1 /* 2 * Copyright (c) 2002, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "code/vmreg.inline.hpp" 27 #include "compiler/oopMap.hpp" 28 #include "memory/resourceArea.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/compile.hpp" 32 #include "opto/machnode.hpp" 33 #include "opto/matcher.hpp" 34 #include "opto/output.hpp" 35 #include "opto/phase.hpp" 36 #include "opto/regalloc.hpp" 37 #include "opto/rootnode.hpp" 38 #include "utilities/align.hpp" 39 40 // The functions in this file builds OopMaps after all scheduling is done. 41 // 42 // OopMaps contain a list of all registers and stack-slots containing oops (so 43 // they can be updated by GC). OopMaps also contain a list of derived-pointer 44 // base-pointer pairs. When the base is moved, the derived pointer moves to 45 // follow it. Finally, any registers holding callee-save values are also 46 // recorded. These might contain oops, but only the caller knows. 47 // 48 // BuildOopMaps implements a simple forward reaching-defs solution. At each 49 // GC point we'll have the reaching-def Nodes. If the reaching Nodes are 50 // typed as pointers (no offset), then they are oops. Pointers+offsets are 51 // derived pointers, and bases can be found from them. Finally, we'll also 52 // track reaching callee-save values. Note that a copy of a callee-save value 53 // "kills" it's source, so that only 1 copy of a callee-save value is alive at 54 // a time. 55 // 56 // We run a simple bitvector liveness pass to help trim out dead oops. Due to 57 // irreducible loops, we can have a reaching def of an oop that only reaches 58 // along one path and no way to know if it's valid or not on the other path. 59 // The bitvectors are quite dense and the liveness pass is fast. 60 // 61 // At GC points, we consult this information to build OopMaps. All reaching 62 // defs typed as oops are added to the OopMap. Only 1 instance of a 63 // callee-save register can be recorded. For derived pointers, we'll have to 64 // find and record the register holding the base. 65 // 66 // The reaching def's is a simple 1-pass worklist approach. I tried a clever 67 // breadth-first approach but it was worse (showed O(n^2) in the 68 // pick-next-block code). 69 // 70 // The relevant data is kept in a struct of arrays (it could just as well be 71 // an array of structs, but the struct-of-arrays is generally a little more 72 // efficient). The arrays are indexed by register number (including 73 // stack-slots as registers) and so is bounded by 200 to 300 elements in 74 // practice. One array will map to a reaching def Node (or NULL for 75 // conflict/dead). The other array will map to a callee-saved register or 76 // OptoReg::Bad for not-callee-saved. 77 78 79 // Structure to pass around 80 struct OopFlow : public ResourceObj { 81 short *_callees; // Array mapping register to callee-saved 82 Node **_defs; // array mapping register to reaching def 83 // or NULL if dead/conflict 84 // OopFlow structs, when not being actively modified, describe the _end_ of 85 // this block. 86 Block *_b; // Block for this struct 87 OopFlow *_next; // Next free OopFlow 88 // or NULL if dead/conflict 89 Compile* C; 90 91 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs), 92 _b(NULL), _next(NULL), C(c) { } 93 94 // Given reaching-defs for this block start, compute it for this block end 95 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ); 96 97 // Merge these two OopFlows into the 'this' pointer. 98 void merge( OopFlow *flow, int max_reg ); 99 100 // Copy a 'flow' over an existing flow 101 void clone( OopFlow *flow, int max_size); 102 103 // Make a new OopFlow from scratch 104 static OopFlow *make( Arena *A, int max_size, Compile* C ); 105 106 // Build an oopmap from the current flow info 107 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ); 108 }; 109 110 // Given reaching-defs for this block start, compute it for this block end 111 void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) { 112 113 for( uint i=0; i<_b->number_of_nodes(); i++ ) { 114 Node *n = _b->get_node(i); 115 116 if( n->jvms() ) { // Build an OopMap here? 117 JVMState *jvms = n->jvms(); 118 // no map needed for leaf calls 119 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) { 120 int *live = (int*) (*safehash)[n]; 121 assert( live, "must find live" ); 122 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) ); 123 } 124 } 125 126 // Assign new reaching def's. 127 // Note that I padded the _defs and _callees arrays so it's legal 128 // to index at _defs[OptoReg::Bad]. 129 OptoReg::Name first = regalloc->get_reg_first(n); 130 OptoReg::Name second = regalloc->get_reg_second(n); 131 _defs[first] = n; 132 _defs[second] = n; 133 134 // Pass callee-save info around copies 135 int idx = n->is_Copy(); 136 if( idx ) { // Copies move callee-save info 137 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx)); 138 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx)); 139 int tmp_first = _callees[old_first]; 140 int tmp_second = _callees[old_second]; 141 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location 142 _callees[old_second] = OptoReg::Bad; 143 _callees[first] = tmp_first; 144 _callees[second] = tmp_second; 145 } else if( n->is_Phi() ) { // Phis do not mod callee-saves 146 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" ); 147 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" ); 148 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" ); 149 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" ); 150 } else { 151 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value 152 _callees[second] = OptoReg::Bad; 153 154 // Find base case for callee saves 155 if( n->is_Proj() && n->in(0)->is_Start() ) { 156 if( OptoReg::is_reg(first) && 157 regalloc->_matcher.is_save_on_entry(first) ) 158 _callees[first] = first; 159 if( OptoReg::is_reg(second) && 160 regalloc->_matcher.is_save_on_entry(second) ) 161 _callees[second] = second; 162 } 163 } 164 } 165 } 166 167 // Merge the given flow into the 'this' flow 168 void OopFlow::merge( OopFlow *flow, int max_reg ) { 169 assert( _b == NULL, "merging into a happy flow" ); 170 assert( flow->_b, "this flow is still alive" ); 171 assert( flow != this, "no self flow" ); 172 173 // Do the merge. If there are any differences, drop to 'bottom' which 174 // is OptoReg::Bad or NULL depending. 175 for( int i=0; i<max_reg; i++ ) { 176 // Merge the callee-save's 177 if( _callees[i] != flow->_callees[i] ) 178 _callees[i] = OptoReg::Bad; 179 // Merge the reaching defs 180 if( _defs[i] != flow->_defs[i] ) 181 _defs[i] = NULL; 182 } 183 184 } 185 186 void OopFlow::clone( OopFlow *flow, int max_size ) { 187 _b = flow->_b; 188 memcpy( _callees, flow->_callees, sizeof(short)*max_size); 189 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size); 190 } 191 192 OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) { 193 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1); 194 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1); 195 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) ); 196 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C); 197 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" ); 198 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" ); 199 return flow; 200 } 201 202 static int get_live_bit( int *live, int reg ) { 203 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); } 204 static void set_live_bit( int *live, int reg ) { 205 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); } 206 static void clr_live_bit( int *live, int reg ) { 207 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); } 208 209 // Build an oopmap from the current flow info 210 OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) { 211 int framesize = regalloc->_framesize; 212 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP); 213 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0()); 214 memset(dup_check,0,OptoReg::stack0()) ); 215 216 OopMap *omap = new OopMap( framesize, max_inarg_slot ); 217 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL; 218 JVMState* jvms = n->jvms(); 219 220 // For all registers do... 221 for( int reg=0; reg<max_reg; reg++ ) { 222 if( get_live_bit(live,reg) == 0 ) 223 continue; // Ignore if not live 224 225 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit 226 // register in that case we'll get an non-concrete register for the second 227 // half. We only need to tell the map the register once! 228 // 229 // However for the moment we disable this change and leave things as they 230 // were. 231 232 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot); 233 234 if (false && r->is_reg() && !r->is_concrete()) { 235 continue; 236 } 237 238 // See if dead (no reaching def). 239 Node *def = _defs[reg]; // Get reaching def 240 assert( def, "since live better have reaching def" ); 241 242 // Classify the reaching def as oop, derived, callee-save, dead, or other 243 const Type *t = def->bottom_type(); 244 if( t->isa_oop_ptr() || // Oop or derived? 245 (C->do_stack_allocation() && t->isa_rawptr() && def->is_BoxLock())) { // consider stack oops too 246 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); 247 #ifdef _LP64 248 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers. 249 // Make sure both are record from the same reaching def, but do not 250 // put both into the oopmap. 251 if( (reg&1) == 1 ) { // High half of oop-pair? 252 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" ); 253 continue; // Do not record high parts in oopmap 254 } 255 #endif 256 257 // Check for a legal reg name in the oopMap and bailout if it is not. 258 if (!omap->legal_vm_reg_name(r)) { 259 regalloc->C->record_method_not_compilable("illegal oopMap register name"); 260 continue; 261 } 262 if( t->is_ptr()->_offset == 0 ) { // Not derived? 263 if( mcall ) { 264 // Outgoing argument GC mask responsibility belongs to the callee, 265 // not the caller. Inspect the inputs to the call, to see if 266 // this live-range is one of them. 267 uint cnt = mcall->tf()->domain()->cnt(); 268 uint j; 269 for( j = TypeFunc::Parms; j < cnt; j++) 270 if( mcall->in(j) == def ) 271 break; // reaching def is an argument oop 272 if( j < cnt ) // arg oops dont go in GC map 273 continue; // Continue on to the next register 274 } 275 omap->set_oop(r); 276 } else { // Else it's derived. 277 // Find the base of the derived value. 278 uint i; 279 // Fast, common case, scan 280 for( i = jvms->oopoff(); i < n->req(); i+=2 ) 281 if( n->in(i) == def ) break; // Common case 282 if( i == n->req() ) { // Missed, try a more generous scan 283 // Scan again, but this time peek through copies 284 for( i = jvms->oopoff(); i < n->req(); i+=2 ) { 285 Node *m = n->in(i); // Get initial derived value 286 while( 1 ) { 287 Node *d = def; // Get initial reaching def 288 while( 1 ) { // Follow copies of reaching def to end 289 if( m == d ) goto found; // breaks 3 loops 290 int idx = d->is_Copy(); 291 if( !idx ) break; 292 d = d->in(idx); // Link through copy 293 } 294 int idx = m->is_Copy(); 295 if( !idx ) break; 296 m = m->in(idx); 297 } 298 } 299 guarantee( 0, "must find derived/base pair" ); 300 } 301 found: ; 302 Node *base = n->in(i+1); // Base is other half of pair 303 int breg = regalloc->get_reg_first(base); 304 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot); 305 306 // I record liveness at safepoints BEFORE I make the inputs 307 // live. This is because argument oops are NOT live at a 308 // safepoint (or at least they cannot appear in the oopmap). 309 // Thus bases of base/derived pairs might not be in the 310 // liveness data but they need to appear in the oopmap. 311 if( get_live_bit(live,breg) == 0 ) {// Not live? 312 // Flag it, so next derived pointer won't re-insert into oopmap 313 set_live_bit(live,breg); 314 // Already missed our turn? 315 if( breg < reg ) { 316 if (b->is_stack() || b->is_concrete() || true ) { 317 omap->set_oop( b); 318 } 319 } 320 } 321 if (b->is_stack() || b->is_concrete() || true ) { 322 omap->set_derived_oop( r, b); 323 } 324 } 325 326 } else if( t->isa_narrowoop() ) { 327 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); 328 // Check for a legal reg name in the oopMap and bailout if it is not. 329 if (!omap->legal_vm_reg_name(r)) { 330 regalloc->C->record_method_not_compilable("illegal oopMap register name"); 331 continue; 332 } 333 if( mcall ) { 334 // Outgoing argument GC mask responsibility belongs to the callee, 335 // not the caller. Inspect the inputs to the call, to see if 336 // this live-range is one of them. 337 uint cnt = mcall->tf()->domain()->cnt(); 338 uint j; 339 for( j = TypeFunc::Parms; j < cnt; j++) 340 if( mcall->in(j) == def ) 341 break; // reaching def is an argument oop 342 if( j < cnt ) // arg oops dont go in GC map 343 continue; // Continue on to the next register 344 } 345 omap->set_narrowoop(r); 346 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save? 347 // It's a callee-save value 348 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" ); 349 debug_only( dup_check[_callees[reg]]=1; ) 350 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg])); 351 if ( callee->is_concrete() || true ) { 352 omap->set_callee_saved( r, callee); 353 } 354 355 } else { 356 // Other - some reaching non-oop value 357 #ifdef ASSERT 358 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) { 359 def->dump(); 360 n->dump(); 361 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint"); 362 } 363 #endif 364 } 365 366 } 367 368 #ifdef ASSERT 369 /* Nice, Intel-only assert 370 int cnt_callee_saves=0; 371 int reg2 = 0; 372 while (OptoReg::is_reg(reg2)) { 373 if( dup_check[reg2] != 0) cnt_callee_saves++; 374 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" ); 375 reg2++; 376 } 377 */ 378 #endif 379 380 #ifdef ASSERT 381 for( OopMapStream oms1(omap); !oms1.is_done(); oms1.next()) { 382 OopMapValue omv1 = oms1.current(); 383 if (omv1.type() != OopMapValue::derived_oop_value) { 384 continue; 385 } 386 bool found = false; 387 for( OopMapStream oms2(omap); !oms2.is_done(); oms2.next()) { 388 OopMapValue omv2 = oms2.current(); 389 if (omv2.type() != OopMapValue::oop_value) { 390 continue; 391 } 392 if( omv1.content_reg() == omv2.reg() ) { 393 found = true; 394 break; 395 } 396 } 397 assert( found, "derived with no base in oopmap" ); 398 } 399 #endif 400 401 return omap; 402 } 403 404 // Compute backwards liveness on registers 405 static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) { 406 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints); 407 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints]; 408 Node* root = cfg->get_root_node(); 409 // On CISC platforms, get the node representing the stack pointer that regalloc 410 // used for spills 411 Node *fp = NodeSentinel; 412 if (UseCISCSpill && root->req() > 1) { 413 fp = root->in(1)->in(TypeFunc::FramePtr); 414 } 415 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt)); 416 // Push preds onto worklist 417 for (uint i = 1; i < root->req(); i++) { 418 Block* block = cfg->get_block_for_node(root->in(i)); 419 worklist->push(block); 420 } 421 422 // ZKM.jar includes tiny infinite loops which are unreached from below. 423 // If we missed any blocks, we'll retry here after pushing all missed 424 // blocks on the worklist. Normally this outer loop never trips more 425 // than once. 426 while (1) { 427 428 while( worklist->size() ) { // Standard worklist algorithm 429 Block *b = worklist->rpop(); 430 431 // Copy first successor into my tmp_live space 432 int s0num = b->_succs[0]->_pre_order; 433 int *t = &live[s0num*max_reg_ints]; 434 for( int i=0; i<max_reg_ints; i++ ) 435 tmp_live[i] = t[i]; 436 437 // OR in the remaining live registers 438 for( uint j=1; j<b->_num_succs; j++ ) { 439 uint sjnum = b->_succs[j]->_pre_order; 440 int *t = &live[sjnum*max_reg_ints]; 441 for( int i=0; i<max_reg_ints; i++ ) 442 tmp_live[i] |= t[i]; 443 } 444 445 // Now walk tmp_live up the block backwards, computing live 446 for( int k=b->number_of_nodes()-1; k>=0; k-- ) { 447 Node *n = b->get_node(k); 448 // KILL def'd bits 449 int first = regalloc->get_reg_first(n); 450 int second = regalloc->get_reg_second(n); 451 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first); 452 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second); 453 454 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL; 455 456 // Check if m is potentially a CISC alternate instruction (i.e, possibly 457 // synthesized by RegAlloc from a conventional instruction and a 458 // spilled input) 459 bool is_cisc_alternate = false; 460 if (UseCISCSpill && m) { 461 is_cisc_alternate = m->is_cisc_alternate(); 462 } 463 464 // GEN use'd bits 465 for( uint l=1; l<n->req(); l++ ) { 466 Node *def = n->in(l); 467 assert(def != 0, "input edge required"); 468 int first = regalloc->get_reg_first(def); 469 int second = regalloc->get_reg_second(def); 470 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first); 471 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second); 472 // If we use the stack pointer in a cisc-alternative instruction, 473 // check for use as a memory operand. Then reconstruct the RegName 474 // for this stack location, and set the appropriate bit in the 475 // live vector 4987749. 476 if (is_cisc_alternate && def == fp) { 477 const TypePtr *adr_type = NULL; 478 intptr_t offset; 479 const Node* base = m->get_base_and_disp(offset, adr_type); 480 if (base == NodeSentinel) { 481 // Machnode has multiple memory inputs. We are unable to reason 482 // with these, but are presuming (with trepidation) that not any of 483 // them are oops. This can be fixed by making get_base_and_disp() 484 // look at a specific input instead of all inputs. 485 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input"); 486 } else if (base != fp || offset == Type::OffsetBot) { 487 // Do nothing: the fp operand is either not from a memory use 488 // (base == NULL) OR the fp is used in a non-memory context 489 // (base is some other register) OR the offset is not constant, 490 // so it is not a stack slot. 491 } else { 492 assert(offset >= 0, "unexpected negative offset"); 493 offset -= (offset % jintSize); // count the whole word 494 int stack_reg = regalloc->offset2reg(offset); 495 if (OptoReg::is_stack(stack_reg)) { 496 set_live_bit(tmp_live, stack_reg); 497 } else { 498 assert(false, "stack_reg not on stack?"); 499 } 500 } 501 } 502 } 503 504 if( n->jvms() ) { // Record liveness at safepoint 505 506 // This placement of this stanza means inputs to calls are 507 // considered live at the callsite's OopMap. Argument oops are 508 // hence live, but NOT included in the oopmap. See cutout in 509 // build_oop_map. Debug oops are live (and in OopMap). 510 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints); 511 for( int l=0; l<max_reg_ints; l++ ) 512 n_live[l] = tmp_live[l]; 513 safehash->Insert(n,n_live); 514 } 515 516 } 517 518 // Now at block top, see if we have any changes. If so, propagate 519 // to prior blocks. 520 int *old_live = &live[b->_pre_order*max_reg_ints]; 521 int l; 522 for( l=0; l<max_reg_ints; l++ ) 523 if( tmp_live[l] != old_live[l] ) 524 break; 525 if( l<max_reg_ints ) { // Change! 526 // Copy in new value 527 for( l=0; l<max_reg_ints; l++ ) 528 old_live[l] = tmp_live[l]; 529 // Push preds onto worklist 530 for (l = 1; l < (int)b->num_preds(); l++) { 531 Block* block = cfg->get_block_for_node(b->pred(l)); 532 worklist->push(block); 533 } 534 } 535 } 536 537 // Scan for any missing safepoints. Happens to infinite loops 538 // ala ZKM.jar 539 uint i; 540 for (i = 1; i < cfg->number_of_blocks(); i++) { 541 Block* block = cfg->get_block(i); 542 uint j; 543 for (j = 1; j < block->number_of_nodes(); j++) { 544 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == NULL) { 545 break; 546 } 547 } 548 if (j < block->number_of_nodes()) { 549 break; 550 } 551 } 552 if (i == cfg->number_of_blocks()) { 553 break; // Got 'em all 554 } 555 556 if (PrintOpto && Verbose) { 557 tty->print_cr("retripping live calc"); 558 } 559 560 // Force the issue (expensively): recheck everybody 561 for (i = 1; i < cfg->number_of_blocks(); i++) { 562 worklist->push(cfg->get_block(i)); 563 } 564 } 565 } 566 567 // Collect GC mask info - where are all the OOPs? 568 void PhaseOutput::BuildOopMaps() { 569 Compile::TracePhase tp("bldOopMaps", &timers[_t_buildOopMaps]); 570 // Can't resource-mark because I need to leave all those OopMaps around, 571 // or else I need to resource-mark some arena other than the default. 572 // ResourceMark rm; // Reclaim all OopFlows when done 573 int max_reg = C->regalloc()->_max_reg; // Current array extent 574 575 Arena *A = Thread::current()->resource_area(); 576 Block_List worklist; // Worklist of pending blocks 577 578 int max_reg_ints = align_up(max_reg, BitsPerInt)>>LogBitsPerInt; 579 Dict *safehash = NULL; // Used for assert only 580 // Compute a backwards liveness per register. Needs a bitarray of 581 // #blocks x (#registers, rounded up to ints) 582 safehash = new Dict(cmpkey,hashkey,A); 583 do_liveness( C->regalloc(), C->cfg(), &worklist, max_reg_ints, A, safehash ); 584 OopFlow *free_list = NULL; // Free, unused 585 586 // Array mapping blocks to completed oopflows 587 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, C->cfg()->number_of_blocks()); 588 memset( flows, 0, C->cfg()->number_of_blocks() * sizeof(OopFlow*) ); 589 590 591 // Do the first block 'by hand' to prime the worklist 592 Block *entry = C->cfg()->get_block(1); 593 OopFlow *rootflow = OopFlow::make(A,max_reg,C); 594 // Initialize to 'bottom' (not 'top') 595 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) ); 596 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) ); 597 flows[entry->_pre_order] = rootflow; 598 599 // Do the first block 'by hand' to prime the worklist 600 rootflow->_b = entry; 601 rootflow->compute_reach( C->regalloc(), max_reg, safehash ); 602 for( uint i=0; i<entry->_num_succs; i++ ) 603 worklist.push(entry->_succs[i]); 604 605 // Now worklist contains blocks which have some, but perhaps not all, 606 // predecessors visited. 607 while( worklist.size() ) { 608 // Scan for a block with all predecessors visited, or any randoms slob 609 // otherwise. All-preds-visited order allows me to recycle OopFlow 610 // structures rapidly and cut down on the memory footprint. 611 // Note: not all predecessors might be visited yet (must happen for 612 // irreducible loops). This is OK, since every live value must have the 613 // SAME reaching def for the block, so any reaching def is OK. 614 uint i; 615 616 Block *b = worklist.pop(); 617 // Ignore root block 618 if (b == C->cfg()->get_root_block()) { 619 continue; 620 } 621 // Block is already done? Happens if block has several predecessors, 622 // he can get on the worklist more than once. 623 if( flows[b->_pre_order] ) continue; 624 625 // If this block has a visited predecessor AND that predecessor has this 626 // last block as his only undone child, we can move the OopFlow from the 627 // pred to this block. Otherwise we have to grab a new OopFlow. 628 OopFlow *flow = NULL; // Flag for finding optimized flow 629 Block *pred = (Block*)((intptr_t)0xdeadbeef); 630 // Scan this block's preds to find a done predecessor 631 for (uint j = 1; j < b->num_preds(); j++) { 632 Block* p = C->cfg()->get_block_for_node(b->pred(j)); 633 OopFlow *p_flow = flows[p->_pre_order]; 634 if( p_flow ) { // Predecessor is done 635 assert( p_flow->_b == p, "cross check" ); 636 pred = p; // Record some predecessor 637 // If all successors of p are done except for 'b', then we can carry 638 // p_flow forward to 'b' without copying, otherwise we have to draw 639 // from the free_list and clone data. 640 uint k; 641 for( k=0; k<p->_num_succs; k++ ) 642 if( !flows[p->_succs[k]->_pre_order] && 643 p->_succs[k] != b ) 644 break; 645 646 // Either carry-forward the now-unused OopFlow for b's use 647 // or draw a new one from the free list 648 if( k==p->_num_succs ) { 649 flow = p_flow; 650 break; // Found an ideal pred, use him 651 } 652 } 653 } 654 655 if( flow ) { 656 // We have an OopFlow that's the last-use of a predecessor. 657 // Carry it forward. 658 } else { // Draw a new OopFlow from the freelist 659 if( !free_list ) 660 free_list = OopFlow::make(A,max_reg,C); 661 flow = free_list; 662 assert( flow->_b == NULL, "oopFlow is not free" ); 663 free_list = flow->_next; 664 flow->_next = NULL; 665 666 // Copy/clone over the data 667 flow->clone(flows[pred->_pre_order], max_reg); 668 } 669 670 // Mark flow for block. Blocks can only be flowed over once, 671 // because after the first time they are guarded from entering 672 // this code again. 673 assert( flow->_b == pred, "have some prior flow" ); 674 flow->_b = NULL; 675 676 // Now push flow forward 677 flows[b->_pre_order] = flow;// Mark flow for this block 678 flow->_b = b; 679 flow->compute_reach( C->regalloc(), max_reg, safehash ); 680 681 // Now push children onto worklist 682 for( i=0; i<b->_num_succs; i++ ) 683 worklist.push(b->_succs[i]); 684 685 } 686 }