1 /* 2 * Copyright (c) 1997, 2017, 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 #include "utilities/align.hpp" 44 #if INCLUDE_ZGC 45 #include "gc/z/zBarrierSetRuntime.hpp" 46 #endif // INCLUDE_ZGC 47 48 OptoReg::Name OptoReg::c_frame_pointer; 49 50 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf]; 51 RegMask Matcher::mreg2regmask[_last_Mach_Reg]; 52 RegMask Matcher::STACK_ONLY_mask; 53 RegMask Matcher::c_frame_ptr_mask; 54 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE; 55 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE; 56 57 //---------------------------Matcher------------------------------------------- 58 Matcher::Matcher() 59 : PhaseTransform( Phase::Ins_Select ), 60 _states_arena(Chunk::medium_size, mtCompiler), 61 _visited(&_states_arena), 62 _shared(&_states_arena), 63 _dontcare(&_states_arena), 64 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp), 65 _swallowed(swallowed), 66 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE), 67 _end_inst_chain_rule(_END_INST_CHAIN_RULE), 68 _must_clone(must_clone), 69 _shared_nodes(C->comp_arena()), 70 #ifdef ASSERT 71 _old2new_map(C->comp_arena()), 72 _new2old_map(C->comp_arena()), 73 #endif 74 _allocation_started(false), 75 _ruleName(ruleName), 76 _register_save_policy(register_save_policy), 77 _c_reg_save_policy(c_reg_save_policy), 78 _register_save_type(register_save_type) { 79 C->set_matcher(this); 80 81 idealreg2spillmask [Op_RegI] = NULL; 82 idealreg2spillmask [Op_RegN] = NULL; 83 idealreg2spillmask [Op_RegL] = NULL; 84 idealreg2spillmask [Op_RegF] = NULL; 85 idealreg2spillmask [Op_RegD] = NULL; 86 idealreg2spillmask [Op_RegP] = NULL; 87 idealreg2spillmask [Op_VecS] = NULL; 88 idealreg2spillmask [Op_VecD] = NULL; 89 idealreg2spillmask [Op_VecX] = NULL; 90 idealreg2spillmask [Op_VecY] = NULL; 91 idealreg2spillmask [Op_VecZ] = NULL; 92 idealreg2spillmask [Op_RegFlags] = NULL; 93 94 idealreg2debugmask [Op_RegI] = NULL; 95 idealreg2debugmask [Op_RegN] = NULL; 96 idealreg2debugmask [Op_RegL] = NULL; 97 idealreg2debugmask [Op_RegF] = NULL; 98 idealreg2debugmask [Op_RegD] = NULL; 99 idealreg2debugmask [Op_RegP] = NULL; 100 idealreg2debugmask [Op_VecS] = NULL; 101 idealreg2debugmask [Op_VecD] = NULL; 102 idealreg2debugmask [Op_VecX] = NULL; 103 idealreg2debugmask [Op_VecY] = NULL; 104 idealreg2debugmask [Op_VecZ] = NULL; 105 idealreg2debugmask [Op_RegFlags] = NULL; 106 107 idealreg2mhdebugmask[Op_RegI] = NULL; 108 idealreg2mhdebugmask[Op_RegN] = NULL; 109 idealreg2mhdebugmask[Op_RegL] = NULL; 110 idealreg2mhdebugmask[Op_RegF] = NULL; 111 idealreg2mhdebugmask[Op_RegD] = NULL; 112 idealreg2mhdebugmask[Op_RegP] = NULL; 113 idealreg2mhdebugmask[Op_VecS] = NULL; 114 idealreg2mhdebugmask[Op_VecD] = NULL; 115 idealreg2mhdebugmask[Op_VecX] = NULL; 116 idealreg2mhdebugmask[Op_VecY] = NULL; 117 idealreg2mhdebugmask[Op_VecZ] = NULL; 118 idealreg2mhdebugmask[Op_RegFlags] = NULL; 119 120 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node 121 } 122 123 //------------------------------warp_incoming_stk_arg------------------------ 124 // This warps a VMReg into an OptoReg::Name 125 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) { 126 OptoReg::Name warped; 127 if( reg->is_stack() ) { // Stack slot argument? 128 warped = OptoReg::add(_old_SP, reg->reg2stack() ); 129 warped = OptoReg::add(warped, C->out_preserve_stack_slots()); 130 if( warped >= _in_arg_limit ) 131 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen 132 if (!RegMask::can_represent_arg(warped)) { 133 // the compiler cannot represent this method's calling sequence 134 C->record_method_not_compilable("unsupported incoming calling sequence"); 135 return OptoReg::Bad; 136 } 137 return warped; 138 } 139 return OptoReg::as_OptoReg(reg); 140 } 141 142 //---------------------------compute_old_SP------------------------------------ 143 OptoReg::Name Compile::compute_old_SP() { 144 int fixed = fixed_slots(); 145 int preserve = in_preserve_stack_slots(); 146 return OptoReg::stack2reg(align_up(fixed + preserve, (int)Matcher::stack_alignment_in_slots())); 147 } 148 149 150 151 #ifdef ASSERT 152 void Matcher::verify_new_nodes_only(Node* xroot) { 153 // Make sure that the new graph only references new nodes 154 ResourceMark rm; 155 Unique_Node_List worklist; 156 VectorSet visited(Thread::current()->resource_area()); 157 worklist.push(xroot); 158 while (worklist.size() > 0) { 159 Node* n = worklist.pop(); 160 visited <<= n->_idx; 161 assert(C->node_arena()->contains(n), "dead node"); 162 for (uint j = 0; j < n->req(); j++) { 163 Node* in = n->in(j); 164 if (in != NULL) { 165 assert(C->node_arena()->contains(in), "dead node"); 166 if (!visited.test(in->_idx)) { 167 worklist.push(in); 168 } 169 } 170 } 171 } 172 } 173 #endif 174 175 // Array of RegMask, one per returned values (value type instances can 176 // be returned as multiple return values, one per field) 177 RegMask* Matcher::return_values_mask(const TypeTuple *range) { 178 uint cnt = range->cnt() - TypeFunc::Parms; 179 if (cnt == 0) { 180 return NULL; 181 } 182 RegMask* mask = NEW_RESOURCE_ARRAY(RegMask, cnt); 183 184 if (!ValueTypeReturnedAsFields) { 185 // Get ideal-register return type 186 uint ireg = range->field_at(TypeFunc::Parms)->ideal_reg(); 187 // Get machine return register 188 OptoRegPair regs = return_value(ireg, false); 189 190 // And mask for same 191 mask[0].Clear(); 192 mask[0].Insert(regs.first()); 193 if (OptoReg::is_valid(regs.second())) { 194 mask[0].Insert(regs.second()); 195 } 196 } else { 197 BasicType* sig_bt = NEW_RESOURCE_ARRAY(BasicType, cnt); 198 VMRegPair* vm_parm_regs = NEW_RESOURCE_ARRAY(VMRegPair, cnt); 199 200 for (uint i = 0; i < cnt; i++) { 201 sig_bt[i] = range->field_at(i+TypeFunc::Parms)->basic_type(); 202 } 203 204 int regs = SharedRuntime::java_return_convention(sig_bt, vm_parm_regs, cnt); 205 assert(regs > 0, "should have been tested during graph construction"); 206 for (uint i = 0; i < cnt; i++) { 207 mask[i].Clear(); 208 209 OptoReg::Name reg1 = OptoReg::as_OptoReg(vm_parm_regs[i].first()); 210 if (OptoReg::is_valid(reg1)) { 211 mask[i].Insert(reg1); 212 } 213 OptoReg::Name reg2 = OptoReg::as_OptoReg(vm_parm_regs[i].second()); 214 if (OptoReg::is_valid(reg2)) { 215 mask[i].Insert(reg2); 216 } 217 } 218 } 219 return mask; 220 } 221 222 //---------------------------match--------------------------------------------- 223 void Matcher::match( ) { 224 if( MaxLabelRootDepth < 100 ) { // Too small? 225 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum"); 226 MaxLabelRootDepth = 100; 227 } 228 // One-time initialization of some register masks. 229 init_spill_mask( C->root()->in(1) ); 230 _return_addr_mask = return_addr(); 231 #ifdef _LP64 232 // Pointers take 2 slots in 64-bit land 233 _return_addr_mask.Insert(OptoReg::add(return_addr(),1)); 234 #endif 235 236 // Map Java-signature return types into return register-value 237 // machine registers. 238 const TypeTuple *range = C->tf()->range_cc(); 239 _return_values_mask = return_values_mask(range); 240 241 // --------------- 242 // Frame Layout 243 244 // Need the method signature to determine the incoming argument types, 245 // because the types determine which registers the incoming arguments are 246 // in, and this affects the matched code. 247 const TypeTuple *domain = C->tf()->domain_cc(); 248 uint argcnt = domain->cnt() - TypeFunc::Parms; 249 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 250 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 251 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt ); 252 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt ); 253 uint i; 254 for( i = 0; i<argcnt; i++ ) { 255 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 256 } 257 258 // Pass array of ideal registers and length to USER code (from the AD file) 259 // that will convert this to an array of register numbers. 260 const StartNode *start = C->start(); 261 start->calling_convention( sig_bt, vm_parm_regs, argcnt ); 262 #ifdef ASSERT 263 // Sanity check users' calling convention. Real handy while trying to 264 // get the initial port correct. 265 { for (uint i = 0; i<argcnt; i++) { 266 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 267 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" ); 268 _parm_regs[i].set_bad(); 269 continue; 270 } 271 VMReg parm_reg = vm_parm_regs[i].first(); 272 assert(parm_reg->is_valid(), "invalid arg?"); 273 if (parm_reg->is_reg()) { 274 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg); 275 assert(can_be_java_arg(opto_parm_reg) || 276 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) || 277 opto_parm_reg == inline_cache_reg(), 278 "parameters in register must be preserved by runtime stubs"); 279 } 280 for (uint j = 0; j < i; j++) { 281 assert(parm_reg != vm_parm_regs[j].first(), 282 "calling conv. must produce distinct regs"); 283 } 284 } 285 } 286 #endif 287 288 // Do some initial frame layout. 289 290 // Compute the old incoming SP (may be called FP) as 291 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2. 292 _old_SP = C->compute_old_SP(); 293 assert( is_even(_old_SP), "must be even" ); 294 295 // Compute highest incoming stack argument as 296 // _old_SP + out_preserve_stack_slots + incoming argument size. 297 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 298 assert( is_even(_in_arg_limit), "out_preserve must be even" ); 299 for( i = 0; i < argcnt; i++ ) { 300 // Permit args to have no register 301 _calling_convention_mask[i].Clear(); 302 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 303 continue; 304 } 305 // calling_convention returns stack arguments as a count of 306 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to 307 // the allocators point of view, taking into account all the 308 // preserve area, locks & pad2. 309 310 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first()); 311 if( OptoReg::is_valid(reg1)) 312 _calling_convention_mask[i].Insert(reg1); 313 314 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second()); 315 if( OptoReg::is_valid(reg2)) 316 _calling_convention_mask[i].Insert(reg2); 317 318 // Saved biased stack-slot register number 319 _parm_regs[i].set_pair(reg2, reg1); 320 } 321 322 // Finally, make sure the incoming arguments take up an even number of 323 // words, in case the arguments or locals need to contain doubleword stack 324 // slots. The rest of the system assumes that stack slot pairs (in 325 // particular, in the spill area) which look aligned will in fact be 326 // aligned relative to the stack pointer in the target machine. Double 327 // stack slots will always be allocated aligned. 328 _new_SP = OptoReg::Name(align_up(_in_arg_limit, (int)RegMask::SlotsPerLong)); 329 330 // Compute highest outgoing stack argument as 331 // _new_SP + out_preserve_stack_slots + max(outgoing argument size). 332 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 333 assert( is_even(_out_arg_limit), "out_preserve must be even" ); 334 335 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) { 336 // the compiler cannot represent this method's calling sequence 337 C->record_method_not_compilable("must be able to represent all call arguments in reg mask"); 338 } 339 340 if (C->failing()) return; // bailed out on incoming arg failure 341 342 // --------------- 343 // Collect roots of matcher trees. Every node for which 344 // _shared[_idx] is cleared is guaranteed to not be shared, and thus 345 // can be a valid interior of some tree. 346 find_shared( C->root() ); 347 find_shared( C->top() ); 348 349 C->print_method(PHASE_BEFORE_MATCHING); 350 351 // Create new ideal node ConP #NULL even if it does exist in old space 352 // to avoid false sharing if the corresponding mach node is not used. 353 // The corresponding mach node is only used in rare cases for derived 354 // pointers. 355 Node* new_ideal_null = ConNode::make(TypePtr::NULL_PTR); 356 357 // Swap out to old-space; emptying new-space 358 Arena *old = C->node_arena()->move_contents(C->old_arena()); 359 360 // Save debug and profile information for nodes in old space: 361 _old_node_note_array = C->node_note_array(); 362 if (_old_node_note_array != NULL) { 363 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*> 364 (C->comp_arena(), _old_node_note_array->length(), 365 0, NULL)); 366 } 367 368 // Pre-size the new_node table to avoid the need for range checks. 369 grow_new_node_array(C->unique()); 370 371 // Reset node counter so MachNodes start with _idx at 0 372 int live_nodes = C->live_nodes(); 373 C->set_unique(0); 374 C->reset_dead_node_list(); 375 376 // Recursively match trees from old space into new space. 377 // Correct leaves of new-space Nodes; they point to old-space. 378 _visited.Clear(); // Clear visit bits for xform call 379 C->set_cached_top_node(xform( C->top(), live_nodes )); 380 if (!C->failing()) { 381 Node* xroot = xform( C->root(), 1 ); 382 if (xroot == NULL) { 383 Matcher::soft_match_failure(); // recursive matching process failed 384 C->record_method_not_compilable("instruction match failed"); 385 } else { 386 // During matching shared constants were attached to C->root() 387 // because xroot wasn't available yet, so transfer the uses to 388 // the xroot. 389 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { 390 Node* n = C->root()->fast_out(j); 391 if (C->node_arena()->contains(n)) { 392 assert(n->in(0) == C->root(), "should be control user"); 393 n->set_req(0, xroot); 394 --j; 395 --jmax; 396 } 397 } 398 399 // Generate new mach node for ConP #NULL 400 assert(new_ideal_null != NULL, "sanity"); 401 _mach_null = match_tree(new_ideal_null); 402 // Don't set control, it will confuse GCM since there are no uses. 403 // The control will be set when this node is used first time 404 // in find_base_for_derived(). 405 assert(_mach_null != NULL, ""); 406 407 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL); 408 409 #ifdef ASSERT 410 verify_new_nodes_only(xroot); 411 #endif 412 } 413 } 414 if (C->top() == NULL || C->root() == NULL) { 415 C->record_method_not_compilable("graph lost"); // %%% cannot happen? 416 } 417 if (C->failing()) { 418 // delete old; 419 old->destruct_contents(); 420 return; 421 } 422 assert( C->top(), "" ); 423 assert( C->root(), "" ); 424 validate_null_checks(); 425 426 // Now smoke old-space 427 NOT_DEBUG( old->destruct_contents() ); 428 429 // ------------------------ 430 // Set up save-on-entry registers 431 Fixup_Save_On_Entry( ); 432 } 433 434 435 //------------------------------Fixup_Save_On_Entry---------------------------- 436 // The stated purpose of this routine is to take care of save-on-entry 437 // registers. However, the overall goal of the Match phase is to convert into 438 // machine-specific instructions which have RegMasks to guide allocation. 439 // So what this procedure really does is put a valid RegMask on each input 440 // to the machine-specific variations of all Return, TailCall and Halt 441 // instructions. It also adds edgs to define the save-on-entry values (and of 442 // course gives them a mask). 443 444 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) { 445 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size ); 446 // Do all the pre-defined register masks 447 rms[TypeFunc::Control ] = RegMask::Empty; 448 rms[TypeFunc::I_O ] = RegMask::Empty; 449 rms[TypeFunc::Memory ] = RegMask::Empty; 450 rms[TypeFunc::ReturnAdr] = ret_adr; 451 rms[TypeFunc::FramePtr ] = fp; 452 return rms; 453 } 454 455 //---------------------------init_first_stack_mask----------------------------- 456 // Create the initial stack mask used by values spilling to the stack. 457 // Disallow any debug info in outgoing argument areas by setting the 458 // initial mask accordingly. 459 void Matcher::init_first_stack_mask() { 460 461 // Allocate storage for spill masks as masks for the appropriate load type. 462 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * (3*6+5)); 463 464 idealreg2spillmask [Op_RegN] = &rms[0]; 465 idealreg2spillmask [Op_RegI] = &rms[1]; 466 idealreg2spillmask [Op_RegL] = &rms[2]; 467 idealreg2spillmask [Op_RegF] = &rms[3]; 468 idealreg2spillmask [Op_RegD] = &rms[4]; 469 idealreg2spillmask [Op_RegP] = &rms[5]; 470 471 idealreg2debugmask [Op_RegN] = &rms[6]; 472 idealreg2debugmask [Op_RegI] = &rms[7]; 473 idealreg2debugmask [Op_RegL] = &rms[8]; 474 idealreg2debugmask [Op_RegF] = &rms[9]; 475 idealreg2debugmask [Op_RegD] = &rms[10]; 476 idealreg2debugmask [Op_RegP] = &rms[11]; 477 478 idealreg2mhdebugmask[Op_RegN] = &rms[12]; 479 idealreg2mhdebugmask[Op_RegI] = &rms[13]; 480 idealreg2mhdebugmask[Op_RegL] = &rms[14]; 481 idealreg2mhdebugmask[Op_RegF] = &rms[15]; 482 idealreg2mhdebugmask[Op_RegD] = &rms[16]; 483 idealreg2mhdebugmask[Op_RegP] = &rms[17]; 484 485 idealreg2spillmask [Op_VecS] = &rms[18]; 486 idealreg2spillmask [Op_VecD] = &rms[19]; 487 idealreg2spillmask [Op_VecX] = &rms[20]; 488 idealreg2spillmask [Op_VecY] = &rms[21]; 489 idealreg2spillmask [Op_VecZ] = &rms[22]; 490 491 OptoReg::Name i; 492 493 // At first, start with the empty mask 494 C->FIRST_STACK_mask().Clear(); 495 496 // Check if method has a reserved entry in the argument stack area that 497 // should not be used for spilling because it holds the return address. 498 OptoRegPair res_entry; 499 if (C->needs_stack_repair()) { 500 int res_idx = C->get_res_entry()._offset; 501 res_entry = _parm_regs[res_idx]; 502 } 503 504 // Add in the incoming argument area 505 OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 506 for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) { 507 if (i == res_entry.first() || i == res_entry.second()) { 508 continue; // Skip reserved slot to avoid spilling 509 } 510 C->FIRST_STACK_mask().Insert(i); 511 } 512 // Add in all bits past the outgoing argument area 513 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)), 514 "must be able to represent all call arguments in reg mask"); 515 OptoReg::Name init = _out_arg_limit; 516 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) { 517 C->FIRST_STACK_mask().Insert(i); 518 } 519 // Finally, set the "infinite stack" bit. 520 C->FIRST_STACK_mask().set_AllStack(); 521 522 // Make spill masks. Registers for their class, plus FIRST_STACK_mask. 523 RegMask aligned_stack_mask = C->FIRST_STACK_mask(); 524 // Keep spill masks aligned. 525 aligned_stack_mask.clear_to_pairs(); 526 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 527 528 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP]; 529 #ifdef _LP64 530 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN]; 531 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask()); 532 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask); 533 #else 534 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask()); 535 #endif 536 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI]; 537 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask()); 538 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL]; 539 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask); 540 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF]; 541 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask()); 542 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD]; 543 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask); 544 545 if (Matcher::vector_size_supported(T_BYTE,4)) { 546 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS]; 547 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask()); 548 } 549 if (Matcher::vector_size_supported(T_FLOAT,2)) { 550 // For VecD we need dual alignment and 8 bytes (2 slots) for spills. 551 // RA guarantees such alignment since it is needed for Double and Long values. 552 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD]; 553 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask); 554 } 555 if (Matcher::vector_size_supported(T_FLOAT,4)) { 556 // For VecX we need quadro alignment and 16 bytes (4 slots) for spills. 557 // 558 // RA can use input arguments stack slots for spills but until RA 559 // we don't know frame size and offset of input arg stack slots. 560 // 561 // Exclude last input arg stack slots to avoid spilling vectors there 562 // otherwise vector spills could stomp over stack slots in caller frame. 563 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 564 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) { 565 aligned_stack_mask.Remove(in); 566 in = OptoReg::add(in, -1); 567 } 568 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX); 569 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 570 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX]; 571 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask); 572 } 573 if (Matcher::vector_size_supported(T_FLOAT,8)) { 574 // For VecY we need octo alignment and 32 bytes (8 slots) for spills. 575 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 576 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) { 577 aligned_stack_mask.Remove(in); 578 in = OptoReg::add(in, -1); 579 } 580 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY); 581 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 582 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY]; 583 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask); 584 } 585 if (Matcher::vector_size_supported(T_FLOAT,16)) { 586 // For VecZ we need enough alignment and 64 bytes (16 slots) for spills. 587 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 588 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecZ); k++) { 589 aligned_stack_mask.Remove(in); 590 in = OptoReg::add(in, -1); 591 } 592 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecZ); 593 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 594 *idealreg2spillmask[Op_VecZ] = *idealreg2regmask[Op_VecZ]; 595 idealreg2spillmask[Op_VecZ]->OR(aligned_stack_mask); 596 } 597 if (UseFPUForSpilling) { 598 // This mask logic assumes that the spill operations are 599 // symmetric and that the registers involved are the same size. 600 // On sparc for instance we may have to use 64 bit moves will 601 // kill 2 registers when used with F0-F31. 602 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]); 603 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]); 604 #ifdef _LP64 605 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]); 606 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 607 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 608 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]); 609 #else 610 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]); 611 #ifdef ARM 612 // ARM has support for moving 64bit values between a pair of 613 // integer registers and a double register 614 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 615 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 616 #endif 617 #endif 618 } 619 620 // Make up debug masks. Any spill slot plus callee-save registers. 621 // Caller-save registers are assumed to be trashable by the various 622 // inline-cache fixup routines. 623 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN]; 624 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI]; 625 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL]; 626 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF]; 627 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD]; 628 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP]; 629 630 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN]; 631 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI]; 632 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL]; 633 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF]; 634 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD]; 635 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP]; 636 637 // Prevent stub compilations from attempting to reference 638 // callee-saved registers from debug info 639 bool exclude_soe = !Compile::current()->is_method_compilation(); 640 641 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 642 // registers the caller has to save do not work 643 if( _register_save_policy[i] == 'C' || 644 _register_save_policy[i] == 'A' || 645 (_register_save_policy[i] == 'E' && exclude_soe) ) { 646 idealreg2debugmask [Op_RegN]->Remove(i); 647 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call 648 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug 649 idealreg2debugmask [Op_RegF]->Remove(i); // masks 650 idealreg2debugmask [Op_RegD]->Remove(i); 651 idealreg2debugmask [Op_RegP]->Remove(i); 652 653 idealreg2mhdebugmask[Op_RegN]->Remove(i); 654 idealreg2mhdebugmask[Op_RegI]->Remove(i); 655 idealreg2mhdebugmask[Op_RegL]->Remove(i); 656 idealreg2mhdebugmask[Op_RegF]->Remove(i); 657 idealreg2mhdebugmask[Op_RegD]->Remove(i); 658 idealreg2mhdebugmask[Op_RegP]->Remove(i); 659 } 660 } 661 662 // Subtract the register we use to save the SP for MethodHandle 663 // invokes to from the debug mask. 664 const RegMask save_mask = method_handle_invoke_SP_save_mask(); 665 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask); 666 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask); 667 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask); 668 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask); 669 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask); 670 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask); 671 } 672 673 //---------------------------is_save_on_entry---------------------------------- 674 bool Matcher::is_save_on_entry( int reg ) { 675 return 676 _register_save_policy[reg] == 'E' || 677 _register_save_policy[reg] == 'A' || // Save-on-entry register? 678 // Also save argument registers in the trampolining stubs 679 (C->save_argument_registers() && is_spillable_arg(reg)); 680 } 681 682 //---------------------------Fixup_Save_On_Entry------------------------------- 683 void Matcher::Fixup_Save_On_Entry( ) { 684 init_first_stack_mask(); 685 686 Node *root = C->root(); // Short name for root 687 // Count number of save-on-entry registers. 688 uint soe_cnt = number_of_saved_registers(); 689 uint i; 690 691 // Find the procedure Start Node 692 StartNode *start = C->start(); 693 assert( start, "Expect a start node" ); 694 695 // Save argument registers in the trampolining stubs 696 if( C->save_argument_registers() ) 697 for( i = 0; i < _last_Mach_Reg; i++ ) 698 if( is_spillable_arg(i) ) 699 soe_cnt++; 700 701 // Input RegMask array shared by all Returns. 702 // The type for doubles and longs has a count of 2, but 703 // there is only 1 returned value 704 uint ret_edge_cnt = C->tf()->range_cc()->cnt(); 705 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 706 for (i = TypeFunc::Parms; i < ret_edge_cnt; i++) { 707 ret_rms[i] = _return_values_mask[i-TypeFunc::Parms]; 708 } 709 710 // Input RegMask array shared by all Rethrows. 711 uint reth_edge_cnt = TypeFunc::Parms+1; 712 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 713 // Rethrow takes exception oop only, but in the argument 0 slot. 714 OptoReg::Name reg = find_receiver(false); 715 if (reg >= 0) { 716 reth_rms[TypeFunc::Parms] = mreg2regmask[reg]; 717 #ifdef _LP64 718 // Need two slots for ptrs in 64-bit land 719 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(reg), 1)); 720 #endif 721 } 722 723 // Input RegMask array shared by all TailCalls 724 uint tail_call_edge_cnt = TypeFunc::Parms+2; 725 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 726 727 // Input RegMask array shared by all TailJumps 728 uint tail_jump_edge_cnt = TypeFunc::Parms+2; 729 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 730 731 // TailCalls have 2 returned values (target & moop), whose masks come 732 // from the usual MachNode/MachOper mechanism. Find a sample 733 // TailCall to extract these masks and put the correct masks into 734 // the tail_call_rms array. 735 for( i=1; i < root->req(); i++ ) { 736 MachReturnNode *m = root->in(i)->as_MachReturn(); 737 if( m->ideal_Opcode() == Op_TailCall ) { 738 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 739 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 740 break; 741 } 742 } 743 744 // TailJumps have 2 returned values (target & ex_oop), whose masks come 745 // from the usual MachNode/MachOper mechanism. Find a sample 746 // TailJump to extract these masks and put the correct masks into 747 // the tail_jump_rms array. 748 for( i=1; i < root->req(); i++ ) { 749 MachReturnNode *m = root->in(i)->as_MachReturn(); 750 if( m->ideal_Opcode() == Op_TailJump ) { 751 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 752 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 753 break; 754 } 755 } 756 757 // Input RegMask array shared by all Halts 758 uint halt_edge_cnt = TypeFunc::Parms; 759 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 760 761 // Capture the return input masks into each exit flavor 762 for( i=1; i < root->req(); i++ ) { 763 MachReturnNode *exit = root->in(i)->as_MachReturn(); 764 switch( exit->ideal_Opcode() ) { 765 case Op_Return : exit->_in_rms = ret_rms; break; 766 case Op_Rethrow : exit->_in_rms = reth_rms; break; 767 case Op_TailCall : exit->_in_rms = tail_call_rms; break; 768 case Op_TailJump : exit->_in_rms = tail_jump_rms; break; 769 case Op_Halt : exit->_in_rms = halt_rms; break; 770 default : ShouldNotReachHere(); 771 } 772 } 773 774 // Next unused projection number from Start. 775 int proj_cnt = C->tf()->domain_cc()->cnt(); 776 777 // Do all the save-on-entry registers. Make projections from Start for 778 // them, and give them a use at the exit points. To the allocator, they 779 // look like incoming register arguments. 780 for( i = 0; i < _last_Mach_Reg; i++ ) { 781 if( is_save_on_entry(i) ) { 782 783 // Add the save-on-entry to the mask array 784 ret_rms [ ret_edge_cnt] = mreg2regmask[i]; 785 reth_rms [ reth_edge_cnt] = mreg2regmask[i]; 786 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i]; 787 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i]; 788 // Halts need the SOE registers, but only in the stack as debug info. 789 // A just-prior uncommon-trap or deoptimization will use the SOE regs. 790 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]]; 791 792 Node *mproj; 793 794 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's 795 // into a single RegD. 796 if( (i&1) == 0 && 797 _register_save_type[i ] == Op_RegF && 798 _register_save_type[i+1] == Op_RegF && 799 is_save_on_entry(i+1) ) { 800 // Add other bit for double 801 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 802 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 803 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 804 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 805 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 806 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD ); 807 proj_cnt += 2; // Skip 2 for doubles 808 } 809 else if( (i&1) == 1 && // Else check for high half of double 810 _register_save_type[i-1] == Op_RegF && 811 _register_save_type[i ] == Op_RegF && 812 is_save_on_entry(i-1) ) { 813 ret_rms [ ret_edge_cnt] = RegMask::Empty; 814 reth_rms [ reth_edge_cnt] = RegMask::Empty; 815 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 816 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 817 halt_rms [ halt_edge_cnt] = RegMask::Empty; 818 mproj = C->top(); 819 } 820 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's 821 // into a single RegL. 822 else if( (i&1) == 0 && 823 _register_save_type[i ] == Op_RegI && 824 _register_save_type[i+1] == Op_RegI && 825 is_save_on_entry(i+1) ) { 826 // Add other bit for long 827 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 828 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 829 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 830 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 831 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 832 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL ); 833 proj_cnt += 2; // Skip 2 for longs 834 } 835 else if( (i&1) == 1 && // Else check for high half of long 836 _register_save_type[i-1] == Op_RegI && 837 _register_save_type[i ] == Op_RegI && 838 is_save_on_entry(i-1) ) { 839 ret_rms [ ret_edge_cnt] = RegMask::Empty; 840 reth_rms [ reth_edge_cnt] = RegMask::Empty; 841 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 842 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 843 halt_rms [ halt_edge_cnt] = RegMask::Empty; 844 mproj = C->top(); 845 } else { 846 // Make a projection for it off the Start 847 mproj = new MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] ); 848 } 849 850 ret_edge_cnt ++; 851 reth_edge_cnt ++; 852 tail_call_edge_cnt ++; 853 tail_jump_edge_cnt ++; 854 halt_edge_cnt ++; 855 856 // Add a use of the SOE register to all exit paths 857 for( uint j=1; j < root->req(); j++ ) 858 root->in(j)->add_req(mproj); 859 } // End of if a save-on-entry register 860 } // End of for all machine registers 861 } 862 863 //------------------------------init_spill_mask-------------------------------- 864 void Matcher::init_spill_mask( Node *ret ) { 865 if( idealreg2regmask[Op_RegI] ) return; // One time only init 866 867 OptoReg::c_frame_pointer = c_frame_pointer(); 868 c_frame_ptr_mask = c_frame_pointer(); 869 #ifdef _LP64 870 // pointers are twice as big 871 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1)); 872 #endif 873 874 // Start at OptoReg::stack0() 875 STACK_ONLY_mask.Clear(); 876 OptoReg::Name init = OptoReg::stack2reg(0); 877 // STACK_ONLY_mask is all stack bits 878 OptoReg::Name i; 879 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) 880 STACK_ONLY_mask.Insert(i); 881 // Also set the "infinite stack" bit. 882 STACK_ONLY_mask.set_AllStack(); 883 884 // Copy the register names over into the shared world 885 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 886 // SharedInfo::regName[i] = regName[i]; 887 // Handy RegMasks per machine register 888 mreg2regmask[i].Insert(i); 889 } 890 891 // Grab the Frame Pointer 892 Node *fp = ret->in(TypeFunc::FramePtr); 893 Node *mem = ret->in(TypeFunc::Memory); 894 const TypePtr* atp = TypePtr::BOTTOM; 895 // Share frame pointer while making spill ops 896 set_shared(fp); 897 898 // Compute generic short-offset Loads 899 #ifdef _LP64 900 MachNode *spillCP = match_tree(new LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 901 #endif 902 MachNode *spillI = match_tree(new LoadINode(NULL,mem,fp,atp,TypeInt::INT,MemNode::unordered)); 903 MachNode *spillL = match_tree(new LoadLNode(NULL,mem,fp,atp,TypeLong::LONG,MemNode::unordered, LoadNode::DependsOnlyOnTest, false)); 904 MachNode *spillF = match_tree(new LoadFNode(NULL,mem,fp,atp,Type::FLOAT,MemNode::unordered)); 905 MachNode *spillD = match_tree(new LoadDNode(NULL,mem,fp,atp,Type::DOUBLE,MemNode::unordered)); 906 MachNode *spillP = match_tree(new LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 907 assert(spillI != NULL && spillL != NULL && spillF != NULL && 908 spillD != NULL && spillP != NULL, ""); 909 // Get the ADLC notion of the right regmask, for each basic type. 910 #ifdef _LP64 911 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask(); 912 #endif 913 idealreg2regmask[Op_RegI] = &spillI->out_RegMask(); 914 idealreg2regmask[Op_RegL] = &spillL->out_RegMask(); 915 idealreg2regmask[Op_RegF] = &spillF->out_RegMask(); 916 idealreg2regmask[Op_RegD] = &spillD->out_RegMask(); 917 idealreg2regmask[Op_RegP] = &spillP->out_RegMask(); 918 919 // Vector regmasks. 920 if (Matcher::vector_size_supported(T_BYTE,4)) { 921 TypeVect::VECTS = TypeVect::make(T_BYTE, 4); 922 MachNode *spillVectS = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTS)); 923 idealreg2regmask[Op_VecS] = &spillVectS->out_RegMask(); 924 } 925 if (Matcher::vector_size_supported(T_FLOAT,2)) { 926 MachNode *spillVectD = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTD)); 927 idealreg2regmask[Op_VecD] = &spillVectD->out_RegMask(); 928 } 929 if (Matcher::vector_size_supported(T_FLOAT,4)) { 930 MachNode *spillVectX = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTX)); 931 idealreg2regmask[Op_VecX] = &spillVectX->out_RegMask(); 932 } 933 if (Matcher::vector_size_supported(T_FLOAT,8)) { 934 MachNode *spillVectY = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTY)); 935 idealreg2regmask[Op_VecY] = &spillVectY->out_RegMask(); 936 } 937 if (Matcher::vector_size_supported(T_FLOAT,16)) { 938 MachNode *spillVectZ = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTZ)); 939 idealreg2regmask[Op_VecZ] = &spillVectZ->out_RegMask(); 940 } 941 } 942 943 #ifdef ASSERT 944 static void match_alias_type(Compile* C, Node* n, Node* m) { 945 if (!VerifyAliases) return; // do not go looking for trouble by default 946 const TypePtr* nat = n->adr_type(); 947 const TypePtr* mat = m->adr_type(); 948 int nidx = C->get_alias_index(nat); 949 int midx = C->get_alias_index(mat); 950 // Detune the assert for cases like (AndI 0xFF (LoadB p)). 951 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) { 952 for (uint i = 1; i < n->req(); i++) { 953 Node* n1 = n->in(i); 954 const TypePtr* n1at = n1->adr_type(); 955 if (n1at != NULL) { 956 nat = n1at; 957 nidx = C->get_alias_index(n1at); 958 } 959 } 960 } 961 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases: 962 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) { 963 switch (n->Opcode()) { 964 case Op_PrefetchAllocation: 965 nidx = Compile::AliasIdxRaw; 966 nat = TypeRawPtr::BOTTOM; 967 break; 968 } 969 } 970 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) { 971 switch (n->Opcode()) { 972 case Op_ClearArray: 973 midx = Compile::AliasIdxRaw; 974 mat = TypeRawPtr::BOTTOM; 975 break; 976 } 977 } 978 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) { 979 switch (n->Opcode()) { 980 case Op_Return: 981 case Op_Rethrow: 982 case Op_Halt: 983 case Op_TailCall: 984 case Op_TailJump: 985 nidx = Compile::AliasIdxBot; 986 nat = TypePtr::BOTTOM; 987 break; 988 } 989 } 990 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) { 991 switch (n->Opcode()) { 992 case Op_StrComp: 993 case Op_StrEquals: 994 case Op_StrIndexOf: 995 case Op_StrIndexOfChar: 996 case Op_AryEq: 997 case Op_HasNegatives: 998 case Op_MemBarVolatile: 999 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type? 1000 case Op_StrInflatedCopy: 1001 case Op_StrCompressedCopy: 1002 case Op_OnSpinWait: 1003 case Op_EncodeISOArray: 1004 nidx = Compile::AliasIdxTop; 1005 nat = NULL; 1006 break; 1007 } 1008 } 1009 if (nidx != midx) { 1010 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) { 1011 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx); 1012 n->dump(); 1013 m->dump(); 1014 } 1015 assert(C->subsume_loads() && C->must_alias(nat, midx), 1016 "must not lose alias info when matching"); 1017 } 1018 } 1019 #endif 1020 1021 //------------------------------xform------------------------------------------ 1022 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine 1023 // Node in new-space. Given a new-space Node, recursively walk his children. 1024 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; } 1025 Node *Matcher::xform( Node *n, int max_stack ) { 1026 // Use one stack to keep both: child's node/state and parent's node/index 1027 MStack mstack(max_stack * 2 * 2); // usually: C->live_nodes() * 2 * 2 1028 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root 1029 while (mstack.is_nonempty()) { 1030 C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions"); 1031 if (C->failing()) return NULL; 1032 n = mstack.node(); // Leave node on stack 1033 Node_State nstate = mstack.state(); 1034 if (nstate == Visit) { 1035 mstack.set_state(Post_Visit); 1036 Node *oldn = n; 1037 // Old-space or new-space check 1038 if (!C->node_arena()->contains(n)) { 1039 // Old space! 1040 Node* m; 1041 if (has_new_node(n)) { // Not yet Label/Reduced 1042 m = new_node(n); 1043 } else { 1044 if (!is_dontcare(n)) { // Matcher can match this guy 1045 // Calls match special. They match alone with no children. 1046 // Their children, the incoming arguments, match normally. 1047 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n); 1048 if (C->failing()) return NULL; 1049 if (m == NULL) { Matcher::soft_match_failure(); return NULL; } 1050 if (n->is_MemBar()) { 1051 m->as_MachMemBar()->set_adr_type(n->adr_type()); 1052 } 1053 } else { // Nothing the matcher cares about 1054 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Multi()) { // Projections? 1055 // Convert to machine-dependent projection 1056 RegMask* mask = NULL; 1057 if (n->in(0)->is_Call()) { 1058 mask = return_values_mask(n->in(0)->as_Call()->tf()->range_cc()); 1059 } 1060 m = n->in(0)->as_Multi()->match(n->as_Proj(), this, mask); 1061 #ifdef ASSERT 1062 _new2old_map.map(m->_idx, n); 1063 #endif 1064 if (m->in(0) != NULL) // m might be top 1065 collect_null_checks(m, n); 1066 } else { // Else just a regular 'ol guy 1067 m = n->clone(); // So just clone into new-space 1068 #ifdef ASSERT 1069 _new2old_map.map(m->_idx, n); 1070 #endif 1071 // Def-Use edges will be added incrementally as Uses 1072 // of this node are matched. 1073 assert(m->outcnt() == 0, "no Uses of this clone yet"); 1074 } 1075 } 1076 1077 set_new_node(n, m); // Map old to new 1078 if (_old_node_note_array != NULL) { 1079 Node_Notes* nn = C->locate_node_notes(_old_node_note_array, 1080 n->_idx); 1081 C->set_node_notes_at(m->_idx, nn); 1082 } 1083 debug_only(match_alias_type(C, n, m)); 1084 } 1085 n = m; // n is now a new-space node 1086 mstack.set_node(n); 1087 } 1088 1089 // New space! 1090 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty()) 1091 1092 int i; 1093 // Put precedence edges on stack first (match them last). 1094 for (i = oldn->req(); (uint)i < oldn->len(); i++) { 1095 Node *m = oldn->in(i); 1096 if (m == NULL) break; 1097 // set -1 to call add_prec() instead of set_req() during Step1 1098 mstack.push(m, Visit, n, -1); 1099 } 1100 1101 // Handle precedence edges for interior nodes 1102 for (i = n->len()-1; (uint)i >= n->req(); i--) { 1103 Node *m = n->in(i); 1104 if (m == NULL || C->node_arena()->contains(m)) continue; 1105 n->rm_prec(i); 1106 // set -1 to call add_prec() instead of set_req() during Step1 1107 mstack.push(m, Visit, n, -1); 1108 } 1109 1110 // For constant debug info, I'd rather have unmatched constants. 1111 int cnt = n->req(); 1112 JVMState* jvms = n->jvms(); 1113 int debug_cnt = jvms ? jvms->debug_start() : cnt; 1114 1115 // Now do only debug info. Clone constants rather than matching. 1116 // Constants are represented directly in the debug info without 1117 // the need for executable machine instructions. 1118 // Monitor boxes are also represented directly. 1119 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do 1120 Node *m = n->in(i); // Get input 1121 int op = m->Opcode(); 1122 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites"); 1123 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass || 1124 op == Op_ConF || op == Op_ConD || op == Op_ConL 1125 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp 1126 ) { 1127 m = m->clone(); 1128 #ifdef ASSERT 1129 _new2old_map.map(m->_idx, n); 1130 #endif 1131 mstack.push(m, Post_Visit, n, i); // Don't need to visit 1132 mstack.push(m->in(0), Visit, m, 0); 1133 } else { 1134 mstack.push(m, Visit, n, i); 1135 } 1136 } 1137 1138 // And now walk his children, and convert his inputs to new-space. 1139 for( ; i >= 0; --i ) { // For all normal inputs do 1140 Node *m = n->in(i); // Get input 1141 if(m != NULL) 1142 mstack.push(m, Visit, n, i); 1143 } 1144 1145 } 1146 else if (nstate == Post_Visit) { 1147 // Set xformed input 1148 Node *p = mstack.parent(); 1149 if (p != NULL) { // root doesn't have parent 1150 int i = (int)mstack.index(); 1151 if (i >= 0) 1152 p->set_req(i, n); // required input 1153 else if (i == -1) 1154 p->add_prec(n); // precedence input 1155 else 1156 ShouldNotReachHere(); 1157 } 1158 mstack.pop(); // remove processed node from stack 1159 } 1160 else { 1161 ShouldNotReachHere(); 1162 } 1163 } // while (mstack.is_nonempty()) 1164 return n; // Return new-space Node 1165 } 1166 1167 //------------------------------warp_outgoing_stk_arg------------------------ 1168 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) { 1169 // Convert outgoing argument location to a pre-biased stack offset 1170 if (reg->is_stack()) { 1171 OptoReg::Name warped = reg->reg2stack(); 1172 // Adjust the stack slot offset to be the register number used 1173 // by the allocator. 1174 warped = OptoReg::add(begin_out_arg_area, warped); 1175 // Keep track of the largest numbered stack slot used for an arg. 1176 // Largest used slot per call-site indicates the amount of stack 1177 // that is killed by the call. 1178 if( warped >= out_arg_limit_per_call ) 1179 out_arg_limit_per_call = OptoReg::add(warped,1); 1180 if (!RegMask::can_represent_arg(warped)) { 1181 C->record_method_not_compilable("unsupported calling sequence"); 1182 return OptoReg::Bad; 1183 } 1184 return warped; 1185 } 1186 return OptoReg::as_OptoReg(reg); 1187 } 1188 1189 1190 //------------------------------match_sfpt------------------------------------- 1191 // Helper function to match call instructions. Calls match special. 1192 // They match alone with no children. Their children, the incoming 1193 // arguments, match normally. 1194 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) { 1195 MachSafePointNode *msfpt = NULL; 1196 MachCallNode *mcall = NULL; 1197 uint cnt; 1198 // Split out case for SafePoint vs Call 1199 CallNode *call; 1200 const TypeTuple *domain; 1201 ciMethod* method = NULL; 1202 bool is_method_handle_invoke = false; // for special kill effects 1203 if( sfpt->is_Call() ) { 1204 call = sfpt->as_Call(); 1205 domain = call->tf()->domain_cc(); 1206 cnt = domain->cnt(); 1207 1208 // Match just the call, nothing else 1209 MachNode *m = match_tree(call); 1210 if (C->failing()) return NULL; 1211 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; } 1212 1213 // Copy data from the Ideal SafePoint to the machine version 1214 mcall = m->as_MachCall(); 1215 1216 mcall->set_tf( call->tf()); 1217 mcall->set_entry_point(call->entry_point()); 1218 mcall->set_cnt( call->cnt()); 1219 1220 if( mcall->is_MachCallJava() ) { 1221 MachCallJavaNode *mcall_java = mcall->as_MachCallJava(); 1222 const CallJavaNode *call_java = call->as_CallJava(); 1223 method = call_java->method(); 1224 mcall_java->_method = method; 1225 mcall_java->_bci = call_java->_bci; 1226 mcall_java->_optimized_virtual = call_java->is_optimized_virtual(); 1227 is_method_handle_invoke = call_java->is_method_handle_invoke(); 1228 mcall_java->_method_handle_invoke = is_method_handle_invoke; 1229 mcall_java->_override_symbolic_info = call_java->override_symbolic_info(); 1230 if (is_method_handle_invoke) { 1231 C->set_has_method_handle_invokes(true); 1232 } 1233 if( mcall_java->is_MachCallStaticJava() ) 1234 mcall_java->as_MachCallStaticJava()->_name = 1235 call_java->as_CallStaticJava()->_name; 1236 if( mcall_java->is_MachCallDynamicJava() ) 1237 mcall_java->as_MachCallDynamicJava()->_vtable_index = 1238 call_java->as_CallDynamicJava()->_vtable_index; 1239 } 1240 else if( mcall->is_MachCallRuntime() ) { 1241 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name; 1242 } 1243 msfpt = mcall; 1244 } 1245 // This is a non-call safepoint 1246 else { 1247 call = NULL; 1248 domain = NULL; 1249 MachNode *mn = match_tree(sfpt); 1250 if (C->failing()) return NULL; 1251 msfpt = mn->as_MachSafePoint(); 1252 cnt = TypeFunc::Parms; 1253 } 1254 1255 // Advertise the correct memory effects (for anti-dependence computation). 1256 msfpt->set_adr_type(sfpt->adr_type()); 1257 1258 // Allocate a private array of RegMasks. These RegMasks are not shared. 1259 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt ); 1260 // Empty them all. 1261 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt ); 1262 1263 // Do all the pre-defined non-Empty register masks 1264 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask; 1265 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask; 1266 1267 // Place first outgoing argument can possibly be put. 1268 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 1269 assert( is_even(begin_out_arg_area), "" ); 1270 // Compute max outgoing register number per call site. 1271 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area; 1272 // Calls to C may hammer extra stack slots above and beyond any arguments. 1273 // These are usually backing store for register arguments for varargs. 1274 if( call != NULL && call->is_CallRuntime() ) 1275 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed()); 1276 1277 1278 // Do the normal argument list (parameters) register masks 1279 // Null entry point is a special cast where the target of the call 1280 // is in a register. 1281 int adj = (call != NULL && call->entry_point() == NULL) ? 1 : 0; 1282 int argcnt = cnt - TypeFunc::Parms - adj; 1283 if( argcnt > 0 ) { // Skip it all if we have no args 1284 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 1285 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 1286 int i; 1287 for( i = 0; i < argcnt; i++ ) { 1288 sig_bt[i] = domain->field_at(i+TypeFunc::Parms+adj)->basic_type(); 1289 } 1290 // V-call to pick proper calling convention 1291 call->calling_convention( sig_bt, parm_regs, argcnt ); 1292 1293 #ifdef ASSERT 1294 // Sanity check users' calling convention. Really handy during 1295 // the initial porting effort. Fairly expensive otherwise. 1296 { for (int i = 0; i<argcnt; i++) { 1297 if( !parm_regs[i].first()->is_valid() && 1298 !parm_regs[i].second()->is_valid() ) continue; 1299 VMReg reg1 = parm_regs[i].first(); 1300 VMReg reg2 = parm_regs[i].second(); 1301 for (int j = 0; j < i; j++) { 1302 if( !parm_regs[j].first()->is_valid() && 1303 !parm_regs[j].second()->is_valid() ) continue; 1304 VMReg reg3 = parm_regs[j].first(); 1305 VMReg reg4 = parm_regs[j].second(); 1306 if( !reg1->is_valid() ) { 1307 assert( !reg2->is_valid(), "valid halvsies" ); 1308 } else if( !reg3->is_valid() ) { 1309 assert( !reg4->is_valid(), "valid halvsies" ); 1310 } else { 1311 assert( reg1 != reg2, "calling conv. must produce distinct regs"); 1312 assert( reg1 != reg3, "calling conv. must produce distinct regs"); 1313 assert( reg1 != reg4, "calling conv. must produce distinct regs"); 1314 assert( reg2 != reg3, "calling conv. must produce distinct regs"); 1315 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs"); 1316 assert( reg3 != reg4, "calling conv. must produce distinct regs"); 1317 } 1318 } 1319 } 1320 } 1321 #endif 1322 1323 // Visit each argument. Compute its outgoing register mask. 1324 // Return results now can have 2 bits returned. 1325 // Compute max over all outgoing arguments both per call-site 1326 // and over the entire method. 1327 for( i = 0; i < argcnt; i++ ) { 1328 // Address of incoming argument mask to fill in 1329 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms+adj]; 1330 if( !parm_regs[i].first()->is_valid() && 1331 !parm_regs[i].second()->is_valid() ) { 1332 continue; // Avoid Halves 1333 } 1334 // Grab first register, adjust stack slots and insert in mask. 1335 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call ); 1336 if (OptoReg::is_valid(reg1)) { 1337 rm->Insert( reg1 ); 1338 } 1339 // Grab second register (if any), adjust stack slots and insert in mask. 1340 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call ); 1341 if (OptoReg::is_valid(reg2)) { 1342 rm->Insert( reg2 ); 1343 } 1344 } // End of for all arguments 1345 1346 // Compute number of stack slots needed to restore stack in case of 1347 // Pascal-style argument popping. 1348 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area; 1349 } 1350 1351 // Compute the max stack slot killed by any call. These will not be 1352 // available for debug info, and will be used to adjust FIRST_STACK_mask 1353 // after all call sites have been visited. 1354 if( _out_arg_limit < out_arg_limit_per_call) 1355 _out_arg_limit = out_arg_limit_per_call; 1356 1357 if (mcall) { 1358 // Kill the outgoing argument area, including any non-argument holes and 1359 // any legacy C-killed slots. Use Fat-Projections to do the killing. 1360 // Since the max-per-method covers the max-per-call-site and debug info 1361 // is excluded on the max-per-method basis, debug info cannot land in 1362 // this killed area. 1363 uint r_cnt = mcall->tf()->range_sig()->cnt(); 1364 MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj ); 1365 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) { 1366 C->record_method_not_compilable("unsupported outgoing calling sequence"); 1367 } else { 1368 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++) 1369 proj->_rout.Insert(OptoReg::Name(i)); 1370 } 1371 if (proj->_rout.is_NotEmpty()) { 1372 push_projection(proj); 1373 } 1374 } 1375 // Transfer the safepoint information from the call to the mcall 1376 // Move the JVMState list 1377 msfpt->set_jvms(sfpt->jvms()); 1378 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) { 1379 jvms->set_map(sfpt); 1380 } 1381 1382 // Debug inputs begin just after the last incoming parameter 1383 assert((mcall == NULL) || (mcall->jvms() == NULL) || 1384 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain_cc()->cnt()), ""); 1385 1386 // Move the OopMap 1387 msfpt->_oop_map = sfpt->_oop_map; 1388 1389 // Add additional edges. 1390 if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) { 1391 // For these calls we can not add MachConstantBase in expand(), as the 1392 // ins are not complete then. 1393 msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node()); 1394 if (msfpt->jvms() && 1395 msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) { 1396 // We added an edge before jvms, so we must adapt the position of the ins. 1397 msfpt->jvms()->adapt_position(+1); 1398 } 1399 } 1400 1401 // Registers killed by the call are set in the local scheduling pass 1402 // of Global Code Motion. 1403 return msfpt; 1404 } 1405 1406 //---------------------------match_tree---------------------------------------- 1407 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part 1408 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for 1409 // making GotoNodes while building the CFG and in init_spill_mask() to identify 1410 // a Load's result RegMask for memoization in idealreg2regmask[] 1411 MachNode *Matcher::match_tree( const Node *n ) { 1412 assert( n->Opcode() != Op_Phi, "cannot match" ); 1413 assert( !n->is_block_start(), "cannot match" ); 1414 // Set the mark for all locally allocated State objects. 1415 // When this call returns, the _states_arena arena will be reset 1416 // freeing all State objects. 1417 ResourceMark rm( &_states_arena ); 1418 1419 LabelRootDepth = 0; 1420 1421 // StoreNodes require their Memory input to match any LoadNodes 1422 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ; 1423 #ifdef ASSERT 1424 Node* save_mem_node = _mem_node; 1425 _mem_node = n->is_Store() ? (Node*)n : NULL; 1426 #endif 1427 // State object for root node of match tree 1428 // Allocate it on _states_arena - stack allocation can cause stack overflow. 1429 State *s = new (&_states_arena) State; 1430 s->_kids[0] = NULL; 1431 s->_kids[1] = NULL; 1432 s->_leaf = (Node*)n; 1433 // Label the input tree, allocating labels from top-level arena 1434 Label_Root( n, s, n->in(0), mem ); 1435 if (C->failing()) return NULL; 1436 1437 // The minimum cost match for the whole tree is found at the root State 1438 uint mincost = max_juint; 1439 uint cost = max_juint; 1440 uint i; 1441 for( i = 0; i < NUM_OPERANDS; i++ ) { 1442 if( s->valid(i) && // valid entry and 1443 s->_cost[i] < cost && // low cost and 1444 s->_rule[i] >= NUM_OPERANDS ) // not an operand 1445 cost = s->_cost[mincost=i]; 1446 } 1447 if (mincost == max_juint) { 1448 #ifndef PRODUCT 1449 tty->print("No matching rule for:"); 1450 s->dump(); 1451 #endif 1452 Matcher::soft_match_failure(); 1453 return NULL; 1454 } 1455 // Reduce input tree based upon the state labels to machine Nodes 1456 MachNode *m = ReduceInst( s, s->_rule[mincost], mem ); 1457 #ifdef ASSERT 1458 _old2new_map.map(n->_idx, m); 1459 _new2old_map.map(m->_idx, (Node*)n); 1460 #endif 1461 1462 // Add any Matcher-ignored edges 1463 uint cnt = n->req(); 1464 uint start = 1; 1465 if( mem != (Node*)1 ) start = MemNode::Memory+1; 1466 if( n->is_AddP() ) { 1467 assert( mem == (Node*)1, "" ); 1468 start = AddPNode::Base+1; 1469 } 1470 for( i = start; i < cnt; i++ ) { 1471 if( !n->match_edge(i) ) { 1472 if( i < m->req() ) 1473 m->ins_req( i, n->in(i) ); 1474 else 1475 m->add_req( n->in(i) ); 1476 } 1477 } 1478 1479 debug_only( _mem_node = save_mem_node; ) 1480 return m; 1481 } 1482 1483 1484 //------------------------------match_into_reg--------------------------------- 1485 // Choose to either match this Node in a register or part of the current 1486 // match tree. Return true for requiring a register and false for matching 1487 // as part of the current match tree. 1488 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) { 1489 1490 const Type *t = m->bottom_type(); 1491 1492 if (t->singleton()) { 1493 // Never force constants into registers. Allow them to match as 1494 // constants or registers. Copies of the same value will share 1495 // the same register. See find_shared_node. 1496 return false; 1497 } else { // Not a constant 1498 // Stop recursion if they have different Controls. 1499 Node* m_control = m->in(0); 1500 // Control of load's memory can post-dominates load's control. 1501 // So use it since load can't float above its memory. 1502 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL; 1503 if (control && m_control && control != m_control && control != mem_control) { 1504 1505 // Actually, we can live with the most conservative control we 1506 // find, if it post-dominates the others. This allows us to 1507 // pick up load/op/store trees where the load can float a little 1508 // above the store. 1509 Node *x = control; 1510 const uint max_scan = 6; // Arbitrary scan cutoff 1511 uint j; 1512 for (j=0; j<max_scan; j++) { 1513 if (x->is_Region()) // Bail out at merge points 1514 return true; 1515 x = x->in(0); 1516 if (x == m_control) // Does 'control' post-dominate 1517 break; // m->in(0)? If so, we can use it 1518 if (x == mem_control) // Does 'control' post-dominate 1519 break; // mem_control? If so, we can use it 1520 } 1521 if (j == max_scan) // No post-domination before scan end? 1522 return true; // Then break the match tree up 1523 } 1524 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) || 1525 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) { 1526 // These are commonly used in address expressions and can 1527 // efficiently fold into them on X64 in some cases. 1528 return false; 1529 } 1530 } 1531 1532 // Not forceable cloning. If shared, put it into a register. 1533 return shared; 1534 } 1535 1536 1537 //------------------------------Instruction Selection-------------------------- 1538 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match 1539 // ideal nodes to machine instructions. Trees are delimited by shared Nodes, 1540 // things the Matcher does not match (e.g., Memory), and things with different 1541 // Controls (hence forced into different blocks). We pass in the Control 1542 // selected for this entire State tree. 1543 1544 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the 1545 // Store and the Load must have identical Memories (as well as identical 1546 // pointers). Since the Matcher does not have anything for Memory (and 1547 // does not handle DAGs), I have to match the Memory input myself. If the 1548 // Tree root is a Store, I require all Loads to have the identical memory. 1549 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){ 1550 // Since Label_Root is a recursive function, its possible that we might run 1551 // out of stack space. See bugs 6272980 & 6227033 for more info. 1552 LabelRootDepth++; 1553 if (LabelRootDepth > MaxLabelRootDepth) { 1554 C->record_method_not_compilable("Out of stack space, increase MaxLabelRootDepth"); 1555 return NULL; 1556 } 1557 uint care = 0; // Edges matcher cares about 1558 uint cnt = n->req(); 1559 uint i = 0; 1560 1561 // Examine children for memory state 1562 // Can only subsume a child into your match-tree if that child's memory state 1563 // is not modified along the path to another input. 1564 // It is unsafe even if the other inputs are separate roots. 1565 Node *input_mem = NULL; 1566 for( i = 1; i < cnt; i++ ) { 1567 if( !n->match_edge(i) ) continue; 1568 Node *m = n->in(i); // Get ith input 1569 assert( m, "expect non-null children" ); 1570 if( m->is_Load() ) { 1571 if( input_mem == NULL ) { 1572 input_mem = m->in(MemNode::Memory); 1573 } else if( input_mem != m->in(MemNode::Memory) ) { 1574 input_mem = NodeSentinel; 1575 } 1576 } 1577 } 1578 1579 for( i = 1; i < cnt; i++ ){// For my children 1580 if( !n->match_edge(i) ) continue; 1581 Node *m = n->in(i); // Get ith input 1582 // Allocate states out of a private arena 1583 State *s = new (&_states_arena) State; 1584 svec->_kids[care++] = s; 1585 assert( care <= 2, "binary only for now" ); 1586 1587 // Recursively label the State tree. 1588 s->_kids[0] = NULL; 1589 s->_kids[1] = NULL; 1590 s->_leaf = m; 1591 1592 // Check for leaves of the State Tree; things that cannot be a part of 1593 // the current tree. If it finds any, that value is matched as a 1594 // register operand. If not, then the normal matching is used. 1595 if( match_into_reg(n, m, control, i, is_shared(m)) || 1596 // 1597 // Stop recursion if this is LoadNode and the root of this tree is a 1598 // StoreNode and the load & store have different memories. 1599 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) || 1600 // Can NOT include the match of a subtree when its memory state 1601 // is used by any of the other subtrees 1602 (input_mem == NodeSentinel) ) { 1603 // Print when we exclude matching due to different memory states at input-loads 1604 if (PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel) 1605 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem)) { 1606 tty->print_cr("invalid input_mem"); 1607 } 1608 // Switch to a register-only opcode; this value must be in a register 1609 // and cannot be subsumed as part of a larger instruction. 1610 s->DFA( m->ideal_reg(), m ); 1611 1612 } else { 1613 // If match tree has no control and we do, adopt it for entire tree 1614 if( control == NULL && m->in(0) != NULL && m->req() > 1 ) 1615 control = m->in(0); // Pick up control 1616 // Else match as a normal part of the match tree. 1617 control = Label_Root(m,s,control,mem); 1618 if (C->failing()) return NULL; 1619 } 1620 } 1621 1622 1623 // Call DFA to match this node, and return 1624 svec->DFA( n->Opcode(), n ); 1625 1626 #ifdef ASSERT 1627 uint x; 1628 for( x = 0; x < _LAST_MACH_OPER; x++ ) 1629 if( svec->valid(x) ) 1630 break; 1631 1632 if (x >= _LAST_MACH_OPER) { 1633 n->dump(); 1634 svec->dump(); 1635 assert( false, "bad AD file" ); 1636 } 1637 #endif 1638 return control; 1639 } 1640 1641 1642 // Con nodes reduced using the same rule can share their MachNode 1643 // which reduces the number of copies of a constant in the final 1644 // program. The register allocator is free to split uses later to 1645 // split live ranges. 1646 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) { 1647 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL; 1648 1649 // See if this Con has already been reduced using this rule. 1650 if (_shared_nodes.Size() <= leaf->_idx) return NULL; 1651 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx); 1652 if (last != NULL && rule == last->rule()) { 1653 // Don't expect control change for DecodeN 1654 if (leaf->is_DecodeNarrowPtr()) 1655 return last; 1656 // Get the new space root. 1657 Node* xroot = new_node(C->root()); 1658 if (xroot == NULL) { 1659 // This shouldn't happen give the order of matching. 1660 return NULL; 1661 } 1662 1663 // Shared constants need to have their control be root so they 1664 // can be scheduled properly. 1665 Node* control = last->in(0); 1666 if (control != xroot) { 1667 if (control == NULL || control == C->root()) { 1668 last->set_req(0, xroot); 1669 } else { 1670 assert(false, "unexpected control"); 1671 return NULL; 1672 } 1673 } 1674 return last; 1675 } 1676 return NULL; 1677 } 1678 1679 1680 //------------------------------ReduceInst------------------------------------- 1681 // Reduce a State tree (with given Control) into a tree of MachNodes. 1682 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into 1683 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes. 1684 // Each MachNode has a number of complicated MachOper operands; each 1685 // MachOper also covers a further tree of Ideal Nodes. 1686 1687 // The root of the Ideal match tree is always an instruction, so we enter 1688 // the recursion here. After building the MachNode, we need to recurse 1689 // the tree checking for these cases: 1690 // (1) Child is an instruction - 1691 // Build the instruction (recursively), add it as an edge. 1692 // Build a simple operand (register) to hold the result of the instruction. 1693 // (2) Child is an interior part of an instruction - 1694 // Skip over it (do nothing) 1695 // (3) Child is the start of a operand - 1696 // Build the operand, place it inside the instruction 1697 // Call ReduceOper. 1698 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) { 1699 assert( rule >= NUM_OPERANDS, "called with operand rule" ); 1700 1701 MachNode* shared_node = find_shared_node(s->_leaf, rule); 1702 if (shared_node != NULL) { 1703 return shared_node; 1704 } 1705 1706 // Build the object to represent this state & prepare for recursive calls 1707 MachNode *mach = s->MachNodeGenerator(rule); 1708 guarantee(mach != NULL, "Missing MachNode"); 1709 mach->_opnds[0] = s->MachOperGenerator(_reduceOp[rule]); 1710 assert( mach->_opnds[0] != NULL, "Missing result operand" ); 1711 Node *leaf = s->_leaf; 1712 // Check for instruction or instruction chain rule 1713 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) { 1714 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf), 1715 "duplicating node that's already been matched"); 1716 // Instruction 1717 mach->add_req( leaf->in(0) ); // Set initial control 1718 // Reduce interior of complex instruction 1719 ReduceInst_Interior( s, rule, mem, mach, 1 ); 1720 } else { 1721 // Instruction chain rules are data-dependent on their inputs 1722 mach->add_req(0); // Set initial control to none 1723 ReduceInst_Chain_Rule( s, rule, mem, mach ); 1724 } 1725 1726 // If a Memory was used, insert a Memory edge 1727 if( mem != (Node*)1 ) { 1728 mach->ins_req(MemNode::Memory,mem); 1729 #ifdef ASSERT 1730 // Verify adr type after matching memory operation 1731 const MachOper* oper = mach->memory_operand(); 1732 if (oper != NULL && oper != (MachOper*)-1) { 1733 // It has a unique memory operand. Find corresponding ideal mem node. 1734 Node* m = NULL; 1735 if (leaf->is_Mem()) { 1736 m = leaf; 1737 } else { 1738 m = _mem_node; 1739 assert(m != NULL && m->is_Mem(), "expecting memory node"); 1740 } 1741 const Type* mach_at = mach->adr_type(); 1742 // DecodeN node consumed by an address may have different type 1743 // than its input. Don't compare types for such case. 1744 if (m->adr_type() != mach_at && 1745 (m->in(MemNode::Address)->is_DecodeNarrowPtr() || 1746 (m->in(MemNode::Address)->is_AddP() && 1747 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()) || 1748 (m->in(MemNode::Address)->is_AddP() && 1749 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() && 1750 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr()))) { 1751 mach_at = m->adr_type(); 1752 } 1753 if (m->adr_type() != mach_at) { 1754 m->dump(); 1755 tty->print_cr("mach:"); 1756 mach->dump(1); 1757 } 1758 assert(m->adr_type() == mach_at, "matcher should not change adr type"); 1759 } 1760 #endif 1761 } 1762 1763 // If the _leaf is an AddP, insert the base edge 1764 if (leaf->is_AddP()) { 1765 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base)); 1766 } 1767 1768 uint number_of_projections_prior = number_of_projections(); 1769 1770 // Perform any 1-to-many expansions required 1771 MachNode *ex = mach->Expand(s, _projection_list, mem); 1772 if (ex != mach) { 1773 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match"); 1774 if( ex->in(1)->is_Con() ) 1775 ex->in(1)->set_req(0, C->root()); 1776 // Remove old node from the graph 1777 for( uint i=0; i<mach->req(); i++ ) { 1778 mach->set_req(i,NULL); 1779 } 1780 #ifdef ASSERT 1781 _new2old_map.map(ex->_idx, s->_leaf); 1782 #endif 1783 } 1784 1785 // PhaseChaitin::fixup_spills will sometimes generate spill code 1786 // via the matcher. By the time, nodes have been wired into the CFG, 1787 // and any further nodes generated by expand rules will be left hanging 1788 // in space, and will not get emitted as output code. Catch this. 1789 // Also, catch any new register allocation constraints ("projections") 1790 // generated belatedly during spill code generation. 1791 if (_allocation_started) { 1792 guarantee(ex == mach, "no expand rules during spill generation"); 1793 guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation"); 1794 } 1795 1796 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) { 1797 // Record the con for sharing 1798 _shared_nodes.map(leaf->_idx, ex); 1799 } 1800 1801 return ex; 1802 } 1803 1804 void Matcher::handle_precedence_edges(Node* n, MachNode *mach) { 1805 for (uint i = n->req(); i < n->len(); i++) { 1806 if (n->in(i) != NULL) { 1807 mach->add_prec(n->in(i)); 1808 } 1809 } 1810 } 1811 1812 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) { 1813 // 'op' is what I am expecting to receive 1814 int op = _leftOp[rule]; 1815 // Operand type to catch childs result 1816 // This is what my child will give me. 1817 int opnd_class_instance = s->_rule[op]; 1818 // Choose between operand class or not. 1819 // This is what I will receive. 1820 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op; 1821 // New rule for child. Chase operand classes to get the actual rule. 1822 int newrule = s->_rule[catch_op]; 1823 1824 if( newrule < NUM_OPERANDS ) { 1825 // Chain from operand or operand class, may be output of shared node 1826 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS, 1827 "Bad AD file: Instruction chain rule must chain from operand"); 1828 // Insert operand into array of operands for this instruction 1829 mach->_opnds[1] = s->MachOperGenerator(opnd_class_instance); 1830 1831 ReduceOper( s, newrule, mem, mach ); 1832 } else { 1833 // Chain from the result of an instruction 1834 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand"); 1835 mach->_opnds[1] = s->MachOperGenerator(_reduceOp[catch_op]); 1836 Node *mem1 = (Node*)1; 1837 debug_only(Node *save_mem_node = _mem_node;) 1838 mach->add_req( ReduceInst(s, newrule, mem1) ); 1839 debug_only(_mem_node = save_mem_node;) 1840 } 1841 return; 1842 } 1843 1844 1845 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) { 1846 handle_precedence_edges(s->_leaf, mach); 1847 1848 if( s->_leaf->is_Load() ) { 1849 Node *mem2 = s->_leaf->in(MemNode::Memory); 1850 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" ); 1851 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;) 1852 mem = mem2; 1853 } 1854 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) { 1855 if( mach->in(0) == NULL ) 1856 mach->set_req(0, s->_leaf->in(0)); 1857 } 1858 1859 // Now recursively walk the state tree & add operand list. 1860 for( uint i=0; i<2; i++ ) { // binary tree 1861 State *newstate = s->_kids[i]; 1862 if( newstate == NULL ) break; // Might only have 1 child 1863 // 'op' is what I am expecting to receive 1864 int op; 1865 if( i == 0 ) { 1866 op = _leftOp[rule]; 1867 } else { 1868 op = _rightOp[rule]; 1869 } 1870 // Operand type to catch childs result 1871 // This is what my child will give me. 1872 int opnd_class_instance = newstate->_rule[op]; 1873 // Choose between operand class or not. 1874 // This is what I will receive. 1875 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op; 1876 // New rule for child. Chase operand classes to get the actual rule. 1877 int newrule = newstate->_rule[catch_op]; 1878 1879 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction? 1880 // Operand/operandClass 1881 // Insert operand into array of operands for this instruction 1882 mach->_opnds[num_opnds++] = newstate->MachOperGenerator(opnd_class_instance); 1883 ReduceOper( newstate, newrule, mem, mach ); 1884 1885 } else { // Child is internal operand or new instruction 1886 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction? 1887 // internal operand --> call ReduceInst_Interior 1888 // Interior of complex instruction. Do nothing but recurse. 1889 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds ); 1890 } else { 1891 // instruction --> call build operand( ) to catch result 1892 // --> ReduceInst( newrule ) 1893 mach->_opnds[num_opnds++] = s->MachOperGenerator(_reduceOp[catch_op]); 1894 Node *mem1 = (Node*)1; 1895 debug_only(Node *save_mem_node = _mem_node;) 1896 mach->add_req( ReduceInst( newstate, newrule, mem1 ) ); 1897 debug_only(_mem_node = save_mem_node;) 1898 } 1899 } 1900 assert( mach->_opnds[num_opnds-1], "" ); 1901 } 1902 return num_opnds; 1903 } 1904 1905 // This routine walks the interior of possible complex operands. 1906 // At each point we check our children in the match tree: 1907 // (1) No children - 1908 // We are a leaf; add _leaf field as an input to the MachNode 1909 // (2) Child is an internal operand - 1910 // Skip over it ( do nothing ) 1911 // (3) Child is an instruction - 1912 // Call ReduceInst recursively and 1913 // and instruction as an input to the MachNode 1914 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) { 1915 assert( rule < _LAST_MACH_OPER, "called with operand rule" ); 1916 State *kid = s->_kids[0]; 1917 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" ); 1918 1919 // Leaf? And not subsumed? 1920 if( kid == NULL && !_swallowed[rule] ) { 1921 mach->add_req( s->_leaf ); // Add leaf pointer 1922 return; // Bail out 1923 } 1924 1925 if( s->_leaf->is_Load() ) { 1926 assert( mem == (Node*)1, "multiple Memories being matched at once?" ); 1927 mem = s->_leaf->in(MemNode::Memory); 1928 debug_only(_mem_node = s->_leaf;) 1929 } 1930 1931 handle_precedence_edges(s->_leaf, mach); 1932 1933 if( s->_leaf->in(0) && s->_leaf->req() > 1) { 1934 if( !mach->in(0) ) 1935 mach->set_req(0,s->_leaf->in(0)); 1936 else { 1937 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" ); 1938 } 1939 } 1940 1941 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree 1942 int newrule; 1943 if( i == 0) 1944 newrule = kid->_rule[_leftOp[rule]]; 1945 else 1946 newrule = kid->_rule[_rightOp[rule]]; 1947 1948 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction? 1949 // Internal operand; recurse but do nothing else 1950 ReduceOper( kid, newrule, mem, mach ); 1951 1952 } else { // Child is a new instruction 1953 // Reduce the instruction, and add a direct pointer from this 1954 // machine instruction to the newly reduced one. 1955 Node *mem1 = (Node*)1; 1956 debug_only(Node *save_mem_node = _mem_node;) 1957 mach->add_req( ReduceInst( kid, newrule, mem1 ) ); 1958 debug_only(_mem_node = save_mem_node;) 1959 } 1960 } 1961 } 1962 1963 1964 // ------------------------------------------------------------------------- 1965 // Java-Java calling convention 1966 // (what you use when Java calls Java) 1967 1968 //------------------------------find_receiver---------------------------------- 1969 // For a given signature, return the OptoReg for parameter 0. 1970 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) { 1971 VMRegPair regs; 1972 BasicType sig_bt = T_OBJECT; 1973 calling_convention(&sig_bt, ®s, 1, is_outgoing); 1974 // Return argument 0 register. In the LP64 build pointers 1975 // take 2 registers, but the VM wants only the 'main' name. 1976 return OptoReg::as_OptoReg(regs.first()); 1977 } 1978 1979 // This function identifies sub-graphs in which a 'load' node is 1980 // input to two different nodes, and such that it can be matched 1981 // with BMI instructions like blsi, blsr, etc. 1982 // Example : for b = -a[i] & a[i] can be matched to blsi r32, m32. 1983 // The graph is (AndL (SubL Con0 LoadL*) LoadL*), where LoadL* 1984 // refers to the same node. 1985 #ifdef X86 1986 // Match the generic fused operations pattern (op1 (op2 Con{ConType} mop) mop) 1987 // This is a temporary solution until we make DAGs expressible in ADL. 1988 template<typename ConType> 1989 class FusedPatternMatcher { 1990 Node* _op1_node; 1991 Node* _mop_node; 1992 int _con_op; 1993 1994 static int match_next(Node* n, int next_op, int next_op_idx) { 1995 if (n->in(1) == NULL || n->in(2) == NULL) { 1996 return -1; 1997 } 1998 1999 if (next_op_idx == -1) { // n is commutative, try rotations 2000 if (n->in(1)->Opcode() == next_op) { 2001 return 1; 2002 } else if (n->in(2)->Opcode() == next_op) { 2003 return 2; 2004 } 2005 } else { 2006 assert(next_op_idx > 0 && next_op_idx <= 2, "Bad argument index"); 2007 if (n->in(next_op_idx)->Opcode() == next_op) { 2008 return next_op_idx; 2009 } 2010 } 2011 return -1; 2012 } 2013 public: 2014 FusedPatternMatcher(Node* op1_node, Node *mop_node, int con_op) : 2015 _op1_node(op1_node), _mop_node(mop_node), _con_op(con_op) { } 2016 2017 bool match(int op1, int op1_op2_idx, // op1 and the index of the op1->op2 edge, -1 if op1 is commutative 2018 int op2, int op2_con_idx, // op2 and the index of the op2->con edge, -1 if op2 is commutative 2019 typename ConType::NativeType con_value) { 2020 if (_op1_node->Opcode() != op1) { 2021 return false; 2022 } 2023 if (_mop_node->outcnt() > 2) { 2024 return false; 2025 } 2026 op1_op2_idx = match_next(_op1_node, op2, op1_op2_idx); 2027 if (op1_op2_idx == -1) { 2028 return false; 2029 } 2030 // Memory operation must be the other edge 2031 int op1_mop_idx = (op1_op2_idx & 1) + 1; 2032 2033 // Check that the mop node is really what we want 2034 if (_op1_node->in(op1_mop_idx) == _mop_node) { 2035 Node *op2_node = _op1_node->in(op1_op2_idx); 2036 if (op2_node->outcnt() > 1) { 2037 return false; 2038 } 2039 assert(op2_node->Opcode() == op2, "Should be"); 2040 op2_con_idx = match_next(op2_node, _con_op, op2_con_idx); 2041 if (op2_con_idx == -1) { 2042 return false; 2043 } 2044 // Memory operation must be the other edge 2045 int op2_mop_idx = (op2_con_idx & 1) + 1; 2046 // Check that the memory operation is the same node 2047 if (op2_node->in(op2_mop_idx) == _mop_node) { 2048 // Now check the constant 2049 const Type* con_type = op2_node->in(op2_con_idx)->bottom_type(); 2050 if (con_type != Type::TOP && ConType::as_self(con_type)->get_con() == con_value) { 2051 return true; 2052 } 2053 } 2054 } 2055 return false; 2056 } 2057 }; 2058 2059 2060 bool Matcher::is_bmi_pattern(Node *n, Node *m) { 2061 if (n != NULL && m != NULL) { 2062 if (m->Opcode() == Op_LoadI) { 2063 FusedPatternMatcher<TypeInt> bmii(n, m, Op_ConI); 2064 return bmii.match(Op_AndI, -1, Op_SubI, 1, 0) || 2065 bmii.match(Op_AndI, -1, Op_AddI, -1, -1) || 2066 bmii.match(Op_XorI, -1, Op_AddI, -1, -1); 2067 } else if (m->Opcode() == Op_LoadL) { 2068 FusedPatternMatcher<TypeLong> bmil(n, m, Op_ConL); 2069 return bmil.match(Op_AndL, -1, Op_SubL, 1, 0) || 2070 bmil.match(Op_AndL, -1, Op_AddL, -1, -1) || 2071 bmil.match(Op_XorL, -1, Op_AddL, -1, -1); 2072 } 2073 } 2074 return false; 2075 } 2076 #endif // X86 2077 2078 bool Matcher::clone_base_plus_offset_address(AddPNode* m, Matcher::MStack& mstack, VectorSet& address_visited) { 2079 Node *off = m->in(AddPNode::Offset); 2080 if (off->is_Con()) { 2081 address_visited.test_set(m->_idx); // Flag as address_visited 2082 mstack.push(m->in(AddPNode::Address), Pre_Visit); 2083 // Clone X+offset as it also folds into most addressing expressions 2084 mstack.push(off, Visit); 2085 mstack.push(m->in(AddPNode::Base), Pre_Visit); 2086 return true; 2087 } 2088 return false; 2089 } 2090 2091 // A method-klass-holder may be passed in the inline_cache_reg 2092 // and then expanded into the inline_cache_reg and a method_oop register 2093 // defined in ad_<arch>.cpp 2094 2095 //------------------------------find_shared------------------------------------ 2096 // Set bits if Node is shared or otherwise a root 2097 void Matcher::find_shared( Node *n ) { 2098 // Allocate stack of size C->live_nodes() * 2 to avoid frequent realloc 2099 MStack mstack(C->live_nodes() * 2); 2100 // Mark nodes as address_visited if they are inputs to an address expression 2101 VectorSet address_visited(Thread::current()->resource_area()); 2102 mstack.push(n, Visit); // Don't need to pre-visit root node 2103 while (mstack.is_nonempty()) { 2104 n = mstack.node(); // Leave node on stack 2105 Node_State nstate = mstack.state(); 2106 uint nop = n->Opcode(); 2107 if (nstate == Pre_Visit) { 2108 if (address_visited.test(n->_idx)) { // Visited in address already? 2109 // Flag as visited and shared now. 2110 set_visited(n); 2111 } 2112 if (is_visited(n)) { // Visited already? 2113 // Node is shared and has no reason to clone. Flag it as shared. 2114 // This causes it to match into a register for the sharing. 2115 set_shared(n); // Flag as shared and 2116 mstack.pop(); // remove node from stack 2117 continue; 2118 } 2119 nstate = Visit; // Not already visited; so visit now 2120 } 2121 if (nstate == Visit) { 2122 mstack.set_state(Post_Visit); 2123 set_visited(n); // Flag as visited now 2124 bool mem_op = false; 2125 int mem_addr_idx = MemNode::Address; 2126 2127 switch( nop ) { // Handle some opcodes special 2128 case Op_Phi: // Treat Phis as shared roots 2129 case Op_Parm: 2130 case Op_Proj: // All handled specially during matching 2131 case Op_SafePointScalarObject: 2132 set_shared(n); 2133 set_dontcare(n); 2134 break; 2135 case Op_If: 2136 case Op_CountedLoopEnd: 2137 mstack.set_state(Alt_Post_Visit); // Alternative way 2138 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps 2139 // with matching cmp/branch in 1 instruction. The Matcher needs the 2140 // Bool and CmpX side-by-side, because it can only get at constants 2141 // that are at the leaves of Match trees, and the Bool's condition acts 2142 // as a constant here. 2143 mstack.push(n->in(1), Visit); // Clone the Bool 2144 mstack.push(n->in(0), Pre_Visit); // Visit control input 2145 continue; // while (mstack.is_nonempty()) 2146 case Op_ConvI2D: // These forms efficiently match with a prior 2147 case Op_ConvI2F: // Load but not a following Store 2148 if( n->in(1)->is_Load() && // Prior load 2149 n->outcnt() == 1 && // Not already shared 2150 n->unique_out()->is_Store() ) // Following store 2151 set_shared(n); // Force it to be a root 2152 break; 2153 case Op_ReverseBytesI: 2154 case Op_ReverseBytesL: 2155 if( n->in(1)->is_Load() && // Prior load 2156 n->outcnt() == 1 ) // Not already shared 2157 set_shared(n); // Force it to be a root 2158 break; 2159 case Op_BoxLock: // Cant match until we get stack-regs in ADLC 2160 case Op_IfFalse: 2161 case Op_IfTrue: 2162 case Op_MachProj: 2163 case Op_MergeMem: 2164 case Op_Catch: 2165 case Op_CatchProj: 2166 case Op_CProj: 2167 case Op_JumpProj: 2168 case Op_JProj: 2169 case Op_NeverBranch: 2170 set_dontcare(n); 2171 break; 2172 case Op_Jump: 2173 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared) 2174 mstack.push(n->in(0), Pre_Visit); // Visit Control input 2175 continue; // while (mstack.is_nonempty()) 2176 case Op_StrComp: 2177 case Op_StrEquals: 2178 case Op_StrIndexOf: 2179 case Op_StrIndexOfChar: 2180 case Op_AryEq: 2181 case Op_HasNegatives: 2182 case Op_StrInflatedCopy: 2183 case Op_StrCompressedCopy: 2184 case Op_EncodeISOArray: 2185 case Op_FmaD: 2186 case Op_FmaF: 2187 case Op_FmaVD: 2188 case Op_FmaVF: 2189 set_shared(n); // Force result into register (it will be anyways) 2190 break; 2191 case Op_ConP: { // Convert pointers above the centerline to NUL 2192 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2193 const TypePtr* tp = tn->type()->is_ptr(); 2194 if (tp->_ptr == TypePtr::AnyNull) { 2195 tn->set_type(TypePtr::NULL_PTR); 2196 } 2197 break; 2198 } 2199 case Op_ConN: { // Convert narrow pointers above the centerline to NUL 2200 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2201 const TypePtr* tp = tn->type()->make_ptr(); 2202 if (tp && tp->_ptr == TypePtr::AnyNull) { 2203 tn->set_type(TypeNarrowOop::NULL_PTR); 2204 } 2205 break; 2206 } 2207 case Op_Binary: // These are introduced in the Post_Visit state. 2208 ShouldNotReachHere(); 2209 break; 2210 case Op_ClearArray: 2211 case Op_SafePoint: 2212 mem_op = true; 2213 break; 2214 #if INCLUDE_ZGC 2215 case Op_CallLeaf: 2216 if (UseZGC) { 2217 if (n->as_Call()->entry_point() == ZBarrierSetRuntime::load_barrier_on_oop_field_preloaded_addr() || 2218 n->as_Call()->entry_point() == ZBarrierSetRuntime::load_barrier_on_weak_oop_field_preloaded_addr()) { 2219 mem_op = true; 2220 mem_addr_idx = TypeFunc::Parms+1; 2221 } 2222 break; 2223 } 2224 #endif 2225 default: 2226 if( n->is_Store() ) { 2227 // Do match stores, despite no ideal reg 2228 mem_op = true; 2229 break; 2230 } 2231 if( n->is_Mem() ) { // Loads and LoadStores 2232 mem_op = true; 2233 // Loads must be root of match tree due to prior load conflict 2234 if( C->subsume_loads() == false ) 2235 set_shared(n); 2236 } 2237 // Fall into default case 2238 if( !n->ideal_reg() ) 2239 set_dontcare(n); // Unmatchable Nodes 2240 } // end_switch 2241 2242 for(int i = n->req() - 1; i >= 0; --i) { // For my children 2243 Node *m = n->in(i); // Get ith input 2244 if (m == NULL) continue; // Ignore NULLs 2245 uint mop = m->Opcode(); 2246 2247 // Must clone all producers of flags, or we will not match correctly. 2248 // Suppose a compare setting int-flags is shared (e.g., a switch-tree) 2249 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags 2250 // are also there, so we may match a float-branch to int-flags and 2251 // expect the allocator to haul the flags from the int-side to the 2252 // fp-side. No can do. 2253 if( _must_clone[mop] ) { 2254 mstack.push(m, Visit); 2255 continue; // for(int i = ...) 2256 } 2257 2258 if( mop == Op_AddP && m->in(AddPNode::Base)->is_DecodeNarrowPtr()) { 2259 // Bases used in addresses must be shared but since 2260 // they are shared through a DecodeN they may appear 2261 // to have a single use so force sharing here. 2262 set_shared(m->in(AddPNode::Base)->in(1)); 2263 } 2264 2265 // if 'n' and 'm' are part of a graph for BMI instruction, clone this node. 2266 #ifdef X86 2267 if (UseBMI1Instructions && is_bmi_pattern(n, m)) { 2268 mstack.push(m, Visit); 2269 continue; 2270 } 2271 #endif 2272 2273 // Clone addressing expressions as they are "free" in memory access instructions 2274 if (mem_op && i == mem_addr_idx && mop == Op_AddP && 2275 // When there are other uses besides address expressions 2276 // put it on stack and mark as shared. 2277 !is_visited(m)) { 2278 // Some inputs for address expression are not put on stack 2279 // to avoid marking them as shared and forcing them into register 2280 // if they are used only in address expressions. 2281 // But they should be marked as shared if there are other uses 2282 // besides address expressions. 2283 2284 if (clone_address_expressions(m->as_AddP(), mstack, address_visited)) { 2285 continue; 2286 } 2287 } // if( mem_op && 2288 mstack.push(m, Pre_Visit); 2289 } // for(int i = ...) 2290 } 2291 else if (nstate == Alt_Post_Visit) { 2292 mstack.pop(); // Remove node from stack 2293 // We cannot remove the Cmp input from the Bool here, as the Bool may be 2294 // shared and all users of the Bool need to move the Cmp in parallel. 2295 // This leaves both the Bool and the If pointing at the Cmp. To 2296 // prevent the Matcher from trying to Match the Cmp along both paths 2297 // BoolNode::match_edge always returns a zero. 2298 2299 // We reorder the Op_If in a pre-order manner, so we can visit without 2300 // accidentally sharing the Cmp (the Bool and the If make 2 users). 2301 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool 2302 } 2303 else if (nstate == Post_Visit) { 2304 mstack.pop(); // Remove node from stack 2305 2306 // Now hack a few special opcodes 2307 switch( n->Opcode() ) { // Handle some opcodes special 2308 case Op_StorePConditional: 2309 case Op_StoreIConditional: 2310 case Op_StoreLConditional: 2311 case Op_CompareAndExchangeB: 2312 case Op_CompareAndExchangeS: 2313 case Op_CompareAndExchangeI: 2314 case Op_CompareAndExchangeL: 2315 case Op_CompareAndExchangeP: 2316 case Op_CompareAndExchangeN: 2317 case Op_WeakCompareAndSwapB: 2318 case Op_WeakCompareAndSwapS: 2319 case Op_WeakCompareAndSwapI: 2320 case Op_WeakCompareAndSwapL: 2321 case Op_WeakCompareAndSwapP: 2322 case Op_WeakCompareAndSwapN: 2323 case Op_CompareAndSwapB: 2324 case Op_CompareAndSwapS: 2325 case Op_CompareAndSwapI: 2326 case Op_CompareAndSwapL: 2327 case Op_CompareAndSwapP: 2328 case Op_CompareAndSwapN: { // Convert trinary to binary-tree 2329 Node *newval = n->in(MemNode::ValueIn ); 2330 Node *oldval = n->in(LoadStoreConditionalNode::ExpectedIn); 2331 Node *pair = new BinaryNode( oldval, newval ); 2332 n->set_req(MemNode::ValueIn,pair); 2333 n->del_req(LoadStoreConditionalNode::ExpectedIn); 2334 break; 2335 } 2336 case Op_CMoveD: // Convert trinary to binary-tree 2337 case Op_CMoveF: 2338 case Op_CMoveI: 2339 case Op_CMoveL: 2340 case Op_CMoveN: 2341 case Op_CMoveP: 2342 case Op_CMoveVF: 2343 case Op_CMoveVD: { 2344 // Restructure into a binary tree for Matching. It's possible that 2345 // we could move this code up next to the graph reshaping for IfNodes 2346 // or vice-versa, but I do not want to debug this for Ladybird. 2347 // 10/2/2000 CNC. 2348 Node *pair1 = new BinaryNode(n->in(1),n->in(1)->in(1)); 2349 n->set_req(1,pair1); 2350 Node *pair2 = new BinaryNode(n->in(2),n->in(3)); 2351 n->set_req(2,pair2); 2352 n->del_req(3); 2353 break; 2354 } 2355 case Op_LoopLimit: { 2356 Node *pair1 = new BinaryNode(n->in(1),n->in(2)); 2357 n->set_req(1,pair1); 2358 n->set_req(2,n->in(3)); 2359 n->del_req(3); 2360 break; 2361 } 2362 case Op_StrEquals: 2363 case Op_StrIndexOfChar: { 2364 Node *pair1 = new BinaryNode(n->in(2),n->in(3)); 2365 n->set_req(2,pair1); 2366 n->set_req(3,n->in(4)); 2367 n->del_req(4); 2368 break; 2369 } 2370 case Op_StrComp: 2371 case Op_StrIndexOf: { 2372 Node *pair1 = new BinaryNode(n->in(2),n->in(3)); 2373 n->set_req(2,pair1); 2374 Node *pair2 = new BinaryNode(n->in(4),n->in(5)); 2375 n->set_req(3,pair2); 2376 n->del_req(5); 2377 n->del_req(4); 2378 break; 2379 } 2380 case Op_StrCompressedCopy: 2381 case Op_StrInflatedCopy: 2382 case Op_EncodeISOArray: { 2383 // Restructure into a binary tree for Matching. 2384 Node* pair = new BinaryNode(n->in(3), n->in(4)); 2385 n->set_req(3, pair); 2386 n->del_req(4); 2387 break; 2388 } 2389 case Op_FmaD: 2390 case Op_FmaF: 2391 case Op_FmaVD: 2392 case Op_FmaVF: { 2393 // Restructure into a binary tree for Matching. 2394 Node* pair = new BinaryNode(n->in(1), n->in(2)); 2395 n->set_req(2, pair); 2396 n->set_req(1, n->in(3)); 2397 n->del_req(3); 2398 break; 2399 } 2400 case Op_ClearArray: { 2401 Node* pair = new BinaryNode(n->in(2), n->in(3)); 2402 n->set_req(2, pair); 2403 n->set_req(3, n->in(4)); 2404 n->del_req(4); 2405 break; 2406 } 2407 default: 2408 break; 2409 } 2410 } 2411 else { 2412 ShouldNotReachHere(); 2413 } 2414 } // end of while (mstack.is_nonempty()) 2415 } 2416 2417 #ifdef ASSERT 2418 // machine-independent root to machine-dependent root 2419 void Matcher::dump_old2new_map() { 2420 _old2new_map.dump(); 2421 } 2422 #endif 2423 2424 //---------------------------collect_null_checks------------------------------- 2425 // Find null checks in the ideal graph; write a machine-specific node for 2426 // it. Used by later implicit-null-check handling. Actually collects 2427 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal 2428 // value being tested. 2429 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) { 2430 Node *iff = proj->in(0); 2431 if( iff->Opcode() == Op_If ) { 2432 // During matching If's have Bool & Cmp side-by-side 2433 BoolNode *b = iff->in(1)->as_Bool(); 2434 Node *cmp = iff->in(2); 2435 int opc = cmp->Opcode(); 2436 if (opc != Op_CmpP && opc != Op_CmpN) return; 2437 2438 const Type* ct = cmp->in(2)->bottom_type(); 2439 if (ct == TypePtr::NULL_PTR || 2440 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) { 2441 2442 bool push_it = false; 2443 if( proj->Opcode() == Op_IfTrue ) { 2444 #ifndef PRODUCT 2445 extern int all_null_checks_found; 2446 all_null_checks_found++; 2447 #endif 2448 if( b->_test._test == BoolTest::ne ) { 2449 push_it = true; 2450 } 2451 } else { 2452 assert( proj->Opcode() == Op_IfFalse, "" ); 2453 if( b->_test._test == BoolTest::eq ) { 2454 push_it = true; 2455 } 2456 } 2457 if( push_it ) { 2458 _null_check_tests.push(proj); 2459 Node* val = cmp->in(1); 2460 #ifdef _LP64 2461 if (val->bottom_type()->isa_narrowoop() && 2462 !Matcher::narrow_oop_use_complex_address()) { 2463 // 2464 // Look for DecodeN node which should be pinned to orig_proj. 2465 // On platforms (Sparc) which can not handle 2 adds 2466 // in addressing mode we have to keep a DecodeN node and 2467 // use it to do implicit NULL check in address. 2468 // 2469 // DecodeN node was pinned to non-null path (orig_proj) during 2470 // CastPP transformation in final_graph_reshaping_impl(). 2471 // 2472 uint cnt = orig_proj->outcnt(); 2473 for (uint i = 0; i < orig_proj->outcnt(); i++) { 2474 Node* d = orig_proj->raw_out(i); 2475 if (d->is_DecodeN() && d->in(1) == val) { 2476 val = d; 2477 val->set_req(0, NULL); // Unpin now. 2478 // Mark this as special case to distinguish from 2479 // a regular case: CmpP(DecodeN, NULL). 2480 val = (Node*)(((intptr_t)val) | 1); 2481 break; 2482 } 2483 } 2484 } 2485 #endif 2486 _null_check_tests.push(val); 2487 } 2488 } 2489 } 2490 } 2491 2492 //---------------------------validate_null_checks------------------------------ 2493 // Its possible that the value being NULL checked is not the root of a match 2494 // tree. If so, I cannot use the value in an implicit null check. 2495 void Matcher::validate_null_checks( ) { 2496 uint cnt = _null_check_tests.size(); 2497 for( uint i=0; i < cnt; i+=2 ) { 2498 Node *test = _null_check_tests[i]; 2499 Node *val = _null_check_tests[i+1]; 2500 bool is_decoden = ((intptr_t)val) & 1; 2501 val = (Node*)(((intptr_t)val) & ~1); 2502 if (has_new_node(val)) { 2503 Node* new_val = new_node(val); 2504 if (is_decoden) { 2505 assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity"); 2506 // Note: new_val may have a control edge if 2507 // the original ideal node DecodeN was matched before 2508 // it was unpinned in Matcher::collect_null_checks(). 2509 // Unpin the mach node and mark it. 2510 new_val->set_req(0, NULL); 2511 new_val = (Node*)(((intptr_t)new_val) | 1); 2512 } 2513 // Is a match-tree root, so replace with the matched value 2514 _null_check_tests.map(i+1, new_val); 2515 } else { 2516 // Yank from candidate list 2517 _null_check_tests.map(i+1,_null_check_tests[--cnt]); 2518 _null_check_tests.map(i,_null_check_tests[--cnt]); 2519 _null_check_tests.pop(); 2520 _null_check_tests.pop(); 2521 i-=2; 2522 } 2523 } 2524 } 2525 2526 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or 2527 // atomic instruction acting as a store_load barrier without any 2528 // intervening volatile load, and thus we don't need a barrier here. 2529 // We retain the Node to act as a compiler ordering barrier. 2530 bool Matcher::post_store_load_barrier(const Node* vmb) { 2531 Compile* C = Compile::current(); 2532 assert(vmb->is_MemBar(), ""); 2533 assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, ""); 2534 const MemBarNode* membar = vmb->as_MemBar(); 2535 2536 // Get the Ideal Proj node, ctrl, that can be used to iterate forward 2537 Node* ctrl = NULL; 2538 for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) { 2539 Node* p = membar->fast_out(i); 2540 assert(p->is_Proj(), "only projections here"); 2541 if ((p->as_Proj()->_con == TypeFunc::Control) && 2542 !C->node_arena()->contains(p)) { // Unmatched old-space only 2543 ctrl = p; 2544 break; 2545 } 2546 } 2547 assert((ctrl != NULL), "missing control projection"); 2548 2549 for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) { 2550 Node *x = ctrl->fast_out(j); 2551 int xop = x->Opcode(); 2552 2553 // We don't need current barrier if we see another or a lock 2554 // before seeing volatile load. 2555 // 2556 // Op_Fastunlock previously appeared in the Op_* list below. 2557 // With the advent of 1-0 lock operations we're no longer guaranteed 2558 // that a monitor exit operation contains a serializing instruction. 2559 2560 if (xop == Op_MemBarVolatile || 2561 xop == Op_CompareAndExchangeB || 2562 xop == Op_CompareAndExchangeS || 2563 xop == Op_CompareAndExchangeI || 2564 xop == Op_CompareAndExchangeL || 2565 xop == Op_CompareAndExchangeP || 2566 xop == Op_CompareAndExchangeN || 2567 xop == Op_WeakCompareAndSwapB || 2568 xop == Op_WeakCompareAndSwapS || 2569 xop == Op_WeakCompareAndSwapL || 2570 xop == Op_WeakCompareAndSwapP || 2571 xop == Op_WeakCompareAndSwapN || 2572 xop == Op_WeakCompareAndSwapI || 2573 xop == Op_CompareAndSwapB || 2574 xop == Op_CompareAndSwapS || 2575 xop == Op_CompareAndSwapL || 2576 xop == Op_CompareAndSwapP || 2577 xop == Op_CompareAndSwapN || 2578 xop == Op_CompareAndSwapI) { 2579 return true; 2580 } 2581 2582 // Op_FastLock previously appeared in the Op_* list above. 2583 // With biased locking we're no longer guaranteed that a monitor 2584 // enter operation contains a serializing instruction. 2585 if ((xop == Op_FastLock) && !UseBiasedLocking) { 2586 return true; 2587 } 2588 2589 if (x->is_MemBar()) { 2590 // We must retain this membar if there is an upcoming volatile 2591 // load, which will be followed by acquire membar. 2592 if (xop == Op_MemBarAcquire || xop == Op_LoadFence) { 2593 return false; 2594 } else { 2595 // For other kinds of barriers, check by pretending we 2596 // are them, and seeing if we can be removed. 2597 return post_store_load_barrier(x->as_MemBar()); 2598 } 2599 } 2600 2601 // probably not necessary to check for these 2602 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) { 2603 return false; 2604 } 2605 } 2606 return false; 2607 } 2608 2609 // Check whether node n is a branch to an uncommon trap that we could 2610 // optimize as test with very high branch costs in case of going to 2611 // the uncommon trap. The code must be able to be recompiled to use 2612 // a cheaper test. 2613 bool Matcher::branches_to_uncommon_trap(const Node *n) { 2614 // Don't do it for natives, adapters, or runtime stubs 2615 Compile *C = Compile::current(); 2616 if (!C->is_method_compilation()) return false; 2617 2618 assert(n->is_If(), "You should only call this on if nodes."); 2619 IfNode *ifn = n->as_If(); 2620 2621 Node *ifFalse = NULL; 2622 for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) { 2623 if (ifn->fast_out(i)->is_IfFalse()) { 2624 ifFalse = ifn->fast_out(i); 2625 break; 2626 } 2627 } 2628 assert(ifFalse, "An If should have an ifFalse. Graph is broken."); 2629 2630 Node *reg = ifFalse; 2631 int cnt = 4; // We must protect against cycles. Limit to 4 iterations. 2632 // Alternatively use visited set? Seems too expensive. 2633 while (reg != NULL && cnt > 0) { 2634 CallNode *call = NULL; 2635 RegionNode *nxt_reg = NULL; 2636 for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) { 2637 Node *o = reg->fast_out(i); 2638 if (o->is_Call()) { 2639 call = o->as_Call(); 2640 } 2641 if (o->is_Region()) { 2642 nxt_reg = o->as_Region(); 2643 } 2644 } 2645 2646 if (call && 2647 call->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 2648 const Type* trtype = call->in(TypeFunc::Parms)->bottom_type(); 2649 if (trtype->isa_int() && trtype->is_int()->is_con()) { 2650 jint tr_con = trtype->is_int()->get_con(); 2651 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 2652 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 2653 assert((int)reason < (int)BitsPerInt, "recode bit map"); 2654 2655 if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason) 2656 && action != Deoptimization::Action_none) { 2657 // This uncommon trap is sure to recompile, eventually. 2658 // When that happens, C->too_many_traps will prevent 2659 // this transformation from happening again. 2660 return true; 2661 } 2662 } 2663 } 2664 2665 reg = nxt_reg; 2666 cnt--; 2667 } 2668 2669 return false; 2670 } 2671 2672 //============================================================================= 2673 //---------------------------State--------------------------------------------- 2674 State::State(void) { 2675 #ifdef ASSERT 2676 _id = 0; 2677 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2678 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2679 //memset(_cost, -1, sizeof(_cost)); 2680 //memset(_rule, -1, sizeof(_rule)); 2681 #endif 2682 memset(_valid, 0, sizeof(_valid)); 2683 } 2684 2685 #ifdef ASSERT 2686 State::~State() { 2687 _id = 99; 2688 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2689 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2690 memset(_cost, -3, sizeof(_cost)); 2691 memset(_rule, -3, sizeof(_rule)); 2692 } 2693 #endif 2694 2695 #ifndef PRODUCT 2696 //---------------------------dump---------------------------------------------- 2697 void State::dump() { 2698 tty->print("\n"); 2699 dump(0); 2700 } 2701 2702 void State::dump(int depth) { 2703 for( int j = 0; j < depth; j++ ) 2704 tty->print(" "); 2705 tty->print("--N: "); 2706 _leaf->dump(); 2707 uint i; 2708 for( i = 0; i < _LAST_MACH_OPER; i++ ) 2709 // Check for valid entry 2710 if( valid(i) ) { 2711 for( int j = 0; j < depth; j++ ) 2712 tty->print(" "); 2713 assert(_cost[i] != max_juint, "cost must be a valid value"); 2714 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule"); 2715 tty->print_cr("%s %d %s", 2716 ruleName[i], _cost[i], ruleName[_rule[i]] ); 2717 } 2718 tty->cr(); 2719 2720 for( i=0; i<2; i++ ) 2721 if( _kids[i] ) 2722 _kids[i]->dump(depth+1); 2723 } 2724 #endif