1 /* 2 * Copyright (c) 1997, 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 "memory/allocation.inline.hpp" 27 #include "memory/resourceArea.hpp" 28 #include "opto/ad.hpp" 29 #include "opto/addnode.hpp" 30 #include "opto/callnode.hpp" 31 #include "opto/idealGraphPrinter.hpp" 32 #include "opto/matcher.hpp" 33 #include "opto/memnode.hpp" 34 #include "opto/movenode.hpp" 35 #include "opto/opcodes.hpp" 36 #include "opto/regmask.hpp" 37 #include "opto/rootnode.hpp" 38 #include "opto/runtime.hpp" 39 #include "opto/type.hpp" 40 #include "opto/vectornode.hpp" 41 #include "runtime/os.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 44 OptoReg::Name OptoReg::c_frame_pointer; 45 46 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf]; 47 RegMask Matcher::mreg2regmask[_last_Mach_Reg]; 48 RegMask Matcher::STACK_ONLY_mask; 49 RegMask Matcher::c_frame_ptr_mask; 50 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE; 51 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE; 52 53 //---------------------------Matcher------------------------------------------- 54 Matcher::Matcher() 55 : PhaseTransform( Phase::Ins_Select ), 56 #ifdef ASSERT 57 _old2new_map(C->comp_arena()), 58 _new2old_map(C->comp_arena()), 59 #endif 60 _shared_nodes(C->comp_arena()), 61 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp), 62 _swallowed(swallowed), 63 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE), 64 _end_inst_chain_rule(_END_INST_CHAIN_RULE), 65 _must_clone(must_clone), 66 _register_save_policy(register_save_policy), 67 _c_reg_save_policy(c_reg_save_policy), 68 _register_save_type(register_save_type), 69 _ruleName(ruleName), 70 _allocation_started(false), 71 _states_arena(Chunk::medium_size), 72 _visited(&_states_arena), 73 _shared(&_states_arena), 74 _dontcare(&_states_arena) { 75 C->set_matcher(this); 76 77 idealreg2spillmask [Op_RegI] = NULL; 78 idealreg2spillmask [Op_RegN] = NULL; 79 idealreg2spillmask [Op_RegL] = NULL; 80 idealreg2spillmask [Op_RegF] = NULL; 81 idealreg2spillmask [Op_RegD] = NULL; 82 idealreg2spillmask [Op_RegP] = NULL; 83 idealreg2spillmask [Op_VecS] = NULL; 84 idealreg2spillmask [Op_VecD] = NULL; 85 idealreg2spillmask [Op_VecX] = NULL; 86 idealreg2spillmask [Op_VecY] = NULL; 87 idealreg2spillmask [Op_VecZ] = NULL; 88 89 idealreg2debugmask [Op_RegI] = NULL; 90 idealreg2debugmask [Op_RegN] = NULL; 91 idealreg2debugmask [Op_RegL] = NULL; 92 idealreg2debugmask [Op_RegF] = NULL; 93 idealreg2debugmask [Op_RegD] = NULL; 94 idealreg2debugmask [Op_RegP] = NULL; 95 idealreg2debugmask [Op_VecS] = NULL; 96 idealreg2debugmask [Op_VecD] = NULL; 97 idealreg2debugmask [Op_VecX] = NULL; 98 idealreg2debugmask [Op_VecY] = NULL; 99 idealreg2debugmask [Op_VecZ] = NULL; 100 101 idealreg2mhdebugmask[Op_RegI] = NULL; 102 idealreg2mhdebugmask[Op_RegN] = NULL; 103 idealreg2mhdebugmask[Op_RegL] = NULL; 104 idealreg2mhdebugmask[Op_RegF] = NULL; 105 idealreg2mhdebugmask[Op_RegD] = NULL; 106 idealreg2mhdebugmask[Op_RegP] = NULL; 107 idealreg2mhdebugmask[Op_VecS] = NULL; 108 idealreg2mhdebugmask[Op_VecD] = NULL; 109 idealreg2mhdebugmask[Op_VecX] = NULL; 110 idealreg2mhdebugmask[Op_VecY] = NULL; 111 idealreg2mhdebugmask[Op_VecZ] = NULL; 112 113 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node 114 } 115 116 //------------------------------warp_incoming_stk_arg------------------------ 117 // This warps a VMReg into an OptoReg::Name 118 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) { 119 OptoReg::Name warped; 120 if( reg->is_stack() ) { // Stack slot argument? 121 warped = OptoReg::add(_old_SP, reg->reg2stack() ); 122 warped = OptoReg::add(warped, C->out_preserve_stack_slots()); 123 if( warped >= _in_arg_limit ) 124 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen 125 if (!RegMask::can_represent_arg(warped)) { 126 // the compiler cannot represent this method's calling sequence 127 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence"); 128 return OptoReg::Bad; 129 } 130 return warped; 131 } 132 return OptoReg::as_OptoReg(reg); 133 } 134 135 //---------------------------compute_old_SP------------------------------------ 136 OptoReg::Name Compile::compute_old_SP() { 137 int fixed = fixed_slots(); 138 int preserve = in_preserve_stack_slots(); 139 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots())); 140 } 141 142 143 144 #ifdef ASSERT 145 void Matcher::verify_new_nodes_only(Node* xroot) { 146 // Make sure that the new graph only references new nodes 147 ResourceMark rm; 148 Unique_Node_List worklist; 149 VectorSet visited(Thread::current()->resource_area()); 150 worklist.push(xroot); 151 while (worklist.size() > 0) { 152 Node* n = worklist.pop(); 153 visited <<= n->_idx; 154 assert(C->node_arena()->contains(n), "dead node"); 155 for (uint j = 0; j < n->req(); j++) { 156 Node* in = n->in(j); 157 if (in != NULL) { 158 assert(C->node_arena()->contains(in), "dead node"); 159 if (!visited.test(in->_idx)) { 160 worklist.push(in); 161 } 162 } 163 } 164 } 165 } 166 #endif 167 168 169 //---------------------------match--------------------------------------------- 170 void Matcher::match( ) { 171 if( MaxLabelRootDepth < 100 ) { // Too small? 172 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum"); 173 MaxLabelRootDepth = 100; 174 } 175 // One-time initialization of some register masks. 176 init_spill_mask( C->root()->in(1) ); 177 _return_addr_mask = return_addr(); 178 #ifdef _LP64 179 // Pointers take 2 slots in 64-bit land 180 _return_addr_mask.Insert(OptoReg::add(return_addr(),1)); 181 #endif 182 183 // Map a Java-signature return type into return register-value 184 // machine registers for 0, 1 and 2 returned values. 185 const TypeTuple *range = C->tf()->range(); 186 if( range->cnt() > TypeFunc::Parms ) { // If not a void function 187 // Get ideal-register return type 188 int ireg = range->field_at(TypeFunc::Parms)->ideal_reg(); 189 // Get machine return register 190 uint sop = C->start()->Opcode(); 191 OptoRegPair regs = return_value(ireg, false); 192 193 // And mask for same 194 _return_value_mask = RegMask(regs.first()); 195 if( OptoReg::is_valid(regs.second()) ) 196 _return_value_mask.Insert(regs.second()); 197 } 198 199 // --------------- 200 // Frame Layout 201 202 // Need the method signature to determine the incoming argument types, 203 // because the types determine which registers the incoming arguments are 204 // in, and this affects the matched code. 205 const TypeTuple *domain = C->tf()->domain(); 206 uint argcnt = domain->cnt() - TypeFunc::Parms; 207 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 208 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 209 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt ); 210 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt ); 211 uint i; 212 for( i = 0; i<argcnt; i++ ) { 213 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 214 } 215 216 // Pass array of ideal registers and length to USER code (from the AD file) 217 // that will convert this to an array of register numbers. 218 const StartNode *start = C->start(); 219 start->calling_convention( sig_bt, vm_parm_regs, argcnt ); 220 #ifdef ASSERT 221 // Sanity check users' calling convention. Real handy while trying to 222 // get the initial port correct. 223 { for (uint i = 0; i<argcnt; i++) { 224 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 225 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" ); 226 _parm_regs[i].set_bad(); 227 continue; 228 } 229 VMReg parm_reg = vm_parm_regs[i].first(); 230 assert(parm_reg->is_valid(), "invalid arg?"); 231 if (parm_reg->is_reg()) { 232 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg); 233 assert(can_be_java_arg(opto_parm_reg) || 234 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) || 235 opto_parm_reg == inline_cache_reg(), 236 "parameters in register must be preserved by runtime stubs"); 237 } 238 for (uint j = 0; j < i; j++) { 239 assert(parm_reg != vm_parm_regs[j].first(), 240 "calling conv. must produce distinct regs"); 241 } 242 } 243 } 244 #endif 245 246 // Do some initial frame layout. 247 248 // Compute the old incoming SP (may be called FP) as 249 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2. 250 _old_SP = C->compute_old_SP(); 251 assert( is_even(_old_SP), "must be even" ); 252 253 // Compute highest incoming stack argument as 254 // _old_SP + out_preserve_stack_slots + incoming argument size. 255 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 256 assert( is_even(_in_arg_limit), "out_preserve must be even" ); 257 for( i = 0; i < argcnt; i++ ) { 258 // Permit args to have no register 259 _calling_convention_mask[i].Clear(); 260 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 261 continue; 262 } 263 // calling_convention returns stack arguments as a count of 264 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to 265 // the allocators point of view, taking into account all the 266 // preserve area, locks & pad2. 267 268 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first()); 269 if( OptoReg::is_valid(reg1)) 270 _calling_convention_mask[i].Insert(reg1); 271 272 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second()); 273 if( OptoReg::is_valid(reg2)) 274 _calling_convention_mask[i].Insert(reg2); 275 276 // Saved biased stack-slot register number 277 _parm_regs[i].set_pair(reg2, reg1); 278 } 279 280 // Finally, make sure the incoming arguments take up an even number of 281 // words, in case the arguments or locals need to contain doubleword stack 282 // slots. The rest of the system assumes that stack slot pairs (in 283 // particular, in the spill area) which look aligned will in fact be 284 // aligned relative to the stack pointer in the target machine. Double 285 // stack slots will always be allocated aligned. 286 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong)); 287 288 // Compute highest outgoing stack argument as 289 // _new_SP + out_preserve_stack_slots + max(outgoing argument size). 290 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 291 assert( is_even(_out_arg_limit), "out_preserve must be even" ); 292 293 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) { 294 // the compiler cannot represent this method's calling sequence 295 C->record_method_not_compilable("must be able to represent all call arguments in reg mask"); 296 } 297 298 if (C->failing()) return; // bailed out on incoming arg failure 299 300 // --------------- 301 // Collect roots of matcher trees. Every node for which 302 // _shared[_idx] is cleared is guaranteed to not be shared, and thus 303 // can be a valid interior of some tree. 304 find_shared( C->root() ); 305 find_shared( C->top() ); 306 307 C->print_method(PHASE_BEFORE_MATCHING); 308 309 // Create new ideal node ConP #NULL even if it does exist in old space 310 // to avoid false sharing if the corresponding mach node is not used. 311 // The corresponding mach node is only used in rare cases for derived 312 // pointers. 313 Node* new_ideal_null = ConNode::make(TypePtr::NULL_PTR); 314 315 // Swap out to old-space; emptying new-space 316 Arena *old = C->node_arena()->move_contents(C->old_arena()); 317 318 // Save debug and profile information for nodes in old space: 319 _old_node_note_array = C->node_note_array(); 320 if (_old_node_note_array != NULL) { 321 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*> 322 (C->comp_arena(), _old_node_note_array->length(), 323 0, NULL)); 324 } 325 326 // Pre-size the new_node table to avoid the need for range checks. 327 grow_new_node_array(C->unique()); 328 329 // Reset node counter so MachNodes start with _idx at 0 330 int live_nodes = C->live_nodes(); 331 C->set_unique(0); 332 C->reset_dead_node_list(); 333 334 // Recursively match trees from old space into new space. 335 // Correct leaves of new-space Nodes; they point to old-space. 336 _visited.Clear(); // Clear visit bits for xform call 337 C->set_cached_top_node(xform( C->top(), live_nodes )); 338 if (!C->failing()) { 339 Node* xroot = xform( C->root(), 1 ); 340 if (xroot == NULL) { 341 Matcher::soft_match_failure(); // recursive matching process failed 342 C->record_method_not_compilable("instruction match failed"); 343 } else { 344 // During matching shared constants were attached to C->root() 345 // because xroot wasn't available yet, so transfer the uses to 346 // the xroot. 347 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { 348 Node* n = C->root()->fast_out(j); 349 if (C->node_arena()->contains(n)) { 350 assert(n->in(0) == C->root(), "should be control user"); 351 n->set_req(0, xroot); 352 --j; 353 --jmax; 354 } 355 } 356 357 // Generate new mach node for ConP #NULL 358 assert(new_ideal_null != NULL, "sanity"); 359 _mach_null = match_tree(new_ideal_null); 360 // Don't set control, it will confuse GCM since there are no uses. 361 // The control will be set when this node is used first time 362 // in find_base_for_derived(). 363 assert(_mach_null != NULL, ""); 364 365 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL); 366 367 #ifdef ASSERT 368 verify_new_nodes_only(xroot); 369 #endif 370 } 371 } 372 if (C->top() == NULL || C->root() == NULL) { 373 C->record_method_not_compilable("graph lost"); // %%% cannot happen? 374 } 375 if (C->failing()) { 376 // delete old; 377 old->destruct_contents(); 378 return; 379 } 380 assert( C->top(), "" ); 381 assert( C->root(), "" ); 382 validate_null_checks(); 383 384 // Now smoke old-space 385 NOT_DEBUG( old->destruct_contents() ); 386 387 // ------------------------ 388 // Set up save-on-entry registers 389 Fixup_Save_On_Entry( ); 390 } 391 392 393 //------------------------------Fixup_Save_On_Entry---------------------------- 394 // The stated purpose of this routine is to take care of save-on-entry 395 // registers. However, the overall goal of the Match phase is to convert into 396 // machine-specific instructions which have RegMasks to guide allocation. 397 // So what this procedure really does is put a valid RegMask on each input 398 // to the machine-specific variations of all Return, TailCall and Halt 399 // instructions. It also adds edgs to define the save-on-entry values (and of 400 // course gives them a mask). 401 402 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) { 403 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size ); 404 // Do all the pre-defined register masks 405 rms[TypeFunc::Control ] = RegMask::Empty; 406 rms[TypeFunc::I_O ] = RegMask::Empty; 407 rms[TypeFunc::Memory ] = RegMask::Empty; 408 rms[TypeFunc::ReturnAdr] = ret_adr; 409 rms[TypeFunc::FramePtr ] = fp; 410 return rms; 411 } 412 413 //---------------------------init_first_stack_mask----------------------------- 414 // Create the initial stack mask used by values spilling to the stack. 415 // Disallow any debug info in outgoing argument areas by setting the 416 // initial mask accordingly. 417 void Matcher::init_first_stack_mask() { 418 419 // Allocate storage for spill masks as masks for the appropriate load type. 420 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * (3*6+5)); 421 422 idealreg2spillmask [Op_RegN] = &rms[0]; 423 idealreg2spillmask [Op_RegI] = &rms[1]; 424 idealreg2spillmask [Op_RegL] = &rms[2]; 425 idealreg2spillmask [Op_RegF] = &rms[3]; 426 idealreg2spillmask [Op_RegD] = &rms[4]; 427 idealreg2spillmask [Op_RegP] = &rms[5]; 428 429 idealreg2debugmask [Op_RegN] = &rms[6]; 430 idealreg2debugmask [Op_RegI] = &rms[7]; 431 idealreg2debugmask [Op_RegL] = &rms[8]; 432 idealreg2debugmask [Op_RegF] = &rms[9]; 433 idealreg2debugmask [Op_RegD] = &rms[10]; 434 idealreg2debugmask [Op_RegP] = &rms[11]; 435 436 idealreg2mhdebugmask[Op_RegN] = &rms[12]; 437 idealreg2mhdebugmask[Op_RegI] = &rms[13]; 438 idealreg2mhdebugmask[Op_RegL] = &rms[14]; 439 idealreg2mhdebugmask[Op_RegF] = &rms[15]; 440 idealreg2mhdebugmask[Op_RegD] = &rms[16]; 441 idealreg2mhdebugmask[Op_RegP] = &rms[17]; 442 443 idealreg2spillmask [Op_VecS] = &rms[18]; 444 idealreg2spillmask [Op_VecD] = &rms[19]; 445 idealreg2spillmask [Op_VecX] = &rms[20]; 446 idealreg2spillmask [Op_VecY] = &rms[21]; 447 idealreg2spillmask [Op_VecZ] = &rms[22]; 448 449 OptoReg::Name i; 450 451 // At first, start with the empty mask 452 C->FIRST_STACK_mask().Clear(); 453 454 // Add in the incoming argument area 455 OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 456 for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) { 457 C->FIRST_STACK_mask().Insert(i); 458 } 459 // Add in all bits past the outgoing argument area 460 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)), 461 "must be able to represent all call arguments in reg mask"); 462 OptoReg::Name init = _out_arg_limit; 463 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) { 464 C->FIRST_STACK_mask().Insert(i); 465 } 466 // Finally, set the "infinite stack" bit. 467 C->FIRST_STACK_mask().set_AllStack(); 468 469 // Make spill masks. Registers for their class, plus FIRST_STACK_mask. 470 RegMask aligned_stack_mask = C->FIRST_STACK_mask(); 471 // Keep spill masks aligned. 472 aligned_stack_mask.clear_to_pairs(); 473 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 474 475 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP]; 476 #ifdef _LP64 477 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN]; 478 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask()); 479 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask); 480 #else 481 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask()); 482 #endif 483 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI]; 484 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask()); 485 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL]; 486 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask); 487 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF]; 488 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask()); 489 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD]; 490 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask); 491 492 if (Matcher::vector_size_supported(T_BYTE,4)) { 493 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS]; 494 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask()); 495 } 496 if (Matcher::vector_size_supported(T_FLOAT,2)) { 497 // For VecD we need dual alignment and 8 bytes (2 slots) for spills. 498 // RA guarantees such alignment since it is needed for Double and Long values. 499 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD]; 500 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask); 501 } 502 if (Matcher::vector_size_supported(T_FLOAT,4)) { 503 // For VecX we need quadro alignment and 16 bytes (4 slots) for spills. 504 // 505 // RA can use input arguments stack slots for spills but until RA 506 // we don't know frame size and offset of input arg stack slots. 507 // 508 // Exclude last input arg stack slots to avoid spilling vectors there 509 // otherwise vector spills could stomp over stack slots in caller frame. 510 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 511 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) { 512 aligned_stack_mask.Remove(in); 513 in = OptoReg::add(in, -1); 514 } 515 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX); 516 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 517 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX]; 518 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask); 519 } 520 if (Matcher::vector_size_supported(T_FLOAT,8)) { 521 // For VecY we need octo alignment and 32 bytes (8 slots) for spills. 522 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 523 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) { 524 aligned_stack_mask.Remove(in); 525 in = OptoReg::add(in, -1); 526 } 527 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY); 528 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 529 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY]; 530 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask); 531 } 532 if (Matcher::vector_size_supported(T_FLOAT,16)) { 533 // For VecZ we need enough alignment and 64 bytes (16 slots) for spills. 534 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 535 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecZ); k++) { 536 aligned_stack_mask.Remove(in); 537 in = OptoReg::add(in, -1); 538 } 539 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecZ); 540 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 541 *idealreg2spillmask[Op_VecZ] = *idealreg2regmask[Op_VecZ]; 542 idealreg2spillmask[Op_VecZ]->OR(aligned_stack_mask); 543 } 544 if (UseFPUForSpilling) { 545 // This mask logic assumes that the spill operations are 546 // symmetric and that the registers involved are the same size. 547 // On sparc for instance we may have to use 64 bit moves will 548 // kill 2 registers when used with F0-F31. 549 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]); 550 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]); 551 #ifdef _LP64 552 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]); 553 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 554 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 555 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]); 556 #else 557 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]); 558 #ifdef ARM 559 // ARM has support for moving 64bit values between a pair of 560 // integer registers and a double register 561 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 562 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 563 #endif 564 #endif 565 } 566 567 // Make up debug masks. Any spill slot plus callee-save registers. 568 // Caller-save registers are assumed to be trashable by the various 569 // inline-cache fixup routines. 570 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN]; 571 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI]; 572 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL]; 573 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF]; 574 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD]; 575 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP]; 576 577 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN]; 578 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI]; 579 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL]; 580 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF]; 581 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD]; 582 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP]; 583 584 // Prevent stub compilations from attempting to reference 585 // callee-saved registers from debug info 586 bool exclude_soe = !Compile::current()->is_method_compilation(); 587 588 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 589 // registers the caller has to save do not work 590 if( _register_save_policy[i] == 'C' || 591 _register_save_policy[i] == 'A' || 592 (_register_save_policy[i] == 'E' && exclude_soe) ) { 593 idealreg2debugmask [Op_RegN]->Remove(i); 594 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call 595 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug 596 idealreg2debugmask [Op_RegF]->Remove(i); // masks 597 idealreg2debugmask [Op_RegD]->Remove(i); 598 idealreg2debugmask [Op_RegP]->Remove(i); 599 600 idealreg2mhdebugmask[Op_RegN]->Remove(i); 601 idealreg2mhdebugmask[Op_RegI]->Remove(i); 602 idealreg2mhdebugmask[Op_RegL]->Remove(i); 603 idealreg2mhdebugmask[Op_RegF]->Remove(i); 604 idealreg2mhdebugmask[Op_RegD]->Remove(i); 605 idealreg2mhdebugmask[Op_RegP]->Remove(i); 606 } 607 } 608 609 // Subtract the register we use to save the SP for MethodHandle 610 // invokes to from the debug mask. 611 const RegMask save_mask = method_handle_invoke_SP_save_mask(); 612 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask); 613 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask); 614 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask); 615 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask); 616 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask); 617 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask); 618 } 619 620 //---------------------------is_save_on_entry---------------------------------- 621 bool Matcher::is_save_on_entry( int reg ) { 622 return 623 _register_save_policy[reg] == 'E' || 624 _register_save_policy[reg] == 'A' || // Save-on-entry register? 625 // Also save argument registers in the trampolining stubs 626 (C->save_argument_registers() && is_spillable_arg(reg)); 627 } 628 629 //---------------------------Fixup_Save_On_Entry------------------------------- 630 void Matcher::Fixup_Save_On_Entry( ) { 631 init_first_stack_mask(); 632 633 Node *root = C->root(); // Short name for root 634 // Count number of save-on-entry registers. 635 uint soe_cnt = number_of_saved_registers(); 636 uint i; 637 638 // Find the procedure Start Node 639 StartNode *start = C->start(); 640 assert( start, "Expect a start node" ); 641 642 // Save argument registers in the trampolining stubs 643 if( C->save_argument_registers() ) 644 for( i = 0; i < _last_Mach_Reg; i++ ) 645 if( is_spillable_arg(i) ) 646 soe_cnt++; 647 648 // Input RegMask array shared by all Returns. 649 // The type for doubles and longs has a count of 2, but 650 // there is only 1 returned value 651 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1); 652 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 653 // Returns have 0 or 1 returned values depending on call signature. 654 // Return register is specified by return_value in the AD file. 655 if (ret_edge_cnt > TypeFunc::Parms) 656 ret_rms[TypeFunc::Parms+0] = _return_value_mask; 657 658 // Input RegMask array shared by all Rethrows. 659 uint reth_edge_cnt = TypeFunc::Parms+1; 660 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 661 // Rethrow takes exception oop only, but in the argument 0 slot. 662 OptoReg::Name reg = find_receiver(false); 663 if (reg >= 0) { 664 reth_rms[TypeFunc::Parms] = mreg2regmask[reg]; 665 #ifdef _LP64 666 // Need two slots for ptrs in 64-bit land 667 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1)); 668 #endif 669 } 670 671 // Input RegMask array shared by all TailCalls 672 uint tail_call_edge_cnt = TypeFunc::Parms+2; 673 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 674 675 // Input RegMask array shared by all TailJumps 676 uint tail_jump_edge_cnt = TypeFunc::Parms+2; 677 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 678 679 // TailCalls have 2 returned values (target & moop), whose masks come 680 // from the usual MachNode/MachOper mechanism. Find a sample 681 // TailCall to extract these masks and put the correct masks into 682 // the tail_call_rms array. 683 for( i=1; i < root->req(); i++ ) { 684 MachReturnNode *m = root->in(i)->as_MachReturn(); 685 if( m->ideal_Opcode() == Op_TailCall ) { 686 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 687 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 688 break; 689 } 690 } 691 692 // TailJumps have 2 returned values (target & ex_oop), whose masks come 693 // from the usual MachNode/MachOper mechanism. Find a sample 694 // TailJump to extract these masks and put the correct masks into 695 // the tail_jump_rms array. 696 for( i=1; i < root->req(); i++ ) { 697 MachReturnNode *m = root->in(i)->as_MachReturn(); 698 if( m->ideal_Opcode() == Op_TailJump ) { 699 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 700 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 701 break; 702 } 703 } 704 705 // Input RegMask array shared by all Halts 706 uint halt_edge_cnt = TypeFunc::Parms; 707 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 708 709 // Capture the return input masks into each exit flavor 710 for( i=1; i < root->req(); i++ ) { 711 MachReturnNode *exit = root->in(i)->as_MachReturn(); 712 switch( exit->ideal_Opcode() ) { 713 case Op_Return : exit->_in_rms = ret_rms; break; 714 case Op_Rethrow : exit->_in_rms = reth_rms; break; 715 case Op_TailCall : exit->_in_rms = tail_call_rms; break; 716 case Op_TailJump : exit->_in_rms = tail_jump_rms; break; 717 case Op_Halt : exit->_in_rms = halt_rms; break; 718 default : ShouldNotReachHere(); 719 } 720 } 721 722 // Next unused projection number from Start. 723 int proj_cnt = C->tf()->domain()->cnt(); 724 725 // Do all the save-on-entry registers. Make projections from Start for 726 // them, and give them a use at the exit points. To the allocator, they 727 // look like incoming register arguments. 728 for( i = 0; i < _last_Mach_Reg; i++ ) { 729 if( is_save_on_entry(i) ) { 730 731 // Add the save-on-entry to the mask array 732 ret_rms [ ret_edge_cnt] = mreg2regmask[i]; 733 reth_rms [ reth_edge_cnt] = mreg2regmask[i]; 734 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i]; 735 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i]; 736 // Halts need the SOE registers, but only in the stack as debug info. 737 // A just-prior uncommon-trap or deoptimization will use the SOE regs. 738 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]]; 739 740 Node *mproj; 741 742 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's 743 // into a single RegD. 744 if( (i&1) == 0 && 745 _register_save_type[i ] == Op_RegF && 746 _register_save_type[i+1] == Op_RegF && 747 is_save_on_entry(i+1) ) { 748 // Add other bit for double 749 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 750 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 751 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 752 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 753 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 754 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD ); 755 proj_cnt += 2; // Skip 2 for doubles 756 } 757 else if( (i&1) == 1 && // Else check for high half of double 758 _register_save_type[i-1] == Op_RegF && 759 _register_save_type[i ] == Op_RegF && 760 is_save_on_entry(i-1) ) { 761 ret_rms [ ret_edge_cnt] = RegMask::Empty; 762 reth_rms [ reth_edge_cnt] = RegMask::Empty; 763 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 764 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 765 halt_rms [ halt_edge_cnt] = RegMask::Empty; 766 mproj = C->top(); 767 } 768 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's 769 // into a single RegL. 770 else if( (i&1) == 0 && 771 _register_save_type[i ] == Op_RegI && 772 _register_save_type[i+1] == Op_RegI && 773 is_save_on_entry(i+1) ) { 774 // Add other bit for long 775 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 776 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 777 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 778 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 779 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 780 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL ); 781 proj_cnt += 2; // Skip 2 for longs 782 } 783 else if( (i&1) == 1 && // Else check for high half of long 784 _register_save_type[i-1] == Op_RegI && 785 _register_save_type[i ] == Op_RegI && 786 is_save_on_entry(i-1) ) { 787 ret_rms [ ret_edge_cnt] = RegMask::Empty; 788 reth_rms [ reth_edge_cnt] = RegMask::Empty; 789 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 790 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 791 halt_rms [ halt_edge_cnt] = RegMask::Empty; 792 mproj = C->top(); 793 } else { 794 // Make a projection for it off the Start 795 mproj = new MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] ); 796 } 797 798 ret_edge_cnt ++; 799 reth_edge_cnt ++; 800 tail_call_edge_cnt ++; 801 tail_jump_edge_cnt ++; 802 halt_edge_cnt ++; 803 804 // Add a use of the SOE register to all exit paths 805 for( uint j=1; j < root->req(); j++ ) 806 root->in(j)->add_req(mproj); 807 } // End of if a save-on-entry register 808 } // End of for all machine registers 809 } 810 811 //------------------------------init_spill_mask-------------------------------- 812 void Matcher::init_spill_mask( Node *ret ) { 813 if( idealreg2regmask[Op_RegI] ) return; // One time only init 814 815 OptoReg::c_frame_pointer = c_frame_pointer(); 816 c_frame_ptr_mask = c_frame_pointer(); 817 #ifdef _LP64 818 // pointers are twice as big 819 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1)); 820 #endif 821 822 // Start at OptoReg::stack0() 823 STACK_ONLY_mask.Clear(); 824 OptoReg::Name init = OptoReg::stack2reg(0); 825 // STACK_ONLY_mask is all stack bits 826 OptoReg::Name i; 827 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) 828 STACK_ONLY_mask.Insert(i); 829 // Also set the "infinite stack" bit. 830 STACK_ONLY_mask.set_AllStack(); 831 832 // Copy the register names over into the shared world 833 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 834 // SharedInfo::regName[i] = regName[i]; 835 // Handy RegMasks per machine register 836 mreg2regmask[i].Insert(i); 837 } 838 839 // Grab the Frame Pointer 840 Node *fp = ret->in(TypeFunc::FramePtr); 841 Node *mem = ret->in(TypeFunc::Memory); 842 const TypePtr* atp = TypePtr::BOTTOM; 843 // Share frame pointer while making spill ops 844 set_shared(fp); 845 846 // Compute generic short-offset Loads 847 #ifdef _LP64 848 MachNode *spillCP = match_tree(new LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 849 #endif 850 MachNode *spillI = match_tree(new LoadINode(NULL,mem,fp,atp,TypeInt::INT,MemNode::unordered)); 851 MachNode *spillL = match_tree(new LoadLNode(NULL,mem,fp,atp,TypeLong::LONG,MemNode::unordered, LoadNode::DependsOnlyOnTest, false)); 852 MachNode *spillF = match_tree(new LoadFNode(NULL,mem,fp,atp,Type::FLOAT,MemNode::unordered)); 853 MachNode *spillD = match_tree(new LoadDNode(NULL,mem,fp,atp,Type::DOUBLE,MemNode::unordered)); 854 MachNode *spillP = match_tree(new LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 855 assert(spillI != NULL && spillL != NULL && spillF != NULL && 856 spillD != NULL && spillP != NULL, ""); 857 // Get the ADLC notion of the right regmask, for each basic type. 858 #ifdef _LP64 859 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask(); 860 #endif 861 idealreg2regmask[Op_RegI] = &spillI->out_RegMask(); 862 idealreg2regmask[Op_RegL] = &spillL->out_RegMask(); 863 idealreg2regmask[Op_RegF] = &spillF->out_RegMask(); 864 idealreg2regmask[Op_RegD] = &spillD->out_RegMask(); 865 idealreg2regmask[Op_RegP] = &spillP->out_RegMask(); 866 867 // Vector regmasks. 868 if (Matcher::vector_size_supported(T_BYTE,4)) { 869 TypeVect::VECTS = TypeVect::make(T_BYTE, 4); 870 MachNode *spillVectS = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTS)); 871 idealreg2regmask[Op_VecS] = &spillVectS->out_RegMask(); 872 } 873 if (Matcher::vector_size_supported(T_FLOAT,2)) { 874 MachNode *spillVectD = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTD)); 875 idealreg2regmask[Op_VecD] = &spillVectD->out_RegMask(); 876 } 877 if (Matcher::vector_size_supported(T_FLOAT,4)) { 878 MachNode *spillVectX = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTX)); 879 idealreg2regmask[Op_VecX] = &spillVectX->out_RegMask(); 880 } 881 if (Matcher::vector_size_supported(T_FLOAT,8)) { 882 MachNode *spillVectY = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTY)); 883 idealreg2regmask[Op_VecY] = &spillVectY->out_RegMask(); 884 } 885 if (Matcher::vector_size_supported(T_FLOAT,16)) { 886 MachNode *spillVectZ = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTZ)); 887 idealreg2regmask[Op_VecZ] = &spillVectZ->out_RegMask(); 888 } 889 } 890 891 #ifdef ASSERT 892 static void match_alias_type(Compile* C, Node* n, Node* m) { 893 if (!VerifyAliases) return; // do not go looking for trouble by default 894 const TypePtr* nat = n->adr_type(); 895 const TypePtr* mat = m->adr_type(); 896 int nidx = C->get_alias_index(nat); 897 int midx = C->get_alias_index(mat); 898 // Detune the assert for cases like (AndI 0xFF (LoadB p)). 899 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) { 900 for (uint i = 1; i < n->req(); i++) { 901 Node* n1 = n->in(i); 902 const TypePtr* n1at = n1->adr_type(); 903 if (n1at != NULL) { 904 nat = n1at; 905 nidx = C->get_alias_index(n1at); 906 } 907 } 908 } 909 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases: 910 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) { 911 switch (n->Opcode()) { 912 case Op_PrefetchAllocation: 913 nidx = Compile::AliasIdxRaw; 914 nat = TypeRawPtr::BOTTOM; 915 break; 916 } 917 } 918 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) { 919 switch (n->Opcode()) { 920 case Op_ClearArray: 921 midx = Compile::AliasIdxRaw; 922 mat = TypeRawPtr::BOTTOM; 923 break; 924 } 925 } 926 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) { 927 switch (n->Opcode()) { 928 case Op_Return: 929 case Op_Rethrow: 930 case Op_Halt: 931 case Op_TailCall: 932 case Op_TailJump: 933 nidx = Compile::AliasIdxBot; 934 nat = TypePtr::BOTTOM; 935 break; 936 } 937 } 938 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) { 939 switch (n->Opcode()) { 940 case Op_StrComp: 941 case Op_StrEquals: 942 case Op_StrIndexOf: 943 case Op_StrIndexOfChar: 944 case Op_AryEq: 945 case Op_HasNegatives: 946 case Op_MemBarVolatile: 947 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type? 948 case Op_StrInflatedCopy: 949 case Op_StrCompressedCopy: 950 case Op_OnSpinWait: 951 case Op_EncodeISOArray: 952 nidx = Compile::AliasIdxTop; 953 nat = NULL; 954 break; 955 } 956 } 957 if (nidx != midx) { 958 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) { 959 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx); 960 n->dump(); 961 m->dump(); 962 } 963 assert(C->subsume_loads() && C->must_alias(nat, midx), 964 "must not lose alias info when matching"); 965 } 966 } 967 #endif 968 969 970 //------------------------------MStack----------------------------------------- 971 // State and MStack class used in xform() and find_shared() iterative methods. 972 enum Node_State { Pre_Visit, // node has to be pre-visited 973 Visit, // visit node 974 Post_Visit, // post-visit node 975 Alt_Post_Visit // alternative post-visit path 976 }; 977 978 class MStack: public Node_Stack { 979 public: 980 MStack(int size) : Node_Stack(size) { } 981 982 void push(Node *n, Node_State ns) { 983 Node_Stack::push(n, (uint)ns); 984 } 985 void push(Node *n, Node_State ns, Node *parent, int indx) { 986 ++_inode_top; 987 if ((_inode_top + 1) >= _inode_max) grow(); 988 _inode_top->node = parent; 989 _inode_top->indx = (uint)indx; 990 ++_inode_top; 991 _inode_top->node = n; 992 _inode_top->indx = (uint)ns; 993 } 994 Node *parent() { 995 pop(); 996 return node(); 997 } 998 Node_State state() const { 999 return (Node_State)index(); 1000 } 1001 void set_state(Node_State ns) { 1002 set_index((uint)ns); 1003 } 1004 }; 1005 1006 1007 //------------------------------xform------------------------------------------ 1008 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine 1009 // Node in new-space. Given a new-space Node, recursively walk his children. 1010 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; } 1011 Node *Matcher::xform( Node *n, int max_stack ) { 1012 // Use one stack to keep both: child's node/state and parent's node/index 1013 MStack mstack(max_stack * 2 * 2); // usually: C->live_nodes() * 2 * 2 1014 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root 1015 1016 while (mstack.is_nonempty()) { 1017 C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions"); 1018 if (C->failing()) return NULL; 1019 n = mstack.node(); // Leave node on stack 1020 Node_State nstate = mstack.state(); 1021 if (nstate == Visit) { 1022 mstack.set_state(Post_Visit); 1023 Node *oldn = n; 1024 // Old-space or new-space check 1025 if (!C->node_arena()->contains(n)) { 1026 // Old space! 1027 Node* m; 1028 if (has_new_node(n)) { // Not yet Label/Reduced 1029 m = new_node(n); 1030 } else { 1031 if (!is_dontcare(n)) { // Matcher can match this guy 1032 // Calls match special. They match alone with no children. 1033 // Their children, the incoming arguments, match normally. 1034 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n); 1035 if (C->failing()) return NULL; 1036 if (m == NULL) { Matcher::soft_match_failure(); return NULL; } 1037 } else { // Nothing the matcher cares about 1038 if( n->is_Proj() && n->in(0)->is_Multi()) { // Projections? 1039 // Convert to machine-dependent projection 1040 m = n->in(0)->as_Multi()->match( n->as_Proj(), this ); 1041 #ifdef ASSERT 1042 _new2old_map.map(m->_idx, n); 1043 #endif 1044 if (m->in(0) != NULL) // m might be top 1045 collect_null_checks(m, n); 1046 } else { // Else just a regular 'ol guy 1047 m = n->clone(); // So just clone into new-space 1048 #ifdef ASSERT 1049 _new2old_map.map(m->_idx, n); 1050 #endif 1051 // Def-Use edges will be added incrementally as Uses 1052 // of this node are matched. 1053 assert(m->outcnt() == 0, "no Uses of this clone yet"); 1054 } 1055 } 1056 1057 set_new_node(n, m); // Map old to new 1058 if (_old_node_note_array != NULL) { 1059 Node_Notes* nn = C->locate_node_notes(_old_node_note_array, 1060 n->_idx); 1061 C->set_node_notes_at(m->_idx, nn); 1062 } 1063 debug_only(match_alias_type(C, n, m)); 1064 } 1065 n = m; // n is now a new-space node 1066 mstack.set_node(n); 1067 } 1068 1069 // New space! 1070 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty()) 1071 1072 int i; 1073 // Put precedence edges on stack first (match them last). 1074 for (i = oldn->req(); (uint)i < oldn->len(); i++) { 1075 Node *m = oldn->in(i); 1076 if (m == NULL) break; 1077 // set -1 to call add_prec() instead of set_req() during Step1 1078 mstack.push(m, Visit, n, -1); 1079 } 1080 1081 // Handle precedence edges for interior nodes 1082 for (i = n->len()-1; (uint)i >= n->req(); i--) { 1083 Node *m = n->in(i); 1084 if (m == NULL || C->node_arena()->contains(m)) continue; 1085 n->rm_prec(i); 1086 // set -1 to call add_prec() instead of set_req() during Step1 1087 mstack.push(m, Visit, n, -1); 1088 } 1089 1090 // For constant debug info, I'd rather have unmatched constants. 1091 int cnt = n->req(); 1092 JVMState* jvms = n->jvms(); 1093 int debug_cnt = jvms ? jvms->debug_start() : cnt; 1094 1095 // Now do only debug info. Clone constants rather than matching. 1096 // Constants are represented directly in the debug info without 1097 // the need for executable machine instructions. 1098 // Monitor boxes are also represented directly. 1099 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do 1100 Node *m = n->in(i); // Get input 1101 int op = m->Opcode(); 1102 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites"); 1103 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass || 1104 op == Op_ConF || op == Op_ConD || op == Op_ConL 1105 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp 1106 ) { 1107 m = m->clone(); 1108 #ifdef ASSERT 1109 _new2old_map.map(m->_idx, n); 1110 #endif 1111 mstack.push(m, Post_Visit, n, i); // Don't need to visit 1112 mstack.push(m->in(0), Visit, m, 0); 1113 } else { 1114 mstack.push(m, Visit, n, i); 1115 } 1116 } 1117 1118 // And now walk his children, and convert his inputs to new-space. 1119 for( ; i >= 0; --i ) { // For all normal inputs do 1120 Node *m = n->in(i); // Get input 1121 if(m != NULL) 1122 mstack.push(m, Visit, n, i); 1123 } 1124 1125 } 1126 else if (nstate == Post_Visit) { 1127 // Set xformed input 1128 Node *p = mstack.parent(); 1129 if (p != NULL) { // root doesn't have parent 1130 int i = (int)mstack.index(); 1131 if (i >= 0) 1132 p->set_req(i, n); // required input 1133 else if (i == -1) 1134 p->add_prec(n); // precedence input 1135 else 1136 ShouldNotReachHere(); 1137 } 1138 mstack.pop(); // remove processed node from stack 1139 } 1140 else { 1141 ShouldNotReachHere(); 1142 } 1143 } // while (mstack.is_nonempty()) 1144 return n; // Return new-space Node 1145 } 1146 1147 //------------------------------warp_outgoing_stk_arg------------------------ 1148 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) { 1149 // Convert outgoing argument location to a pre-biased stack offset 1150 if (reg->is_stack()) { 1151 OptoReg::Name warped = reg->reg2stack(); 1152 // Adjust the stack slot offset to be the register number used 1153 // by the allocator. 1154 warped = OptoReg::add(begin_out_arg_area, warped); 1155 // Keep track of the largest numbered stack slot used for an arg. 1156 // Largest used slot per call-site indicates the amount of stack 1157 // that is killed by the call. 1158 if( warped >= out_arg_limit_per_call ) 1159 out_arg_limit_per_call = OptoReg::add(warped,1); 1160 if (!RegMask::can_represent_arg(warped)) { 1161 C->record_method_not_compilable_all_tiers("unsupported calling sequence"); 1162 return OptoReg::Bad; 1163 } 1164 return warped; 1165 } 1166 return OptoReg::as_OptoReg(reg); 1167 } 1168 1169 1170 //------------------------------match_sfpt------------------------------------- 1171 // Helper function to match call instructions. Calls match special. 1172 // They match alone with no children. Their children, the incoming 1173 // arguments, match normally. 1174 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) { 1175 MachSafePointNode *msfpt = NULL; 1176 MachCallNode *mcall = NULL; 1177 uint cnt; 1178 // Split out case for SafePoint vs Call 1179 CallNode *call; 1180 const TypeTuple *domain; 1181 ciMethod* method = NULL; 1182 bool is_method_handle_invoke = false; // for special kill effects 1183 if( sfpt->is_Call() ) { 1184 call = sfpt->as_Call(); 1185 domain = call->tf()->domain(); 1186 cnt = domain->cnt(); 1187 1188 // Match just the call, nothing else 1189 MachNode *m = match_tree(call); 1190 if (C->failing()) return NULL; 1191 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; } 1192 1193 // Copy data from the Ideal SafePoint to the machine version 1194 mcall = m->as_MachCall(); 1195 1196 mcall->set_tf( call->tf()); 1197 mcall->set_entry_point(call->entry_point()); 1198 mcall->set_cnt( call->cnt()); 1199 1200 if( mcall->is_MachCallJava() ) { 1201 MachCallJavaNode *mcall_java = mcall->as_MachCallJava(); 1202 const CallJavaNode *call_java = call->as_CallJava(); 1203 method = call_java->method(); 1204 mcall_java->_method = method; 1205 mcall_java->_bci = call_java->_bci; 1206 mcall_java->_optimized_virtual = call_java->is_optimized_virtual(); 1207 is_method_handle_invoke = call_java->is_method_handle_invoke(); 1208 mcall_java->_method_handle_invoke = is_method_handle_invoke; 1209 mcall_java->_override_symbolic_info = call_java->override_symbolic_info(); 1210 if (is_method_handle_invoke) { 1211 C->set_has_method_handle_invokes(true); 1212 } 1213 if( mcall_java->is_MachCallStaticJava() ) 1214 mcall_java->as_MachCallStaticJava()->_name = 1215 call_java->as_CallStaticJava()->_name; 1216 if( mcall_java->is_MachCallDynamicJava() ) 1217 mcall_java->as_MachCallDynamicJava()->_vtable_index = 1218 call_java->as_CallDynamicJava()->_vtable_index; 1219 } 1220 else if( mcall->is_MachCallRuntime() ) { 1221 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name; 1222 } 1223 msfpt = mcall; 1224 } 1225 // This is a non-call safepoint 1226 else { 1227 call = NULL; 1228 domain = NULL; 1229 MachNode *mn = match_tree(sfpt); 1230 if (C->failing()) return NULL; 1231 msfpt = mn->as_MachSafePoint(); 1232 cnt = TypeFunc::Parms; 1233 } 1234 1235 // Advertise the correct memory effects (for anti-dependence computation). 1236 msfpt->set_adr_type(sfpt->adr_type()); 1237 1238 // Allocate a private array of RegMasks. These RegMasks are not shared. 1239 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt ); 1240 // Empty them all. 1241 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt ); 1242 1243 // Do all the pre-defined non-Empty register masks 1244 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask; 1245 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask; 1246 1247 // Place first outgoing argument can possibly be put. 1248 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 1249 assert( is_even(begin_out_arg_area), "" ); 1250 // Compute max outgoing register number per call site. 1251 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area; 1252 // Calls to C may hammer extra stack slots above and beyond any arguments. 1253 // These are usually backing store for register arguments for varargs. 1254 if( call != NULL && call->is_CallRuntime() ) 1255 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed()); 1256 1257 1258 // Do the normal argument list (parameters) register masks 1259 int argcnt = cnt - TypeFunc::Parms; 1260 if( argcnt > 0 ) { // Skip it all if we have no args 1261 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 1262 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 1263 int i; 1264 for( i = 0; i < argcnt; i++ ) { 1265 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 1266 } 1267 // V-call to pick proper calling convention 1268 call->calling_convention( sig_bt, parm_regs, argcnt ); 1269 1270 #ifdef ASSERT 1271 // Sanity check users' calling convention. Really handy during 1272 // the initial porting effort. Fairly expensive otherwise. 1273 { for (int i = 0; i<argcnt; i++) { 1274 if( !parm_regs[i].first()->is_valid() && 1275 !parm_regs[i].second()->is_valid() ) continue; 1276 VMReg reg1 = parm_regs[i].first(); 1277 VMReg reg2 = parm_regs[i].second(); 1278 for (int j = 0; j < i; j++) { 1279 if( !parm_regs[j].first()->is_valid() && 1280 !parm_regs[j].second()->is_valid() ) continue; 1281 VMReg reg3 = parm_regs[j].first(); 1282 VMReg reg4 = parm_regs[j].second(); 1283 if( !reg1->is_valid() ) { 1284 assert( !reg2->is_valid(), "valid halvsies" ); 1285 } else if( !reg3->is_valid() ) { 1286 assert( !reg4->is_valid(), "valid halvsies" ); 1287 } else { 1288 assert( reg1 != reg2, "calling conv. must produce distinct regs"); 1289 assert( reg1 != reg3, "calling conv. must produce distinct regs"); 1290 assert( reg1 != reg4, "calling conv. must produce distinct regs"); 1291 assert( reg2 != reg3, "calling conv. must produce distinct regs"); 1292 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs"); 1293 assert( reg3 != reg4, "calling conv. must produce distinct regs"); 1294 } 1295 } 1296 } 1297 } 1298 #endif 1299 1300 // Visit each argument. Compute its outgoing register mask. 1301 // Return results now can have 2 bits returned. 1302 // Compute max over all outgoing arguments both per call-site 1303 // and over the entire method. 1304 for( i = 0; i < argcnt; i++ ) { 1305 // Address of incoming argument mask to fill in 1306 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms]; 1307 if( !parm_regs[i].first()->is_valid() && 1308 !parm_regs[i].second()->is_valid() ) { 1309 continue; // Avoid Halves 1310 } 1311 // Grab first register, adjust stack slots and insert in mask. 1312 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call ); 1313 if (OptoReg::is_valid(reg1)) 1314 rm->Insert( reg1 ); 1315 // Grab second register (if any), adjust stack slots and insert in mask. 1316 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call ); 1317 if (OptoReg::is_valid(reg2)) 1318 rm->Insert( reg2 ); 1319 } // End of for all arguments 1320 1321 // Compute number of stack slots needed to restore stack in case of 1322 // Pascal-style argument popping. 1323 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area; 1324 } 1325 1326 // Compute the max stack slot killed by any call. These will not be 1327 // available for debug info, and will be used to adjust FIRST_STACK_mask 1328 // after all call sites have been visited. 1329 if( _out_arg_limit < out_arg_limit_per_call) 1330 _out_arg_limit = out_arg_limit_per_call; 1331 1332 if (mcall) { 1333 // Kill the outgoing argument area, including any non-argument holes and 1334 // any legacy C-killed slots. Use Fat-Projections to do the killing. 1335 // Since the max-per-method covers the max-per-call-site and debug info 1336 // is excluded on the max-per-method basis, debug info cannot land in 1337 // this killed area. 1338 uint r_cnt = mcall->tf()->range()->cnt(); 1339 MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj ); 1340 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) { 1341 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence"); 1342 } else { 1343 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++) 1344 proj->_rout.Insert(OptoReg::Name(i)); 1345 } 1346 if (proj->_rout.is_NotEmpty()) { 1347 push_projection(proj); 1348 } 1349 } 1350 // Transfer the safepoint information from the call to the mcall 1351 // Move the JVMState list 1352 msfpt->set_jvms(sfpt->jvms()); 1353 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) { 1354 jvms->set_map(sfpt); 1355 } 1356 1357 // Debug inputs begin just after the last incoming parameter 1358 assert((mcall == NULL) || (mcall->jvms() == NULL) || 1359 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), ""); 1360 1361 // Move the OopMap 1362 msfpt->_oop_map = sfpt->_oop_map; 1363 1364 // Add additional edges. 1365 if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) { 1366 // For these calls we can not add MachConstantBase in expand(), as the 1367 // ins are not complete then. 1368 msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node()); 1369 if (msfpt->jvms() && 1370 msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) { 1371 // We added an edge before jvms, so we must adapt the position of the ins. 1372 msfpt->jvms()->adapt_position(+1); 1373 } 1374 } 1375 1376 // Registers killed by the call are set in the local scheduling pass 1377 // of Global Code Motion. 1378 return msfpt; 1379 } 1380 1381 //---------------------------match_tree---------------------------------------- 1382 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part 1383 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for 1384 // making GotoNodes while building the CFG and in init_spill_mask() to identify 1385 // a Load's result RegMask for memoization in idealreg2regmask[] 1386 MachNode *Matcher::match_tree( const Node *n ) { 1387 assert( n->Opcode() != Op_Phi, "cannot match" ); 1388 assert( !n->is_block_start(), "cannot match" ); 1389 // Set the mark for all locally allocated State objects. 1390 // When this call returns, the _states_arena arena will be reset 1391 // freeing all State objects. 1392 ResourceMark rm( &_states_arena ); 1393 1394 LabelRootDepth = 0; 1395 1396 // StoreNodes require their Memory input to match any LoadNodes 1397 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ; 1398 #ifdef ASSERT 1399 Node* save_mem_node = _mem_node; 1400 _mem_node = n->is_Store() ? (Node*)n : NULL; 1401 #endif 1402 // State object for root node of match tree 1403 // Allocate it on _states_arena - stack allocation can cause stack overflow. 1404 State *s = new (&_states_arena) State; 1405 s->_kids[0] = NULL; 1406 s->_kids[1] = NULL; 1407 s->_leaf = (Node*)n; 1408 // Label the input tree, allocating labels from top-level arena 1409 Label_Root( n, s, n->in(0), mem ); 1410 if (C->failing()) return NULL; 1411 1412 // The minimum cost match for the whole tree is found at the root State 1413 uint mincost = max_juint; 1414 uint cost = max_juint; 1415 uint i; 1416 for( i = 0; i < NUM_OPERANDS; i++ ) { 1417 if( s->valid(i) && // valid entry and 1418 s->_cost[i] < cost && // low cost and 1419 s->_rule[i] >= NUM_OPERANDS ) // not an operand 1420 cost = s->_cost[mincost=i]; 1421 } 1422 if (mincost == max_juint) { 1423 #ifndef PRODUCT 1424 tty->print("No matching rule for:"); 1425 s->dump(); 1426 #endif 1427 Matcher::soft_match_failure(); 1428 return NULL; 1429 } 1430 // Reduce input tree based upon the state labels to machine Nodes 1431 MachNode *m = ReduceInst( s, s->_rule[mincost], mem ); 1432 #ifdef ASSERT 1433 _old2new_map.map(n->_idx, m); 1434 _new2old_map.map(m->_idx, (Node*)n); 1435 #endif 1436 1437 // Add any Matcher-ignored edges 1438 uint cnt = n->req(); 1439 uint start = 1; 1440 if( mem != (Node*)1 ) start = MemNode::Memory+1; 1441 if( n->is_AddP() ) { 1442 assert( mem == (Node*)1, "" ); 1443 start = AddPNode::Base+1; 1444 } 1445 for( i = start; i < cnt; i++ ) { 1446 if( !n->match_edge(i) ) { 1447 if( i < m->req() ) 1448 m->ins_req( i, n->in(i) ); 1449 else 1450 m->add_req( n->in(i) ); 1451 } 1452 } 1453 1454 debug_only( _mem_node = save_mem_node; ) 1455 return m; 1456 } 1457 1458 1459 //------------------------------match_into_reg--------------------------------- 1460 // Choose to either match this Node in a register or part of the current 1461 // match tree. Return true for requiring a register and false for matching 1462 // as part of the current match tree. 1463 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) { 1464 1465 const Type *t = m->bottom_type(); 1466 1467 if (t->singleton()) { 1468 // Never force constants into registers. Allow them to match as 1469 // constants or registers. Copies of the same value will share 1470 // the same register. See find_shared_node. 1471 return false; 1472 } else { // Not a constant 1473 // Stop recursion if they have different Controls. 1474 Node* m_control = m->in(0); 1475 // Control of load's memory can post-dominates load's control. 1476 // So use it since load can't float above its memory. 1477 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL; 1478 if (control && m_control && control != m_control && control != mem_control) { 1479 1480 // Actually, we can live with the most conservative control we 1481 // find, if it post-dominates the others. This allows us to 1482 // pick up load/op/store trees where the load can float a little 1483 // above the store. 1484 Node *x = control; 1485 const uint max_scan = 6; // Arbitrary scan cutoff 1486 uint j; 1487 for (j=0; j<max_scan; j++) { 1488 if (x->is_Region()) // Bail out at merge points 1489 return true; 1490 x = x->in(0); 1491 if (x == m_control) // Does 'control' post-dominate 1492 break; // m->in(0)? If so, we can use it 1493 if (x == mem_control) // Does 'control' post-dominate 1494 break; // mem_control? If so, we can use it 1495 } 1496 if (j == max_scan) // No post-domination before scan end? 1497 return true; // Then break the match tree up 1498 } 1499 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) || 1500 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) { 1501 // These are commonly used in address expressions and can 1502 // efficiently fold into them on X64 in some cases. 1503 return false; 1504 } 1505 } 1506 1507 // Not forceable cloning. If shared, put it into a register. 1508 return shared; 1509 } 1510 1511 1512 //------------------------------Instruction Selection-------------------------- 1513 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match 1514 // ideal nodes to machine instructions. Trees are delimited by shared Nodes, 1515 // things the Matcher does not match (e.g., Memory), and things with different 1516 // Controls (hence forced into different blocks). We pass in the Control 1517 // selected for this entire State tree. 1518 1519 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the 1520 // Store and the Load must have identical Memories (as well as identical 1521 // pointers). Since the Matcher does not have anything for Memory (and 1522 // does not handle DAGs), I have to match the Memory input myself. If the 1523 // Tree root is a Store, I require all Loads to have the identical memory. 1524 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){ 1525 // Since Label_Root is a recursive function, its possible that we might run 1526 // out of stack space. See bugs 6272980 & 6227033 for more info. 1527 LabelRootDepth++; 1528 if (LabelRootDepth > MaxLabelRootDepth) { 1529 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth"); 1530 return NULL; 1531 } 1532 uint care = 0; // Edges matcher cares about 1533 uint cnt = n->req(); 1534 uint i = 0; 1535 1536 // Examine children for memory state 1537 // Can only subsume a child into your match-tree if that child's memory state 1538 // is not modified along the path to another input. 1539 // It is unsafe even if the other inputs are separate roots. 1540 Node *input_mem = NULL; 1541 for( i = 1; i < cnt; i++ ) { 1542 if( !n->match_edge(i) ) continue; 1543 Node *m = n->in(i); // Get ith input 1544 assert( m, "expect non-null children" ); 1545 if( m->is_Load() ) { 1546 if( input_mem == NULL ) { 1547 input_mem = m->in(MemNode::Memory); 1548 } else if( input_mem != m->in(MemNode::Memory) ) { 1549 input_mem = NodeSentinel; 1550 } 1551 } 1552 } 1553 1554 for( i = 1; i < cnt; i++ ){// For my children 1555 if( !n->match_edge(i) ) continue; 1556 Node *m = n->in(i); // Get ith input 1557 // Allocate states out of a private arena 1558 State *s = new (&_states_arena) State; 1559 svec->_kids[care++] = s; 1560 assert( care <= 2, "binary only for now" ); 1561 1562 // Recursively label the State tree. 1563 s->_kids[0] = NULL; 1564 s->_kids[1] = NULL; 1565 s->_leaf = m; 1566 1567 // Check for leaves of the State Tree; things that cannot be a part of 1568 // the current tree. If it finds any, that value is matched as a 1569 // register operand. If not, then the normal matching is used. 1570 if( match_into_reg(n, m, control, i, is_shared(m)) || 1571 // 1572 // Stop recursion if this is LoadNode and the root of this tree is a 1573 // StoreNode and the load & store have different memories. 1574 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) || 1575 // Can NOT include the match of a subtree when its memory state 1576 // is used by any of the other subtrees 1577 (input_mem == NodeSentinel) ) { 1578 // Print when we exclude matching due to different memory states at input-loads 1579 if (PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel) 1580 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem)) { 1581 tty->print_cr("invalid input_mem"); 1582 } 1583 // Switch to a register-only opcode; this value must be in a register 1584 // and cannot be subsumed as part of a larger instruction. 1585 s->DFA( m->ideal_reg(), m ); 1586 1587 } else { 1588 // If match tree has no control and we do, adopt it for entire tree 1589 if( control == NULL && m->in(0) != NULL && m->req() > 1 ) 1590 control = m->in(0); // Pick up control 1591 // Else match as a normal part of the match tree. 1592 control = Label_Root(m,s,control,mem); 1593 if (C->failing()) return NULL; 1594 } 1595 } 1596 1597 1598 // Call DFA to match this node, and return 1599 svec->DFA( n->Opcode(), n ); 1600 1601 #ifdef ASSERT 1602 uint x; 1603 for( x = 0; x < _LAST_MACH_OPER; x++ ) 1604 if( svec->valid(x) ) 1605 break; 1606 1607 if (x >= _LAST_MACH_OPER) { 1608 n->dump(); 1609 svec->dump(); 1610 assert( false, "bad AD file" ); 1611 } 1612 #endif 1613 return control; 1614 } 1615 1616 1617 // Con nodes reduced using the same rule can share their MachNode 1618 // which reduces the number of copies of a constant in the final 1619 // program. The register allocator is free to split uses later to 1620 // split live ranges. 1621 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) { 1622 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL; 1623 1624 // See if this Con has already been reduced using this rule. 1625 if (_shared_nodes.Size() <= leaf->_idx) return NULL; 1626 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx); 1627 if (last != NULL && rule == last->rule()) { 1628 // Don't expect control change for DecodeN 1629 if (leaf->is_DecodeNarrowPtr()) 1630 return last; 1631 // Get the new space root. 1632 Node* xroot = new_node(C->root()); 1633 if (xroot == NULL) { 1634 // This shouldn't happen give the order of matching. 1635 return NULL; 1636 } 1637 1638 // Shared constants need to have their control be root so they 1639 // can be scheduled properly. 1640 Node* control = last->in(0); 1641 if (control != xroot) { 1642 if (control == NULL || control == C->root()) { 1643 last->set_req(0, xroot); 1644 } else { 1645 assert(false, "unexpected control"); 1646 return NULL; 1647 } 1648 } 1649 return last; 1650 } 1651 return NULL; 1652 } 1653 1654 1655 //------------------------------ReduceInst------------------------------------- 1656 // Reduce a State tree (with given Control) into a tree of MachNodes. 1657 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into 1658 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes. 1659 // Each MachNode has a number of complicated MachOper operands; each 1660 // MachOper also covers a further tree of Ideal Nodes. 1661 1662 // The root of the Ideal match tree is always an instruction, so we enter 1663 // the recursion here. After building the MachNode, we need to recurse 1664 // the tree checking for these cases: 1665 // (1) Child is an instruction - 1666 // Build the instruction (recursively), add it as an edge. 1667 // Build a simple operand (register) to hold the result of the instruction. 1668 // (2) Child is an interior part of an instruction - 1669 // Skip over it (do nothing) 1670 // (3) Child is the start of a operand - 1671 // Build the operand, place it inside the instruction 1672 // Call ReduceOper. 1673 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) { 1674 assert( rule >= NUM_OPERANDS, "called with operand rule" ); 1675 1676 MachNode* shared_node = find_shared_node(s->_leaf, rule); 1677 if (shared_node != NULL) { 1678 return shared_node; 1679 } 1680 1681 // Build the object to represent this state & prepare for recursive calls 1682 MachNode *mach = s->MachNodeGenerator(rule); 1683 mach->_opnds[0] = s->MachOperGenerator(_reduceOp[rule]); 1684 assert( mach->_opnds[0] != NULL, "Missing result operand" ); 1685 Node *leaf = s->_leaf; 1686 // Check for instruction or instruction chain rule 1687 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) { 1688 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf), 1689 "duplicating node that's already been matched"); 1690 // Instruction 1691 mach->add_req( leaf->in(0) ); // Set initial control 1692 // Reduce interior of complex instruction 1693 ReduceInst_Interior( s, rule, mem, mach, 1 ); 1694 } else { 1695 // Instruction chain rules are data-dependent on their inputs 1696 mach->add_req(0); // Set initial control to none 1697 ReduceInst_Chain_Rule( s, rule, mem, mach ); 1698 } 1699 1700 // If a Memory was used, insert a Memory edge 1701 if( mem != (Node*)1 ) { 1702 mach->ins_req(MemNode::Memory,mem); 1703 #ifdef ASSERT 1704 // Verify adr type after matching memory operation 1705 const MachOper* oper = mach->memory_operand(); 1706 if (oper != NULL && oper != (MachOper*)-1) { 1707 // It has a unique memory operand. Find corresponding ideal mem node. 1708 Node* m = NULL; 1709 if (leaf->is_Mem()) { 1710 m = leaf; 1711 } else { 1712 m = _mem_node; 1713 assert(m != NULL && m->is_Mem(), "expecting memory node"); 1714 } 1715 const Type* mach_at = mach->adr_type(); 1716 // DecodeN node consumed by an address may have different type 1717 // then its input. Don't compare types for such case. 1718 if (m->adr_type() != mach_at && 1719 (m->in(MemNode::Address)->is_DecodeNarrowPtr() || 1720 m->in(MemNode::Address)->is_AddP() && 1721 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr() || 1722 m->in(MemNode::Address)->is_AddP() && 1723 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() && 1724 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr())) { 1725 mach_at = m->adr_type(); 1726 } 1727 if (m->adr_type() != mach_at) { 1728 m->dump(); 1729 tty->print_cr("mach:"); 1730 mach->dump(1); 1731 } 1732 assert(m->adr_type() == mach_at, "matcher should not change adr type"); 1733 } 1734 #endif 1735 } 1736 1737 // If the _leaf is an AddP, insert the base edge 1738 if (leaf->is_AddP()) { 1739 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base)); 1740 } 1741 1742 uint number_of_projections_prior = number_of_projections(); 1743 1744 // Perform any 1-to-many expansions required 1745 MachNode *ex = mach->Expand(s, _projection_list, mem); 1746 if (ex != mach) { 1747 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match"); 1748 if( ex->in(1)->is_Con() ) 1749 ex->in(1)->set_req(0, C->root()); 1750 // Remove old node from the graph 1751 for( uint i=0; i<mach->req(); i++ ) { 1752 mach->set_req(i,NULL); 1753 } 1754 #ifdef ASSERT 1755 _new2old_map.map(ex->_idx, s->_leaf); 1756 #endif 1757 } 1758 1759 // PhaseChaitin::fixup_spills will sometimes generate spill code 1760 // via the matcher. By the time, nodes have been wired into the CFG, 1761 // and any further nodes generated by expand rules will be left hanging 1762 // in space, and will not get emitted as output code. Catch this. 1763 // Also, catch any new register allocation constraints ("projections") 1764 // generated belatedly during spill code generation. 1765 if (_allocation_started) { 1766 guarantee(ex == mach, "no expand rules during spill generation"); 1767 guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation"); 1768 } 1769 1770 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) { 1771 // Record the con for sharing 1772 _shared_nodes.map(leaf->_idx, ex); 1773 } 1774 1775 return ex; 1776 } 1777 1778 void Matcher::handle_precedence_edges(Node* n, MachNode *mach) { 1779 for (uint i = n->req(); i < n->len(); i++) { 1780 if (n->in(i) != NULL) { 1781 mach->add_prec(n->in(i)); 1782 } 1783 } 1784 } 1785 1786 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) { 1787 // 'op' is what I am expecting to receive 1788 int op = _leftOp[rule]; 1789 // Operand type to catch childs result 1790 // This is what my child will give me. 1791 int opnd_class_instance = s->_rule[op]; 1792 // Choose between operand class or not. 1793 // This is what I will receive. 1794 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op; 1795 // New rule for child. Chase operand classes to get the actual rule. 1796 int newrule = s->_rule[catch_op]; 1797 1798 if( newrule < NUM_OPERANDS ) { 1799 // Chain from operand or operand class, may be output of shared node 1800 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS, 1801 "Bad AD file: Instruction chain rule must chain from operand"); 1802 // Insert operand into array of operands for this instruction 1803 mach->_opnds[1] = s->MachOperGenerator(opnd_class_instance); 1804 1805 ReduceOper( s, newrule, mem, mach ); 1806 } else { 1807 // Chain from the result of an instruction 1808 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand"); 1809 mach->_opnds[1] = s->MachOperGenerator(_reduceOp[catch_op]); 1810 Node *mem1 = (Node*)1; 1811 debug_only(Node *save_mem_node = _mem_node;) 1812 mach->add_req( ReduceInst(s, newrule, mem1) ); 1813 debug_only(_mem_node = save_mem_node;) 1814 } 1815 return; 1816 } 1817 1818 1819 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) { 1820 handle_precedence_edges(s->_leaf, mach); 1821 1822 if( s->_leaf->is_Load() ) { 1823 Node *mem2 = s->_leaf->in(MemNode::Memory); 1824 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" ); 1825 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;) 1826 mem = mem2; 1827 } 1828 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) { 1829 if( mach->in(0) == NULL ) 1830 mach->set_req(0, s->_leaf->in(0)); 1831 } 1832 1833 // Now recursively walk the state tree & add operand list. 1834 for( uint i=0; i<2; i++ ) { // binary tree 1835 State *newstate = s->_kids[i]; 1836 if( newstate == NULL ) break; // Might only have 1 child 1837 // 'op' is what I am expecting to receive 1838 int op; 1839 if( i == 0 ) { 1840 op = _leftOp[rule]; 1841 } else { 1842 op = _rightOp[rule]; 1843 } 1844 // Operand type to catch childs result 1845 // This is what my child will give me. 1846 int opnd_class_instance = newstate->_rule[op]; 1847 // Choose between operand class or not. 1848 // This is what I will receive. 1849 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op; 1850 // New rule for child. Chase operand classes to get the actual rule. 1851 int newrule = newstate->_rule[catch_op]; 1852 1853 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction? 1854 // Operand/operandClass 1855 // Insert operand into array of operands for this instruction 1856 mach->_opnds[num_opnds++] = newstate->MachOperGenerator(opnd_class_instance); 1857 ReduceOper( newstate, newrule, mem, mach ); 1858 1859 } else { // Child is internal operand or new instruction 1860 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction? 1861 // internal operand --> call ReduceInst_Interior 1862 // Interior of complex instruction. Do nothing but recurse. 1863 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds ); 1864 } else { 1865 // instruction --> call build operand( ) to catch result 1866 // --> ReduceInst( newrule ) 1867 mach->_opnds[num_opnds++] = s->MachOperGenerator(_reduceOp[catch_op]); 1868 Node *mem1 = (Node*)1; 1869 debug_only(Node *save_mem_node = _mem_node;) 1870 mach->add_req( ReduceInst( newstate, newrule, mem1 ) ); 1871 debug_only(_mem_node = save_mem_node;) 1872 } 1873 } 1874 assert( mach->_opnds[num_opnds-1], "" ); 1875 } 1876 return num_opnds; 1877 } 1878 1879 // This routine walks the interior of possible complex operands. 1880 // At each point we check our children in the match tree: 1881 // (1) No children - 1882 // We are a leaf; add _leaf field as an input to the MachNode 1883 // (2) Child is an internal operand - 1884 // Skip over it ( do nothing ) 1885 // (3) Child is an instruction - 1886 // Call ReduceInst recursively and 1887 // and instruction as an input to the MachNode 1888 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) { 1889 assert( rule < _LAST_MACH_OPER, "called with operand rule" ); 1890 State *kid = s->_kids[0]; 1891 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" ); 1892 1893 // Leaf? And not subsumed? 1894 if( kid == NULL && !_swallowed[rule] ) { 1895 mach->add_req( s->_leaf ); // Add leaf pointer 1896 return; // Bail out 1897 } 1898 1899 if( s->_leaf->is_Load() ) { 1900 assert( mem == (Node*)1, "multiple Memories being matched at once?" ); 1901 mem = s->_leaf->in(MemNode::Memory); 1902 debug_only(_mem_node = s->_leaf;) 1903 } 1904 1905 handle_precedence_edges(s->_leaf, mach); 1906 1907 if( s->_leaf->in(0) && s->_leaf->req() > 1) { 1908 if( !mach->in(0) ) 1909 mach->set_req(0,s->_leaf->in(0)); 1910 else { 1911 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" ); 1912 } 1913 } 1914 1915 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree 1916 int newrule; 1917 if( i == 0) 1918 newrule = kid->_rule[_leftOp[rule]]; 1919 else 1920 newrule = kid->_rule[_rightOp[rule]]; 1921 1922 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction? 1923 // Internal operand; recurse but do nothing else 1924 ReduceOper( kid, newrule, mem, mach ); 1925 1926 } else { // Child is a new instruction 1927 // Reduce the instruction, and add a direct pointer from this 1928 // machine instruction to the newly reduced one. 1929 Node *mem1 = (Node*)1; 1930 debug_only(Node *save_mem_node = _mem_node;) 1931 mach->add_req( ReduceInst( kid, newrule, mem1 ) ); 1932 debug_only(_mem_node = save_mem_node;) 1933 } 1934 } 1935 } 1936 1937 1938 // ------------------------------------------------------------------------- 1939 // Java-Java calling convention 1940 // (what you use when Java calls Java) 1941 1942 //------------------------------find_receiver---------------------------------- 1943 // For a given signature, return the OptoReg for parameter 0. 1944 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) { 1945 VMRegPair regs; 1946 BasicType sig_bt = T_OBJECT; 1947 calling_convention(&sig_bt, ®s, 1, is_outgoing); 1948 // Return argument 0 register. In the LP64 build pointers 1949 // take 2 registers, but the VM wants only the 'main' name. 1950 return OptoReg::as_OptoReg(regs.first()); 1951 } 1952 1953 // This function identifies sub-graphs in which a 'load' node is 1954 // input to two different nodes, and such that it can be matched 1955 // with BMI instructions like blsi, blsr, etc. 1956 // Example : for b = -a[i] & a[i] can be matched to blsi r32, m32. 1957 // The graph is (AndL (SubL Con0 LoadL*) LoadL*), where LoadL* 1958 // refers to the same node. 1959 #ifdef X86 1960 // Match the generic fused operations pattern (op1 (op2 Con{ConType} mop) mop) 1961 // This is a temporary solution until we make DAGs expressible in ADL. 1962 template<typename ConType> 1963 class FusedPatternMatcher { 1964 Node* _op1_node; 1965 Node* _mop_node; 1966 int _con_op; 1967 1968 static int match_next(Node* n, int next_op, int next_op_idx) { 1969 if (n->in(1) == NULL || n->in(2) == NULL) { 1970 return -1; 1971 } 1972 1973 if (next_op_idx == -1) { // n is commutative, try rotations 1974 if (n->in(1)->Opcode() == next_op) { 1975 return 1; 1976 } else if (n->in(2)->Opcode() == next_op) { 1977 return 2; 1978 } 1979 } else { 1980 assert(next_op_idx > 0 && next_op_idx <= 2, "Bad argument index"); 1981 if (n->in(next_op_idx)->Opcode() == next_op) { 1982 return next_op_idx; 1983 } 1984 } 1985 return -1; 1986 } 1987 public: 1988 FusedPatternMatcher(Node* op1_node, Node *mop_node, int con_op) : 1989 _op1_node(op1_node), _mop_node(mop_node), _con_op(con_op) { } 1990 1991 bool match(int op1, int op1_op2_idx, // op1 and the index of the op1->op2 edge, -1 if op1 is commutative 1992 int op2, int op2_con_idx, // op2 and the index of the op2->con edge, -1 if op2 is commutative 1993 typename ConType::NativeType con_value) { 1994 if (_op1_node->Opcode() != op1) { 1995 return false; 1996 } 1997 if (_mop_node->outcnt() > 2) { 1998 return false; 1999 } 2000 op1_op2_idx = match_next(_op1_node, op2, op1_op2_idx); 2001 if (op1_op2_idx == -1) { 2002 return false; 2003 } 2004 // Memory operation must be the other edge 2005 int op1_mop_idx = (op1_op2_idx & 1) + 1; 2006 2007 // Check that the mop node is really what we want 2008 if (_op1_node->in(op1_mop_idx) == _mop_node) { 2009 Node *op2_node = _op1_node->in(op1_op2_idx); 2010 if (op2_node->outcnt() > 1) { 2011 return false; 2012 } 2013 assert(op2_node->Opcode() == op2, "Should be"); 2014 op2_con_idx = match_next(op2_node, _con_op, op2_con_idx); 2015 if (op2_con_idx == -1) { 2016 return false; 2017 } 2018 // Memory operation must be the other edge 2019 int op2_mop_idx = (op2_con_idx & 1) + 1; 2020 // Check that the memory operation is the same node 2021 if (op2_node->in(op2_mop_idx) == _mop_node) { 2022 // Now check the constant 2023 const Type* con_type = op2_node->in(op2_con_idx)->bottom_type(); 2024 if (con_type != Type::TOP && ConType::as_self(con_type)->get_con() == con_value) { 2025 return true; 2026 } 2027 } 2028 } 2029 return false; 2030 } 2031 }; 2032 2033 2034 bool Matcher::is_bmi_pattern(Node *n, Node *m) { 2035 if (n != NULL && m != NULL) { 2036 if (m->Opcode() == Op_LoadI) { 2037 FusedPatternMatcher<TypeInt> bmii(n, m, Op_ConI); 2038 return bmii.match(Op_AndI, -1, Op_SubI, 1, 0) || 2039 bmii.match(Op_AndI, -1, Op_AddI, -1, -1) || 2040 bmii.match(Op_XorI, -1, Op_AddI, -1, -1); 2041 } else if (m->Opcode() == Op_LoadL) { 2042 FusedPatternMatcher<TypeLong> bmil(n, m, Op_ConL); 2043 return bmil.match(Op_AndL, -1, Op_SubL, 1, 0) || 2044 bmil.match(Op_AndL, -1, Op_AddL, -1, -1) || 2045 bmil.match(Op_XorL, -1, Op_AddL, -1, -1); 2046 } 2047 } 2048 return false; 2049 } 2050 #endif // X86 2051 2052 // A method-klass-holder may be passed in the inline_cache_reg 2053 // and then expanded into the inline_cache_reg and a method_oop register 2054 // defined in ad_<arch>.cpp 2055 2056 // Check for shift by small constant as well 2057 static bool clone_shift(Node* shift, Matcher* matcher, MStack& mstack, VectorSet& address_visited) { 2058 if (shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() && 2059 shift->in(2)->get_int() <= 3 && 2060 // Are there other uses besides address expressions? 2061 !matcher->is_visited(shift)) { 2062 address_visited.set(shift->_idx); // Flag as address_visited 2063 mstack.push(shift->in(2), Visit); 2064 Node *conv = shift->in(1); 2065 #ifdef _LP64 2066 // Allow Matcher to match the rule which bypass 2067 // ConvI2L operation for an array index on LP64 2068 // if the index value is positive. 2069 if (conv->Opcode() == Op_ConvI2L && 2070 conv->as_Type()->type()->is_long()->_lo >= 0 && 2071 // Are there other uses besides address expressions? 2072 !matcher->is_visited(conv)) { 2073 address_visited.set(conv->_idx); // Flag as address_visited 2074 mstack.push(conv->in(1), Pre_Visit); 2075 } else 2076 #endif 2077 mstack.push(conv, Pre_Visit); 2078 return true; 2079 } 2080 return false; 2081 } 2082 2083 2084 //------------------------------find_shared------------------------------------ 2085 // Set bits if Node is shared or otherwise a root 2086 void Matcher::find_shared( Node *n ) { 2087 // Allocate stack of size C->live_nodes() * 2 to avoid frequent realloc 2088 MStack mstack(C->live_nodes() * 2); 2089 // Mark nodes as address_visited if they are inputs to an address expression 2090 VectorSet address_visited(Thread::current()->resource_area()); 2091 mstack.push(n, Visit); // Don't need to pre-visit root node 2092 while (mstack.is_nonempty()) { 2093 n = mstack.node(); // Leave node on stack 2094 Node_State nstate = mstack.state(); 2095 uint nop = n->Opcode(); 2096 if (nstate == Pre_Visit) { 2097 if (address_visited.test(n->_idx)) { // Visited in address already? 2098 // Flag as visited and shared now. 2099 set_visited(n); 2100 } 2101 if (is_visited(n)) { // Visited already? 2102 // Node is shared and has no reason to clone. Flag it as shared. 2103 // This causes it to match into a register for the sharing. 2104 set_shared(n); // Flag as shared and 2105 mstack.pop(); // remove node from stack 2106 continue; 2107 } 2108 nstate = Visit; // Not already visited; so visit now 2109 } 2110 if (nstate == Visit) { 2111 mstack.set_state(Post_Visit); 2112 set_visited(n); // Flag as visited now 2113 bool mem_op = false; 2114 2115 switch( nop ) { // Handle some opcodes special 2116 case Op_Phi: // Treat Phis as shared roots 2117 case Op_Parm: 2118 case Op_Proj: // All handled specially during matching 2119 case Op_SafePointScalarObject: 2120 set_shared(n); 2121 set_dontcare(n); 2122 break; 2123 case Op_If: 2124 case Op_CountedLoopEnd: 2125 mstack.set_state(Alt_Post_Visit); // Alternative way 2126 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps 2127 // with matching cmp/branch in 1 instruction. The Matcher needs the 2128 // Bool and CmpX side-by-side, because it can only get at constants 2129 // that are at the leaves of Match trees, and the Bool's condition acts 2130 // as a constant here. 2131 mstack.push(n->in(1), Visit); // Clone the Bool 2132 mstack.push(n->in(0), Pre_Visit); // Visit control input 2133 continue; // while (mstack.is_nonempty()) 2134 case Op_ConvI2D: // These forms efficiently match with a prior 2135 case Op_ConvI2F: // Load but not a following Store 2136 if( n->in(1)->is_Load() && // Prior load 2137 n->outcnt() == 1 && // Not already shared 2138 n->unique_out()->is_Store() ) // Following store 2139 set_shared(n); // Force it to be a root 2140 break; 2141 case Op_ReverseBytesI: 2142 case Op_ReverseBytesL: 2143 if( n->in(1)->is_Load() && // Prior load 2144 n->outcnt() == 1 ) // Not already shared 2145 set_shared(n); // Force it to be a root 2146 break; 2147 case Op_BoxLock: // Cant match until we get stack-regs in ADLC 2148 case Op_IfFalse: 2149 case Op_IfTrue: 2150 case Op_MachProj: 2151 case Op_MergeMem: 2152 case Op_Catch: 2153 case Op_CatchProj: 2154 case Op_CProj: 2155 case Op_JumpProj: 2156 case Op_JProj: 2157 case Op_NeverBranch: 2158 set_dontcare(n); 2159 break; 2160 case Op_Jump: 2161 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared) 2162 mstack.push(n->in(0), Pre_Visit); // Visit Control input 2163 continue; // while (mstack.is_nonempty()) 2164 case Op_StrComp: 2165 case Op_StrEquals: 2166 case Op_StrIndexOf: 2167 case Op_StrIndexOfChar: 2168 case Op_AryEq: 2169 case Op_HasNegatives: 2170 case Op_StrInflatedCopy: 2171 case Op_StrCompressedCopy: 2172 case Op_EncodeISOArray: 2173 set_shared(n); // Force result into register (it will be anyways) 2174 break; 2175 case Op_ConP: { // Convert pointers above the centerline to NUL 2176 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2177 const TypePtr* tp = tn->type()->is_ptr(); 2178 if (tp->_ptr == TypePtr::AnyNull) { 2179 tn->set_type(TypePtr::NULL_PTR); 2180 } 2181 break; 2182 } 2183 case Op_ConN: { // Convert narrow pointers above the centerline to NUL 2184 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2185 const TypePtr* tp = tn->type()->make_ptr(); 2186 if (tp && tp->_ptr == TypePtr::AnyNull) { 2187 tn->set_type(TypeNarrowOop::NULL_PTR); 2188 } 2189 break; 2190 } 2191 case Op_Binary: // These are introduced in the Post_Visit state. 2192 ShouldNotReachHere(); 2193 break; 2194 case Op_ClearArray: 2195 case Op_SafePoint: 2196 mem_op = true; 2197 break; 2198 default: 2199 if( n->is_Store() ) { 2200 // Do match stores, despite no ideal reg 2201 mem_op = true; 2202 break; 2203 } 2204 if( n->is_Mem() ) { // Loads and LoadStores 2205 mem_op = true; 2206 // Loads must be root of match tree due to prior load conflict 2207 if( C->subsume_loads() == false ) 2208 set_shared(n); 2209 } 2210 // Fall into default case 2211 if( !n->ideal_reg() ) 2212 set_dontcare(n); // Unmatchable Nodes 2213 } // end_switch 2214 2215 for(int i = n->req() - 1; i >= 0; --i) { // For my children 2216 Node *m = n->in(i); // Get ith input 2217 if (m == NULL) continue; // Ignore NULLs 2218 uint mop = m->Opcode(); 2219 2220 // Must clone all producers of flags, or we will not match correctly. 2221 // Suppose a compare setting int-flags is shared (e.g., a switch-tree) 2222 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags 2223 // are also there, so we may match a float-branch to int-flags and 2224 // expect the allocator to haul the flags from the int-side to the 2225 // fp-side. No can do. 2226 if( _must_clone[mop] ) { 2227 mstack.push(m, Visit); 2228 continue; // for(int i = ...) 2229 } 2230 2231 if( mop == Op_AddP && m->in(AddPNode::Base)->is_DecodeNarrowPtr()) { 2232 // Bases used in addresses must be shared but since 2233 // they are shared through a DecodeN they may appear 2234 // to have a single use so force sharing here. 2235 set_shared(m->in(AddPNode::Base)->in(1)); 2236 } 2237 2238 // if 'n' and 'm' are part of a graph for BMI instruction, clone this node. 2239 #ifdef X86 2240 if (UseBMI1Instructions && is_bmi_pattern(n, m)) { 2241 mstack.push(m, Visit); 2242 continue; 2243 } 2244 #endif 2245 2246 // Clone addressing expressions as they are "free" in memory access instructions 2247 if (mem_op && i == MemNode::Address && mop == Op_AddP && 2248 // When there are other uses besides address expressions 2249 // put it on stack and mark as shared. 2250 !is_visited(m)) { 2251 // Some inputs for address expression are not put on stack 2252 // to avoid marking them as shared and forcing them into register 2253 // if they are used only in address expressions. 2254 // But they should be marked as shared if there are other uses 2255 // besides address expressions. 2256 2257 Node *off = m->in(AddPNode::Offset); 2258 if (off->is_Con()) { 2259 address_visited.test_set(m->_idx); // Flag as address_visited 2260 Node *adr = m->in(AddPNode::Address); 2261 2262 // Intel, ARM and friends can handle 2 adds in addressing mode 2263 if( clone_shift_expressions && adr->is_AddP() && 2264 // AtomicAdd is not an addressing expression. 2265 // Cheap to find it by looking for screwy base. 2266 !adr->in(AddPNode::Base)->is_top() && 2267 // Are there other uses besides address expressions? 2268 !is_visited(adr) ) { 2269 address_visited.set(adr->_idx); // Flag as address_visited 2270 Node *shift = adr->in(AddPNode::Offset); 2271 if (!clone_shift(shift, this, mstack, address_visited)) { 2272 mstack.push(shift, Pre_Visit); 2273 } 2274 mstack.push(adr->in(AddPNode::Address), Pre_Visit); 2275 mstack.push(adr->in(AddPNode::Base), Pre_Visit); 2276 } else { // Sparc, Alpha, PPC and friends 2277 mstack.push(adr, Pre_Visit); 2278 } 2279 2280 // Clone X+offset as it also folds into most addressing expressions 2281 mstack.push(off, Visit); 2282 mstack.push(m->in(AddPNode::Base), Pre_Visit); 2283 continue; // for(int i = ...) 2284 } else if (clone_shift_expressions && 2285 clone_shift(off, this, mstack, address_visited)) { 2286 address_visited.test_set(m->_idx); // Flag as address_visited 2287 mstack.push(m->in(AddPNode::Address), Pre_Visit); 2288 mstack.push(m->in(AddPNode::Base), Pre_Visit); 2289 continue; 2290 } // if( off->is_Con() ) 2291 } // if( mem_op && 2292 mstack.push(m, Pre_Visit); 2293 } // for(int i = ...) 2294 } 2295 else if (nstate == Alt_Post_Visit) { 2296 mstack.pop(); // Remove node from stack 2297 // We cannot remove the Cmp input from the Bool here, as the Bool may be 2298 // shared and all users of the Bool need to move the Cmp in parallel. 2299 // This leaves both the Bool and the If pointing at the Cmp. To 2300 // prevent the Matcher from trying to Match the Cmp along both paths 2301 // BoolNode::match_edge always returns a zero. 2302 2303 // We reorder the Op_If in a pre-order manner, so we can visit without 2304 // accidentally sharing the Cmp (the Bool and the If make 2 users). 2305 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool 2306 } 2307 else if (nstate == Post_Visit) { 2308 mstack.pop(); // Remove node from stack 2309 2310 // Now hack a few special opcodes 2311 switch( n->Opcode() ) { // Handle some opcodes special 2312 case Op_StorePConditional: 2313 case Op_StoreIConditional: 2314 case Op_StoreLConditional: 2315 case Op_CompareAndExchangeI: 2316 case Op_CompareAndExchangeL: 2317 case Op_CompareAndExchangeP: 2318 case Op_CompareAndExchangeN: 2319 case Op_WeakCompareAndSwapI: 2320 case Op_WeakCompareAndSwapL: 2321 case Op_WeakCompareAndSwapP: 2322 case Op_WeakCompareAndSwapN: 2323 case Op_CompareAndSwapI: 2324 case Op_CompareAndSwapL: 2325 case Op_CompareAndSwapP: 2326 case Op_CompareAndSwapN: { // Convert trinary to binary-tree 2327 Node *newval = n->in(MemNode::ValueIn ); 2328 Node *oldval = n->in(LoadStoreConditionalNode::ExpectedIn); 2329 Node *pair = new BinaryNode( oldval, newval ); 2330 n->set_req(MemNode::ValueIn,pair); 2331 n->del_req(LoadStoreConditionalNode::ExpectedIn); 2332 break; 2333 } 2334 case Op_CMoveD: // Convert trinary to binary-tree 2335 case Op_CMoveF: 2336 case Op_CMoveI: 2337 case Op_CMoveL: 2338 case Op_CMoveN: 2339 case Op_CMoveP: 2340 case Op_CMoveVD: { 2341 // Restructure into a binary tree for Matching. It's possible that 2342 // we could move this code up next to the graph reshaping for IfNodes 2343 // or vice-versa, but I do not want to debug this for Ladybird. 2344 // 10/2/2000 CNC. 2345 Node *pair1 = new BinaryNode(n->in(1),n->in(1)->in(1)); 2346 n->set_req(1,pair1); 2347 Node *pair2 = new BinaryNode(n->in(2),n->in(3)); 2348 n->set_req(2,pair2); 2349 n->del_req(3); 2350 break; 2351 } 2352 case Op_LoopLimit: { 2353 Node *pair1 = new BinaryNode(n->in(1),n->in(2)); 2354 n->set_req(1,pair1); 2355 n->set_req(2,n->in(3)); 2356 n->del_req(3); 2357 break; 2358 } 2359 case Op_StrEquals: 2360 case Op_StrIndexOfChar: { 2361 Node *pair1 = new BinaryNode(n->in(2),n->in(3)); 2362 n->set_req(2,pair1); 2363 n->set_req(3,n->in(4)); 2364 n->del_req(4); 2365 break; 2366 } 2367 case Op_StrComp: 2368 case Op_StrIndexOf: { 2369 Node *pair1 = new BinaryNode(n->in(2),n->in(3)); 2370 n->set_req(2,pair1); 2371 Node *pair2 = new BinaryNode(n->in(4),n->in(5)); 2372 n->set_req(3,pair2); 2373 n->del_req(5); 2374 n->del_req(4); 2375 break; 2376 } 2377 case Op_StrCompressedCopy: 2378 case Op_StrInflatedCopy: 2379 case Op_EncodeISOArray: { 2380 // Restructure into a binary tree for Matching. 2381 Node* pair = new BinaryNode(n->in(3), n->in(4)); 2382 n->set_req(3, pair); 2383 n->del_req(4); 2384 break; 2385 } 2386 default: 2387 break; 2388 } 2389 } 2390 else { 2391 ShouldNotReachHere(); 2392 } 2393 } // end of while (mstack.is_nonempty()) 2394 } 2395 2396 #ifdef ASSERT 2397 // machine-independent root to machine-dependent root 2398 void Matcher::dump_old2new_map() { 2399 _old2new_map.dump(); 2400 } 2401 #endif 2402 2403 //---------------------------collect_null_checks------------------------------- 2404 // Find null checks in the ideal graph; write a machine-specific node for 2405 // it. Used by later implicit-null-check handling. Actually collects 2406 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal 2407 // value being tested. 2408 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) { 2409 Node *iff = proj->in(0); 2410 if( iff->Opcode() == Op_If ) { 2411 // During matching If's have Bool & Cmp side-by-side 2412 BoolNode *b = iff->in(1)->as_Bool(); 2413 Node *cmp = iff->in(2); 2414 int opc = cmp->Opcode(); 2415 if (opc != Op_CmpP && opc != Op_CmpN) return; 2416 2417 const Type* ct = cmp->in(2)->bottom_type(); 2418 if (ct == TypePtr::NULL_PTR || 2419 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) { 2420 2421 bool push_it = false; 2422 if( proj->Opcode() == Op_IfTrue ) { 2423 #ifndef PRODUCT 2424 extern int all_null_checks_found; 2425 all_null_checks_found++; 2426 #endif 2427 if( b->_test._test == BoolTest::ne ) { 2428 push_it = true; 2429 } 2430 } else { 2431 assert( proj->Opcode() == Op_IfFalse, "" ); 2432 if( b->_test._test == BoolTest::eq ) { 2433 push_it = true; 2434 } 2435 } 2436 if( push_it ) { 2437 _null_check_tests.push(proj); 2438 Node* val = cmp->in(1); 2439 #ifdef _LP64 2440 if (val->bottom_type()->isa_narrowoop() && 2441 !Matcher::narrow_oop_use_complex_address()) { 2442 // 2443 // Look for DecodeN node which should be pinned to orig_proj. 2444 // On platforms (Sparc) which can not handle 2 adds 2445 // in addressing mode we have to keep a DecodeN node and 2446 // use it to do implicit NULL check in address. 2447 // 2448 // DecodeN node was pinned to non-null path (orig_proj) during 2449 // CastPP transformation in final_graph_reshaping_impl(). 2450 // 2451 uint cnt = orig_proj->outcnt(); 2452 for (uint i = 0; i < orig_proj->outcnt(); i++) { 2453 Node* d = orig_proj->raw_out(i); 2454 if (d->is_DecodeN() && d->in(1) == val) { 2455 val = d; 2456 val->set_req(0, NULL); // Unpin now. 2457 // Mark this as special case to distinguish from 2458 // a regular case: CmpP(DecodeN, NULL). 2459 val = (Node*)(((intptr_t)val) | 1); 2460 break; 2461 } 2462 } 2463 } 2464 #endif 2465 _null_check_tests.push(val); 2466 } 2467 } 2468 } 2469 } 2470 2471 //---------------------------validate_null_checks------------------------------ 2472 // Its possible that the value being NULL checked is not the root of a match 2473 // tree. If so, I cannot use the value in an implicit null check. 2474 void Matcher::validate_null_checks( ) { 2475 uint cnt = _null_check_tests.size(); 2476 for( uint i=0; i < cnt; i+=2 ) { 2477 Node *test = _null_check_tests[i]; 2478 Node *val = _null_check_tests[i+1]; 2479 bool is_decoden = ((intptr_t)val) & 1; 2480 val = (Node*)(((intptr_t)val) & ~1); 2481 if (has_new_node(val)) { 2482 Node* new_val = new_node(val); 2483 if (is_decoden) { 2484 assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity"); 2485 // Note: new_val may have a control edge if 2486 // the original ideal node DecodeN was matched before 2487 // it was unpinned in Matcher::collect_null_checks(). 2488 // Unpin the mach node and mark it. 2489 new_val->set_req(0, NULL); 2490 new_val = (Node*)(((intptr_t)new_val) | 1); 2491 } 2492 // Is a match-tree root, so replace with the matched value 2493 _null_check_tests.map(i+1, new_val); 2494 } else { 2495 // Yank from candidate list 2496 _null_check_tests.map(i+1,_null_check_tests[--cnt]); 2497 _null_check_tests.map(i,_null_check_tests[--cnt]); 2498 _null_check_tests.pop(); 2499 _null_check_tests.pop(); 2500 i-=2; 2501 } 2502 } 2503 } 2504 2505 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or 2506 // atomic instruction acting as a store_load barrier without any 2507 // intervening volatile load, and thus we don't need a barrier here. 2508 // We retain the Node to act as a compiler ordering barrier. 2509 bool Matcher::post_store_load_barrier(const Node* vmb) { 2510 Compile* C = Compile::current(); 2511 assert(vmb->is_MemBar(), ""); 2512 assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, ""); 2513 const MemBarNode* membar = vmb->as_MemBar(); 2514 2515 // Get the Ideal Proj node, ctrl, that can be used to iterate forward 2516 Node* ctrl = NULL; 2517 for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) { 2518 Node* p = membar->fast_out(i); 2519 assert(p->is_Proj(), "only projections here"); 2520 if ((p->as_Proj()->_con == TypeFunc::Control) && 2521 !C->node_arena()->contains(p)) { // Unmatched old-space only 2522 ctrl = p; 2523 break; 2524 } 2525 } 2526 assert((ctrl != NULL), "missing control projection"); 2527 2528 for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) { 2529 Node *x = ctrl->fast_out(j); 2530 int xop = x->Opcode(); 2531 2532 // We don't need current barrier if we see another or a lock 2533 // before seeing volatile load. 2534 // 2535 // Op_Fastunlock previously appeared in the Op_* list below. 2536 // With the advent of 1-0 lock operations we're no longer guaranteed 2537 // that a monitor exit operation contains a serializing instruction. 2538 2539 if (xop == Op_MemBarVolatile || 2540 xop == Op_CompareAndExchangeI || 2541 xop == Op_CompareAndExchangeL || 2542 xop == Op_CompareAndExchangeP || 2543 xop == Op_CompareAndExchangeN || 2544 xop == Op_WeakCompareAndSwapL || 2545 xop == Op_WeakCompareAndSwapP || 2546 xop == Op_WeakCompareAndSwapN || 2547 xop == Op_WeakCompareAndSwapI || 2548 xop == Op_CompareAndSwapL || 2549 xop == Op_CompareAndSwapP || 2550 xop == Op_CompareAndSwapN || 2551 xop == Op_CompareAndSwapI) { 2552 return true; 2553 } 2554 2555 // Op_FastLock previously appeared in the Op_* list above. 2556 // With biased locking we're no longer guaranteed that a monitor 2557 // enter operation contains a serializing instruction. 2558 if ((xop == Op_FastLock) && !UseBiasedLocking) { 2559 return true; 2560 } 2561 2562 if (x->is_MemBar()) { 2563 // We must retain this membar if there is an upcoming volatile 2564 // load, which will be followed by acquire membar. 2565 if (xop == Op_MemBarAcquire || xop == Op_LoadFence) { 2566 return false; 2567 } else { 2568 // For other kinds of barriers, check by pretending we 2569 // are them, and seeing if we can be removed. 2570 return post_store_load_barrier(x->as_MemBar()); 2571 } 2572 } 2573 2574 // probably not necessary to check for these 2575 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) { 2576 return false; 2577 } 2578 } 2579 return false; 2580 } 2581 2582 // Check whether node n is a branch to an uncommon trap that we could 2583 // optimize as test with very high branch costs in case of going to 2584 // the uncommon trap. The code must be able to be recompiled to use 2585 // a cheaper test. 2586 bool Matcher::branches_to_uncommon_trap(const Node *n) { 2587 // Don't do it for natives, adapters, or runtime stubs 2588 Compile *C = Compile::current(); 2589 if (!C->is_method_compilation()) return false; 2590 2591 assert(n->is_If(), "You should only call this on if nodes."); 2592 IfNode *ifn = n->as_If(); 2593 2594 Node *ifFalse = NULL; 2595 for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) { 2596 if (ifn->fast_out(i)->is_IfFalse()) { 2597 ifFalse = ifn->fast_out(i); 2598 break; 2599 } 2600 } 2601 assert(ifFalse, "An If should have an ifFalse. Graph is broken."); 2602 2603 Node *reg = ifFalse; 2604 int cnt = 4; // We must protect against cycles. Limit to 4 iterations. 2605 // Alternatively use visited set? Seems too expensive. 2606 while (reg != NULL && cnt > 0) { 2607 CallNode *call = NULL; 2608 RegionNode *nxt_reg = NULL; 2609 for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) { 2610 Node *o = reg->fast_out(i); 2611 if (o->is_Call()) { 2612 call = o->as_Call(); 2613 } 2614 if (o->is_Region()) { 2615 nxt_reg = o->as_Region(); 2616 } 2617 } 2618 2619 if (call && 2620 call->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 2621 const Type* trtype = call->in(TypeFunc::Parms)->bottom_type(); 2622 if (trtype->isa_int() && trtype->is_int()->is_con()) { 2623 jint tr_con = trtype->is_int()->get_con(); 2624 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 2625 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 2626 assert((int)reason < (int)BitsPerInt, "recode bit map"); 2627 2628 if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason) 2629 && action != Deoptimization::Action_none) { 2630 // This uncommon trap is sure to recompile, eventually. 2631 // When that happens, C->too_many_traps will prevent 2632 // this transformation from happening again. 2633 return true; 2634 } 2635 } 2636 } 2637 2638 reg = nxt_reg; 2639 cnt--; 2640 } 2641 2642 return false; 2643 } 2644 2645 //============================================================================= 2646 //---------------------------State--------------------------------------------- 2647 State::State(void) { 2648 #ifdef ASSERT 2649 _id = 0; 2650 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2651 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2652 //memset(_cost, -1, sizeof(_cost)); 2653 //memset(_rule, -1, sizeof(_rule)); 2654 #endif 2655 memset(_valid, 0, sizeof(_valid)); 2656 } 2657 2658 #ifdef ASSERT 2659 State::~State() { 2660 _id = 99; 2661 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2662 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2663 memset(_cost, -3, sizeof(_cost)); 2664 memset(_rule, -3, sizeof(_rule)); 2665 } 2666 #endif 2667 2668 #ifndef PRODUCT 2669 //---------------------------dump---------------------------------------------- 2670 void State::dump() { 2671 tty->print("\n"); 2672 dump(0); 2673 } 2674 2675 void State::dump(int depth) { 2676 for( int j = 0; j < depth; j++ ) 2677 tty->print(" "); 2678 tty->print("--N: "); 2679 _leaf->dump(); 2680 uint i; 2681 for( i = 0; i < _LAST_MACH_OPER; i++ ) 2682 // Check for valid entry 2683 if( valid(i) ) { 2684 for( int j = 0; j < depth; j++ ) 2685 tty->print(" "); 2686 assert(_cost[i] != max_juint, "cost must be a valid value"); 2687 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule"); 2688 tty->print_cr("%s %d %s", 2689 ruleName[i], _cost[i], ruleName[_rule[i]] ); 2690 } 2691 tty->cr(); 2692 2693 for( i=0; i<2; i++ ) 2694 if( _kids[i] ) 2695 _kids[i]->dump(depth+1); 2696 } 2697 #endif