1 /* 2 * Copyright (c) 1997, 2013, 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 #ifndef SHARE_VM_OPTO_MATCHER_HPP 26 #define SHARE_VM_OPTO_MATCHER_HPP 27 28 #include "libadt/vectset.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/node.hpp" 31 #include "opto/phaseX.hpp" 32 #include "opto/regmask.hpp" 33 34 class Compile; 35 class Node; 36 class MachNode; 37 class MachTypeNode; 38 class MachOper; 39 40 //---------------------------Matcher------------------------------------------- 41 class Matcher : public PhaseTransform { 42 friend class VMStructs; 43 // Private arena of State objects 44 ResourceArea _states_arena; 45 46 VectorSet _visited; // Visit bits 47 48 // Used to control the Label pass 49 VectorSet _shared; // Shared Ideal Node 50 VectorSet _dontcare; // Nothing the matcher cares about 51 52 // Private methods which perform the actual matching and reduction 53 // Walks the label tree, generating machine nodes 54 MachNode *ReduceInst( State *s, int rule, Node *&mem); 55 void ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach); 56 uint ReduceInst_Interior(State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds); 57 void ReduceOper( State *s, int newrule, Node *&mem, MachNode *mach ); 58 59 // If this node already matched using "rule", return the MachNode for it. 60 MachNode* find_shared_node(Node* n, uint rule); 61 62 // Convert a dense opcode number to an expanded rule number 63 const int *_reduceOp; 64 const int *_leftOp; 65 const int *_rightOp; 66 67 // Map dense opcode number to info on when rule is swallowed constant. 68 const bool *_swallowed; 69 70 // Map dense rule number to determine if this is an instruction chain rule 71 const uint _begin_inst_chain_rule; 72 const uint _end_inst_chain_rule; 73 74 // We want to clone constants and possible CmpI-variants. 75 // If we do not clone CmpI, then we can have many instances of 76 // condition codes alive at once. This is OK on some chips and 77 // bad on others. Hence the machine-dependent table lookup. 78 const char *_must_clone; 79 80 // Find shared Nodes, or Nodes that otherwise are Matcher roots 81 void find_shared( Node *n ); 82 83 // Debug and profile information for nodes in old space: 84 GrowableArray<Node_Notes*>* _old_node_note_array; 85 86 // Node labeling iterator for instruction selection 87 Node *Label_Root( const Node *n, State *svec, Node *control, const Node *mem ); 88 89 Node *transform( Node *dummy ); 90 91 Node_List _projection_list; // For Machine nodes killing many values 92 93 Node_Array _shared_nodes; 94 95 debug_only(Node_Array _old2new_map;) // Map roots of ideal-trees to machine-roots 96 debug_only(Node_Array _new2old_map;) // Maps machine nodes back to ideal 97 98 // Accessors for the inherited field PhaseTransform::_nodes: 99 void grow_new_node_array(uint idx_limit) { 100 _nodes.map(idx_limit-1, NULL); 101 } 102 bool has_new_node(const Node* n) const { 103 return _nodes.at(n->_idx) != NULL; 104 } 105 Node* new_node(const Node* n) const { 106 assert(has_new_node(n), "set before get"); 107 return _nodes.at(n->_idx); 108 } 109 void set_new_node(const Node* n, Node *nn) { 110 assert(!has_new_node(n), "set only once"); 111 _nodes.map(n->_idx, nn); 112 } 113 114 #ifdef ASSERT 115 // Make sure only new nodes are reachable from this node 116 void verify_new_nodes_only(Node* root); 117 118 Node* _mem_node; // Ideal memory node consumed by mach node 119 #endif 120 121 // Mach node for ConP #NULL 122 MachNode* _mach_null; 123 124 public: 125 int LabelRootDepth; 126 // Convert ideal machine register to a register mask for spill-loads 127 static const RegMask *idealreg2regmask[]; 128 RegMask *idealreg2spillmask [_last_machine_leaf]; 129 RegMask *idealreg2debugmask [_last_machine_leaf]; 130 RegMask *idealreg2mhdebugmask[_last_machine_leaf]; 131 void init_spill_mask( Node *ret ); 132 // Convert machine register number to register mask 133 static uint mreg2regmask_max; 134 static RegMask mreg2regmask[]; 135 static RegMask STACK_ONLY_mask; 136 137 MachNode* mach_null() const { return _mach_null; } 138 139 bool is_shared( Node *n ) { return _shared.test(n->_idx) != 0; } 140 void set_shared( Node *n ) { _shared.set(n->_idx); } 141 bool is_visited( Node *n ) { return _visited.test(n->_idx) != 0; } 142 void set_visited( Node *n ) { _visited.set(n->_idx); } 143 bool is_dontcare( Node *n ) { return _dontcare.test(n->_idx) != 0; } 144 void set_dontcare( Node *n ) { _dontcare.set(n->_idx); } 145 146 // Mode bit to tell DFA and expand rules whether we are running after 147 // (or during) register selection. Usually, the matcher runs before, 148 // but it will also get called to generate post-allocation spill code. 149 // In this situation, it is a deadly error to attempt to allocate more 150 // temporary registers. 151 bool _allocation_started; 152 153 // Machine register names 154 static const char *regName[]; 155 // Machine register encodings 156 static const unsigned char _regEncode[]; 157 // Machine Node names 158 const char **_ruleName; 159 // Rules that are cheaper to rematerialize than to spill 160 static const uint _begin_rematerialize; 161 static const uint _end_rematerialize; 162 163 // An array of chars, from 0 to _last_Mach_Reg. 164 // No Save = 'N' (for register windows) 165 // Save on Entry = 'E' 166 // Save on Call = 'C' 167 // Always Save = 'A' (same as SOE + SOC) 168 const char *_register_save_policy; 169 const char *_c_reg_save_policy; 170 // Convert a machine register to a machine register type, so-as to 171 // properly match spill code. 172 const int *_register_save_type; 173 // Maps from machine register to boolean; true if machine register can 174 // be holding a call argument in some signature. 175 static bool can_be_java_arg( int reg ); 176 // Maps from machine register to boolean; true if machine register holds 177 // a spillable argument. 178 static bool is_spillable_arg( int reg ); 179 180 // List of IfFalse or IfTrue Nodes that indicate a taken null test. 181 // List is valid in the post-matching space. 182 Node_List _null_check_tests; 183 void collect_null_checks( Node *proj, Node *orig_proj ); 184 void validate_null_checks( ); 185 186 Matcher(); 187 188 // Get a projection node at position pos 189 Node* get_projection(uint pos) { 190 return _projection_list[pos]; 191 } 192 193 // Push a projection node onto the projection list 194 void push_projection(Node* node) { 195 _projection_list.push(node); 196 } 197 198 Node* pop_projection() { 199 return _projection_list.pop(); 200 } 201 202 // Number of nodes in the projection list 203 uint number_of_projections() const { 204 return _projection_list.size(); 205 } 206 207 // Select instructions for entire method 208 void match(); 209 210 // Helper for match 211 OptoReg::Name warp_incoming_stk_arg( VMReg reg ); 212 213 // Transform, then walk. Does implicit DCE while walking. 214 // Name changed from "transform" to avoid it being virtual. 215 Node *xform( Node *old_space_node, int Nodes ); 216 217 // Match a single Ideal Node - turn it into a 1-Node tree; Label & Reduce. 218 MachNode *match_tree( const Node *n ); 219 MachNode *match_sfpt( SafePointNode *sfpt ); 220 // Helper for match_sfpt 221 OptoReg::Name warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call ); 222 223 // Initialize first stack mask and related masks. 224 void init_first_stack_mask(); 225 226 // If we should save-on-entry this register 227 bool is_save_on_entry( int reg ); 228 229 // Fixup the save-on-entry registers 230 void Fixup_Save_On_Entry( ); 231 232 // --- Frame handling --- 233 234 // Register number of the stack slot corresponding to the incoming SP. 235 // Per the Big Picture in the AD file, it is: 236 // SharedInfo::stack0 + locks + in_preserve_stack_slots + pad2. 237 OptoReg::Name _old_SP; 238 239 // Register number of the stack slot corresponding to the highest incoming 240 // argument on the stack. Per the Big Picture in the AD file, it is: 241 // _old_SP + out_preserve_stack_slots + incoming argument size. 242 OptoReg::Name _in_arg_limit; 243 244 // Register number of the stack slot corresponding to the new SP. 245 // Per the Big Picture in the AD file, it is: 246 // _in_arg_limit + pad0 247 OptoReg::Name _new_SP; 248 249 // Register number of the stack slot corresponding to the highest outgoing 250 // argument on the stack. Per the Big Picture in the AD file, it is: 251 // _new_SP + max outgoing arguments of all calls 252 OptoReg::Name _out_arg_limit; 253 254 OptoRegPair *_parm_regs; // Array of machine registers per argument 255 RegMask *_calling_convention_mask; // Array of RegMasks per argument 256 257 // Does matcher have a match rule for this ideal node? 258 static const bool has_match_rule(int opcode); 259 static const bool _hasMatchRule[_last_opcode]; 260 261 // Does matcher have a match rule for this ideal node and is the 262 // predicate (if there is one) true? 263 // NOTE: If this function is used more commonly in the future, ADLC 264 // should generate this one. 265 static const bool match_rule_supported(int opcode); 266 267 // Used to determine if we have fast l2f conversion 268 // USII has it, USIII doesn't 269 static const bool convL2FSupported(void); 270 271 // Vector width in bytes 272 static const int vector_width_in_bytes(BasicType bt); 273 274 // Limits on vector size (number of elements). 275 static const int max_vector_size(const BasicType bt); 276 static const int min_vector_size(const BasicType bt); 277 static const bool vector_size_supported(const BasicType bt, int size) { 278 return (Matcher::max_vector_size(bt) >= size && 279 Matcher::min_vector_size(bt) <= size); 280 } 281 282 // Vector ideal reg 283 static const int vector_ideal_reg(int len); 284 static const int vector_shift_count_ideal_reg(int len); 285 286 // CPU supports misaligned vectors store/load. 287 static const bool misaligned_vectors_ok(); 288 289 // Should original key array reference be passed to AES stubs 290 static const bool pass_original_key_for_aes(); 291 292 // Used to determine a "low complexity" 64-bit constant. (Zero is simple.) 293 // The standard of comparison is one (StoreL ConL) vs. two (StoreI ConI). 294 // Depends on the details of 64-bit constant generation on the CPU. 295 static const bool isSimpleConstant64(jlong con); 296 297 // These calls are all generated by the ADLC 298 299 // TRUE - grows up, FALSE - grows down (Intel) 300 virtual bool stack_direction() const; 301 302 // Java-Java calling convention 303 // (what you use when Java calls Java) 304 305 // Alignment of stack in bytes, standard Intel word alignment is 4. 306 // Sparc probably wants at least double-word (8). 307 static uint stack_alignment_in_bytes(); 308 // Alignment of stack, measured in stack slots. 309 // The size of stack slots is defined by VMRegImpl::stack_slot_size. 310 static uint stack_alignment_in_slots() { 311 return stack_alignment_in_bytes() / (VMRegImpl::stack_slot_size); 312 } 313 314 // Array mapping arguments to registers. Argument 0 is usually the 'this' 315 // pointer. Registers can include stack-slots and regular registers. 316 static void calling_convention( BasicType *, VMRegPair *, uint len, bool is_outgoing ); 317 318 // Convert a sig into a calling convention register layout 319 // and find interesting things about it. 320 static OptoReg::Name find_receiver( bool is_outgoing ); 321 // Return address register. On Intel it is a stack-slot. On PowerPC 322 // it is the Link register. On Sparc it is r31? 323 virtual OptoReg::Name return_addr() const; 324 RegMask _return_addr_mask; 325 // Return value register. On Intel it is EAX. On Sparc i0/o0. 326 static OptoRegPair return_value(int ideal_reg, bool is_outgoing); 327 static OptoRegPair c_return_value(int ideal_reg, bool is_outgoing); 328 RegMask _return_value_mask; 329 // Inline Cache Register 330 static OptoReg::Name inline_cache_reg(); 331 static int inline_cache_reg_encode(); 332 333 // Register for DIVI projection of divmodI 334 static RegMask divI_proj_mask(); 335 // Register for MODI projection of divmodI 336 static RegMask modI_proj_mask(); 337 338 // Register for DIVL projection of divmodL 339 static RegMask divL_proj_mask(); 340 // Register for MODL projection of divmodL 341 static RegMask modL_proj_mask(); 342 343 // Use hardware DIV instruction when it is faster than 344 // a code which use multiply for division by constant. 345 static bool use_asm_for_ldiv_by_con( jlong divisor ); 346 347 static const RegMask method_handle_invoke_SP_save_mask(); 348 349 // Java-Interpreter calling convention 350 // (what you use when calling between compiled-Java and Interpreted-Java 351 352 // Number of callee-save + always-save registers 353 // Ignores frame pointer and "special" registers 354 static int number_of_saved_registers(); 355 356 // The Method-klass-holder may be passed in the inline_cache_reg 357 // and then expanded into the inline_cache_reg and a method_oop register 358 359 static OptoReg::Name interpreter_method_oop_reg(); 360 static int interpreter_method_oop_reg_encode(); 361 362 static OptoReg::Name compiler_method_oop_reg(); 363 static const RegMask &compiler_method_oop_reg_mask(); 364 static int compiler_method_oop_reg_encode(); 365 366 // Interpreter's Frame Pointer Register 367 static OptoReg::Name interpreter_frame_pointer_reg(); 368 369 // Java-Native calling convention 370 // (what you use when intercalling between Java and C++ code) 371 372 // Array mapping arguments to registers. Argument 0 is usually the 'this' 373 // pointer. Registers can include stack-slots and regular registers. 374 static void c_calling_convention( BasicType*, VMRegPair *, uint ); 375 // Frame pointer. The frame pointer is kept at the base of the stack 376 // and so is probably the stack pointer for most machines. On Intel 377 // it is ESP. On the PowerPC it is R1. On Sparc it is SP. 378 OptoReg::Name c_frame_pointer() const; 379 static RegMask c_frame_ptr_mask; 380 381 // !!!!! Special stuff for building ScopeDescs 382 virtual int regnum_to_fpu_offset(int regnum); 383 384 // Is this branch offset small enough to be addressed by a short branch? 385 bool is_short_branch_offset(int rule, int br_size, int offset); 386 387 // Optional scaling for the parameter to the ClearArray/CopyArray node. 388 static const bool init_array_count_is_in_bytes; 389 390 // Threshold small size (in bytes) for a ClearArray/CopyArray node. 391 // Anything this size or smaller may get converted to discrete scalar stores. 392 static const int init_array_short_size; 393 394 // Some hardware needs 2 CMOV's for longs. 395 static const int long_cmove_cost(); 396 397 // Some hardware have expensive CMOV for float and double. 398 static const int float_cmove_cost(); 399 400 // Should the Matcher clone shifts on addressing modes, expecting them to 401 // be subsumed into complex addressing expressions or compute them into 402 // registers? True for Intel but false for most RISCs 403 static const bool clone_shift_expressions; 404 405 static bool narrow_oop_use_complex_address(); 406 static bool narrow_klass_use_complex_address(); 407 408 // Generate implicit null check for narrow oops if it can fold 409 // into address expression (x64). 410 // 411 // [R12 + narrow_oop_reg<<3 + offset] // fold into address expression 412 // NullCheck narrow_oop_reg 413 // 414 // When narrow oops can't fold into address expression (Sparc) and 415 // base is not null use decode_not_null and normal implicit null check. 416 // Note, decode_not_null node can be used here since it is referenced 417 // only on non null path but it requires special handling, see 418 // collect_null_checks(): 419 // 420 // decode_not_null narrow_oop_reg, oop_reg // 'shift' and 'add base' 421 // [oop_reg + offset] 422 // NullCheck oop_reg 423 // 424 // With Zero base and when narrow oops can not fold into address 425 // expression use normal implicit null check since only shift 426 // is needed to decode narrow oop. 427 // 428 // decode narrow_oop_reg, oop_reg // only 'shift' 429 // [oop_reg + offset] 430 // NullCheck oop_reg 431 // 432 inline static bool gen_narrow_oop_implicit_null_checks() { 433 return Universe::narrow_oop_use_implicit_null_checks() && 434 (narrow_oop_use_complex_address() || 435 Universe::narrow_oop_base() != NULL); 436 } 437 438 // Is it better to copy float constants, or load them directly from memory? 439 // Intel can load a float constant from a direct address, requiring no 440 // extra registers. Most RISCs will have to materialize an address into a 441 // register first, so they may as well materialize the constant immediately. 442 static const bool rematerialize_float_constants; 443 444 // If CPU can load and store mis-aligned doubles directly then no fixup is 445 // needed. Else we split the double into 2 integer pieces and move it 446 // piece-by-piece. Only happens when passing doubles into C code or when 447 // calling i2c adapters as the Java calling convention forces doubles to be 448 // aligned. 449 static const bool misaligned_doubles_ok; 450 451 // Does the CPU require postalloc expand (see block.cpp for description of 452 // postalloc expand)? 453 static const bool require_postalloc_expand; 454 455 // Perform a platform dependent implicit null fixup. This is needed 456 // on windows95 to take care of some unusual register constraints. 457 void pd_implicit_null_fixup(MachNode *load, uint idx); 458 459 // Advertise here if the CPU requires explicit rounding operations 460 // to implement the UseStrictFP mode. 461 static const bool strict_fp_requires_explicit_rounding; 462 463 // Are floats conerted to double when stored to stack during deoptimization? 464 static bool float_in_double(); 465 // Do ints take an entire long register or just half? 466 static const bool int_in_long; 467 468 // Do the processor's shift instructions only use the low 5/6 bits 469 // of the count for 32/64 bit ints? If not we need to do the masking 470 // ourselves. 471 static const bool need_masked_shift_count; 472 473 // This routine is run whenever a graph fails to match. 474 // If it returns, the compiler should bailout to interpreter without error. 475 // In non-product mode, SoftMatchFailure is false to detect non-canonical 476 // graphs. Print a message and exit. 477 static void soft_match_failure() { 478 if( SoftMatchFailure ) return; 479 else { fatal("SoftMatchFailure is not allowed except in product"); } 480 } 481 482 // Check for a following volatile memory barrier without an 483 // intervening load and thus we don't need a barrier here. We 484 // retain the Node to act as a compiler ordering barrier. 485 static bool post_store_load_barrier(const Node* mb); 486 487 // Does n lead to an uncommon trap that can cause deoptimization? 488 static bool branches_to_uncommon_trap(const Node *n); 489 490 #ifdef ASSERT 491 void dump_old2new_map(); // machine-independent to machine-dependent 492 493 Node* find_old_node(Node* new_node) { 494 return _new2old_map[new_node->_idx]; 495 } 496 #endif 497 }; 498 499 #endif // SHARE_VM_OPTO_MATCHER_HPP