1 /*
   2  * Copyright (c) 1997, 2015, 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 
  44 public:
  45 
  46   // State and MStack class used in xform() and find_shared() iterative methods.
  47   enum Node_State { Pre_Visit,  // node has to be pre-visited
  48                     Visit,  // visit node
  49                     Post_Visit,  // post-visit node
  50                     Alt_Post_Visit   // alternative post-visit path
  51   };
  52 
  53   class MStack: public Node_Stack {
  54   public:
  55     MStack(int size) : Node_Stack(size) { }
  56 
  57     void push(Node *n, Node_State ns) {
  58       Node_Stack::push(n, (uint)ns);
  59     }
  60     void push(Node *n, Node_State ns, Node *parent, int indx) {
  61       ++_inode_top;
  62       if ((_inode_top + 1) >= _inode_max) grow();
  63       _inode_top->node = parent;
  64       _inode_top->indx = (uint)indx;
  65       ++_inode_top;
  66       _inode_top->node = n;
  67       _inode_top->indx = (uint)ns;
  68     }
  69     Node *parent() {
  70       pop();
  71       return node();
  72     }
  73     Node_State state() const {
  74       return (Node_State)index();
  75     }
  76     void set_state(Node_State ns) {
  77       set_index((uint)ns);
  78     }
  79   };
  80 
  81 private:
  82   // Private arena of State objects
  83   ResourceArea _states_arena;
  84 
  85   VectorSet   _visited;         // Visit bits
  86 
  87   // Used to control the Label pass
  88   VectorSet   _shared;          // Shared Ideal Node
  89   VectorSet   _dontcare;        // Nothing the matcher cares about
  90 
  91   // Private methods which perform the actual matching and reduction
  92   // Walks the label tree, generating machine nodes
  93   MachNode *ReduceInst( State *s, int rule, Node *&mem);
  94   void ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach);
  95   uint ReduceInst_Interior(State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds);
  96   void ReduceOper( State *s, int newrule, Node *&mem, MachNode *mach );
  97 
  98   // If this node already matched using "rule", return the MachNode for it.
  99   MachNode* find_shared_node(Node* n, uint rule);
 100 
 101   // Convert a dense opcode number to an expanded rule number
 102   const int *_reduceOp;
 103   const int *_leftOp;
 104   const int *_rightOp;
 105 
 106   // Map dense opcode number to info on when rule is swallowed constant.
 107   const bool *_swallowed;
 108 
 109   // Map dense rule number to determine if this is an instruction chain rule
 110   const uint _begin_inst_chain_rule;
 111   const uint _end_inst_chain_rule;
 112 
 113   // We want to clone constants and possible CmpI-variants.
 114   // If we do not clone CmpI, then we can have many instances of
 115   // condition codes alive at once.  This is OK on some chips and
 116   // bad on others.  Hence the machine-dependent table lookup.
 117   const char *_must_clone;
 118 
 119   // Find shared Nodes, or Nodes that otherwise are Matcher roots
 120   void find_shared( Node *n );
 121 #ifdef X86
 122   bool is_bmi_pattern(Node *n, Node *m);
 123 #endif
 124 
 125   // Debug and profile information for nodes in old space:
 126   GrowableArray<Node_Notes*>* _old_node_note_array;
 127 
 128   // Node labeling iterator for instruction selection
 129   Node *Label_Root( const Node *n, State *svec, Node *control, const Node *mem );
 130 
 131   Node *transform( Node *dummy );
 132 
 133   Node_List _projection_list;        // For Machine nodes killing many values
 134 
 135   Node_Array _shared_nodes;
 136 
 137   debug_only(Node_Array _old2new_map;)   // Map roots of ideal-trees to machine-roots
 138   debug_only(Node_Array _new2old_map;)   // Maps machine nodes back to ideal
 139 
 140   // Accessors for the inherited field PhaseTransform::_nodes:
 141   void   grow_new_node_array(uint idx_limit) {
 142     _nodes.map(idx_limit-1, NULL);
 143   }
 144   bool    has_new_node(const Node* n) const {
 145     return _nodes.at(n->_idx) != NULL;
 146   }
 147   Node*       new_node(const Node* n) const {
 148     assert(has_new_node(n), "set before get");
 149     return _nodes.at(n->_idx);
 150   }
 151   void    set_new_node(const Node* n, Node *nn) {
 152     assert(!has_new_node(n), "set only once");
 153     _nodes.map(n->_idx, nn);
 154   }
 155 
 156 #ifdef ASSERT
 157   // Make sure only new nodes are reachable from this node
 158   void verify_new_nodes_only(Node* root);
 159 
 160   Node* _mem_node;   // Ideal memory node consumed by mach node
 161 #endif
 162 
 163   // Mach node for ConP #NULL
 164   MachNode* _mach_null;
 165 
 166   void handle_precedence_edges(Node* n, MachNode *mach);
 167 
 168 public:
 169   int LabelRootDepth;
 170   // Convert ideal machine register to a register mask for spill-loads
 171   static const RegMask *idealreg2regmask[];
 172   RegMask *idealreg2spillmask  [_last_machine_leaf];
 173   RegMask *idealreg2debugmask  [_last_machine_leaf];
 174   RegMask *idealreg2mhdebugmask[_last_machine_leaf];
 175   void init_spill_mask( Node *ret );
 176   // Convert machine register number to register mask
 177   static uint mreg2regmask_max;
 178   static RegMask mreg2regmask[];
 179   static RegMask STACK_ONLY_mask;
 180 
 181   MachNode* mach_null() const { return _mach_null; }
 182 
 183   bool    is_shared( Node *n ) { return _shared.test(n->_idx) != 0; }
 184   void   set_shared( Node *n ) {  _shared.set(n->_idx); }
 185   bool   is_visited( Node *n ) { return _visited.test(n->_idx) != 0; }
 186   void  set_visited( Node *n ) { _visited.set(n->_idx); }
 187   bool  is_dontcare( Node *n ) { return _dontcare.test(n->_idx) != 0; }
 188   void set_dontcare( Node *n ) {  _dontcare.set(n->_idx); }
 189 
 190   // Mode bit to tell DFA and expand rules whether we are running after
 191   // (or during) register selection.  Usually, the matcher runs before,
 192   // but it will also get called to generate post-allocation spill code.
 193   // In this situation, it is a deadly error to attempt to allocate more
 194   // temporary registers.
 195   bool _allocation_started;
 196 
 197   // Machine register names
 198   static const char *regName[];
 199   // Machine register encodings
 200   static const unsigned char _regEncode[];
 201   // Machine Node names
 202   const char **_ruleName;
 203   // Rules that are cheaper to rematerialize than to spill
 204   static const uint _begin_rematerialize;
 205   static const uint _end_rematerialize;
 206 
 207   // An array of chars, from 0 to _last_Mach_Reg.
 208   // No Save       = 'N' (for register windows)
 209   // Save on Entry = 'E'
 210   // Save on Call  = 'C'
 211   // Always Save   = 'A' (same as SOE + SOC)
 212   const char *_register_save_policy;
 213   const char *_c_reg_save_policy;
 214   // Convert a machine register to a machine register type, so-as to
 215   // properly match spill code.
 216   const int *_register_save_type;
 217   // Maps from machine register to boolean; true if machine register can
 218   // be holding a call argument in some signature.
 219   static bool can_be_java_arg( int reg );
 220   // Maps from machine register to boolean; true if machine register holds
 221   // a spillable argument.
 222   static bool is_spillable_arg( int reg );
 223 
 224   // List of IfFalse or IfTrue Nodes that indicate a taken null test.
 225   // List is valid in the post-matching space.
 226   Node_List _null_check_tests;
 227   void collect_null_checks( Node *proj, Node *orig_proj );
 228   void validate_null_checks( );
 229 
 230   Matcher();
 231 
 232   // Get a projection node at position pos
 233   Node* get_projection(uint pos) {
 234     return _projection_list[pos];
 235   }
 236 
 237   // Push a projection node onto the projection list
 238   void push_projection(Node* node) {
 239     _projection_list.push(node);
 240   }
 241 
 242   Node* pop_projection() {
 243     return _projection_list.pop();
 244   }
 245 
 246   // Number of nodes in the projection list
 247   uint number_of_projections() const {
 248     return _projection_list.size();
 249   }
 250 
 251   // Select instructions for entire method
 252   void match();
 253 
 254   // Helper for match
 255   OptoReg::Name warp_incoming_stk_arg( VMReg reg );
 256 
 257   // Transform, then walk.  Does implicit DCE while walking.
 258   // Name changed from "transform" to avoid it being virtual.
 259   Node *xform( Node *old_space_node, int Nodes );
 260 
 261   // Match a single Ideal Node - turn it into a 1-Node tree; Label & Reduce.
 262   MachNode *match_tree( const Node *n );
 263   MachNode *match_sfpt( SafePointNode *sfpt );
 264   // Helper for match_sfpt
 265   OptoReg::Name warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call );
 266 
 267   // Initialize first stack mask and related masks.
 268   void init_first_stack_mask();
 269 
 270   // If we should save-on-entry this register
 271   bool is_save_on_entry( int reg );
 272 
 273   // Fixup the save-on-entry registers
 274   void Fixup_Save_On_Entry( );
 275 
 276   // --- Frame handling ---
 277 
 278   // Register number of the stack slot corresponding to the incoming SP.
 279   // Per the Big Picture in the AD file, it is:
 280   //   SharedInfo::stack0 + locks + in_preserve_stack_slots + pad2.
 281   OptoReg::Name _old_SP;
 282 
 283   // Register number of the stack slot corresponding to the highest incoming
 284   // argument on the stack.  Per the Big Picture in the AD file, it is:
 285   //   _old_SP + out_preserve_stack_slots + incoming argument size.
 286   OptoReg::Name _in_arg_limit;
 287 
 288   // Register number of the stack slot corresponding to the new SP.
 289   // Per the Big Picture in the AD file, it is:
 290   //   _in_arg_limit + pad0
 291   OptoReg::Name _new_SP;
 292 
 293   // Register number of the stack slot corresponding to the highest outgoing
 294   // argument on the stack.  Per the Big Picture in the AD file, it is:
 295   //   _new_SP + max outgoing arguments of all calls
 296   OptoReg::Name _out_arg_limit;
 297 
 298   OptoRegPair *_parm_regs;        // Array of machine registers per argument
 299   RegMask *_calling_convention_mask; // Array of RegMasks per argument
 300 
 301   // Does matcher have a match rule for this ideal node?
 302   static const bool has_match_rule(int opcode);
 303   static const bool _hasMatchRule[_last_opcode];
 304 
 305   // Does matcher have a match rule for this ideal node and is the
 306   // predicate (if there is one) true?
 307   // NOTE: If this function is used more commonly in the future, ADLC
 308   // should generate this one.
 309   static const bool match_rule_supported(int opcode);
 310 
 311   // identify extra cases that we might want to provide match rules for
 312   // e.g. Op_ vector nodes and other intrinsics while guarding with vlen
 313   static const bool match_rule_supported_vector(int opcode, int vlen);
 314 
 315   // Some microarchitectures have mask registers used on vectors
 316   static const bool has_predicated_vectors(void);
 317 
 318   // Some uarchs have different sized float register resources
 319   static const int float_pressure(int default_pressure_threshold);
 320 
 321   // Used to determine if we have fast l2f conversion
 322   // USII has it, USIII doesn't
 323   static const bool convL2FSupported(void);
 324 
 325   // Vector width in bytes
 326   static const int vector_width_in_bytes(BasicType bt);
 327 
 328   // Limits on vector size (number of elements).
 329   static const int max_vector_size(const BasicType bt);
 330   static const int min_vector_size(const BasicType bt);
 331   static const bool vector_size_supported(const BasicType bt, int size) {
 332     return (Matcher::max_vector_size(bt) >= size &&
 333             Matcher::min_vector_size(bt) <= size);
 334   }
 335 
 336   // Vector ideal reg
 337   static const int vector_ideal_reg(int len);
 338   static const int vector_shift_count_ideal_reg(int len);
 339 
 340   // CPU supports misaligned vectors store/load.
 341   static const bool misaligned_vectors_ok();
 342 
 343   // Should original key array reference be passed to AES stubs
 344   static const bool pass_original_key_for_aes();
 345 
 346   // Used to determine a "low complexity" 64-bit constant.  (Zero is simple.)
 347   // The standard of comparison is one (StoreL ConL) vs. two (StoreI ConI).
 348   // Depends on the details of 64-bit constant generation on the CPU.
 349   static const bool isSimpleConstant64(jlong con);
 350 
 351   // These calls are all generated by the ADLC
 352 
 353   // TRUE - grows up, FALSE - grows down (Intel)
 354   virtual bool stack_direction() const;
 355 
 356   // Java-Java calling convention
 357   // (what you use when Java calls Java)
 358 
 359   // Alignment of stack in bytes, standard Intel word alignment is 4.
 360   // Sparc probably wants at least double-word (8).
 361   static uint stack_alignment_in_bytes();
 362   // Alignment of stack, measured in stack slots.
 363   // The size of stack slots is defined by VMRegImpl::stack_slot_size.
 364   static uint stack_alignment_in_slots() {
 365     return stack_alignment_in_bytes() / (VMRegImpl::stack_slot_size);
 366   }
 367 
 368   // Array mapping arguments to registers.  Argument 0 is usually the 'this'
 369   // pointer.  Registers can include stack-slots and regular registers.
 370   static void calling_convention( BasicType *, VMRegPair *, uint len, bool is_outgoing );
 371 
 372   // Convert a sig into a calling convention register layout
 373   // and find interesting things about it.
 374   static OptoReg::Name  find_receiver( bool is_outgoing );
 375   // Return address register.  On Intel it is a stack-slot.  On PowerPC
 376   // it is the Link register.  On Sparc it is r31?
 377   virtual OptoReg::Name return_addr() const;
 378   RegMask              _return_addr_mask;
 379   // Return value register.  On Intel it is EAX.  On Sparc i0/o0.
 380   static OptoRegPair   return_value(int ideal_reg, bool is_outgoing);
 381   static OptoRegPair c_return_value(int ideal_reg, bool is_outgoing);
 382   RegMask                     _return_value_mask;
 383   // Inline Cache Register
 384   static OptoReg::Name  inline_cache_reg();
 385   static int            inline_cache_reg_encode();
 386 
 387   // Register for DIVI projection of divmodI
 388   static RegMask divI_proj_mask();
 389   // Register for MODI projection of divmodI
 390   static RegMask modI_proj_mask();
 391 
 392   // Register for DIVL projection of divmodL
 393   static RegMask divL_proj_mask();
 394   // Register for MODL projection of divmodL
 395   static RegMask modL_proj_mask();
 396 
 397   // Use hardware DIV instruction when it is faster than
 398   // a code which use multiply for division by constant.
 399   static bool use_asm_for_ldiv_by_con( jlong divisor );
 400 
 401   static const RegMask method_handle_invoke_SP_save_mask();
 402 
 403   // Java-Interpreter calling convention
 404   // (what you use when calling between compiled-Java and Interpreted-Java
 405 
 406   // Number of callee-save + always-save registers
 407   // Ignores frame pointer and "special" registers
 408   static int  number_of_saved_registers();
 409 
 410   // The Method-klass-holder may be passed in the inline_cache_reg
 411   // and then expanded into the inline_cache_reg and a method_oop register
 412 
 413   static OptoReg::Name  interpreter_method_oop_reg();
 414   static int            interpreter_method_oop_reg_encode();
 415 
 416   static OptoReg::Name  compiler_method_oop_reg();
 417   static const RegMask &compiler_method_oop_reg_mask();
 418   static int            compiler_method_oop_reg_encode();
 419 
 420   // Interpreter's Frame Pointer Register
 421   static OptoReg::Name  interpreter_frame_pointer_reg();
 422 
 423   // Java-Native calling convention
 424   // (what you use when intercalling between Java and C++ code)
 425 
 426   // Array mapping arguments to registers.  Argument 0 is usually the 'this'
 427   // pointer.  Registers can include stack-slots and regular registers.
 428   static void c_calling_convention( BasicType*, VMRegPair *, uint );
 429   // Frame pointer. The frame pointer is kept at the base of the stack
 430   // and so is probably the stack pointer for most machines.  On Intel
 431   // it is ESP.  On the PowerPC it is R1.  On Sparc it is SP.
 432   OptoReg::Name  c_frame_pointer() const;
 433   static RegMask c_frame_ptr_mask;
 434 
 435   // !!!!! Special stuff for building ScopeDescs
 436   virtual int      regnum_to_fpu_offset(int regnum);
 437 
 438   // Is this branch offset small enough to be addressed by a short branch?
 439   bool is_short_branch_offset(int rule, int br_size, int offset);
 440 
 441   // Optional scaling for the parameter to the ClearArray/CopyArray node.
 442   static const bool init_array_count_is_in_bytes;
 443 
 444   // Some hardware needs 2 CMOV's for longs.
 445   static const int long_cmove_cost();
 446 
 447   // Some hardware have expensive CMOV for float and double.
 448   static const int float_cmove_cost();
 449 
 450   // Should the Matcher clone shifts on addressing modes, expecting them to
 451   // be subsumed into complex addressing expressions or compute them into
 452   // registers?  True for Intel but false for most RISCs
 453   bool clone_address_expressions(AddPNode* m, MStack& mstack, VectorSet& address_visited);
 454   // Clone base + offset address expression
 455   bool clone_base_plus_offset_address(AddPNode* m, MStack& mstack, VectorSet& address_visited);
 456 
 457   static bool narrow_oop_use_complex_address();
 458   static bool narrow_klass_use_complex_address();
 459 
 460   // Generate implicit null check for narrow oops if it can fold
 461   // into address expression (x64).
 462   //
 463   // [R12 + narrow_oop_reg<<3 + offset] // fold into address expression
 464   // NullCheck narrow_oop_reg
 465   //
 466   // When narrow oops can't fold into address expression (Sparc) and
 467   // base is not null use decode_not_null and normal implicit null check.
 468   // Note, decode_not_null node can be used here since it is referenced
 469   // only on non null path but it requires special handling, see
 470   // collect_null_checks():
 471   //
 472   // decode_not_null narrow_oop_reg, oop_reg // 'shift' and 'add base'
 473   // [oop_reg + offset]
 474   // NullCheck oop_reg
 475   //
 476   // With Zero base and when narrow oops can not fold into address
 477   // expression use normal implicit null check since only shift
 478   // is needed to decode narrow oop.
 479   //
 480   // decode narrow_oop_reg, oop_reg // only 'shift'
 481   // [oop_reg + offset]
 482   // NullCheck oop_reg
 483   //
 484   inline static bool gen_narrow_oop_implicit_null_checks() {
 485     // Advice matcher to perform null checks on the narrow oop side.
 486     // Implicit checks are not possible on the uncompressed oop side anyway
 487     // (at least not for read accesses).
 488     // Performs significantly better (especially on Power 6).
 489     if (!os::zero_page_read_protected()) {
 490       return true;
 491     }
 492     return Universe::narrow_oop_use_implicit_null_checks() &&
 493            (narrow_oop_use_complex_address() ||
 494             Universe::narrow_oop_base() != NULL);
 495   }
 496 
 497   // Is it better to copy float constants, or load them directly from memory?
 498   // Intel can load a float constant from a direct address, requiring no
 499   // extra registers.  Most RISCs will have to materialize an address into a
 500   // register first, so they may as well materialize the constant immediately.
 501   static const bool rematerialize_float_constants;
 502 
 503   // If CPU can load and store mis-aligned doubles directly then no fixup is
 504   // needed.  Else we split the double into 2 integer pieces and move it
 505   // piece-by-piece.  Only happens when passing doubles into C code or when
 506   // calling i2c adapters as the Java calling convention forces doubles to be
 507   // aligned.
 508   static const bool misaligned_doubles_ok;
 509 
 510   // Does the CPU require postalloc expand (see block.cpp for description of
 511   // postalloc expand)?
 512   static const bool require_postalloc_expand;
 513 
 514   // Perform a platform dependent implicit null fixup.  This is needed
 515   // on windows95 to take care of some unusual register constraints.
 516   void pd_implicit_null_fixup(MachNode *load, uint idx);
 517 
 518   // Advertise here if the CPU requires explicit rounding operations
 519   // to implement the UseStrictFP mode.
 520   static const bool strict_fp_requires_explicit_rounding;
 521 
 522   // Are floats conerted to double when stored to stack during deoptimization?
 523   static bool float_in_double();
 524   // Do ints take an entire long register or just half?
 525   static const bool int_in_long;
 526 
 527   // Do the processor's shift instructions only use the low 5/6 bits
 528   // of the count for 32/64 bit ints? If not we need to do the masking
 529   // ourselves.
 530   static const bool need_masked_shift_count;
 531 
 532   // Whether code generation need accurate ConvI2L types.
 533   static const bool convi2l_type_required;
 534 
 535   // This routine is run whenever a graph fails to match.
 536   // If it returns, the compiler should bailout to interpreter without error.
 537   // In non-product mode, SoftMatchFailure is false to detect non-canonical
 538   // graphs.  Print a message and exit.
 539   static void soft_match_failure() {
 540     if( SoftMatchFailure ) return;
 541     else { fatal("SoftMatchFailure is not allowed except in product"); }
 542   }
 543 
 544   // Check for a following volatile memory barrier without an
 545   // intervening load and thus we don't need a barrier here.  We
 546   // retain the Node to act as a compiler ordering barrier.
 547   static bool post_store_load_barrier(const Node* mb);
 548 
 549   // Does n lead to an uncommon trap that can cause deoptimization?
 550   static bool branches_to_uncommon_trap(const Node *n);
 551 
 552 #ifdef ASSERT
 553   void dump_old2new_map();      // machine-independent to machine-dependent
 554 
 555   Node* find_old_node(Node* new_node) {
 556     return _new2old_map[new_node->_idx];
 557   }
 558 #endif
 559 };
 560 
 561 #endif // SHARE_VM_OPTO_MATCHER_HPP