1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "memory/allocation.inline.hpp" 27 #include "opto/ad.hpp" 28 #include "opto/addnode.hpp" 29 #include "opto/callnode.hpp" 30 #include "opto/idealGraphPrinter.hpp" 31 #include "opto/matcher.hpp" 32 #include "opto/memnode.hpp" 33 #include "opto/movenode.hpp" 34 #include "opto/opcodes.hpp" 35 #include "opto/regmask.hpp" 36 #include "opto/rootnode.hpp" 37 #include "opto/runtime.hpp" 38 #include "opto/type.hpp" 39 #include "opto/vectornode.hpp" 40 #include "runtime/os.hpp" 41 #include "runtime/sharedRuntime.hpp" 42 43 OptoReg::Name OptoReg::c_frame_pointer; 44 45 const RegMask *Matcher::idealreg2regmask[_last_machine_leaf]; 46 RegMask Matcher::mreg2regmask[_last_Mach_Reg]; 47 RegMask Matcher::STACK_ONLY_mask; 48 RegMask Matcher::c_frame_ptr_mask; 49 const uint Matcher::_begin_rematerialize = _BEGIN_REMATERIALIZE; 50 const uint Matcher::_end_rematerialize = _END_REMATERIALIZE; 51 52 //---------------------------Matcher------------------------------------------- 53 Matcher::Matcher() 54 : PhaseTransform( Phase::Ins_Select ), 55 #ifdef ASSERT 56 _old2new_map(C->comp_arena()), 57 _new2old_map(C->comp_arena()), 58 #endif 59 _shared_nodes(C->comp_arena()), 60 _reduceOp(reduceOp), _leftOp(leftOp), _rightOp(rightOp), 61 _swallowed(swallowed), 62 _begin_inst_chain_rule(_BEGIN_INST_CHAIN_RULE), 63 _end_inst_chain_rule(_END_INST_CHAIN_RULE), 64 _must_clone(must_clone), 65 _register_save_policy(register_save_policy), 66 _c_reg_save_policy(c_reg_save_policy), 67 _register_save_type(register_save_type), 68 _ruleName(ruleName), 69 _allocation_started(false), 70 _states_arena(Chunk::medium_size), 71 _visited(&_states_arena), 72 _shared(&_states_arena), 73 _dontcare(&_states_arena) { 74 C->set_matcher(this); 75 76 idealreg2spillmask [Op_RegI] = NULL; 77 idealreg2spillmask [Op_RegN] = NULL; 78 idealreg2spillmask [Op_RegL] = NULL; 79 idealreg2spillmask [Op_RegF] = NULL; 80 idealreg2spillmask [Op_RegD] = NULL; 81 idealreg2spillmask [Op_RegP] = NULL; 82 idealreg2spillmask [Op_VecS] = NULL; 83 idealreg2spillmask [Op_VecD] = NULL; 84 idealreg2spillmask [Op_VecX] = NULL; 85 idealreg2spillmask [Op_VecY] = NULL; 86 87 idealreg2debugmask [Op_RegI] = NULL; 88 idealreg2debugmask [Op_RegN] = NULL; 89 idealreg2debugmask [Op_RegL] = NULL; 90 idealreg2debugmask [Op_RegF] = NULL; 91 idealreg2debugmask [Op_RegD] = NULL; 92 idealreg2debugmask [Op_RegP] = NULL; 93 idealreg2debugmask [Op_VecS] = NULL; 94 idealreg2debugmask [Op_VecD] = NULL; 95 idealreg2debugmask [Op_VecX] = NULL; 96 idealreg2debugmask [Op_VecY] = NULL; 97 98 idealreg2mhdebugmask[Op_RegI] = NULL; 99 idealreg2mhdebugmask[Op_RegN] = NULL; 100 idealreg2mhdebugmask[Op_RegL] = NULL; 101 idealreg2mhdebugmask[Op_RegF] = NULL; 102 idealreg2mhdebugmask[Op_RegD] = NULL; 103 idealreg2mhdebugmask[Op_RegP] = NULL; 104 idealreg2mhdebugmask[Op_VecS] = NULL; 105 idealreg2mhdebugmask[Op_VecD] = NULL; 106 idealreg2mhdebugmask[Op_VecX] = NULL; 107 idealreg2mhdebugmask[Op_VecY] = NULL; 108 109 debug_only(_mem_node = NULL;) // Ideal memory node consumed by mach node 110 } 111 112 //------------------------------warp_incoming_stk_arg------------------------ 113 // This warps a VMReg into an OptoReg::Name 114 OptoReg::Name Matcher::warp_incoming_stk_arg( VMReg reg ) { 115 OptoReg::Name warped; 116 if( reg->is_stack() ) { // Stack slot argument? 117 warped = OptoReg::add(_old_SP, reg->reg2stack() ); 118 warped = OptoReg::add(warped, C->out_preserve_stack_slots()); 119 if( warped >= _in_arg_limit ) 120 _in_arg_limit = OptoReg::add(warped, 1); // Bump max stack slot seen 121 if (!RegMask::can_represent_arg(warped)) { 122 // the compiler cannot represent this method's calling sequence 123 C->record_method_not_compilable_all_tiers("unsupported incoming calling sequence"); 124 return OptoReg::Bad; 125 } 126 return warped; 127 } 128 return OptoReg::as_OptoReg(reg); 129 } 130 131 //---------------------------compute_old_SP------------------------------------ 132 OptoReg::Name Compile::compute_old_SP() { 133 int fixed = fixed_slots(); 134 int preserve = in_preserve_stack_slots(); 135 return OptoReg::stack2reg(round_to(fixed + preserve, Matcher::stack_alignment_in_slots())); 136 } 137 138 139 140 #ifdef ASSERT 141 void Matcher::verify_new_nodes_only(Node* xroot) { 142 // Make sure that the new graph only references new nodes 143 ResourceMark rm; 144 Unique_Node_List worklist; 145 VectorSet visited(Thread::current()->resource_area()); 146 worklist.push(xroot); 147 while (worklist.size() > 0) { 148 Node* n = worklist.pop(); 149 visited <<= n->_idx; 150 assert(C->node_arena()->contains(n), "dead node"); 151 for (uint j = 0; j < n->req(); j++) { 152 Node* in = n->in(j); 153 if (in != NULL) { 154 assert(C->node_arena()->contains(in), "dead node"); 155 if (!visited.test(in->_idx)) { 156 worklist.push(in); 157 } 158 } 159 } 160 } 161 } 162 #endif 163 164 165 //---------------------------match--------------------------------------------- 166 void Matcher::match( ) { 167 if( MaxLabelRootDepth < 100 ) { // Too small? 168 assert(false, "invalid MaxLabelRootDepth, increase it to 100 minimum"); 169 MaxLabelRootDepth = 100; 170 } 171 // One-time initialization of some register masks. 172 init_spill_mask( C->root()->in(1) ); 173 _return_addr_mask = return_addr(); 174 #ifdef _LP64 175 // Pointers take 2 slots in 64-bit land 176 _return_addr_mask.Insert(OptoReg::add(return_addr(),1)); 177 #endif 178 179 // Map a Java-signature return type into return register-value 180 // machine registers for 0, 1 and 2 returned values. 181 const TypeTuple *range = C->tf()->range(); 182 if( range->cnt() > TypeFunc::Parms ) { // If not a void function 183 // Get ideal-register return type 184 int ireg = range->field_at(TypeFunc::Parms)->ideal_reg(); 185 // Get machine return register 186 uint sop = C->start()->Opcode(); 187 OptoRegPair regs = return_value(ireg, false); 188 189 // And mask for same 190 _return_value_mask = RegMask(regs.first()); 191 if( OptoReg::is_valid(regs.second()) ) 192 _return_value_mask.Insert(regs.second()); 193 } 194 195 // --------------- 196 // Frame Layout 197 198 // Need the method signature to determine the incoming argument types, 199 // because the types determine which registers the incoming arguments are 200 // in, and this affects the matched code. 201 const TypeTuple *domain = C->tf()->domain(); 202 uint argcnt = domain->cnt() - TypeFunc::Parms; 203 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 204 VMRegPair *vm_parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 205 _parm_regs = NEW_RESOURCE_ARRAY( OptoRegPair, argcnt ); 206 _calling_convention_mask = NEW_RESOURCE_ARRAY( RegMask, argcnt ); 207 uint i; 208 for( i = 0; i<argcnt; i++ ) { 209 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 210 } 211 212 // Pass array of ideal registers and length to USER code (from the AD file) 213 // that will convert this to an array of register numbers. 214 const StartNode *start = C->start(); 215 start->calling_convention( sig_bt, vm_parm_regs, argcnt ); 216 #ifdef ASSERT 217 // Sanity check users' calling convention. Real handy while trying to 218 // get the initial port correct. 219 { for (uint i = 0; i<argcnt; i++) { 220 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 221 assert(domain->field_at(i+TypeFunc::Parms)==Type::HALF, "only allowed on halve" ); 222 _parm_regs[i].set_bad(); 223 continue; 224 } 225 VMReg parm_reg = vm_parm_regs[i].first(); 226 assert(parm_reg->is_valid(), "invalid arg?"); 227 if (parm_reg->is_reg()) { 228 OptoReg::Name opto_parm_reg = OptoReg::as_OptoReg(parm_reg); 229 assert(can_be_java_arg(opto_parm_reg) || 230 C->stub_function() == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C) || 231 opto_parm_reg == inline_cache_reg(), 232 "parameters in register must be preserved by runtime stubs"); 233 } 234 for (uint j = 0; j < i; j++) { 235 assert(parm_reg != vm_parm_regs[j].first(), 236 "calling conv. must produce distinct regs"); 237 } 238 } 239 } 240 #endif 241 242 // Do some initial frame layout. 243 244 // Compute the old incoming SP (may be called FP) as 245 // OptoReg::stack0() + locks + in_preserve_stack_slots + pad2. 246 _old_SP = C->compute_old_SP(); 247 assert( is_even(_old_SP), "must be even" ); 248 249 // Compute highest incoming stack argument as 250 // _old_SP + out_preserve_stack_slots + incoming argument size. 251 _in_arg_limit = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 252 assert( is_even(_in_arg_limit), "out_preserve must be even" ); 253 for( i = 0; i < argcnt; i++ ) { 254 // Permit args to have no register 255 _calling_convention_mask[i].Clear(); 256 if( !vm_parm_regs[i].first()->is_valid() && !vm_parm_regs[i].second()->is_valid() ) { 257 continue; 258 } 259 // calling_convention returns stack arguments as a count of 260 // slots beyond OptoReg::stack0()/VMRegImpl::stack0. We need to convert this to 261 // the allocators point of view, taking into account all the 262 // preserve area, locks & pad2. 263 264 OptoReg::Name reg1 = warp_incoming_stk_arg(vm_parm_regs[i].first()); 265 if( OptoReg::is_valid(reg1)) 266 _calling_convention_mask[i].Insert(reg1); 267 268 OptoReg::Name reg2 = warp_incoming_stk_arg(vm_parm_regs[i].second()); 269 if( OptoReg::is_valid(reg2)) 270 _calling_convention_mask[i].Insert(reg2); 271 272 // Saved biased stack-slot register number 273 _parm_regs[i].set_pair(reg2, reg1); 274 } 275 276 // Finally, make sure the incoming arguments take up an even number of 277 // words, in case the arguments or locals need to contain doubleword stack 278 // slots. The rest of the system assumes that stack slot pairs (in 279 // particular, in the spill area) which look aligned will in fact be 280 // aligned relative to the stack pointer in the target machine. Double 281 // stack slots will always be allocated aligned. 282 _new_SP = OptoReg::Name(round_to(_in_arg_limit, RegMask::SlotsPerLong)); 283 284 // Compute highest outgoing stack argument as 285 // _new_SP + out_preserve_stack_slots + max(outgoing argument size). 286 _out_arg_limit = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 287 assert( is_even(_out_arg_limit), "out_preserve must be even" ); 288 289 if (!RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1))) { 290 // the compiler cannot represent this method's calling sequence 291 C->record_method_not_compilable("must be able to represent all call arguments in reg mask"); 292 } 293 294 if (C->failing()) return; // bailed out on incoming arg failure 295 296 // --------------- 297 // Collect roots of matcher trees. Every node for which 298 // _shared[_idx] is cleared is guaranteed to not be shared, and thus 299 // can be a valid interior of some tree. 300 find_shared( C->root() ); 301 find_shared( C->top() ); 302 303 C->print_method(PHASE_BEFORE_MATCHING); 304 305 // Create new ideal node ConP #NULL even if it does exist in old space 306 // to avoid false sharing if the corresponding mach node is not used. 307 // The corresponding mach node is only used in rare cases for derived 308 // pointers. 309 Node* new_ideal_null = ConNode::make(C, TypePtr::NULL_PTR); 310 311 // Swap out to old-space; emptying new-space 312 Arena *old = C->node_arena()->move_contents(C->old_arena()); 313 314 // Save debug and profile information for nodes in old space: 315 _old_node_note_array = C->node_note_array(); 316 if (_old_node_note_array != NULL) { 317 C->set_node_note_array(new(C->comp_arena()) GrowableArray<Node_Notes*> 318 (C->comp_arena(), _old_node_note_array->length(), 319 0, NULL)); 320 } 321 322 // Pre-size the new_node table to avoid the need for range checks. 323 grow_new_node_array(C->unique()); 324 325 // Reset node counter so MachNodes start with _idx at 0 326 int nodes = C->unique(); // save value 327 C->set_unique(0); 328 C->reset_dead_node_list(); 329 330 // Recursively match trees from old space into new space. 331 // Correct leaves of new-space Nodes; they point to old-space. 332 _visited.Clear(); // Clear visit bits for xform call 333 C->set_cached_top_node(xform( C->top(), nodes )); 334 if (!C->failing()) { 335 Node* xroot = xform( C->root(), 1 ); 336 if (xroot == NULL) { 337 Matcher::soft_match_failure(); // recursive matching process failed 338 C->record_method_not_compilable("instruction match failed"); 339 } else { 340 // During matching shared constants were attached to C->root() 341 // because xroot wasn't available yet, so transfer the uses to 342 // the xroot. 343 for( DUIterator_Fast jmax, j = C->root()->fast_outs(jmax); j < jmax; j++ ) { 344 Node* n = C->root()->fast_out(j); 345 if (C->node_arena()->contains(n)) { 346 assert(n->in(0) == C->root(), "should be control user"); 347 n->set_req(0, xroot); 348 --j; 349 --jmax; 350 } 351 } 352 353 // Generate new mach node for ConP #NULL 354 assert(new_ideal_null != NULL, "sanity"); 355 _mach_null = match_tree(new_ideal_null); 356 // Don't set control, it will confuse GCM since there are no uses. 357 // The control will be set when this node is used first time 358 // in find_base_for_derived(). 359 assert(_mach_null != NULL, ""); 360 361 C->set_root(xroot->is_Root() ? xroot->as_Root() : NULL); 362 363 #ifdef ASSERT 364 verify_new_nodes_only(xroot); 365 #endif 366 } 367 } 368 if (C->top() == NULL || C->root() == NULL) { 369 C->record_method_not_compilable("graph lost"); // %%% cannot happen? 370 } 371 if (C->failing()) { 372 // delete old; 373 old->destruct_contents(); 374 return; 375 } 376 assert( C->top(), "" ); 377 assert( C->root(), "" ); 378 validate_null_checks(); 379 380 // Now smoke old-space 381 NOT_DEBUG( old->destruct_contents() ); 382 383 // ------------------------ 384 // Set up save-on-entry registers 385 Fixup_Save_On_Entry( ); 386 } 387 388 389 //------------------------------Fixup_Save_On_Entry---------------------------- 390 // The stated purpose of this routine is to take care of save-on-entry 391 // registers. However, the overall goal of the Match phase is to convert into 392 // machine-specific instructions which have RegMasks to guide allocation. 393 // So what this procedure really does is put a valid RegMask on each input 394 // to the machine-specific variations of all Return, TailCall and Halt 395 // instructions. It also adds edgs to define the save-on-entry values (and of 396 // course gives them a mask). 397 398 static RegMask *init_input_masks( uint size, RegMask &ret_adr, RegMask &fp ) { 399 RegMask *rms = NEW_RESOURCE_ARRAY( RegMask, size ); 400 // Do all the pre-defined register masks 401 rms[TypeFunc::Control ] = RegMask::Empty; 402 rms[TypeFunc::I_O ] = RegMask::Empty; 403 rms[TypeFunc::Memory ] = RegMask::Empty; 404 rms[TypeFunc::ReturnAdr] = ret_adr; 405 rms[TypeFunc::FramePtr ] = fp; 406 return rms; 407 } 408 409 //---------------------------init_first_stack_mask----------------------------- 410 // Create the initial stack mask used by values spilling to the stack. 411 // Disallow any debug info in outgoing argument areas by setting the 412 // initial mask accordingly. 413 void Matcher::init_first_stack_mask() { 414 415 // Allocate storage for spill masks as masks for the appropriate load type. 416 RegMask *rms = (RegMask*)C->comp_arena()->Amalloc_D(sizeof(RegMask) * (3*6+4)); 417 418 idealreg2spillmask [Op_RegN] = &rms[0]; 419 idealreg2spillmask [Op_RegI] = &rms[1]; 420 idealreg2spillmask [Op_RegL] = &rms[2]; 421 idealreg2spillmask [Op_RegF] = &rms[3]; 422 idealreg2spillmask [Op_RegD] = &rms[4]; 423 idealreg2spillmask [Op_RegP] = &rms[5]; 424 425 idealreg2debugmask [Op_RegN] = &rms[6]; 426 idealreg2debugmask [Op_RegI] = &rms[7]; 427 idealreg2debugmask [Op_RegL] = &rms[8]; 428 idealreg2debugmask [Op_RegF] = &rms[9]; 429 idealreg2debugmask [Op_RegD] = &rms[10]; 430 idealreg2debugmask [Op_RegP] = &rms[11]; 431 432 idealreg2mhdebugmask[Op_RegN] = &rms[12]; 433 idealreg2mhdebugmask[Op_RegI] = &rms[13]; 434 idealreg2mhdebugmask[Op_RegL] = &rms[14]; 435 idealreg2mhdebugmask[Op_RegF] = &rms[15]; 436 idealreg2mhdebugmask[Op_RegD] = &rms[16]; 437 idealreg2mhdebugmask[Op_RegP] = &rms[17]; 438 439 idealreg2spillmask [Op_VecS] = &rms[18]; 440 idealreg2spillmask [Op_VecD] = &rms[19]; 441 idealreg2spillmask [Op_VecX] = &rms[20]; 442 idealreg2spillmask [Op_VecY] = &rms[21]; 443 444 OptoReg::Name i; 445 446 // At first, start with the empty mask 447 C->FIRST_STACK_mask().Clear(); 448 449 // Add in the incoming argument area 450 OptoReg::Name init_in = OptoReg::add(_old_SP, C->out_preserve_stack_slots()); 451 for (i = init_in; i < _in_arg_limit; i = OptoReg::add(i,1)) { 452 C->FIRST_STACK_mask().Insert(i); 453 } 454 // Add in all bits past the outgoing argument area 455 guarantee(RegMask::can_represent_arg(OptoReg::add(_out_arg_limit,-1)), 456 "must be able to represent all call arguments in reg mask"); 457 OptoReg::Name init = _out_arg_limit; 458 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) { 459 C->FIRST_STACK_mask().Insert(i); 460 } 461 // Finally, set the "infinite stack" bit. 462 C->FIRST_STACK_mask().set_AllStack(); 463 464 // Make spill masks. Registers for their class, plus FIRST_STACK_mask. 465 RegMask aligned_stack_mask = C->FIRST_STACK_mask(); 466 // Keep spill masks aligned. 467 aligned_stack_mask.clear_to_pairs(); 468 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 469 470 *idealreg2spillmask[Op_RegP] = *idealreg2regmask[Op_RegP]; 471 #ifdef _LP64 472 *idealreg2spillmask[Op_RegN] = *idealreg2regmask[Op_RegN]; 473 idealreg2spillmask[Op_RegN]->OR(C->FIRST_STACK_mask()); 474 idealreg2spillmask[Op_RegP]->OR(aligned_stack_mask); 475 #else 476 idealreg2spillmask[Op_RegP]->OR(C->FIRST_STACK_mask()); 477 #endif 478 *idealreg2spillmask[Op_RegI] = *idealreg2regmask[Op_RegI]; 479 idealreg2spillmask[Op_RegI]->OR(C->FIRST_STACK_mask()); 480 *idealreg2spillmask[Op_RegL] = *idealreg2regmask[Op_RegL]; 481 idealreg2spillmask[Op_RegL]->OR(aligned_stack_mask); 482 *idealreg2spillmask[Op_RegF] = *idealreg2regmask[Op_RegF]; 483 idealreg2spillmask[Op_RegF]->OR(C->FIRST_STACK_mask()); 484 *idealreg2spillmask[Op_RegD] = *idealreg2regmask[Op_RegD]; 485 idealreg2spillmask[Op_RegD]->OR(aligned_stack_mask); 486 487 if (Matcher::vector_size_supported(T_BYTE,4)) { 488 *idealreg2spillmask[Op_VecS] = *idealreg2regmask[Op_VecS]; 489 idealreg2spillmask[Op_VecS]->OR(C->FIRST_STACK_mask()); 490 } 491 if (Matcher::vector_size_supported(T_FLOAT,2)) { 492 // For VecD we need dual alignment and 8 bytes (2 slots) for spills. 493 // RA guarantees such alignment since it is needed for Double and Long values. 494 *idealreg2spillmask[Op_VecD] = *idealreg2regmask[Op_VecD]; 495 idealreg2spillmask[Op_VecD]->OR(aligned_stack_mask); 496 } 497 if (Matcher::vector_size_supported(T_FLOAT,4)) { 498 // For VecX we need quadro alignment and 16 bytes (4 slots) for spills. 499 // 500 // RA can use input arguments stack slots for spills but until RA 501 // we don't know frame size and offset of input arg stack slots. 502 // 503 // Exclude last input arg stack slots to avoid spilling vectors there 504 // otherwise vector spills could stomp over stack slots in caller frame. 505 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 506 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecX); k++) { 507 aligned_stack_mask.Remove(in); 508 in = OptoReg::add(in, -1); 509 } 510 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecX); 511 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 512 *idealreg2spillmask[Op_VecX] = *idealreg2regmask[Op_VecX]; 513 idealreg2spillmask[Op_VecX]->OR(aligned_stack_mask); 514 } 515 if (Matcher::vector_size_supported(T_FLOAT,8)) { 516 // For VecY we need octo alignment and 32 bytes (8 slots) for spills. 517 OptoReg::Name in = OptoReg::add(_in_arg_limit, -1); 518 for (int k = 1; (in >= init_in) && (k < RegMask::SlotsPerVecY); k++) { 519 aligned_stack_mask.Remove(in); 520 in = OptoReg::add(in, -1); 521 } 522 aligned_stack_mask.clear_to_sets(RegMask::SlotsPerVecY); 523 assert(aligned_stack_mask.is_AllStack(), "should be infinite stack"); 524 *idealreg2spillmask[Op_VecY] = *idealreg2regmask[Op_VecY]; 525 idealreg2spillmask[Op_VecY]->OR(aligned_stack_mask); 526 } 527 if (UseFPUForSpilling) { 528 // This mask logic assumes that the spill operations are 529 // symmetric and that the registers involved are the same size. 530 // On sparc for instance we may have to use 64 bit moves will 531 // kill 2 registers when used with F0-F31. 532 idealreg2spillmask[Op_RegI]->OR(*idealreg2regmask[Op_RegF]); 533 idealreg2spillmask[Op_RegF]->OR(*idealreg2regmask[Op_RegI]); 534 #ifdef _LP64 535 idealreg2spillmask[Op_RegN]->OR(*idealreg2regmask[Op_RegF]); 536 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 537 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 538 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegD]); 539 #else 540 idealreg2spillmask[Op_RegP]->OR(*idealreg2regmask[Op_RegF]); 541 #ifdef ARM 542 // ARM has support for moving 64bit values between a pair of 543 // integer registers and a double register 544 idealreg2spillmask[Op_RegL]->OR(*idealreg2regmask[Op_RegD]); 545 idealreg2spillmask[Op_RegD]->OR(*idealreg2regmask[Op_RegL]); 546 #endif 547 #endif 548 } 549 550 // Make up debug masks. Any spill slot plus callee-save registers. 551 // Caller-save registers are assumed to be trashable by the various 552 // inline-cache fixup routines. 553 *idealreg2debugmask [Op_RegN]= *idealreg2spillmask[Op_RegN]; 554 *idealreg2debugmask [Op_RegI]= *idealreg2spillmask[Op_RegI]; 555 *idealreg2debugmask [Op_RegL]= *idealreg2spillmask[Op_RegL]; 556 *idealreg2debugmask [Op_RegF]= *idealreg2spillmask[Op_RegF]; 557 *idealreg2debugmask [Op_RegD]= *idealreg2spillmask[Op_RegD]; 558 *idealreg2debugmask [Op_RegP]= *idealreg2spillmask[Op_RegP]; 559 560 *idealreg2mhdebugmask[Op_RegN]= *idealreg2spillmask[Op_RegN]; 561 *idealreg2mhdebugmask[Op_RegI]= *idealreg2spillmask[Op_RegI]; 562 *idealreg2mhdebugmask[Op_RegL]= *idealreg2spillmask[Op_RegL]; 563 *idealreg2mhdebugmask[Op_RegF]= *idealreg2spillmask[Op_RegF]; 564 *idealreg2mhdebugmask[Op_RegD]= *idealreg2spillmask[Op_RegD]; 565 *idealreg2mhdebugmask[Op_RegP]= *idealreg2spillmask[Op_RegP]; 566 567 // Prevent stub compilations from attempting to reference 568 // callee-saved registers from debug info 569 bool exclude_soe = !Compile::current()->is_method_compilation(); 570 571 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 572 // registers the caller has to save do not work 573 if( _register_save_policy[i] == 'C' || 574 _register_save_policy[i] == 'A' || 575 (_register_save_policy[i] == 'E' && exclude_soe) ) { 576 idealreg2debugmask [Op_RegN]->Remove(i); 577 idealreg2debugmask [Op_RegI]->Remove(i); // Exclude save-on-call 578 idealreg2debugmask [Op_RegL]->Remove(i); // registers from debug 579 idealreg2debugmask [Op_RegF]->Remove(i); // masks 580 idealreg2debugmask [Op_RegD]->Remove(i); 581 idealreg2debugmask [Op_RegP]->Remove(i); 582 583 idealreg2mhdebugmask[Op_RegN]->Remove(i); 584 idealreg2mhdebugmask[Op_RegI]->Remove(i); 585 idealreg2mhdebugmask[Op_RegL]->Remove(i); 586 idealreg2mhdebugmask[Op_RegF]->Remove(i); 587 idealreg2mhdebugmask[Op_RegD]->Remove(i); 588 idealreg2mhdebugmask[Op_RegP]->Remove(i); 589 } 590 } 591 592 // Subtract the register we use to save the SP for MethodHandle 593 // invokes to from the debug mask. 594 const RegMask save_mask = method_handle_invoke_SP_save_mask(); 595 idealreg2mhdebugmask[Op_RegN]->SUBTRACT(save_mask); 596 idealreg2mhdebugmask[Op_RegI]->SUBTRACT(save_mask); 597 idealreg2mhdebugmask[Op_RegL]->SUBTRACT(save_mask); 598 idealreg2mhdebugmask[Op_RegF]->SUBTRACT(save_mask); 599 idealreg2mhdebugmask[Op_RegD]->SUBTRACT(save_mask); 600 idealreg2mhdebugmask[Op_RegP]->SUBTRACT(save_mask); 601 } 602 603 //---------------------------is_save_on_entry---------------------------------- 604 bool Matcher::is_save_on_entry( int reg ) { 605 return 606 _register_save_policy[reg] == 'E' || 607 _register_save_policy[reg] == 'A' || // Save-on-entry register? 608 // Also save argument registers in the trampolining stubs 609 (C->save_argument_registers() && is_spillable_arg(reg)); 610 } 611 612 //---------------------------Fixup_Save_On_Entry------------------------------- 613 void Matcher::Fixup_Save_On_Entry( ) { 614 init_first_stack_mask(); 615 616 Node *root = C->root(); // Short name for root 617 // Count number of save-on-entry registers. 618 uint soe_cnt = number_of_saved_registers(); 619 uint i; 620 621 // Find the procedure Start Node 622 StartNode *start = C->start(); 623 assert( start, "Expect a start node" ); 624 625 // Save argument registers in the trampolining stubs 626 if( C->save_argument_registers() ) 627 for( i = 0; i < _last_Mach_Reg; i++ ) 628 if( is_spillable_arg(i) ) 629 soe_cnt++; 630 631 // Input RegMask array shared by all Returns. 632 // The type for doubles and longs has a count of 2, but 633 // there is only 1 returned value 634 uint ret_edge_cnt = TypeFunc::Parms + ((C->tf()->range()->cnt() == TypeFunc::Parms) ? 0 : 1); 635 RegMask *ret_rms = init_input_masks( ret_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 636 // Returns have 0 or 1 returned values depending on call signature. 637 // Return register is specified by return_value in the AD file. 638 if (ret_edge_cnt > TypeFunc::Parms) 639 ret_rms[TypeFunc::Parms+0] = _return_value_mask; 640 641 // Input RegMask array shared by all Rethrows. 642 uint reth_edge_cnt = TypeFunc::Parms+1; 643 RegMask *reth_rms = init_input_masks( reth_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 644 // Rethrow takes exception oop only, but in the argument 0 slot. 645 reth_rms[TypeFunc::Parms] = mreg2regmask[find_receiver(false)]; 646 #ifdef _LP64 647 // Need two slots for ptrs in 64-bit land 648 reth_rms[TypeFunc::Parms].Insert(OptoReg::add(OptoReg::Name(find_receiver(false)),1)); 649 #endif 650 651 // Input RegMask array shared by all TailCalls 652 uint tail_call_edge_cnt = TypeFunc::Parms+2; 653 RegMask *tail_call_rms = init_input_masks( tail_call_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 654 655 // Input RegMask array shared by all TailJumps 656 uint tail_jump_edge_cnt = TypeFunc::Parms+2; 657 RegMask *tail_jump_rms = init_input_masks( tail_jump_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 658 659 // TailCalls have 2 returned values (target & moop), whose masks come 660 // from the usual MachNode/MachOper mechanism. Find a sample 661 // TailCall to extract these masks and put the correct masks into 662 // the tail_call_rms array. 663 for( i=1; i < root->req(); i++ ) { 664 MachReturnNode *m = root->in(i)->as_MachReturn(); 665 if( m->ideal_Opcode() == Op_TailCall ) { 666 tail_call_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 667 tail_call_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 668 break; 669 } 670 } 671 672 // TailJumps have 2 returned values (target & ex_oop), whose masks come 673 // from the usual MachNode/MachOper mechanism. Find a sample 674 // TailJump to extract these masks and put the correct masks into 675 // the tail_jump_rms array. 676 for( i=1; i < root->req(); i++ ) { 677 MachReturnNode *m = root->in(i)->as_MachReturn(); 678 if( m->ideal_Opcode() == Op_TailJump ) { 679 tail_jump_rms[TypeFunc::Parms+0] = m->MachNode::in_RegMask(TypeFunc::Parms+0); 680 tail_jump_rms[TypeFunc::Parms+1] = m->MachNode::in_RegMask(TypeFunc::Parms+1); 681 break; 682 } 683 } 684 685 // Input RegMask array shared by all Halts 686 uint halt_edge_cnt = TypeFunc::Parms; 687 RegMask *halt_rms = init_input_masks( halt_edge_cnt + soe_cnt, _return_addr_mask, c_frame_ptr_mask ); 688 689 // Capture the return input masks into each exit flavor 690 for( i=1; i < root->req(); i++ ) { 691 MachReturnNode *exit = root->in(i)->as_MachReturn(); 692 switch( exit->ideal_Opcode() ) { 693 case Op_Return : exit->_in_rms = ret_rms; break; 694 case Op_Rethrow : exit->_in_rms = reth_rms; break; 695 case Op_TailCall : exit->_in_rms = tail_call_rms; break; 696 case Op_TailJump : exit->_in_rms = tail_jump_rms; break; 697 case Op_Halt : exit->_in_rms = halt_rms; break; 698 default : ShouldNotReachHere(); 699 } 700 } 701 702 // Next unused projection number from Start. 703 int proj_cnt = C->tf()->domain()->cnt(); 704 705 // Do all the save-on-entry registers. Make projections from Start for 706 // them, and give them a use at the exit points. To the allocator, they 707 // look like incoming register arguments. 708 for( i = 0; i < _last_Mach_Reg; i++ ) { 709 if( is_save_on_entry(i) ) { 710 711 // Add the save-on-entry to the mask array 712 ret_rms [ ret_edge_cnt] = mreg2regmask[i]; 713 reth_rms [ reth_edge_cnt] = mreg2regmask[i]; 714 tail_call_rms[tail_call_edge_cnt] = mreg2regmask[i]; 715 tail_jump_rms[tail_jump_edge_cnt] = mreg2regmask[i]; 716 // Halts need the SOE registers, but only in the stack as debug info. 717 // A just-prior uncommon-trap or deoptimization will use the SOE regs. 718 halt_rms [ halt_edge_cnt] = *idealreg2spillmask[_register_save_type[i]]; 719 720 Node *mproj; 721 722 // Is this a RegF low half of a RegD? Double up 2 adjacent RegF's 723 // into a single RegD. 724 if( (i&1) == 0 && 725 _register_save_type[i ] == Op_RegF && 726 _register_save_type[i+1] == Op_RegF && 727 is_save_on_entry(i+1) ) { 728 // Add other bit for double 729 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 730 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 731 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 732 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 733 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 734 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegD ); 735 proj_cnt += 2; // Skip 2 for doubles 736 } 737 else if( (i&1) == 1 && // Else check for high half of double 738 _register_save_type[i-1] == Op_RegF && 739 _register_save_type[i ] == Op_RegF && 740 is_save_on_entry(i-1) ) { 741 ret_rms [ ret_edge_cnt] = RegMask::Empty; 742 reth_rms [ reth_edge_cnt] = RegMask::Empty; 743 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 744 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 745 halt_rms [ halt_edge_cnt] = RegMask::Empty; 746 mproj = C->top(); 747 } 748 // Is this a RegI low half of a RegL? Double up 2 adjacent RegI's 749 // into a single RegL. 750 else if( (i&1) == 0 && 751 _register_save_type[i ] == Op_RegI && 752 _register_save_type[i+1] == Op_RegI && 753 is_save_on_entry(i+1) ) { 754 // Add other bit for long 755 ret_rms [ ret_edge_cnt].Insert(OptoReg::Name(i+1)); 756 reth_rms [ reth_edge_cnt].Insert(OptoReg::Name(i+1)); 757 tail_call_rms[tail_call_edge_cnt].Insert(OptoReg::Name(i+1)); 758 tail_jump_rms[tail_jump_edge_cnt].Insert(OptoReg::Name(i+1)); 759 halt_rms [ halt_edge_cnt].Insert(OptoReg::Name(i+1)); 760 mproj = new MachProjNode( start, proj_cnt, ret_rms[ret_edge_cnt], Op_RegL ); 761 proj_cnt += 2; // Skip 2 for longs 762 } 763 else if( (i&1) == 1 && // Else check for high half of long 764 _register_save_type[i-1] == Op_RegI && 765 _register_save_type[i ] == Op_RegI && 766 is_save_on_entry(i-1) ) { 767 ret_rms [ ret_edge_cnt] = RegMask::Empty; 768 reth_rms [ reth_edge_cnt] = RegMask::Empty; 769 tail_call_rms[tail_call_edge_cnt] = RegMask::Empty; 770 tail_jump_rms[tail_jump_edge_cnt] = RegMask::Empty; 771 halt_rms [ halt_edge_cnt] = RegMask::Empty; 772 mproj = C->top(); 773 } else { 774 // Make a projection for it off the Start 775 mproj = new MachProjNode( start, proj_cnt++, ret_rms[ret_edge_cnt], _register_save_type[i] ); 776 } 777 778 ret_edge_cnt ++; 779 reth_edge_cnt ++; 780 tail_call_edge_cnt ++; 781 tail_jump_edge_cnt ++; 782 halt_edge_cnt ++; 783 784 // Add a use of the SOE register to all exit paths 785 for( uint j=1; j < root->req(); j++ ) 786 root->in(j)->add_req(mproj); 787 } // End of if a save-on-entry register 788 } // End of for all machine registers 789 } 790 791 //------------------------------init_spill_mask-------------------------------- 792 void Matcher::init_spill_mask( Node *ret ) { 793 if( idealreg2regmask[Op_RegI] ) return; // One time only init 794 795 OptoReg::c_frame_pointer = c_frame_pointer(); 796 c_frame_ptr_mask = c_frame_pointer(); 797 #ifdef _LP64 798 // pointers are twice as big 799 c_frame_ptr_mask.Insert(OptoReg::add(c_frame_pointer(),1)); 800 #endif 801 802 // Start at OptoReg::stack0() 803 STACK_ONLY_mask.Clear(); 804 OptoReg::Name init = OptoReg::stack2reg(0); 805 // STACK_ONLY_mask is all stack bits 806 OptoReg::Name i; 807 for (i = init; RegMask::can_represent(i); i = OptoReg::add(i,1)) 808 STACK_ONLY_mask.Insert(i); 809 // Also set the "infinite stack" bit. 810 STACK_ONLY_mask.set_AllStack(); 811 812 // Copy the register names over into the shared world 813 for( i=OptoReg::Name(0); i<OptoReg::Name(_last_Mach_Reg); i = OptoReg::add(i,1) ) { 814 // SharedInfo::regName[i] = regName[i]; 815 // Handy RegMasks per machine register 816 mreg2regmask[i].Insert(i); 817 } 818 819 // Grab the Frame Pointer 820 Node *fp = ret->in(TypeFunc::FramePtr); 821 Node *mem = ret->in(TypeFunc::Memory); 822 const TypePtr* atp = TypePtr::BOTTOM; 823 // Share frame pointer while making spill ops 824 set_shared(fp); 825 826 // Compute generic short-offset Loads 827 #ifdef _LP64 828 MachNode *spillCP = match_tree(new LoadNNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 829 #endif 830 MachNode *spillI = match_tree(new LoadINode(NULL,mem,fp,atp,TypeInt::INT,MemNode::unordered)); 831 MachNode *spillL = match_tree(new LoadLNode(NULL,mem,fp,atp,TypeLong::LONG,MemNode::unordered,false)); 832 MachNode *spillF = match_tree(new LoadFNode(NULL,mem,fp,atp,Type::FLOAT,MemNode::unordered)); 833 MachNode *spillD = match_tree(new LoadDNode(NULL,mem,fp,atp,Type::DOUBLE,MemNode::unordered)); 834 MachNode *spillP = match_tree(new LoadPNode(NULL,mem,fp,atp,TypeInstPtr::BOTTOM,MemNode::unordered)); 835 assert(spillI != NULL && spillL != NULL && spillF != NULL && 836 spillD != NULL && spillP != NULL, ""); 837 // Get the ADLC notion of the right regmask, for each basic type. 838 #ifdef _LP64 839 idealreg2regmask[Op_RegN] = &spillCP->out_RegMask(); 840 #endif 841 idealreg2regmask[Op_RegI] = &spillI->out_RegMask(); 842 idealreg2regmask[Op_RegL] = &spillL->out_RegMask(); 843 idealreg2regmask[Op_RegF] = &spillF->out_RegMask(); 844 idealreg2regmask[Op_RegD] = &spillD->out_RegMask(); 845 idealreg2regmask[Op_RegP] = &spillP->out_RegMask(); 846 847 // Vector regmasks. 848 if (Matcher::vector_size_supported(T_BYTE,4)) { 849 TypeVect::VECTS = TypeVect::make(T_BYTE, 4); 850 MachNode *spillVectS = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTS)); 851 idealreg2regmask[Op_VecS] = &spillVectS->out_RegMask(); 852 } 853 if (Matcher::vector_size_supported(T_FLOAT,2)) { 854 MachNode *spillVectD = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTD)); 855 idealreg2regmask[Op_VecD] = &spillVectD->out_RegMask(); 856 } 857 if (Matcher::vector_size_supported(T_FLOAT,4)) { 858 MachNode *spillVectX = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTX)); 859 idealreg2regmask[Op_VecX] = &spillVectX->out_RegMask(); 860 } 861 if (Matcher::vector_size_supported(T_FLOAT,8)) { 862 MachNode *spillVectY = match_tree(new LoadVectorNode(NULL,mem,fp,atp,TypeVect::VECTY)); 863 idealreg2regmask[Op_VecY] = &spillVectY->out_RegMask(); 864 } 865 } 866 867 #ifdef ASSERT 868 static void match_alias_type(Compile* C, Node* n, Node* m) { 869 if (!VerifyAliases) return; // do not go looking for trouble by default 870 const TypePtr* nat = n->adr_type(); 871 const TypePtr* mat = m->adr_type(); 872 int nidx = C->get_alias_index(nat); 873 int midx = C->get_alias_index(mat); 874 // Detune the assert for cases like (AndI 0xFF (LoadB p)). 875 if (nidx == Compile::AliasIdxTop && midx >= Compile::AliasIdxRaw) { 876 for (uint i = 1; i < n->req(); i++) { 877 Node* n1 = n->in(i); 878 const TypePtr* n1at = n1->adr_type(); 879 if (n1at != NULL) { 880 nat = n1at; 881 nidx = C->get_alias_index(n1at); 882 } 883 } 884 } 885 // %%% Kludgery. Instead, fix ideal adr_type methods for all these cases: 886 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxRaw) { 887 switch (n->Opcode()) { 888 case Op_PrefetchRead: 889 case Op_PrefetchWrite: 890 case Op_PrefetchAllocation: 891 nidx = Compile::AliasIdxRaw; 892 nat = TypeRawPtr::BOTTOM; 893 break; 894 } 895 } 896 if (nidx == Compile::AliasIdxRaw && midx == Compile::AliasIdxTop) { 897 switch (n->Opcode()) { 898 case Op_ClearArray: 899 midx = Compile::AliasIdxRaw; 900 mat = TypeRawPtr::BOTTOM; 901 break; 902 } 903 } 904 if (nidx == Compile::AliasIdxTop && midx == Compile::AliasIdxBot) { 905 switch (n->Opcode()) { 906 case Op_Return: 907 case Op_Rethrow: 908 case Op_Halt: 909 case Op_TailCall: 910 case Op_TailJump: 911 nidx = Compile::AliasIdxBot; 912 nat = TypePtr::BOTTOM; 913 break; 914 } 915 } 916 if (nidx == Compile::AliasIdxBot && midx == Compile::AliasIdxTop) { 917 switch (n->Opcode()) { 918 case Op_StrComp: 919 case Op_StrEquals: 920 case Op_StrIndexOf: 921 case Op_AryEq: 922 case Op_MemBarVolatile: 923 case Op_MemBarCPUOrder: // %%% these ideals should have narrower adr_type? 924 case Op_EncodeISOArray: 925 nidx = Compile::AliasIdxTop; 926 nat = NULL; 927 break; 928 } 929 } 930 if (nidx != midx) { 931 if (PrintOpto || (PrintMiscellaneous && (WizardMode || Verbose))) { 932 tty->print_cr("==== Matcher alias shift %d => %d", nidx, midx); 933 n->dump(); 934 m->dump(); 935 } 936 assert(C->subsume_loads() && C->must_alias(nat, midx), 937 "must not lose alias info when matching"); 938 } 939 } 940 #endif 941 942 943 //------------------------------MStack----------------------------------------- 944 // State and MStack class used in xform() and find_shared() iterative methods. 945 enum Node_State { Pre_Visit, // node has to be pre-visited 946 Visit, // visit node 947 Post_Visit, // post-visit node 948 Alt_Post_Visit // alternative post-visit path 949 }; 950 951 class MStack: public Node_Stack { 952 public: 953 MStack(int size) : Node_Stack(size) { } 954 955 void push(Node *n, Node_State ns) { 956 Node_Stack::push(n, (uint)ns); 957 } 958 void push(Node *n, Node_State ns, Node *parent, int indx) { 959 ++_inode_top; 960 if ((_inode_top + 1) >= _inode_max) grow(); 961 _inode_top->node = parent; 962 _inode_top->indx = (uint)indx; 963 ++_inode_top; 964 _inode_top->node = n; 965 _inode_top->indx = (uint)ns; 966 } 967 Node *parent() { 968 pop(); 969 return node(); 970 } 971 Node_State state() const { 972 return (Node_State)index(); 973 } 974 void set_state(Node_State ns) { 975 set_index((uint)ns); 976 } 977 }; 978 979 980 //------------------------------xform------------------------------------------ 981 // Given a Node in old-space, Match him (Label/Reduce) to produce a machine 982 // Node in new-space. Given a new-space Node, recursively walk his children. 983 Node *Matcher::transform( Node *n ) { ShouldNotCallThis(); return n; } 984 Node *Matcher::xform( Node *n, int max_stack ) { 985 // Use one stack to keep both: child's node/state and parent's node/index 986 MStack mstack(max_stack * 2 * 2); // C->unique() * 2 * 2 987 mstack.push(n, Visit, NULL, -1); // set NULL as parent to indicate root 988 989 while (mstack.is_nonempty()) { 990 C->check_node_count(NodeLimitFudgeFactor, "too many nodes matching instructions"); 991 if (C->failing()) return NULL; 992 n = mstack.node(); // Leave node on stack 993 Node_State nstate = mstack.state(); 994 if (nstate == Visit) { 995 mstack.set_state(Post_Visit); 996 Node *oldn = n; 997 // Old-space or new-space check 998 if (!C->node_arena()->contains(n)) { 999 // Old space! 1000 Node* m; 1001 if (has_new_node(n)) { // Not yet Label/Reduced 1002 m = new_node(n); 1003 } else { 1004 if (!is_dontcare(n)) { // Matcher can match this guy 1005 // Calls match special. They match alone with no children. 1006 // Their children, the incoming arguments, match normally. 1007 m = n->is_SafePoint() ? match_sfpt(n->as_SafePoint()):match_tree(n); 1008 if (C->failing()) return NULL; 1009 if (m == NULL) { Matcher::soft_match_failure(); return NULL; } 1010 } else { // Nothing the matcher cares about 1011 if( n->is_Proj() && n->in(0)->is_Multi()) { // Projections? 1012 // Convert to machine-dependent projection 1013 m = n->in(0)->as_Multi()->match( n->as_Proj(), this ); 1014 #ifdef ASSERT 1015 _new2old_map.map(m->_idx, n); 1016 #endif 1017 if (m->in(0) != NULL) // m might be top 1018 collect_null_checks(m, n); 1019 } else { // Else just a regular 'ol guy 1020 m = n->clone(); // So just clone into new-space 1021 #ifdef ASSERT 1022 _new2old_map.map(m->_idx, n); 1023 #endif 1024 // Def-Use edges will be added incrementally as Uses 1025 // of this node are matched. 1026 assert(m->outcnt() == 0, "no Uses of this clone yet"); 1027 } 1028 } 1029 1030 set_new_node(n, m); // Map old to new 1031 if (_old_node_note_array != NULL) { 1032 Node_Notes* nn = C->locate_node_notes(_old_node_note_array, 1033 n->_idx); 1034 C->set_node_notes_at(m->_idx, nn); 1035 } 1036 debug_only(match_alias_type(C, n, m)); 1037 } 1038 n = m; // n is now a new-space node 1039 mstack.set_node(n); 1040 } 1041 1042 // New space! 1043 if (_visited.test_set(n->_idx)) continue; // while(mstack.is_nonempty()) 1044 1045 int i; 1046 // Put precedence edges on stack first (match them last). 1047 for (i = oldn->req(); (uint)i < oldn->len(); i++) { 1048 Node *m = oldn->in(i); 1049 if (m == NULL) break; 1050 // set -1 to call add_prec() instead of set_req() during Step1 1051 mstack.push(m, Visit, n, -1); 1052 } 1053 1054 // For constant debug info, I'd rather have unmatched constants. 1055 int cnt = n->req(); 1056 JVMState* jvms = n->jvms(); 1057 int debug_cnt = jvms ? jvms->debug_start() : cnt; 1058 1059 // Now do only debug info. Clone constants rather than matching. 1060 // Constants are represented directly in the debug info without 1061 // the need for executable machine instructions. 1062 // Monitor boxes are also represented directly. 1063 for (i = cnt - 1; i >= debug_cnt; --i) { // For all debug inputs do 1064 Node *m = n->in(i); // Get input 1065 int op = m->Opcode(); 1066 assert((op == Op_BoxLock) == jvms->is_monitor_use(i), "boxes only at monitor sites"); 1067 if( op == Op_ConI || op == Op_ConP || op == Op_ConN || op == Op_ConNKlass || 1068 op == Op_ConF || op == Op_ConD || op == Op_ConL 1069 // || op == Op_BoxLock // %%%% enable this and remove (+++) in chaitin.cpp 1070 ) { 1071 m = m->clone(); 1072 #ifdef ASSERT 1073 _new2old_map.map(m->_idx, n); 1074 #endif 1075 mstack.push(m, Post_Visit, n, i); // Don't need to visit 1076 mstack.push(m->in(0), Visit, m, 0); 1077 } else { 1078 mstack.push(m, Visit, n, i); 1079 } 1080 } 1081 1082 // And now walk his children, and convert his inputs to new-space. 1083 for( ; i >= 0; --i ) { // For all normal inputs do 1084 Node *m = n->in(i); // Get input 1085 if(m != NULL) 1086 mstack.push(m, Visit, n, i); 1087 } 1088 1089 } 1090 else if (nstate == Post_Visit) { 1091 // Set xformed input 1092 Node *p = mstack.parent(); 1093 if (p != NULL) { // root doesn't have parent 1094 int i = (int)mstack.index(); 1095 if (i >= 0) 1096 p->set_req(i, n); // required input 1097 else if (i == -1) 1098 p->add_prec(n); // precedence input 1099 else 1100 ShouldNotReachHere(); 1101 } 1102 mstack.pop(); // remove processed node from stack 1103 } 1104 else { 1105 ShouldNotReachHere(); 1106 } 1107 } // while (mstack.is_nonempty()) 1108 return n; // Return new-space Node 1109 } 1110 1111 //------------------------------warp_outgoing_stk_arg------------------------ 1112 OptoReg::Name Matcher::warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ) { 1113 // Convert outgoing argument location to a pre-biased stack offset 1114 if (reg->is_stack()) { 1115 OptoReg::Name warped = reg->reg2stack(); 1116 // Adjust the stack slot offset to be the register number used 1117 // by the allocator. 1118 warped = OptoReg::add(begin_out_arg_area, warped); 1119 // Keep track of the largest numbered stack slot used for an arg. 1120 // Largest used slot per call-site indicates the amount of stack 1121 // that is killed by the call. 1122 if( warped >= out_arg_limit_per_call ) 1123 out_arg_limit_per_call = OptoReg::add(warped,1); 1124 if (!RegMask::can_represent_arg(warped)) { 1125 C->record_method_not_compilable_all_tiers("unsupported calling sequence"); 1126 return OptoReg::Bad; 1127 } 1128 return warped; 1129 } 1130 return OptoReg::as_OptoReg(reg); 1131 } 1132 1133 1134 //------------------------------match_sfpt------------------------------------- 1135 // Helper function to match call instructions. Calls match special. 1136 // They match alone with no children. Their children, the incoming 1137 // arguments, match normally. 1138 MachNode *Matcher::match_sfpt( SafePointNode *sfpt ) { 1139 MachSafePointNode *msfpt = NULL; 1140 MachCallNode *mcall = NULL; 1141 uint cnt; 1142 // Split out case for SafePoint vs Call 1143 CallNode *call; 1144 const TypeTuple *domain; 1145 ciMethod* method = NULL; 1146 bool is_method_handle_invoke = false; // for special kill effects 1147 if( sfpt->is_Call() ) { 1148 call = sfpt->as_Call(); 1149 domain = call->tf()->domain(); 1150 cnt = domain->cnt(); 1151 1152 // Match just the call, nothing else 1153 MachNode *m = match_tree(call); 1154 if (C->failing()) return NULL; 1155 if( m == NULL ) { Matcher::soft_match_failure(); return NULL; } 1156 1157 // Copy data from the Ideal SafePoint to the machine version 1158 mcall = m->as_MachCall(); 1159 1160 mcall->set_tf( call->tf()); 1161 mcall->set_entry_point(call->entry_point()); 1162 mcall->set_cnt( call->cnt()); 1163 1164 if( mcall->is_MachCallJava() ) { 1165 MachCallJavaNode *mcall_java = mcall->as_MachCallJava(); 1166 const CallJavaNode *call_java = call->as_CallJava(); 1167 method = call_java->method(); 1168 mcall_java->_method = method; 1169 mcall_java->_bci = call_java->_bci; 1170 mcall_java->_optimized_virtual = call_java->is_optimized_virtual(); 1171 is_method_handle_invoke = call_java->is_method_handle_invoke(); 1172 mcall_java->_method_handle_invoke = is_method_handle_invoke; 1173 if (is_method_handle_invoke) { 1174 C->set_has_method_handle_invokes(true); 1175 } 1176 if( mcall_java->is_MachCallStaticJava() ) 1177 mcall_java->as_MachCallStaticJava()->_name = 1178 call_java->as_CallStaticJava()->_name; 1179 if( mcall_java->is_MachCallDynamicJava() ) 1180 mcall_java->as_MachCallDynamicJava()->_vtable_index = 1181 call_java->as_CallDynamicJava()->_vtable_index; 1182 } 1183 else if( mcall->is_MachCallRuntime() ) { 1184 mcall->as_MachCallRuntime()->_name = call->as_CallRuntime()->_name; 1185 } 1186 msfpt = mcall; 1187 } 1188 // This is a non-call safepoint 1189 else { 1190 call = NULL; 1191 domain = NULL; 1192 MachNode *mn = match_tree(sfpt); 1193 if (C->failing()) return NULL; 1194 msfpt = mn->as_MachSafePoint(); 1195 cnt = TypeFunc::Parms; 1196 } 1197 1198 // Advertise the correct memory effects (for anti-dependence computation). 1199 msfpt->set_adr_type(sfpt->adr_type()); 1200 1201 // Allocate a private array of RegMasks. These RegMasks are not shared. 1202 msfpt->_in_rms = NEW_RESOURCE_ARRAY( RegMask, cnt ); 1203 // Empty them all. 1204 memset( msfpt->_in_rms, 0, sizeof(RegMask)*cnt ); 1205 1206 // Do all the pre-defined non-Empty register masks 1207 msfpt->_in_rms[TypeFunc::ReturnAdr] = _return_addr_mask; 1208 msfpt->_in_rms[TypeFunc::FramePtr ] = c_frame_ptr_mask; 1209 1210 // Place first outgoing argument can possibly be put. 1211 OptoReg::Name begin_out_arg_area = OptoReg::add(_new_SP, C->out_preserve_stack_slots()); 1212 assert( is_even(begin_out_arg_area), "" ); 1213 // Compute max outgoing register number per call site. 1214 OptoReg::Name out_arg_limit_per_call = begin_out_arg_area; 1215 // Calls to C may hammer extra stack slots above and beyond any arguments. 1216 // These are usually backing store for register arguments for varargs. 1217 if( call != NULL && call->is_CallRuntime() ) 1218 out_arg_limit_per_call = OptoReg::add(out_arg_limit_per_call,C->varargs_C_out_slots_killed()); 1219 1220 1221 // Do the normal argument list (parameters) register masks 1222 int argcnt = cnt - TypeFunc::Parms; 1223 if( argcnt > 0 ) { // Skip it all if we have no args 1224 BasicType *sig_bt = NEW_RESOURCE_ARRAY( BasicType, argcnt ); 1225 VMRegPair *parm_regs = NEW_RESOURCE_ARRAY( VMRegPair, argcnt ); 1226 int i; 1227 for( i = 0; i < argcnt; i++ ) { 1228 sig_bt[i] = domain->field_at(i+TypeFunc::Parms)->basic_type(); 1229 } 1230 // V-call to pick proper calling convention 1231 call->calling_convention( sig_bt, parm_regs, argcnt ); 1232 1233 #ifdef ASSERT 1234 // Sanity check users' calling convention. Really handy during 1235 // the initial porting effort. Fairly expensive otherwise. 1236 { for (int i = 0; i<argcnt; i++) { 1237 if( !parm_regs[i].first()->is_valid() && 1238 !parm_regs[i].second()->is_valid() ) continue; 1239 VMReg reg1 = parm_regs[i].first(); 1240 VMReg reg2 = parm_regs[i].second(); 1241 for (int j = 0; j < i; j++) { 1242 if( !parm_regs[j].first()->is_valid() && 1243 !parm_regs[j].second()->is_valid() ) continue; 1244 VMReg reg3 = parm_regs[j].first(); 1245 VMReg reg4 = parm_regs[j].second(); 1246 if( !reg1->is_valid() ) { 1247 assert( !reg2->is_valid(), "valid halvsies" ); 1248 } else if( !reg3->is_valid() ) { 1249 assert( !reg4->is_valid(), "valid halvsies" ); 1250 } else { 1251 assert( reg1 != reg2, "calling conv. must produce distinct regs"); 1252 assert( reg1 != reg3, "calling conv. must produce distinct regs"); 1253 assert( reg1 != reg4, "calling conv. must produce distinct regs"); 1254 assert( reg2 != reg3, "calling conv. must produce distinct regs"); 1255 assert( reg2 != reg4 || !reg2->is_valid(), "calling conv. must produce distinct regs"); 1256 assert( reg3 != reg4, "calling conv. must produce distinct regs"); 1257 } 1258 } 1259 } 1260 } 1261 #endif 1262 1263 // Visit each argument. Compute its outgoing register mask. 1264 // Return results now can have 2 bits returned. 1265 // Compute max over all outgoing arguments both per call-site 1266 // and over the entire method. 1267 for( i = 0; i < argcnt; i++ ) { 1268 // Address of incoming argument mask to fill in 1269 RegMask *rm = &mcall->_in_rms[i+TypeFunc::Parms]; 1270 if( !parm_regs[i].first()->is_valid() && 1271 !parm_regs[i].second()->is_valid() ) { 1272 continue; // Avoid Halves 1273 } 1274 // Grab first register, adjust stack slots and insert in mask. 1275 OptoReg::Name reg1 = warp_outgoing_stk_arg(parm_regs[i].first(), begin_out_arg_area, out_arg_limit_per_call ); 1276 if (OptoReg::is_valid(reg1)) 1277 rm->Insert( reg1 ); 1278 // Grab second register (if any), adjust stack slots and insert in mask. 1279 OptoReg::Name reg2 = warp_outgoing_stk_arg(parm_regs[i].second(), begin_out_arg_area, out_arg_limit_per_call ); 1280 if (OptoReg::is_valid(reg2)) 1281 rm->Insert( reg2 ); 1282 } // End of for all arguments 1283 1284 // Compute number of stack slots needed to restore stack in case of 1285 // Pascal-style argument popping. 1286 mcall->_argsize = out_arg_limit_per_call - begin_out_arg_area; 1287 } 1288 1289 // Compute the max stack slot killed by any call. These will not be 1290 // available for debug info, and will be used to adjust FIRST_STACK_mask 1291 // after all call sites have been visited. 1292 if( _out_arg_limit < out_arg_limit_per_call) 1293 _out_arg_limit = out_arg_limit_per_call; 1294 1295 if (mcall) { 1296 // Kill the outgoing argument area, including any non-argument holes and 1297 // any legacy C-killed slots. Use Fat-Projections to do the killing. 1298 // Since the max-per-method covers the max-per-call-site and debug info 1299 // is excluded on the max-per-method basis, debug info cannot land in 1300 // this killed area. 1301 uint r_cnt = mcall->tf()->range()->cnt(); 1302 MachProjNode *proj = new MachProjNode( mcall, r_cnt+10000, RegMask::Empty, MachProjNode::fat_proj ); 1303 if (!RegMask::can_represent_arg(OptoReg::Name(out_arg_limit_per_call-1))) { 1304 C->record_method_not_compilable_all_tiers("unsupported outgoing calling sequence"); 1305 } else { 1306 for (int i = begin_out_arg_area; i < out_arg_limit_per_call; i++) 1307 proj->_rout.Insert(OptoReg::Name(i)); 1308 } 1309 if (proj->_rout.is_NotEmpty()) { 1310 push_projection(proj); 1311 } 1312 } 1313 // Transfer the safepoint information from the call to the mcall 1314 // Move the JVMState list 1315 msfpt->set_jvms(sfpt->jvms()); 1316 for (JVMState* jvms = msfpt->jvms(); jvms; jvms = jvms->caller()) { 1317 jvms->set_map(sfpt); 1318 } 1319 1320 // Debug inputs begin just after the last incoming parameter 1321 assert((mcall == NULL) || (mcall->jvms() == NULL) || 1322 (mcall->jvms()->debug_start() + mcall->_jvmadj == mcall->tf()->domain()->cnt()), ""); 1323 1324 // Move the OopMap 1325 msfpt->_oop_map = sfpt->_oop_map; 1326 1327 // Add additional edges. 1328 if (msfpt->mach_constant_base_node_input() != (uint)-1 && !msfpt->is_MachCallLeaf()) { 1329 // For these calls we can not add MachConstantBase in expand(), as the 1330 // ins are not complete then. 1331 msfpt->ins_req(msfpt->mach_constant_base_node_input(), C->mach_constant_base_node()); 1332 if (msfpt->jvms() && 1333 msfpt->mach_constant_base_node_input() <= msfpt->jvms()->debug_start() + msfpt->_jvmadj) { 1334 // We added an edge before jvms, so we must adapt the position of the ins. 1335 msfpt->jvms()->adapt_position(+1); 1336 } 1337 } 1338 1339 // Registers killed by the call are set in the local scheduling pass 1340 // of Global Code Motion. 1341 return msfpt; 1342 } 1343 1344 //---------------------------match_tree---------------------------------------- 1345 // Match a Ideal Node DAG - turn it into a tree; Label & Reduce. Used as part 1346 // of the whole-sale conversion from Ideal to Mach Nodes. Also used for 1347 // making GotoNodes while building the CFG and in init_spill_mask() to identify 1348 // a Load's result RegMask for memoization in idealreg2regmask[] 1349 MachNode *Matcher::match_tree( const Node *n ) { 1350 assert( n->Opcode() != Op_Phi, "cannot match" ); 1351 assert( !n->is_block_start(), "cannot match" ); 1352 // Set the mark for all locally allocated State objects. 1353 // When this call returns, the _states_arena arena will be reset 1354 // freeing all State objects. 1355 ResourceMark rm( &_states_arena ); 1356 1357 LabelRootDepth = 0; 1358 1359 // StoreNodes require their Memory input to match any LoadNodes 1360 Node *mem = n->is_Store() ? n->in(MemNode::Memory) : (Node*)1 ; 1361 #ifdef ASSERT 1362 Node* save_mem_node = _mem_node; 1363 _mem_node = n->is_Store() ? (Node*)n : NULL; 1364 #endif 1365 // State object for root node of match tree 1366 // Allocate it on _states_arena - stack allocation can cause stack overflow. 1367 State *s = new (&_states_arena) State; 1368 s->_kids[0] = NULL; 1369 s->_kids[1] = NULL; 1370 s->_leaf = (Node*)n; 1371 // Label the input tree, allocating labels from top-level arena 1372 Label_Root( n, s, n->in(0), mem ); 1373 if (C->failing()) return NULL; 1374 1375 // The minimum cost match for the whole tree is found at the root State 1376 uint mincost = max_juint; 1377 uint cost = max_juint; 1378 uint i; 1379 for( i = 0; i < NUM_OPERANDS; i++ ) { 1380 if( s->valid(i) && // valid entry and 1381 s->_cost[i] < cost && // low cost and 1382 s->_rule[i] >= NUM_OPERANDS ) // not an operand 1383 cost = s->_cost[mincost=i]; 1384 } 1385 if (mincost == max_juint) { 1386 #ifndef PRODUCT 1387 tty->print("No matching rule for:"); 1388 s->dump(); 1389 #endif 1390 Matcher::soft_match_failure(); 1391 return NULL; 1392 } 1393 // Reduce input tree based upon the state labels to machine Nodes 1394 MachNode *m = ReduceInst( s, s->_rule[mincost], mem ); 1395 #ifdef ASSERT 1396 _old2new_map.map(n->_idx, m); 1397 _new2old_map.map(m->_idx, (Node*)n); 1398 #endif 1399 1400 // Add any Matcher-ignored edges 1401 uint cnt = n->req(); 1402 uint start = 1; 1403 if( mem != (Node*)1 ) start = MemNode::Memory+1; 1404 if( n->is_AddP() ) { 1405 assert( mem == (Node*)1, "" ); 1406 start = AddPNode::Base+1; 1407 } 1408 for( i = start; i < cnt; i++ ) { 1409 if( !n->match_edge(i) ) { 1410 if( i < m->req() ) 1411 m->ins_req( i, n->in(i) ); 1412 else 1413 m->add_req( n->in(i) ); 1414 } 1415 } 1416 1417 debug_only( _mem_node = save_mem_node; ) 1418 return m; 1419 } 1420 1421 1422 //------------------------------match_into_reg--------------------------------- 1423 // Choose to either match this Node in a register or part of the current 1424 // match tree. Return true for requiring a register and false for matching 1425 // as part of the current match tree. 1426 static bool match_into_reg( const Node *n, Node *m, Node *control, int i, bool shared ) { 1427 1428 const Type *t = m->bottom_type(); 1429 1430 if (t->singleton()) { 1431 // Never force constants into registers. Allow them to match as 1432 // constants or registers. Copies of the same value will share 1433 // the same register. See find_shared_node. 1434 return false; 1435 } else { // Not a constant 1436 // Stop recursion if they have different Controls. 1437 Node* m_control = m->in(0); 1438 // Control of load's memory can post-dominates load's control. 1439 // So use it since load can't float above its memory. 1440 Node* mem_control = (m->is_Load()) ? m->in(MemNode::Memory)->in(0) : NULL; 1441 if (control && m_control && control != m_control && control != mem_control) { 1442 1443 // Actually, we can live with the most conservative control we 1444 // find, if it post-dominates the others. This allows us to 1445 // pick up load/op/store trees where the load can float a little 1446 // above the store. 1447 Node *x = control; 1448 const uint max_scan = 6; // Arbitrary scan cutoff 1449 uint j; 1450 for (j=0; j<max_scan; j++) { 1451 if (x->is_Region()) // Bail out at merge points 1452 return true; 1453 x = x->in(0); 1454 if (x == m_control) // Does 'control' post-dominate 1455 break; // m->in(0)? If so, we can use it 1456 if (x == mem_control) // Does 'control' post-dominate 1457 break; // mem_control? If so, we can use it 1458 } 1459 if (j == max_scan) // No post-domination before scan end? 1460 return true; // Then break the match tree up 1461 } 1462 if ((m->is_DecodeN() && Matcher::narrow_oop_use_complex_address()) || 1463 (m->is_DecodeNKlass() && Matcher::narrow_klass_use_complex_address())) { 1464 // These are commonly used in address expressions and can 1465 // efficiently fold into them on X64 in some cases. 1466 return false; 1467 } 1468 } 1469 1470 // Not forceable cloning. If shared, put it into a register. 1471 return shared; 1472 } 1473 1474 1475 //------------------------------Instruction Selection-------------------------- 1476 // Label method walks a "tree" of nodes, using the ADLC generated DFA to match 1477 // ideal nodes to machine instructions. Trees are delimited by shared Nodes, 1478 // things the Matcher does not match (e.g., Memory), and things with different 1479 // Controls (hence forced into different blocks). We pass in the Control 1480 // selected for this entire State tree. 1481 1482 // The Matcher works on Trees, but an Intel add-to-memory requires a DAG: the 1483 // Store and the Load must have identical Memories (as well as identical 1484 // pointers). Since the Matcher does not have anything for Memory (and 1485 // does not handle DAGs), I have to match the Memory input myself. If the 1486 // Tree root is a Store, I require all Loads to have the identical memory. 1487 Node *Matcher::Label_Root( const Node *n, State *svec, Node *control, const Node *mem){ 1488 // Since Label_Root is a recursive function, its possible that we might run 1489 // out of stack space. See bugs 6272980 & 6227033 for more info. 1490 LabelRootDepth++; 1491 if (LabelRootDepth > MaxLabelRootDepth) { 1492 C->record_method_not_compilable_all_tiers("Out of stack space, increase MaxLabelRootDepth"); 1493 return NULL; 1494 } 1495 uint care = 0; // Edges matcher cares about 1496 uint cnt = n->req(); 1497 uint i = 0; 1498 1499 // Examine children for memory state 1500 // Can only subsume a child into your match-tree if that child's memory state 1501 // is not modified along the path to another input. 1502 // It is unsafe even if the other inputs are separate roots. 1503 Node *input_mem = NULL; 1504 for( i = 1; i < cnt; i++ ) { 1505 if( !n->match_edge(i) ) continue; 1506 Node *m = n->in(i); // Get ith input 1507 assert( m, "expect non-null children" ); 1508 if( m->is_Load() ) { 1509 if( input_mem == NULL ) { 1510 input_mem = m->in(MemNode::Memory); 1511 } else if( input_mem != m->in(MemNode::Memory) ) { 1512 input_mem = NodeSentinel; 1513 } 1514 } 1515 } 1516 1517 for( i = 1; i < cnt; i++ ){// For my children 1518 if( !n->match_edge(i) ) continue; 1519 Node *m = n->in(i); // Get ith input 1520 // Allocate states out of a private arena 1521 State *s = new (&_states_arena) State; 1522 svec->_kids[care++] = s; 1523 assert( care <= 2, "binary only for now" ); 1524 1525 // Recursively label the State tree. 1526 s->_kids[0] = NULL; 1527 s->_kids[1] = NULL; 1528 s->_leaf = m; 1529 1530 // Check for leaves of the State Tree; things that cannot be a part of 1531 // the current tree. If it finds any, that value is matched as a 1532 // register operand. If not, then the normal matching is used. 1533 if( match_into_reg(n, m, control, i, is_shared(m)) || 1534 // 1535 // Stop recursion if this is LoadNode and the root of this tree is a 1536 // StoreNode and the load & store have different memories. 1537 ((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) || 1538 // Can NOT include the match of a subtree when its memory state 1539 // is used by any of the other subtrees 1540 (input_mem == NodeSentinel) ) { 1541 #ifndef PRODUCT 1542 // Print when we exclude matching due to different memory states at input-loads 1543 if( PrintOpto && (Verbose && WizardMode) && (input_mem == NodeSentinel) 1544 && !((mem!=(Node*)1) && m->is_Load() && m->in(MemNode::Memory) != mem) ) { 1545 tty->print_cr("invalid input_mem"); 1546 } 1547 #endif 1548 // Switch to a register-only opcode; this value must be in a register 1549 // and cannot be subsumed as part of a larger instruction. 1550 s->DFA( m->ideal_reg(), m ); 1551 1552 } else { 1553 // If match tree has no control and we do, adopt it for entire tree 1554 if( control == NULL && m->in(0) != NULL && m->req() > 1 ) 1555 control = m->in(0); // Pick up control 1556 // Else match as a normal part of the match tree. 1557 control = Label_Root(m,s,control,mem); 1558 if (C->failing()) return NULL; 1559 } 1560 } 1561 1562 1563 // Call DFA to match this node, and return 1564 svec->DFA( n->Opcode(), n ); 1565 1566 #ifdef ASSERT 1567 uint x; 1568 for( x = 0; x < _LAST_MACH_OPER; x++ ) 1569 if( svec->valid(x) ) 1570 break; 1571 1572 if (x >= _LAST_MACH_OPER) { 1573 n->dump(); 1574 svec->dump(); 1575 assert( false, "bad AD file" ); 1576 } 1577 #endif 1578 return control; 1579 } 1580 1581 1582 // Con nodes reduced using the same rule can share their MachNode 1583 // which reduces the number of copies of a constant in the final 1584 // program. The register allocator is free to split uses later to 1585 // split live ranges. 1586 MachNode* Matcher::find_shared_node(Node* leaf, uint rule) { 1587 if (!leaf->is_Con() && !leaf->is_DecodeNarrowPtr()) return NULL; 1588 1589 // See if this Con has already been reduced using this rule. 1590 if (_shared_nodes.Size() <= leaf->_idx) return NULL; 1591 MachNode* last = (MachNode*)_shared_nodes.at(leaf->_idx); 1592 if (last != NULL && rule == last->rule()) { 1593 // Don't expect control change for DecodeN 1594 if (leaf->is_DecodeNarrowPtr()) 1595 return last; 1596 // Get the new space root. 1597 Node* xroot = new_node(C->root()); 1598 if (xroot == NULL) { 1599 // This shouldn't happen give the order of matching. 1600 return NULL; 1601 } 1602 1603 // Shared constants need to have their control be root so they 1604 // can be scheduled properly. 1605 Node* control = last->in(0); 1606 if (control != xroot) { 1607 if (control == NULL || control == C->root()) { 1608 last->set_req(0, xroot); 1609 } else { 1610 assert(false, "unexpected control"); 1611 return NULL; 1612 } 1613 } 1614 return last; 1615 } 1616 return NULL; 1617 } 1618 1619 1620 //------------------------------ReduceInst------------------------------------- 1621 // Reduce a State tree (with given Control) into a tree of MachNodes. 1622 // This routine (and it's cohort ReduceOper) convert Ideal Nodes into 1623 // complicated machine Nodes. Each MachNode covers some tree of Ideal Nodes. 1624 // Each MachNode has a number of complicated MachOper operands; each 1625 // MachOper also covers a further tree of Ideal Nodes. 1626 1627 // The root of the Ideal match tree is always an instruction, so we enter 1628 // the recursion here. After building the MachNode, we need to recurse 1629 // the tree checking for these cases: 1630 // (1) Child is an instruction - 1631 // Build the instruction (recursively), add it as an edge. 1632 // Build a simple operand (register) to hold the result of the instruction. 1633 // (2) Child is an interior part of an instruction - 1634 // Skip over it (do nothing) 1635 // (3) Child is the start of a operand - 1636 // Build the operand, place it inside the instruction 1637 // Call ReduceOper. 1638 MachNode *Matcher::ReduceInst( State *s, int rule, Node *&mem ) { 1639 assert( rule >= NUM_OPERANDS, "called with operand rule" ); 1640 1641 MachNode* shared_node = find_shared_node(s->_leaf, rule); 1642 if (shared_node != NULL) { 1643 return shared_node; 1644 } 1645 1646 // Build the object to represent this state & prepare for recursive calls 1647 MachNode *mach = s->MachNodeGenerator( rule, C ); 1648 mach->_opnds[0] = s->MachOperGenerator( _reduceOp[rule], C ); 1649 assert( mach->_opnds[0] != NULL, "Missing result operand" ); 1650 Node *leaf = s->_leaf; 1651 // Check for instruction or instruction chain rule 1652 if( rule >= _END_INST_CHAIN_RULE || rule < _BEGIN_INST_CHAIN_RULE ) { 1653 assert(C->node_arena()->contains(s->_leaf) || !has_new_node(s->_leaf), 1654 "duplicating node that's already been matched"); 1655 // Instruction 1656 mach->add_req( leaf->in(0) ); // Set initial control 1657 // Reduce interior of complex instruction 1658 ReduceInst_Interior( s, rule, mem, mach, 1 ); 1659 } else { 1660 // Instruction chain rules are data-dependent on their inputs 1661 mach->add_req(0); // Set initial control to none 1662 ReduceInst_Chain_Rule( s, rule, mem, mach ); 1663 } 1664 1665 // If a Memory was used, insert a Memory edge 1666 if( mem != (Node*)1 ) { 1667 mach->ins_req(MemNode::Memory,mem); 1668 #ifdef ASSERT 1669 // Verify adr type after matching memory operation 1670 const MachOper* oper = mach->memory_operand(); 1671 if (oper != NULL && oper != (MachOper*)-1) { 1672 // It has a unique memory operand. Find corresponding ideal mem node. 1673 Node* m = NULL; 1674 if (leaf->is_Mem()) { 1675 m = leaf; 1676 } else { 1677 m = _mem_node; 1678 assert(m != NULL && m->is_Mem(), "expecting memory node"); 1679 } 1680 const Type* mach_at = mach->adr_type(); 1681 // DecodeN node consumed by an address may have different type 1682 // then its input. Don't compare types for such case. 1683 if (m->adr_type() != mach_at && 1684 (m->in(MemNode::Address)->is_DecodeNarrowPtr() || 1685 m->in(MemNode::Address)->is_AddP() && 1686 m->in(MemNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr() || 1687 m->in(MemNode::Address)->is_AddP() && 1688 m->in(MemNode::Address)->in(AddPNode::Address)->is_AddP() && 1689 m->in(MemNode::Address)->in(AddPNode::Address)->in(AddPNode::Address)->is_DecodeNarrowPtr())) { 1690 mach_at = m->adr_type(); 1691 } 1692 if (m->adr_type() != mach_at) { 1693 m->dump(); 1694 tty->print_cr("mach:"); 1695 mach->dump(1); 1696 } 1697 assert(m->adr_type() == mach_at, "matcher should not change adr type"); 1698 } 1699 #endif 1700 } 1701 1702 // If the _leaf is an AddP, insert the base edge 1703 if (leaf->is_AddP()) { 1704 mach->ins_req(AddPNode::Base,leaf->in(AddPNode::Base)); 1705 } 1706 1707 uint number_of_projections_prior = number_of_projections(); 1708 1709 // Perform any 1-to-many expansions required 1710 MachNode *ex = mach->Expand(s, _projection_list, mem); 1711 if (ex != mach) { 1712 assert(ex->ideal_reg() == mach->ideal_reg(), "ideal types should match"); 1713 if( ex->in(1)->is_Con() ) 1714 ex->in(1)->set_req(0, C->root()); 1715 // Remove old node from the graph 1716 for( uint i=0; i<mach->req(); i++ ) { 1717 mach->set_req(i,NULL); 1718 } 1719 #ifdef ASSERT 1720 _new2old_map.map(ex->_idx, s->_leaf); 1721 #endif 1722 } 1723 1724 // PhaseChaitin::fixup_spills will sometimes generate spill code 1725 // via the matcher. By the time, nodes have been wired into the CFG, 1726 // and any further nodes generated by expand rules will be left hanging 1727 // in space, and will not get emitted as output code. Catch this. 1728 // Also, catch any new register allocation constraints ("projections") 1729 // generated belatedly during spill code generation. 1730 if (_allocation_started) { 1731 guarantee(ex == mach, "no expand rules during spill generation"); 1732 guarantee(number_of_projections_prior == number_of_projections(), "no allocation during spill generation"); 1733 } 1734 1735 if (leaf->is_Con() || leaf->is_DecodeNarrowPtr()) { 1736 // Record the con for sharing 1737 _shared_nodes.map(leaf->_idx, ex); 1738 } 1739 1740 return ex; 1741 } 1742 1743 void Matcher::ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach ) { 1744 // 'op' is what I am expecting to receive 1745 int op = _leftOp[rule]; 1746 // Operand type to catch childs result 1747 // This is what my child will give me. 1748 int opnd_class_instance = s->_rule[op]; 1749 // Choose between operand class or not. 1750 // This is what I will receive. 1751 int catch_op = (FIRST_OPERAND_CLASS <= op && op < NUM_OPERANDS) ? opnd_class_instance : op; 1752 // New rule for child. Chase operand classes to get the actual rule. 1753 int newrule = s->_rule[catch_op]; 1754 1755 if( newrule < NUM_OPERANDS ) { 1756 // Chain from operand or operand class, may be output of shared node 1757 assert( 0 <= opnd_class_instance && opnd_class_instance < NUM_OPERANDS, 1758 "Bad AD file: Instruction chain rule must chain from operand"); 1759 // Insert operand into array of operands for this instruction 1760 mach->_opnds[1] = s->MachOperGenerator( opnd_class_instance, C ); 1761 1762 ReduceOper( s, newrule, mem, mach ); 1763 } else { 1764 // Chain from the result of an instruction 1765 assert( newrule >= _LAST_MACH_OPER, "Do NOT chain from internal operand"); 1766 mach->_opnds[1] = s->MachOperGenerator( _reduceOp[catch_op], C ); 1767 Node *mem1 = (Node*)1; 1768 debug_only(Node *save_mem_node = _mem_node;) 1769 mach->add_req( ReduceInst(s, newrule, mem1) ); 1770 debug_only(_mem_node = save_mem_node;) 1771 } 1772 return; 1773 } 1774 1775 1776 uint Matcher::ReduceInst_Interior( State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds ) { 1777 if( s->_leaf->is_Load() ) { 1778 Node *mem2 = s->_leaf->in(MemNode::Memory); 1779 assert( mem == (Node*)1 || mem == mem2, "multiple Memories being matched at once?" ); 1780 debug_only( if( mem == (Node*)1 ) _mem_node = s->_leaf;) 1781 mem = mem2; 1782 } 1783 if( s->_leaf->in(0) != NULL && s->_leaf->req() > 1) { 1784 if( mach->in(0) == NULL ) 1785 mach->set_req(0, s->_leaf->in(0)); 1786 } 1787 1788 // Now recursively walk the state tree & add operand list. 1789 for( uint i=0; i<2; i++ ) { // binary tree 1790 State *newstate = s->_kids[i]; 1791 if( newstate == NULL ) break; // Might only have 1 child 1792 // 'op' is what I am expecting to receive 1793 int op; 1794 if( i == 0 ) { 1795 op = _leftOp[rule]; 1796 } else { 1797 op = _rightOp[rule]; 1798 } 1799 // Operand type to catch childs result 1800 // This is what my child will give me. 1801 int opnd_class_instance = newstate->_rule[op]; 1802 // Choose between operand class or not. 1803 // This is what I will receive. 1804 int catch_op = (op >= FIRST_OPERAND_CLASS && op < NUM_OPERANDS) ? opnd_class_instance : op; 1805 // New rule for child. Chase operand classes to get the actual rule. 1806 int newrule = newstate->_rule[catch_op]; 1807 1808 if( newrule < NUM_OPERANDS ) { // Operand/operandClass or internalOp/instruction? 1809 // Operand/operandClass 1810 // Insert operand into array of operands for this instruction 1811 mach->_opnds[num_opnds++] = newstate->MachOperGenerator( opnd_class_instance, C ); 1812 ReduceOper( newstate, newrule, mem, mach ); 1813 1814 } else { // Child is internal operand or new instruction 1815 if( newrule < _LAST_MACH_OPER ) { // internal operand or instruction? 1816 // internal operand --> call ReduceInst_Interior 1817 // Interior of complex instruction. Do nothing but recurse. 1818 num_opnds = ReduceInst_Interior( newstate, newrule, mem, mach, num_opnds ); 1819 } else { 1820 // instruction --> call build operand( ) to catch result 1821 // --> ReduceInst( newrule ) 1822 mach->_opnds[num_opnds++] = s->MachOperGenerator( _reduceOp[catch_op], C ); 1823 Node *mem1 = (Node*)1; 1824 debug_only(Node *save_mem_node = _mem_node;) 1825 mach->add_req( ReduceInst( newstate, newrule, mem1 ) ); 1826 debug_only(_mem_node = save_mem_node;) 1827 } 1828 } 1829 assert( mach->_opnds[num_opnds-1], "" ); 1830 } 1831 return num_opnds; 1832 } 1833 1834 // This routine walks the interior of possible complex operands. 1835 // At each point we check our children in the match tree: 1836 // (1) No children - 1837 // We are a leaf; add _leaf field as an input to the MachNode 1838 // (2) Child is an internal operand - 1839 // Skip over it ( do nothing ) 1840 // (3) Child is an instruction - 1841 // Call ReduceInst recursively and 1842 // and instruction as an input to the MachNode 1843 void Matcher::ReduceOper( State *s, int rule, Node *&mem, MachNode *mach ) { 1844 assert( rule < _LAST_MACH_OPER, "called with operand rule" ); 1845 State *kid = s->_kids[0]; 1846 assert( kid == NULL || s->_leaf->in(0) == NULL, "internal operands have no control" ); 1847 1848 // Leaf? And not subsumed? 1849 if( kid == NULL && !_swallowed[rule] ) { 1850 mach->add_req( s->_leaf ); // Add leaf pointer 1851 return; // Bail out 1852 } 1853 1854 if( s->_leaf->is_Load() ) { 1855 assert( mem == (Node*)1, "multiple Memories being matched at once?" ); 1856 mem = s->_leaf->in(MemNode::Memory); 1857 debug_only(_mem_node = s->_leaf;) 1858 } 1859 if( s->_leaf->in(0) && s->_leaf->req() > 1) { 1860 if( !mach->in(0) ) 1861 mach->set_req(0,s->_leaf->in(0)); 1862 else { 1863 assert( s->_leaf->in(0) == mach->in(0), "same instruction, differing controls?" ); 1864 } 1865 } 1866 1867 for( uint i=0; kid != NULL && i<2; kid = s->_kids[1], i++ ) { // binary tree 1868 int newrule; 1869 if( i == 0) 1870 newrule = kid->_rule[_leftOp[rule]]; 1871 else 1872 newrule = kid->_rule[_rightOp[rule]]; 1873 1874 if( newrule < _LAST_MACH_OPER ) { // Operand or instruction? 1875 // Internal operand; recurse but do nothing else 1876 ReduceOper( kid, newrule, mem, mach ); 1877 1878 } else { // Child is a new instruction 1879 // Reduce the instruction, and add a direct pointer from this 1880 // machine instruction to the newly reduced one. 1881 Node *mem1 = (Node*)1; 1882 debug_only(Node *save_mem_node = _mem_node;) 1883 mach->add_req( ReduceInst( kid, newrule, mem1 ) ); 1884 debug_only(_mem_node = save_mem_node;) 1885 } 1886 } 1887 } 1888 1889 1890 // ------------------------------------------------------------------------- 1891 // Java-Java calling convention 1892 // (what you use when Java calls Java) 1893 1894 //------------------------------find_receiver---------------------------------- 1895 // For a given signature, return the OptoReg for parameter 0. 1896 OptoReg::Name Matcher::find_receiver( bool is_outgoing ) { 1897 VMRegPair regs; 1898 BasicType sig_bt = T_OBJECT; 1899 calling_convention(&sig_bt, ®s, 1, is_outgoing); 1900 // Return argument 0 register. In the LP64 build pointers 1901 // take 2 registers, but the VM wants only the 'main' name. 1902 return OptoReg::as_OptoReg(regs.first()); 1903 } 1904 1905 // This function identifies sub-graphs in which a 'load' node is 1906 // input to two different nodes, and such that it can be matched 1907 // with BMI instructions like blsi, blsr, etc. 1908 // Example : for b = -a[i] & a[i] can be matched to blsi r32, m32. 1909 // The graph is (AndL (SubL Con0 LoadL*) LoadL*), where LoadL* 1910 // refers to the same node. 1911 #ifdef X86 1912 // Match the generic fused operations pattern (op1 (op2 Con{ConType} mop) mop) 1913 // This is a temporary solution until we make DAGs expressible in ADL. 1914 template<typename ConType> 1915 class FusedPatternMatcher { 1916 Node* _op1_node; 1917 Node* _mop_node; 1918 int _con_op; 1919 1920 static int match_next(Node* n, int next_op, int next_op_idx) { 1921 if (n->in(1) == NULL || n->in(2) == NULL) { 1922 return -1; 1923 } 1924 1925 if (next_op_idx == -1) { // n is commutative, try rotations 1926 if (n->in(1)->Opcode() == next_op) { 1927 return 1; 1928 } else if (n->in(2)->Opcode() == next_op) { 1929 return 2; 1930 } 1931 } else { 1932 assert(next_op_idx > 0 && next_op_idx <= 2, "Bad argument index"); 1933 if (n->in(next_op_idx)->Opcode() == next_op) { 1934 return next_op_idx; 1935 } 1936 } 1937 return -1; 1938 } 1939 public: 1940 FusedPatternMatcher(Node* op1_node, Node *mop_node, int con_op) : 1941 _op1_node(op1_node), _mop_node(mop_node), _con_op(con_op) { } 1942 1943 bool match(int op1, int op1_op2_idx, // op1 and the index of the op1->op2 edge, -1 if op1 is commutative 1944 int op2, int op2_con_idx, // op2 and the index of the op2->con edge, -1 if op2 is commutative 1945 typename ConType::NativeType con_value) { 1946 if (_op1_node->Opcode() != op1) { 1947 return false; 1948 } 1949 if (_mop_node->outcnt() > 2) { 1950 return false; 1951 } 1952 op1_op2_idx = match_next(_op1_node, op2, op1_op2_idx); 1953 if (op1_op2_idx == -1) { 1954 return false; 1955 } 1956 // Memory operation must be the other edge 1957 int op1_mop_idx = (op1_op2_idx & 1) + 1; 1958 1959 // Check that the mop node is really what we want 1960 if (_op1_node->in(op1_mop_idx) == _mop_node) { 1961 Node *op2_node = _op1_node->in(op1_op2_idx); 1962 if (op2_node->outcnt() > 1) { 1963 return false; 1964 } 1965 assert(op2_node->Opcode() == op2, "Should be"); 1966 op2_con_idx = match_next(op2_node, _con_op, op2_con_idx); 1967 if (op2_con_idx == -1) { 1968 return false; 1969 } 1970 // Memory operation must be the other edge 1971 int op2_mop_idx = (op2_con_idx & 1) + 1; 1972 // Check that the memory operation is the same node 1973 if (op2_node->in(op2_mop_idx) == _mop_node) { 1974 // Now check the constant 1975 const Type* con_type = op2_node->in(op2_con_idx)->bottom_type(); 1976 if (con_type != Type::TOP && ConType::as_self(con_type)->get_con() == con_value) { 1977 return true; 1978 } 1979 } 1980 } 1981 return false; 1982 } 1983 }; 1984 1985 1986 bool Matcher::is_bmi_pattern(Node *n, Node *m) { 1987 if (n != NULL && m != NULL) { 1988 if (m->Opcode() == Op_LoadI) { 1989 FusedPatternMatcher<TypeInt> bmii(n, m, Op_ConI); 1990 return bmii.match(Op_AndI, -1, Op_SubI, 1, 0) || 1991 bmii.match(Op_AndI, -1, Op_AddI, -1, -1) || 1992 bmii.match(Op_XorI, -1, Op_AddI, -1, -1); 1993 } else if (m->Opcode() == Op_LoadL) { 1994 FusedPatternMatcher<TypeLong> bmil(n, m, Op_ConL); 1995 return bmil.match(Op_AndL, -1, Op_SubL, 1, 0) || 1996 bmil.match(Op_AndL, -1, Op_AddL, -1, -1) || 1997 bmil.match(Op_XorL, -1, Op_AddL, -1, -1); 1998 } 1999 } 2000 return false; 2001 } 2002 #endif // X86 2003 2004 // A method-klass-holder may be passed in the inline_cache_reg 2005 // and then expanded into the inline_cache_reg and a method_oop register 2006 // defined in ad_<arch>.cpp 2007 2008 2009 //------------------------------find_shared------------------------------------ 2010 // Set bits if Node is shared or otherwise a root 2011 void Matcher::find_shared( Node *n ) { 2012 // Allocate stack of size C->unique() * 2 to avoid frequent realloc 2013 MStack mstack(C->unique() * 2); 2014 // Mark nodes as address_visited if they are inputs to an address expression 2015 VectorSet address_visited(Thread::current()->resource_area()); 2016 mstack.push(n, Visit); // Don't need to pre-visit root node 2017 while (mstack.is_nonempty()) { 2018 n = mstack.node(); // Leave node on stack 2019 Node_State nstate = mstack.state(); 2020 uint nop = n->Opcode(); 2021 if (nstate == Pre_Visit) { 2022 if (address_visited.test(n->_idx)) { // Visited in address already? 2023 // Flag as visited and shared now. 2024 set_visited(n); 2025 } 2026 if (is_visited(n)) { // Visited already? 2027 // Node is shared and has no reason to clone. Flag it as shared. 2028 // This causes it to match into a register for the sharing. 2029 set_shared(n); // Flag as shared and 2030 mstack.pop(); // remove node from stack 2031 continue; 2032 } 2033 nstate = Visit; // Not already visited; so visit now 2034 } 2035 if (nstate == Visit) { 2036 mstack.set_state(Post_Visit); 2037 set_visited(n); // Flag as visited now 2038 bool mem_op = false; 2039 2040 switch( nop ) { // Handle some opcodes special 2041 case Op_Phi: // Treat Phis as shared roots 2042 case Op_Parm: 2043 case Op_Proj: // All handled specially during matching 2044 case Op_SafePointScalarObject: 2045 set_shared(n); 2046 set_dontcare(n); 2047 break; 2048 case Op_If: 2049 case Op_CountedLoopEnd: 2050 mstack.set_state(Alt_Post_Visit); // Alternative way 2051 // Convert (If (Bool (CmpX A B))) into (If (Bool) (CmpX A B)). Helps 2052 // with matching cmp/branch in 1 instruction. The Matcher needs the 2053 // Bool and CmpX side-by-side, because it can only get at constants 2054 // that are at the leaves of Match trees, and the Bool's condition acts 2055 // as a constant here. 2056 mstack.push(n->in(1), Visit); // Clone the Bool 2057 mstack.push(n->in(0), Pre_Visit); // Visit control input 2058 continue; // while (mstack.is_nonempty()) 2059 case Op_ConvI2D: // These forms efficiently match with a prior 2060 case Op_ConvI2F: // Load but not a following Store 2061 if( n->in(1)->is_Load() && // Prior load 2062 n->outcnt() == 1 && // Not already shared 2063 n->unique_out()->is_Store() ) // Following store 2064 set_shared(n); // Force it to be a root 2065 break; 2066 case Op_ReverseBytesI: 2067 case Op_ReverseBytesL: 2068 if( n->in(1)->is_Load() && // Prior load 2069 n->outcnt() == 1 ) // Not already shared 2070 set_shared(n); // Force it to be a root 2071 break; 2072 case Op_BoxLock: // Cant match until we get stack-regs in ADLC 2073 case Op_IfFalse: 2074 case Op_IfTrue: 2075 case Op_MachProj: 2076 case Op_MergeMem: 2077 case Op_Catch: 2078 case Op_CatchProj: 2079 case Op_CProj: 2080 case Op_JumpProj: 2081 case Op_JProj: 2082 case Op_NeverBranch: 2083 set_dontcare(n); 2084 break; 2085 case Op_Jump: 2086 mstack.push(n->in(1), Pre_Visit); // Switch Value (could be shared) 2087 mstack.push(n->in(0), Pre_Visit); // Visit Control input 2088 continue; // while (mstack.is_nonempty()) 2089 case Op_StrComp: 2090 case Op_StrEquals: 2091 case Op_StrIndexOf: 2092 case Op_AryEq: 2093 case Op_EncodeISOArray: 2094 set_shared(n); // Force result into register (it will be anyways) 2095 break; 2096 case Op_ConP: { // Convert pointers above the centerline to NUL 2097 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2098 const TypePtr* tp = tn->type()->is_ptr(); 2099 if (tp->_ptr == TypePtr::AnyNull) { 2100 tn->set_type(TypePtr::NULL_PTR); 2101 } 2102 break; 2103 } 2104 case Op_ConN: { // Convert narrow pointers above the centerline to NUL 2105 TypeNode *tn = n->as_Type(); // Constants derive from type nodes 2106 const TypePtr* tp = tn->type()->make_ptr(); 2107 if (tp && tp->_ptr == TypePtr::AnyNull) { 2108 tn->set_type(TypeNarrowOop::NULL_PTR); 2109 } 2110 break; 2111 } 2112 case Op_Binary: // These are introduced in the Post_Visit state. 2113 ShouldNotReachHere(); 2114 break; 2115 case Op_ClearArray: 2116 case Op_SafePoint: 2117 mem_op = true; 2118 break; 2119 default: 2120 if( n->is_Store() ) { 2121 // Do match stores, despite no ideal reg 2122 mem_op = true; 2123 break; 2124 } 2125 if( n->is_Mem() ) { // Loads and LoadStores 2126 mem_op = true; 2127 // Loads must be root of match tree due to prior load conflict 2128 if( C->subsume_loads() == false ) 2129 set_shared(n); 2130 } 2131 // Fall into default case 2132 if( !n->ideal_reg() ) 2133 set_dontcare(n); // Unmatchable Nodes 2134 } // end_switch 2135 2136 for(int i = n->req() - 1; i >= 0; --i) { // For my children 2137 Node *m = n->in(i); // Get ith input 2138 if (m == NULL) continue; // Ignore NULLs 2139 uint mop = m->Opcode(); 2140 2141 // Must clone all producers of flags, or we will not match correctly. 2142 // Suppose a compare setting int-flags is shared (e.g., a switch-tree) 2143 // then it will match into an ideal Op_RegFlags. Alas, the fp-flags 2144 // are also there, so we may match a float-branch to int-flags and 2145 // expect the allocator to haul the flags from the int-side to the 2146 // fp-side. No can do. 2147 if( _must_clone[mop] ) { 2148 mstack.push(m, Visit); 2149 continue; // for(int i = ...) 2150 } 2151 2152 if( mop == Op_AddP && m->in(AddPNode::Base)->is_DecodeNarrowPtr()) { 2153 // Bases used in addresses must be shared but since 2154 // they are shared through a DecodeN they may appear 2155 // to have a single use so force sharing here. 2156 set_shared(m->in(AddPNode::Base)->in(1)); 2157 } 2158 2159 // if 'n' and 'm' are part of a graph for BMI instruction, clone this node. 2160 #ifdef X86 2161 if (UseBMI1Instructions && is_bmi_pattern(n, m)) { 2162 mstack.push(m, Visit); 2163 continue; 2164 } 2165 #endif 2166 2167 // Clone addressing expressions as they are "free" in memory access instructions 2168 if( mem_op && i == MemNode::Address && mop == Op_AddP ) { 2169 // Some inputs for address expression are not put on stack 2170 // to avoid marking them as shared and forcing them into register 2171 // if they are used only in address expressions. 2172 // But they should be marked as shared if there are other uses 2173 // besides address expressions. 2174 2175 Node *off = m->in(AddPNode::Offset); 2176 if( off->is_Con() && 2177 // When there are other uses besides address expressions 2178 // put it on stack and mark as shared. 2179 !is_visited(m) ) { 2180 address_visited.test_set(m->_idx); // Flag as address_visited 2181 Node *adr = m->in(AddPNode::Address); 2182 2183 // Intel, ARM and friends can handle 2 adds in addressing mode 2184 if( clone_shift_expressions && adr->is_AddP() && 2185 // AtomicAdd is not an addressing expression. 2186 // Cheap to find it by looking for screwy base. 2187 !adr->in(AddPNode::Base)->is_top() && 2188 // Are there other uses besides address expressions? 2189 !is_visited(adr) ) { 2190 address_visited.set(adr->_idx); // Flag as address_visited 2191 Node *shift = adr->in(AddPNode::Offset); 2192 // Check for shift by small constant as well 2193 if( shift->Opcode() == Op_LShiftX && shift->in(2)->is_Con() && 2194 shift->in(2)->get_int() <= 3 && 2195 // Are there other uses besides address expressions? 2196 !is_visited(shift) ) { 2197 address_visited.set(shift->_idx); // Flag as address_visited 2198 mstack.push(shift->in(2), Visit); 2199 Node *conv = shift->in(1); 2200 #ifdef _LP64 2201 // Allow Matcher to match the rule which bypass 2202 // ConvI2L operation for an array index on LP64 2203 // if the index value is positive. 2204 if( conv->Opcode() == Op_ConvI2L && 2205 conv->as_Type()->type()->is_long()->_lo >= 0 && 2206 // Are there other uses besides address expressions? 2207 !is_visited(conv) ) { 2208 address_visited.set(conv->_idx); // Flag as address_visited 2209 mstack.push(conv->in(1), Pre_Visit); 2210 } else 2211 #endif 2212 mstack.push(conv, Pre_Visit); 2213 } else { 2214 mstack.push(shift, Pre_Visit); 2215 } 2216 mstack.push(adr->in(AddPNode::Address), Pre_Visit); 2217 mstack.push(adr->in(AddPNode::Base), Pre_Visit); 2218 } else { // Sparc, Alpha, PPC and friends 2219 mstack.push(adr, Pre_Visit); 2220 } 2221 2222 // Clone X+offset as it also folds into most addressing expressions 2223 mstack.push(off, Visit); 2224 mstack.push(m->in(AddPNode::Base), Pre_Visit); 2225 continue; // for(int i = ...) 2226 } // if( off->is_Con() ) 2227 } // if( mem_op && 2228 mstack.push(m, Pre_Visit); 2229 } // for(int i = ...) 2230 } 2231 else if (nstate == Alt_Post_Visit) { 2232 mstack.pop(); // Remove node from stack 2233 // We cannot remove the Cmp input from the Bool here, as the Bool may be 2234 // shared and all users of the Bool need to move the Cmp in parallel. 2235 // This leaves both the Bool and the If pointing at the Cmp. To 2236 // prevent the Matcher from trying to Match the Cmp along both paths 2237 // BoolNode::match_edge always returns a zero. 2238 2239 // We reorder the Op_If in a pre-order manner, so we can visit without 2240 // accidentally sharing the Cmp (the Bool and the If make 2 users). 2241 n->add_req( n->in(1)->in(1) ); // Add the Cmp next to the Bool 2242 } 2243 else if (nstate == Post_Visit) { 2244 mstack.pop(); // Remove node from stack 2245 2246 // Now hack a few special opcodes 2247 switch( n->Opcode() ) { // Handle some opcodes special 2248 case Op_StorePConditional: 2249 case Op_StoreIConditional: 2250 case Op_StoreLConditional: 2251 case Op_CompareAndSwapI: 2252 case Op_CompareAndSwapL: 2253 case Op_CompareAndSwapP: 2254 case Op_CompareAndSwapN: { // Convert trinary to binary-tree 2255 Node *newval = n->in(MemNode::ValueIn ); 2256 Node *oldval = n->in(LoadStoreConditionalNode::ExpectedIn); 2257 Node *pair = new BinaryNode( oldval, newval ); 2258 n->set_req(MemNode::ValueIn,pair); 2259 n->del_req(LoadStoreConditionalNode::ExpectedIn); 2260 break; 2261 } 2262 case Op_CMoveD: // Convert trinary to binary-tree 2263 case Op_CMoveF: 2264 case Op_CMoveI: 2265 case Op_CMoveL: 2266 case Op_CMoveN: 2267 case Op_CMoveP: { 2268 // Restructure into a binary tree for Matching. It's possible that 2269 // we could move this code up next to the graph reshaping for IfNodes 2270 // or vice-versa, but I do not want to debug this for Ladybird. 2271 // 10/2/2000 CNC. 2272 Node *pair1 = new BinaryNode(n->in(1),n->in(1)->in(1)); 2273 n->set_req(1,pair1); 2274 Node *pair2 = new BinaryNode(n->in(2),n->in(3)); 2275 n->set_req(2,pair2); 2276 n->del_req(3); 2277 break; 2278 } 2279 case Op_LoopLimit: { 2280 Node *pair1 = new BinaryNode(n->in(1),n->in(2)); 2281 n->set_req(1,pair1); 2282 n->set_req(2,n->in(3)); 2283 n->del_req(3); 2284 break; 2285 } 2286 case Op_StrEquals: { 2287 Node *pair1 = new BinaryNode(n->in(2),n->in(3)); 2288 n->set_req(2,pair1); 2289 n->set_req(3,n->in(4)); 2290 n->del_req(4); 2291 break; 2292 } 2293 case Op_StrComp: 2294 case Op_StrIndexOf: { 2295 Node *pair1 = new BinaryNode(n->in(2),n->in(3)); 2296 n->set_req(2,pair1); 2297 Node *pair2 = new BinaryNode(n->in(4),n->in(5)); 2298 n->set_req(3,pair2); 2299 n->del_req(5); 2300 n->del_req(4); 2301 break; 2302 } 2303 case Op_EncodeISOArray: { 2304 // Restructure into a binary tree for Matching. 2305 Node* pair = new BinaryNode(n->in(3), n->in(4)); 2306 n->set_req(3, pair); 2307 n->del_req(4); 2308 break; 2309 } 2310 default: 2311 break; 2312 } 2313 } 2314 else { 2315 ShouldNotReachHere(); 2316 } 2317 } // end of while (mstack.is_nonempty()) 2318 } 2319 2320 #ifdef ASSERT 2321 // machine-independent root to machine-dependent root 2322 void Matcher::dump_old2new_map() { 2323 _old2new_map.dump(); 2324 } 2325 #endif 2326 2327 //---------------------------collect_null_checks------------------------------- 2328 // Find null checks in the ideal graph; write a machine-specific node for 2329 // it. Used by later implicit-null-check handling. Actually collects 2330 // either an IfTrue or IfFalse for the common NOT-null path, AND the ideal 2331 // value being tested. 2332 void Matcher::collect_null_checks( Node *proj, Node *orig_proj ) { 2333 Node *iff = proj->in(0); 2334 if( iff->Opcode() == Op_If ) { 2335 // During matching If's have Bool & Cmp side-by-side 2336 BoolNode *b = iff->in(1)->as_Bool(); 2337 Node *cmp = iff->in(2); 2338 int opc = cmp->Opcode(); 2339 if (opc != Op_CmpP && opc != Op_CmpN) return; 2340 2341 const Type* ct = cmp->in(2)->bottom_type(); 2342 if (ct == TypePtr::NULL_PTR || 2343 (opc == Op_CmpN && ct == TypeNarrowOop::NULL_PTR)) { 2344 2345 bool push_it = false; 2346 if( proj->Opcode() == Op_IfTrue ) { 2347 extern int all_null_checks_found; 2348 all_null_checks_found++; 2349 if( b->_test._test == BoolTest::ne ) { 2350 push_it = true; 2351 } 2352 } else { 2353 assert( proj->Opcode() == Op_IfFalse, "" ); 2354 if( b->_test._test == BoolTest::eq ) { 2355 push_it = true; 2356 } 2357 } 2358 if( push_it ) { 2359 _null_check_tests.push(proj); 2360 Node* val = cmp->in(1); 2361 #ifdef _LP64 2362 if (val->bottom_type()->isa_narrowoop() && 2363 !Matcher::narrow_oop_use_complex_address()) { 2364 // 2365 // Look for DecodeN node which should be pinned to orig_proj. 2366 // On platforms (Sparc) which can not handle 2 adds 2367 // in addressing mode we have to keep a DecodeN node and 2368 // use it to do implicit NULL check in address. 2369 // 2370 // DecodeN node was pinned to non-null path (orig_proj) during 2371 // CastPP transformation in final_graph_reshaping_impl(). 2372 // 2373 uint cnt = orig_proj->outcnt(); 2374 for (uint i = 0; i < orig_proj->outcnt(); i++) { 2375 Node* d = orig_proj->raw_out(i); 2376 if (d->is_DecodeN() && d->in(1) == val) { 2377 val = d; 2378 val->set_req(0, NULL); // Unpin now. 2379 // Mark this as special case to distinguish from 2380 // a regular case: CmpP(DecodeN, NULL). 2381 val = (Node*)(((intptr_t)val) | 1); 2382 break; 2383 } 2384 } 2385 } 2386 #endif 2387 _null_check_tests.push(val); 2388 } 2389 } 2390 } 2391 } 2392 2393 //---------------------------validate_null_checks------------------------------ 2394 // Its possible that the value being NULL checked is not the root of a match 2395 // tree. If so, I cannot use the value in an implicit null check. 2396 void Matcher::validate_null_checks( ) { 2397 uint cnt = _null_check_tests.size(); 2398 for( uint i=0; i < cnt; i+=2 ) { 2399 Node *test = _null_check_tests[i]; 2400 Node *val = _null_check_tests[i+1]; 2401 bool is_decoden = ((intptr_t)val) & 1; 2402 val = (Node*)(((intptr_t)val) & ~1); 2403 if (has_new_node(val)) { 2404 Node* new_val = new_node(val); 2405 if (is_decoden) { 2406 assert(val->is_DecodeNarrowPtr() && val->in(0) == NULL, "sanity"); 2407 // Note: new_val may have a control edge if 2408 // the original ideal node DecodeN was matched before 2409 // it was unpinned in Matcher::collect_null_checks(). 2410 // Unpin the mach node and mark it. 2411 new_val->set_req(0, NULL); 2412 new_val = (Node*)(((intptr_t)new_val) | 1); 2413 } 2414 // Is a match-tree root, so replace with the matched value 2415 _null_check_tests.map(i+1, new_val); 2416 } else { 2417 // Yank from candidate list 2418 _null_check_tests.map(i+1,_null_check_tests[--cnt]); 2419 _null_check_tests.map(i,_null_check_tests[--cnt]); 2420 _null_check_tests.pop(); 2421 _null_check_tests.pop(); 2422 i-=2; 2423 } 2424 } 2425 } 2426 2427 // Used by the DFA in dfa_xxx.cpp. Check for a following barrier or 2428 // atomic instruction acting as a store_load barrier without any 2429 // intervening volatile load, and thus we don't need a barrier here. 2430 // We retain the Node to act as a compiler ordering barrier. 2431 bool Matcher::post_store_load_barrier(const Node* vmb) { 2432 Compile* C = Compile::current(); 2433 assert(vmb->is_MemBar(), ""); 2434 assert(vmb->Opcode() != Op_MemBarAcquire && vmb->Opcode() != Op_LoadFence, ""); 2435 const MemBarNode* membar = vmb->as_MemBar(); 2436 2437 // Get the Ideal Proj node, ctrl, that can be used to iterate forward 2438 Node* ctrl = NULL; 2439 for (DUIterator_Fast imax, i = membar->fast_outs(imax); i < imax; i++) { 2440 Node* p = membar->fast_out(i); 2441 assert(p->is_Proj(), "only projections here"); 2442 if ((p->as_Proj()->_con == TypeFunc::Control) && 2443 !C->node_arena()->contains(p)) { // Unmatched old-space only 2444 ctrl = p; 2445 break; 2446 } 2447 } 2448 assert((ctrl != NULL), "missing control projection"); 2449 2450 for (DUIterator_Fast jmax, j = ctrl->fast_outs(jmax); j < jmax; j++) { 2451 Node *x = ctrl->fast_out(j); 2452 int xop = x->Opcode(); 2453 2454 // We don't need current barrier if we see another or a lock 2455 // before seeing volatile load. 2456 // 2457 // Op_Fastunlock previously appeared in the Op_* list below. 2458 // With the advent of 1-0 lock operations we're no longer guaranteed 2459 // that a monitor exit operation contains a serializing instruction. 2460 2461 if (xop == Op_MemBarVolatile || 2462 xop == Op_CompareAndSwapL || 2463 xop == Op_CompareAndSwapP || 2464 xop == Op_CompareAndSwapN || 2465 xop == Op_CompareAndSwapI) { 2466 return true; 2467 } 2468 2469 // Op_FastLock previously appeared in the Op_* list above. 2470 // With biased locking we're no longer guaranteed that a monitor 2471 // enter operation contains a serializing instruction. 2472 if ((xop == Op_FastLock) && !UseBiasedLocking) { 2473 return true; 2474 } 2475 2476 if (x->is_MemBar()) { 2477 // We must retain this membar if there is an upcoming volatile 2478 // load, which will be followed by acquire membar. 2479 if (xop == Op_MemBarAcquire || xop == Op_LoadFence) { 2480 return false; 2481 } else { 2482 // For other kinds of barriers, check by pretending we 2483 // are them, and seeing if we can be removed. 2484 return post_store_load_barrier(x->as_MemBar()); 2485 } 2486 } 2487 2488 // probably not necessary to check for these 2489 if (x->is_Call() || x->is_SafePoint() || x->is_block_proj()) { 2490 return false; 2491 } 2492 } 2493 return false; 2494 } 2495 2496 // Check whether node n is a branch to an uncommon trap that we could 2497 // optimize as test with very high branch costs in case of going to 2498 // the uncommon trap. The code must be able to be recompiled to use 2499 // a cheaper test. 2500 bool Matcher::branches_to_uncommon_trap(const Node *n) { 2501 // Don't do it for natives, adapters, or runtime stubs 2502 Compile *C = Compile::current(); 2503 if (!C->is_method_compilation()) return false; 2504 2505 assert(n->is_If(), "You should only call this on if nodes."); 2506 IfNode *ifn = n->as_If(); 2507 2508 Node *ifFalse = NULL; 2509 for (DUIterator_Fast imax, i = ifn->fast_outs(imax); i < imax; i++) { 2510 if (ifn->fast_out(i)->is_IfFalse()) { 2511 ifFalse = ifn->fast_out(i); 2512 break; 2513 } 2514 } 2515 assert(ifFalse, "An If should have an ifFalse. Graph is broken."); 2516 2517 Node *reg = ifFalse; 2518 int cnt = 4; // We must protect against cycles. Limit to 4 iterations. 2519 // Alternatively use visited set? Seems too expensive. 2520 while (reg != NULL && cnt > 0) { 2521 CallNode *call = NULL; 2522 RegionNode *nxt_reg = NULL; 2523 for (DUIterator_Fast imax, i = reg->fast_outs(imax); i < imax; i++) { 2524 Node *o = reg->fast_out(i); 2525 if (o->is_Call()) { 2526 call = o->as_Call(); 2527 } 2528 if (o->is_Region()) { 2529 nxt_reg = o->as_Region(); 2530 } 2531 } 2532 2533 if (call && 2534 call->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { 2535 const Type* trtype = call->in(TypeFunc::Parms)->bottom_type(); 2536 if (trtype->isa_int() && trtype->is_int()->is_con()) { 2537 jint tr_con = trtype->is_int()->get_con(); 2538 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); 2539 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); 2540 assert((int)reason < (int)BitsPerInt, "recode bit map"); 2541 2542 if (is_set_nth_bit(C->allowed_deopt_reasons(), (int)reason) 2543 && action != Deoptimization::Action_none) { 2544 // This uncommon trap is sure to recompile, eventually. 2545 // When that happens, C->too_many_traps will prevent 2546 // this transformation from happening again. 2547 return true; 2548 } 2549 } 2550 } 2551 2552 reg = nxt_reg; 2553 cnt--; 2554 } 2555 2556 return false; 2557 } 2558 2559 //============================================================================= 2560 //---------------------------State--------------------------------------------- 2561 State::State(void) { 2562 #ifdef ASSERT 2563 _id = 0; 2564 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2565 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2566 //memset(_cost, -1, sizeof(_cost)); 2567 //memset(_rule, -1, sizeof(_rule)); 2568 #endif 2569 memset(_valid, 0, sizeof(_valid)); 2570 } 2571 2572 #ifdef ASSERT 2573 State::~State() { 2574 _id = 99; 2575 _kids[0] = _kids[1] = (State*)(intptr_t) CONST64(0xcafebabecafebabe); 2576 _leaf = (Node*)(intptr_t) CONST64(0xbaadf00dbaadf00d); 2577 memset(_cost, -3, sizeof(_cost)); 2578 memset(_rule, -3, sizeof(_rule)); 2579 } 2580 #endif 2581 2582 #ifndef PRODUCT 2583 //---------------------------dump---------------------------------------------- 2584 void State::dump() { 2585 tty->print("\n"); 2586 dump(0); 2587 } 2588 2589 void State::dump(int depth) { 2590 for( int j = 0; j < depth; j++ ) 2591 tty->print(" "); 2592 tty->print("--N: "); 2593 _leaf->dump(); 2594 uint i; 2595 for( i = 0; i < _LAST_MACH_OPER; i++ ) 2596 // Check for valid entry 2597 if( valid(i) ) { 2598 for( int j = 0; j < depth; j++ ) 2599 tty->print(" "); 2600 assert(_cost[i] != max_juint, "cost must be a valid value"); 2601 assert(_rule[i] < _last_Mach_Node, "rule[i] must be valid rule"); 2602 tty->print_cr("%s %d %s", 2603 ruleName[i], _cost[i], ruleName[_rule[i]] ); 2604 } 2605 tty->cr(); 2606 2607 for( i=0; i<2; i++ ) 2608 if( _kids[i] ) 2609 _kids[i]->dump(depth+1); 2610 } 2611 #endif