true; 2477 return false; 2478 } 2479 2480 #ifdef ASSERT 2481 void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { 2482 // Check for bad kills 2483 if( OptoReg::is_valid(def) ) { // Ignore stores & control flow 2484 Node *prior_use = _reg_node[def]; 2485 if( prior_use && !edge_from_to(prior_use,n) ) { 2486 tty->print("%s = ",OptoReg::as_VMReg(def)->name()); 2487 n->dump(); 2488 tty->print_cr("..."); 2489 prior_use->dump(); 2490 assert(edge_from_to(prior_use,n),msg); 2491 } 2492 _reg_node.map(def,NULL); // Kill live USEs 2493 } 2494 } 2495 2496 void Scheduling::verify_good_schedule( Block *b, const char *msg ) { 2497 2498 // Zap to something reasonable for the verify code 2499 _reg_node.clear(); 2500 2501 // Walk over the block backwards. Check to make sure each DEF doesn't 2502 // kill a live value (other than the one it's supposed to). Add each 2503 // USE to the live set. 2504 for( uint i = b->number_of_nodes()-1; i >= _bb_start; i-- ) { 2505 Node *n = b->get_node(i); 2506 int n_op = n->Opcode(); 2507 if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { 2508 // Fat-proj kills a slew of registers 2509 RegMask rm = n->out_RegMask();// Make local copy 2510 while( rm.is_NotEmpty() ) { 2511 OptoReg::Name kill = rm.find_first_elem(); 2512 rm.Remove(kill); 2513 verify_do_def( n, kill, msg ); 2514 } 2515 } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes 2516 // Get DEF'd registers the normal way 2517 verify_do_def( n, _regalloc->get_reg_first(n), msg ); 2518 verify_do_def( n, _regalloc->get_reg_second(n), msg ); 2519 } 2520 2521 // Now make all USEs live 2522 for( uint i=1; i<n->req(); i++ ) { 2523 Node *def = n->in(i); 2524 assert(def != 0, "input edge required"); 2525 OptoReg::Name reg_lo = _regalloc->get_reg_first(def); 2526 OptoReg::Name reg_hi = _regalloc->get_reg_second(def); 2527 if( OptoReg::is_valid(reg_lo) ) { 2528 assert(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg); 2529 _reg_node.map(reg_lo,n); 2530 } 2531 if( OptoReg::is_valid(reg_hi) ) { 2532 assert(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg); 2533 _reg_node.map(reg_hi,n); 2534 } 2535 } 2536 2537 } 2538 2539 // Zap to something reasonable for the Antidependence code 2540 _reg_node.clear(); 2541 } 2542 #endif 2543 2544 // Conditionally add precedence edges. Avoid putting edges on Projs. 2545 static void add_prec_edge_from_to( Node *from, Node *to ) { 2546 if( from->is_Proj() ) { // Put precedence edge on Proj's input 2547 assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); 2548 from = from->in(0); 2549 } 2550 if( from != to && // No cycles (for things like LD L0,[L0+4] ) 2551 !edge_from_to( from, to ) ) // Avoid duplicate edge 2552 from->add_prec(to); 2553 } 2554 2555 void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { 2556 if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow 2557 return; 2558 2559 Node *pinch = _reg_node[def_reg]; // Get pinch point 2560 if ((pinch == NULL) || _cfg->get_block_for_node(pinch) != b || // No pinch-point yet? 2561 is_def ) { // Check for a true def (not a kill) 2562 _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point 2563 return; 2564 } 2565 2566 Node *kill = def; // Rename 'def' to more descriptive 'kill' 2567 debug_only( def = (Node*)0xdeadbeef; ) 2568 2569 // After some number of kills there _may_ be a later def 2570 Node *later_def = NULL; 2571 2572 // Finding a kill requires a real pinch-point. 2573 // Check for not already having a pinch-point. 2574 // Pinch points are Op_Node's. 2575 if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? 2576 later_def = pinch; // Must be def/kill as optimistic pinch-point 2577 if ( _pinch_free_list.size() > 0) { 2578 pinch = _pinch_free_list.pop(); 2579 } else { 2580 pinch = new Node(1); // Pinch point to-be 2581 } 2582 if (pinch->_idx >= _regalloc->node_regs_max_index()) { 2583 _cfg->C->record_method_not_compilable("too many D-U pinch points"); 2584 return; 2585 } 2586 _cfg->map_node_to_block(pinch, b); // Pretend it's valid in this block (lazy init) 2587 _reg_node.map(def_reg,pinch); // Record pinch-point 2588 //_regalloc->set_bad(pinch->_idx); // Already initialized this way. 2589 if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill 2590 pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call 2591 add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch 2592 later_def = NULL; // and no later def 2593 } 2594 pinch->set_req(0,later_def); // Hook later def so we can find it 2595 } else { // Else have valid pinch point 2596 if( pinch->in(0) ) // If there is a later-def 2597 later_def = pinch->in(0); // Get it 2598 } 2599 2600 // Add output-dependence edge from later def to kill 2601 if( later_def ) // If there is some original def 2602 add_prec_edge_from_to(later_def,kill); // Add edge from def to kill 2603 2604 // See if current kill is also a use, and so is forced to be the pinch-point. 2605 if( pinch->Opcode() == Op_Node ) { 2606 Node *uses = kill->is_Proj() ? kill->in(0) : kill; 2607 for( uint i=1; i<uses->req(); i++ ) { 2608 if( _regalloc->get_reg_first(uses->in(i)) == def_reg || 2609 _regalloc->get_reg_second(uses->in(i)) == def_reg ) { 2610 // Yes, found a use/kill pinch-point 2611 pinch->set_req(0,NULL); // 2612 pinch->replace_by(kill); // Move anti-dep edges up 2613 pinch = kill; 2614 _reg_node.map(def_reg,pinch); 2615 return; 2616 } 2617 } 2618 } 2619 2620 // Add edge from kill to pinch-point 2621 add_prec_edge_from_to(kill,pinch); 2622 } 2623 2624 void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { 2625 if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow 2626 return; 2627 Node *pinch = _reg_node[use_reg]; // Get pinch point 2628 // Check for no later def_reg/kill in block 2629 if ((pinch != NULL) && _cfg->get_block_for_node(pinch) == b && 2630 // Use has to be block-local as well 2631 _cfg->get_block_for_node(use) == b) { 2632 if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) 2633 pinch->req() == 1 ) { // pinch not yet in block? 2634 pinch->del_req(0); // yank pointer to later-def, also set flag 2635 // Insert the pinch-point in the block just after the last use 2636 b->insert_node(pinch, b->find_node(use) + 1); 2637 _bb_end++; // Increase size scheduled region in block 2638 } 2639 2640 add_prec_edge_from_to(pinch,use); 2641 } 2642 } 2643 2644 // We insert antidependences between the reads and following write of 2645 // allocated registers to prevent illegal code motion. Hopefully, the 2646 // number of added references should be fairly small, especially as we 2647 // are only adding references within the current basic block. 2648 void Scheduling::ComputeRegisterAntidependencies(Block *b) { 2649 2650 #ifdef ASSERT 2651 verify_good_schedule(b,"before block local scheduling"); 2652 #endif 2653 2654 // A valid schedule, for each register independently, is an endless cycle 2655 // of: a def, then some uses (connected to the def by true dependencies), 2656 // then some kills (defs with no uses), finally the cycle repeats with a new 2657 // def. The uses are allowed to float relative to each other, as are the 2658 // kills. No use is allowed to slide past a kill (or def). This requires 2659 // antidependencies between all uses of a single def and all kills that 2660 // follow, up to the next def. More edges are redundant, because later defs 2661 // & kills are already serialized with true or antidependencies. To keep 2662 // the edge count down, we add a 'pinch point' node if there's more than 2663 // one use or more than one kill/def. 2664 2665 // We add dependencies in one bottom-up pass. 2666 2667 // For each instruction we handle it's DEFs/KILLs, then it's USEs. 2668 2669 // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this 2670 // register. If not, we record the DEF/KILL in _reg_node, the 2671 // register-to-def mapping. If there is a prior DEF/KILL, we insert a 2672 // "pinch point", a new Node that's in the graph but not in the block. 2673 // We put edges from the prior and current DEF/KILLs to the pinch point. 2674 // We put the pinch point in _reg_node. If there's already a pinch point 2675 // we merely add an edge from the current DEF/KILL to the pinch point. 2676 2677 // After doing the DEF/KILLs, we handle USEs. For each used register, we 2678 // put an edge from the pinch point to the USE. 2679 2680 // To be expedient, the _reg_node array is pre-allocated for the whole 2681 // compilation. _reg_node is lazily initialized; it either contains a NULL, 2682 // or a valid def/kill/pinch-point, or a leftover node from some prior 2683 // block. Leftover node from some prior block is treated like a NULL (no 2684 // prior def, so no anti-dependence needed). Valid def is distinguished by 2685 // it being in the current block. 2686 bool fat_proj_seen = false; 2687 uint last_safept = _bb_end-1; 2688 Node* end_node = (_bb_end-1 >= _bb_start) ? b->get_node(last_safept) : NULL; 2689 Node* last_safept_node = end_node; 2690 for( uint i = _bb_end-1; i >= _bb_start; i-- ) { 2691 Node *n = b->get_node(i); 2692 int is_def = n->outcnt(); // def if some uses prior to adding precedence edges 2693 if( n->is_MachProj() && n->ideal_reg() == MachProjNode::fat_proj ) { 2694 // Fat-proj kills a slew of registers 2695 // This can add edges to 'n' and obscure whether or not it was a def, 2696 // hence the is_def flag. 2697 fat_proj_seen = true; 2698 RegMask rm = n->out_RegMask();// Make local copy 2699 while( rm.is_NotEmpty() ) { 2700 OptoReg::Name kill = rm.find_first_elem(); 2701 rm.Remove(kill); 2702 anti_do_def( b, n, kill, is_def ); 2703 } 2704 } else { 2705 // Get DEF'd registers the normal way 2706 anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); 2707 anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); 2708 } 2709 2710 // Kill projections on a branch should appear to occur on the 2711 // branch, not afterwards, so grab the masks from the projections 2712 // and process them. 2713 if (n->is_MachBranch() || n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_Jump) { 2714 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2715 Node* use = n->fast_out(i); 2716 if (use->is_Proj()) { 2717 RegMask rm = use->out_RegMask();// Make local copy 2718 while( rm.is_NotEmpty() ) { 2719 OptoReg::Name kill = rm.find_first_elem(); 2720 rm.Remove(kill); 2721 anti_do_def( b, n, kill, false ); 2722 } 2723 } 2724 } 2725 } 2726 2727 // Check each register used by this instruction for a following DEF/KILL 2728 // that must occur afterward and requires an anti-dependence edge. 2729 for( uint j=0; j<n->req(); j++ ) { 2730 Node *def = n->in(j); 2731 if( def ) { 2732 assert( !def->is_MachProj() || def->ideal_reg() != MachProjNode::fat_proj, "" ); 2733 anti_do_use( b, n, _regalloc->get_reg_first(def) ); 2734 anti_do_use( b, n, _regalloc->get_reg_second(def) ); 2735 } 2736 } 2737 // Do not allow defs of new derived values to float above GC 2738 // points unless the base is definitely available at the GC point. 2739 2740 Node *m = b->get_node(i); 2741 2742 // Add precedence edge from following safepoint to use of derived pointer 2743 if( last_safept_node != end_node && 2744 m != last_safept_node) { 2745 for (uint k = 1; k < m->req(); k++) { 2746 const Type *t = m->in(k)->bottom_type(); 2747 if( t->isa_oop_ptr() && 2748 t->is_ptr()->offset() != 0 ) { 2749 last_safept_node->add_prec( m ); 2750 break; 2751 } 2752 } 2753 } 2754 2755 if( n->jvms() ) { // Precedence edge from derived to safept 2756 // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() 2757 if( b->get_node(last_safept) != last_safept_node ) { 2758 last_safept = b->find_node(last_safept_node); 2759 } 2760 for( uint j=last_safept; j > i; j-- ) { 2761 Node *mach = b->get_node(j); 2762 if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) 2763 mach->add_prec( n ); 2764 } 2765 last_safept = i; 2766 last_safept_node = m; 2767 } 2768 } 2769 2770 if (fat_proj_seen) { 2771 // Garbage collect pinch nodes that were not consumed. 2772 // They are usually created by a fat kill MachProj for a call. 2773 garbage_collect_pinch_nodes(); 2774 } 2775 } 2776 2777 // Garbage collect pinch nodes for reuse by other blocks. 2778 // 2779 // The block scheduler's insertion of anti-dependence 2780 // edges creates many pinch nodes when the block contains 2781 // 2 or more Calls. A pinch node is used to prevent a 2782 // combinatorial explosion of edges. If a set of kills for a 2783 // register is anti-dependent on a set of uses (or defs), rather 2784 // than adding an edge in the graph between each pair of kill 2785 // and use (or def), a pinch is inserted between them: 2786 // 2787 // use1 use2 use3 2788 // \ | / 2789 // \ | / 2790 // pinch 2791 // / | \ 2792 // / | \ 2793 // kill1 kill2 kill3 2794 // 2795 // One pinch node is created per register killed when 2796 // the second call is encountered during a backwards pass 2797 // over the block. Most of these pinch nodes are never 2798 // wired into the graph because the register is never 2799 // used or def'ed in the block. 2800 // 2801 void Scheduling::garbage_collect_pinch_nodes() { 2802 #ifndef PRODUCT 2803 if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); 2804 #endif 2805 int trace_cnt = 0; 2806 for (uint k = 0; k < _reg_node.Size(); k++) { 2807 Node* pinch = _reg_node[k]; 2808 if ((pinch != NULL) && pinch->Opcode() == Op_Node && 2809 // no predecence input edges 2810 (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { 2811 cleanup_pinch(pinch); 2812 _pinch_free_list.push(pinch); 2813 _reg_node.map(k, NULL); 2814 #ifndef PRODUCT 2815 if (_cfg->C->trace_opto_output()) { 2816 trace_cnt++; 2817 if (trace_cnt > 40) { 2818 tty->print("\n"); 2819 trace_cnt = 0; 2820 } 2821 tty->print(" %d", pinch->_idx); 2822 } 2823 #endif 2824 } 2825 } 2826 #ifndef PRODUCT 2827 if (_cfg->C->trace_opto_output()) tty->print("\n"); 2828 #endif 2829 } 2830 2831 // Clean up a pinch node for reuse. 2832 void Scheduling::cleanup_pinch( Node *pinch ) { 2833 assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); 2834 2835 for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { 2836 Node* use = pinch->last_out(i); 2837 uint uses_found = 0; 2838 for (uint j = use->req(); j < use->len(); j++) { 2839 if (use->in(j) == pinch) { 2840 use->rm_prec(j); 2841 uses_found++; 2842 } 2843 } 2844 assert(uses_found > 0, "must be a precedence edge"); 2845 i -= uses_found; // we deleted 1 or more copies of this edge 2846 } 2847 // May have a later_def entry 2848 pinch->set_req(0, NULL); 2849 } 2850 2851 #ifndef PRODUCT 2852 2853 void Scheduling::dump_available() const { 2854 tty->print("#Availist "); 2855 for (uint i = 0; i < _available.size(); i++) 2856 tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); 2857 tty->cr(); 2858 } 2859 2860 // Print Scheduling Statistics 2861 void Scheduling::print_statistics() { 2862 // Print the size added by nops for bundling 2863 tty->print("Nops added %d bytes to total of %d bytes", 2864 _total_nop_size, _total_method_size); 2865 if (_total_method_size > 0) 2866 tty->print(", for %.2f%%", 2867 ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); 2868 tty->print("\n"); 2869 2870 // Print the number of branch shadows filled 2871 if (Pipeline::_branch_has_delay_slot) { 2872 tty->print("Of %d branches, %d had unconditional delay slots filled", 2873 _total_branches, _total_unconditional_delays); 2874 if (_total_branches > 0) 2875 tty->print(", for %.2f%%", 2876 ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); 2877 tty->print("\n"); 2878 } 2879 2880 uint total_instructions = 0, total_bundles = 0; 2881 2882 for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { 2883 uint bundle_count = _total_instructions_per_bundle[i]; 2884 total_instructions += bundle_count * i; 2885 total_bundles += bundle_count; 2886 } 2887 2888 if (total_bundles > 0) 2889 tty->print("Average ILP (excluding nops) is %.2f\n", 2890 ((double)total_instructions) / ((double)total_bundles)); 2891 } 2892 #endif