1/* 2 * Copyright (c) 2002, 2016, 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/phase.hpp" 35#include "opto/regalloc.hpp" 36#include "opto/rootnode.hpp" 37 38// The functions in this file builds OopMaps after all scheduling is done. 39// 40// OopMaps contain a list of all registers and stack-slots containing oops (so 41// they can be updated by GC). OopMaps also contain a list of derived-pointer 42// base-pointer pairs. When the base is moved, the derived pointer moves to 43// follow it. Finally, any registers holding callee-save values are also 44// recorded. These might contain oops, but only the caller knows. 45// 46// BuildOopMaps implements a simple forward reaching-defs solution. At each 47// GC point we'll have the reaching-def Nodes. If the reaching Nodes are 48// typed as pointers (no offset), then they are oops. Pointers+offsets are 49// derived pointers, and bases can be found from them. Finally, we'll also 50// track reaching callee-save values. Note that a copy of a callee-save value 51// "kills" it's source, so that only 1 copy of a callee-save value is alive at 52// a time. 53// 54// We run a simple bitvector liveness pass to help trim out dead oops. Due to 55// irreducible loops, we can have a reaching def of an oop that only reaches 56// along one path and no way to know if it's valid or not on the other path. 57// The bitvectors are quite dense and the liveness pass is fast. 58// 59// At GC points, we consult this information to build OopMaps. All reaching 60// defs typed as oops are added to the OopMap. Only 1 instance of a 61// callee-save register can be recorded. For derived pointers, we'll have to 62// find and record the register holding the base. 63// 64// The reaching def's is a simple 1-pass worklist approach. I tried a clever 65// breadth-first approach but it was worse (showed O(n^2) in the 66// pick-next-block code). 67// 68// The relevant data is kept in a struct of arrays (it could just as well be 69// an array of structs, but the struct-of-arrays is generally a little more 70// efficient). The arrays are indexed by register number (including 71// stack-slots as registers) and so is bounded by 200 to 300 elements in 72// practice. One array will map to a reaching def Node (or NULL for 73// conflict/dead). The other array will map to a callee-saved register or 74// OptoReg::Bad for not-callee-saved. 75 76 77// Structure to pass around 78struct OopFlow : public ResourceObj { 79 short *_callees; // Array mapping register to callee-saved 80 Node **_defs; // array mapping register to reaching def 81 // or NULL if dead/conflict 82 // OopFlow structs, when not being actively modified, describe the _end_ of 83 // this block. 84 Block *_b; // Block for this struct 85 OopFlow *_next; // Next free OopFlow 86 // or NULL if dead/conflict 87 Compile* C; 88 89 OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs), 90 _b(NULL), _next(NULL), C(c) { } 91 92 // Given reaching-defs for this block start, compute it for this block end 93 void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ); 94 95 // Merge these two OopFlows into the 'this' pointer. 96 void merge( OopFlow *flow, int max_reg ); 97 98 // Copy a 'flow' over an existing flow 99 void clone( OopFlow *flow, int max_size); 100 101 // Make a new OopFlow from scratch 102 static OopFlow *make( Arena *A, int max_size, Compile* C ); 103 104 // Build an oopmap from the current flow info 105 OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ); 106}; 107 108// Given reaching-defs for this block start, compute it for this block end 109void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) { 110 111 for( uint i=0; i<_b->number_of_nodes(); i++ ) { 112 Node *n = _b->get_node(i); 113 114 if( n->jvms() ) { // Build an OopMap here? 115 JVMState *jvms = n->jvms(); 116 // no map needed for leaf calls 117 if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) { 118 int *live = (int*) (*safehash)[n]; 119 assert( live, "must find live" ); 120 n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) ); 121 } 122 } 123 124 // Assign new reaching def's. 125 // Note that I padded the _defs and _callees arrays so it's legal 126 // to index at _defs[OptoReg::Bad]. 127 OptoReg::Name first = regalloc->get_reg_first(n); 128 OptoReg::Name second = regalloc->get_reg_second(n); 129 _defs[first] = n; 130 _defs[second] = n; 131 132 // Pass callee-save info around copies 133 int idx = n->is_Copy(); 134 if( idx ) { // Copies move callee-save info 135 OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx)); 136 OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx)); 137 int tmp_first = _callees[old_first]; 138 int tmp_second = _callees[old_second]; 139 _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location 140 _callees[old_second] = OptoReg::Bad; 141 _callees[first] = tmp_first; 142 _callees[second] = tmp_second; 143 } else if( n->is_Phi() ) { // Phis do not mod callee-saves 144 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" ); 145 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" ); 146 assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" ); 147 assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" ); 148 } else { 149 _callees[first] = OptoReg::Bad; // No longer holding a callee-save value 150 _callees[second] = OptoReg::Bad; 151 152 // Find base case for callee saves 153 if( n->is_Proj() && n->in(0)->is_Start() ) { 154 if( OptoReg::is_reg(first) && 155 regalloc->_matcher.is_save_on_entry(first) ) 156 _callees[first] = first; 157 if( OptoReg::is_reg(second) && 158 regalloc->_matcher.is_save_on_entry(second) ) 159 _callees[second] = second; 160 } 161 } 162 } 163} 164 165// Merge the given flow into the 'this' flow 166void OopFlow::merge( OopFlow *flow, int max_reg ) { 167 assert( _b == NULL, "merging into a happy flow" ); 168 assert( flow->_b, "this flow is still alive" ); 169 assert( flow != this, "no self flow" ); 170 171 // Do the merge. If there are any differences, drop to 'bottom' which 172 // is OptoReg::Bad or NULL depending. 173 for( int i=0; i<max_reg; i++ ) { 174 // Merge the callee-save's 175 if( _callees[i] != flow->_callees[i] ) 176 _callees[i] = OptoReg::Bad; 177 // Merge the reaching defs 178 if( _defs[i] != flow->_defs[i] ) 179 _defs[i] = NULL; 180 } 181 182} 183 184void OopFlow::clone( OopFlow *flow, int max_size ) { 185 _b = flow->_b; 186 memcpy( _callees, flow->_callees, sizeof(short)*max_size); 187 memcpy( _defs , flow->_defs , sizeof(Node*)*max_size); 188} 189 190OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) { 191 short *callees = NEW_ARENA_ARRAY(A,short,max_size+1); 192 Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1); 193 debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) ); 194 OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C); 195 assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" ); 196 assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" ); 197 return flow; 198} 199 200static int get_live_bit( int *live, int reg ) { 201 return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); } 202static void set_live_bit( int *live, int reg ) { 203 live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); } 204static void clr_live_bit( int *live, int reg ) { 205 live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); } 206 207// Build an oopmap from the current flow info 208OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) { 209 int framesize = regalloc->_framesize; 210 int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP); 211 debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0()); 212 memset(dup_check,0,OptoReg::stack0()) ); 213 214 OopMap *omap = new OopMap( framesize, max_inarg_slot ); 215 MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL; 216 JVMState* jvms = n->jvms(); 217 218 // For all registers do... 219 for( int reg=0; reg<max_reg; reg++ ) { 220 if( get_live_bit(live,reg) == 0 ) 221 continue; // Ignore if not live 222 223 // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit 224 // register in that case we'll get an non-concrete register for the second 225 // half. We only need to tell the map the register once! 226 // 227 // However for the moment we disable this change and leave things as they 228 // were. 229 230 VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot); 231 232 if (false && r->is_reg() && !r->is_concrete()) { 233 continue; 234 } 235 236 // See if dead (no reaching def). 237 Node *def = _defs[reg]; // Get reaching def 238 assert( def, "since live better have reaching def" ); 239 240 // Classify the reaching def as oop, derived, callee-save, dead, or other 241 const Type *t = def->bottom_type(); 242 if( t->isa_oop_ptr() ) { // Oop or derived? 243 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); 244#ifdef _LP64 245 // 64-bit pointers record oop-ishness on 2 aligned adjacent registers. 246 // Make sure both are record from the same reaching def, but do not 247 // put both into the oopmap. 248 if( (reg&1) == 1 ) { // High half of oop-pair? 249 assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" ); 250 continue; // Do not record high parts in oopmap 251 } 252#endif 253 254 // Check for a legal reg name in the oopMap and bailout if it is not. 255 if (!omap->legal_vm_reg_name(r)) { 256 regalloc->C->record_method_not_compilable("illegal oopMap register name"); 257 continue; 258 } 259 if( t->is_ptr()->_offset == 0 ) { // Not derived? 260 if( mcall ) { 261 // Outgoing argument GC mask responsibility belongs to the callee, 262 // not the caller. Inspect the inputs to the call, to see if 263 // this live-range is one of them. 264 uint cnt = mcall->tf()->domain()->cnt(); 265 uint j; 266 for( j = TypeFunc::Parms; j < cnt; j++) 267 if( mcall->in(j) == def ) 268 break; // reaching def is an argument oop 269 if( j < cnt ) // arg oops dont go in GC map 270 continue; // Continue on to the next register 271 } 272 omap->set_oop(r); 273 } else { // Else it's derived. 274 // Find the base of the derived value. 275 uint i; 276 // Fast, common case, scan 277 for( i = jvms->oopoff(); i < n->req(); i+=2 ) 278 if( n->in(i) == def ) break; // Common case 279 if( i == n->req() ) { // Missed, try a more generous scan 280 // Scan again, but this time peek through copies 281 for( i = jvms->oopoff(); i < n->req(); i+=2 ) { 282 Node *m = n->in(i); // Get initial derived value 283 while( 1 ) { 284 Node *d = def; // Get initial reaching def 285 while( 1 ) { // Follow copies of reaching def to end 286 if( m == d ) goto found; // breaks 3 loops 287 int idx = d->is_Copy(); 288 if( !idx ) break; 289 d = d->in(idx); // Link through copy 290 } 291 int idx = m->is_Copy(); 292 if( !idx ) break; 293 m = m->in(idx); 294 } 295 } 296 guarantee( 0, "must find derived/base pair" ); 297 } 298 found: ; 299 Node *base = n->in(i+1); // Base is other half of pair 300 int breg = regalloc->get_reg_first(base); 301 VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot); 302 303 // I record liveness at safepoints BEFORE I make the inputs 304 // live. This is because argument oops are NOT live at a 305 // safepoint (or at least they cannot appear in the oopmap). 306 // Thus bases of base/derived pairs might not be in the 307 // liveness data but they need to appear in the oopmap. 308 if( get_live_bit(live,breg) == 0 ) {// Not live? 309 // Flag it, so next derived pointer won't re-insert into oopmap 310 set_live_bit(live,breg); 311 // Already missed our turn? 312 if( breg < reg ) { 313 if (b->is_stack() || b->is_concrete() || true ) { 314 omap->set_oop( b); 315 } 316 } 317 } 318 if (b->is_stack() || b->is_concrete() || true ) { 319 omap->set_derived_oop( r, b); 320 } 321 } 322 323 } else if( t->isa_narrowoop() ) { 324 assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); 325 // Check for a legal reg name in the oopMap and bailout if it is not. 326 if (!omap->legal_vm_reg_name(r)) { 327 regalloc->C->record_method_not_compilable("illegal oopMap register name"); 328 continue; 329 } 330 if( mcall ) { 331 // Outgoing argument GC mask responsibility belongs to the callee, 332 // not the caller. Inspect the inputs to the call, to see if 333 // this live-range is one of them. 334 uint cnt = mcall->tf()->domain()->cnt(); 335 uint j; 336 for( j = TypeFunc::Parms; j < cnt; j++) 337 if( mcall->in(j) == def ) 338 break; // reaching def is an argument oop 339 if( j < cnt ) // arg oops dont go in GC map 340 continue; // Continue on to the next register 341 } 342 omap->set_narrowoop(r); 343 } else if( OptoReg::is_valid(_callees[reg])) { // callee-save? 344 // It's a callee-save value 345 assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" ); 346 debug_only( dup_check[_callees[reg]]=1; ) 347 VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg])); 348 if ( callee->is_concrete() || true ) { 349 omap->set_callee_saved( r, callee); 350 } 351 352 } else { 353 // Other - some reaching non-oop value 354 omap->set_value( r); 355#ifdef ASSERT 356 if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) { 357 def->dump(); 358 n->dump(); 359 assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint"); 360 } 361#endif 362 } 363 364 } 365 366#ifdef ASSERT 367 /* Nice, Intel-only assert 368 int cnt_callee_saves=0; 369 int reg2 = 0; 370 while (OptoReg::is_reg(reg2)) { 371 if( dup_check[reg2] != 0) cnt_callee_saves++; 372 assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" ); 373 reg2++; 374 } 375 */ 376#endif 377 378#ifdef ASSERT 379 for( OopMapStream oms1(omap, OopMapValue::derived_oop_value); !oms1.is_done(); oms1.next()) { 380 OopMapValue omv1 = oms1.current(); 381 bool found = false; 382 for( OopMapStream oms2(omap,OopMapValue::oop_value); !oms2.is_done(); oms2.next()) { 383 if( omv1.content_reg() == oms2.current().reg() ) { 384 found = true; 385 break; 386 } 387 } 388 assert( found, "derived with no base in oopmap" ); 389 } 390#endif 391 392 return omap; 393} 394 395// Compute backwards liveness on registers 396static void do_liveness(PhaseRegAlloc* regalloc, PhaseCFG* cfg, Block_List* worklist, int max_reg_ints, Arena* A, Dict* safehash) { 397 int* live = NEW_ARENA_ARRAY(A, int, (cfg->number_of_blocks() + 1) * max_reg_ints); 398 int* tmp_live = &live[cfg->number_of_blocks() * max_reg_ints]; 399 Node* root = cfg->get_root_node(); 400 // On CISC platforms, get the node representing the stack pointer that regalloc 401 // used for spills 402 Node *fp = NodeSentinel; 403 if (UseCISCSpill && root->req() > 1) { 404 fp = root->in(1)->in(TypeFunc::FramePtr); 405 } 406 memset(live, 0, cfg->number_of_blocks() * (max_reg_ints << LogBytesPerInt)); 407 // Push preds onto worklist 408 for (uint i = 1; i < root->req(); i++) { 409 Block* block = cfg->get_block_for_node(root->in(i)); 410 worklist->push(block); 411 } 412 413 // ZKM.jar includes tiny infinite loops which are unreached from below. 414 // If we missed any blocks, we'll retry here after pushing all missed 415 // blocks on the worklist. Normally this outer loop never trips more 416 // than once. 417 while (1) { 418 419 while( worklist->size() ) { // Standard worklist algorithm 420 Block *b = worklist->rpop(); 421 422 // Copy first successor into my tmp_live space 423 int s0num = b->_succs[0]->_pre_order; 424 int *t = &live[s0num*max_reg_ints]; 425 for( int i=0; i<max_reg_ints; i++ ) 426 tmp_live[i] = t[i]; 427 428 // OR in the remaining live registers 429 for( uint j=1; j<b->_num_succs; j++ ) { 430 uint sjnum = b->_succs[j]->_pre_order; 431 int *t = &live[sjnum*max_reg_ints]; 432 for( int i=0; i<max_reg_ints; i++ ) 433 tmp_live[i] |= t[i]; 434 } 435 436 // Now walk tmp_live up the block backwards, computing live 437 for( int k=b->number_of_nodes()-1; k>=0; k-- ) { 438 Node *n = b->get_node(k); 439 // KILL def'd bits 440 int first = regalloc->get_reg_first(n); 441 int second = regalloc->get_reg_second(n); 442 if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first); 443 if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second); 444 445 MachNode *m = n->is_Mach() ? n->as_Mach() : NULL; 446 447 // Check if m is potentially a CISC alternate instruction (i.e, possibly 448 // synthesized by RegAlloc from a conventional instruction and a 449 // spilled input) 450 bool is_cisc_alternate = false; 451 if (UseCISCSpill && m) { 452 is_cisc_alternate = m->is_cisc_alternate(); 453 } 454 455 // GEN use'd bits 456 for( uint l=1; l<n->req(); l++ ) { 457 Node *def = n->in(l); 458 assert(def != 0, "input edge required"); 459 int first = regalloc->get_reg_first(def); 460 int second = regalloc->get_reg_second(def); 461 if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first); 462 if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second); 463 // If we use the stack pointer in a cisc-alternative instruction, 464 // check for use as a memory operand. Then reconstruct the RegName 465 // for this stack location, and set the appropriate bit in the 466 // live vector 4987749. 467 if (is_cisc_alternate && def == fp) { 468 const TypePtr *adr_type = NULL; 469 intptr_t offset; 470 const Node* base = m->get_base_and_disp(offset, adr_type); 471 if (base == NodeSentinel) { 472 // Machnode has multiple memory inputs. We are unable to reason 473 // with these, but are presuming (with trepidation) that not any of 474 // them are oops. This can be fixed by making get_base_and_disp() 475 // look at a specific input instead of all inputs. 476 assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input"); 477 } else if (base != fp || offset == Type::OffsetBot) { 478 // Do nothing: the fp operand is either not from a memory use 479 // (base == NULL) OR the fp is used in a non-memory context 480 // (base is some other register) OR the offset is not constant, 481 // so it is not a stack slot. 482 } else { 483 assert(offset >= 0, "unexpected negative offset"); 484 offset -= (offset % jintSize); // count the whole word 485 int stack_reg = regalloc->offset2reg(offset); 486 if (OptoReg::is_stack(stack_reg)) { 487 set_live_bit(tmp_live, stack_reg); 488 } else { 489 assert(false, "stack_reg not on stack?"); 490 } 491 } 492 } 493 } 494 495 if( n->jvms() ) { // Record liveness at safepoint 496 497 // This placement of this stanza means inputs to calls are 498 // considered live at the callsite's OopMap. Argument oops are 499 // hence live, but NOT included in the oopmap. See cutout in 500 // build_oop_map. Debug oops are live (and in OopMap). 501 int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints); 502 for( int l=0; l<max_reg_ints; l++ ) 503 n_live[l] = tmp_live[l]; 504 safehash->Insert(n,n_live); 505 } 506 507 } 508 509 // Now at block top, see if we have any changes. If so, propagate 510 // to prior blocks. 511 int *old_live = &live[b->_pre_order*max_reg_ints]; 512 int l; 513 for( l=0; l<max_reg_ints; l++ ) 514 if( tmp_live[l] != old_live[l] ) 515 break; 516 if( l<max_reg_ints ) { // Change! 517 // Copy in new value 518 for( l=0; l<max_reg_ints; l++ ) 519 old_live[l] = tmp_live[l]; 520 // Push preds onto worklist 521 for (l = 1; l < (int)b->num_preds(); l++) { 522 Block* block = cfg->get_block_for_node(b->pred(l)); 523 worklist->push(block); 524 } 525 } 526 } 527 528 // Scan for any missing safepoints. Happens to infinite loops 529 // ala ZKM.jar 530 uint i; 531 for (i = 1; i < cfg->number_of_blocks(); i++) { 532 Block* block = cfg->get_block(i); 533 uint j; 534 for (j = 1; j < block->number_of_nodes(); j++) { 535 if (block->get_node(j)->jvms() && (*safehash)[block->get_node(j)] == NULL) { 536 break; 537 } 538 } 539 if (j < block->number_of_nodes()) { 540 break; 541 } 542 } 543 if (i == cfg->number_of_blocks()) { 544 break; // Got 'em all 545 } 546 547 if (PrintOpto && Verbose) { 548 tty->print_cr("retripping live calc"); 549 } 550 551 // Force the issue (expensively): recheck everybody 552 for (i = 1; i < cfg->number_of_blocks(); i++) { 553 worklist->push(cfg->get_block(i)); 554 } 555 } 556} 557 558// Collect GC mask info - where are all the OOPs? 559void Compile::BuildOopMaps() { 560 TracePhase tp("bldOopMaps", &timers[_t_buildOopMaps]); 561 // Can't resource-mark because I need to leave all those OopMaps around, 562 // or else I need to resource-mark some arena other than the default. 563 // ResourceMark rm; // Reclaim all OopFlows when done 564 int max_reg = _regalloc->_max_reg; // Current array extent 565 566 Arena *A = Thread::current()->resource_area(); 567 Block_List worklist; // Worklist of pending blocks 568 569 int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt; 570 Dict *safehash = NULL; // Used for assert only 571 // Compute a backwards liveness per register. Needs a bitarray of 572 // #blocks x (#registers, rounded up to ints) 573 safehash = new Dict(cmpkey,hashkey,A); 574 do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash ); 575 OopFlow *free_list = NULL; // Free, unused 576 577 // Array mapping blocks to completed oopflows 578 OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->number_of_blocks()); 579 memset( flows, 0, _cfg->number_of_blocks() * sizeof(OopFlow*) ); 580 581 582 // Do the first block 'by hand' to prime the worklist 583 Block *entry = _cfg->get_block(1); 584 OopFlow *rootflow = OopFlow::make(A,max_reg,this); 585 // Initialize to 'bottom' (not 'top') 586 memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) ); 587 memset( rootflow->_defs , 0, max_reg*sizeof(Node*) ); 588 flows[entry->_pre_order] = rootflow; 589 590 // Do the first block 'by hand' to prime the worklist 591 rootflow->_b = entry; 592 rootflow->compute_reach( _regalloc, max_reg, safehash ); 593 for( uint i=0; i<entry->_num_succs; i++ ) 594 worklist.push(entry->_succs[i]); 595 596 // Now worklist contains blocks which have some, but perhaps not all, 597 // predecessors visited. 598 while( worklist.size() ) { 599 // Scan for a block with all predecessors visited, or any randoms slob 600 // otherwise. All-preds-visited order allows me to recycle OopFlow 601 // structures rapidly and cut down on the memory footprint. 602 // Note: not all predecessors might be visited yet (must happen for 603 // irreducible loops). This is OK, since every live value must have the 604 // SAME reaching def for the block, so any reaching def is OK. 605 uint i; 606 607 Block *b = worklist.pop(); 608 // Ignore root block 609 if (b == _cfg->get_root_block()) { 610 continue; 611 } 612 // Block is already done? Happens if block has several predecessors, 613 // he can get on the worklist more than once. 614 if( flows[b->_pre_order] ) continue; 615 616 // If this block has a visited predecessor AND that predecessor has this 617 // last block as his only undone child, we can move the OopFlow from the 618 // pred to this block. Otherwise we have to grab a new OopFlow. 619 OopFlow *flow = NULL; // Flag for finding optimized flow 620 Block *pred = (Block*)0xdeadbeef; 621 // Scan this block's preds to find a done predecessor 622 for (uint j = 1; j < b->num_preds(); j++) { 623 Block* p = _cfg->get_block_for_node(b->pred(j)); 624 OopFlow *p_flow = flows[p->_pre_order]; 625 if( p_flow ) { // Predecessor is done 626 assert( p_flow->_b == p, "cross check" ); 627 pred = p; // Record some predecessor 628 // If all successors of p are done except for 'b', then we can carry 629 // p_flow forward to 'b' without copying, otherwise we have to draw 630 // from the free_list and clone data. 631 uint k; 632 for( k=0; k<p->_num_succs; k++ ) 633 if( !flows[p->_succs[k]->_pre_order] && 634 p->_succs[k] != b ) 635 break; 636 637 // Either carry-forward the now-unused OopFlow for b's use 638 // or draw a new one from the free list 639 if( k==p->_num_succs ) { 640 flow = p_flow; 641 break; // Found an ideal pred, use him 642 } 643 } 644 } 645 646 if( flow ) { 647 // We have an OopFlow that's the last-use of a predecessor. 648 // Carry it forward. 649 } else { // Draw a new OopFlow from the freelist 650 if( !free_list ) 651 free_list = OopFlow::make(A,max_reg,C); 652 flow = free_list; 653 assert( flow->_b == NULL, "oopFlow is not free" ); 654 free_list = flow->_next; 655 flow->_next = NULL; 656 657 // Copy/clone over the data 658 flow->clone(flows[pred->_pre_order], max_reg); 659 } 660 661 // Mark flow for block. Blocks can only be flowed over once, 662 // because after the first time they are guarded from entering 663 // this code again. 664 assert( flow->_b == pred, "have some prior flow" ); 665 flow->_b = NULL; 666 667 // Now push flow forward 668 flows[b->_pre_order] = flow;// Mark flow for this block 669 flow->_b = b; 670 flow->compute_reach( _regalloc, max_reg, safehash ); 671 672 // Now push children onto worklist 673 for( i=0; i<b->_num_succs; i++ ) 674 worklist.push(b->_succs[i]); 675 676 } 677} 678