1 /*
2 * Copyright (c) 1997, 2020, 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_OPTO_MATCHER_HPP
26 #define SHARE_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 bool find_shared_visit(MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx);
122 void find_shared_post_visit(Node* n, uint opcode);
123
124 bool is_vshift_con_pattern(Node *n, Node *m);
125
126 // Debug and profile information for nodes in old space:
127 GrowableArray<Node_Notes*>* _old_node_note_array;
128
129 // Node labeling iterator for instruction selection
130 Node* Label_Root(const Node* n, State* svec, Node* control, Node*& mem);
131
132 Node *transform( Node *dummy );
133
134 Node_List _projection_list; // For Machine nodes killing many values
135
136 Node_Array _shared_nodes;
137
138 debug_only(Node_Array _old2new_map;) // Map roots of ideal-trees to machine-roots
139 debug_only(Node_Array _new2old_map;) // Maps machine nodes back to ideal
140
141 // Accessors for the inherited field PhaseTransform::_nodes:
142 void grow_new_node_array(uint idx_limit) {
143 _nodes.map(idx_limit-1, NULL);
144 }
145 bool has_new_node(const Node* n) const {
146 return _nodes.at(n->_idx) != NULL;
147 }
148 Node* new_node(const Node* n) const {
149 assert(has_new_node(n), "set before get");
150 return _nodes.at(n->_idx);
151 }
152 void set_new_node(const Node* n, Node *nn) {
153 assert(!has_new_node(n), "set only once");
154 _nodes.map(n->_idx, nn);
155 }
156
157 #ifdef ASSERT
158 // Make sure only new nodes are reachable from this node
159 void verify_new_nodes_only(Node* root);
160
161 Node* _mem_node; // Ideal memory node consumed by mach node
162 #endif
163
164 // Mach node for ConP #NULL
165 MachNode* _mach_null;
166
167 void handle_precedence_edges(Node* n, MachNode *mach);
168
169 public:
170 int LabelRootDepth;
171 // Convert ideal machine register to a register mask for spill-loads
172 static const RegMask *idealreg2regmask[];
173 RegMask *idealreg2spillmask [_last_machine_leaf];
174 RegMask *idealreg2debugmask [_last_machine_leaf];
175 RegMask *idealreg2mhdebugmask[_last_machine_leaf];
176 void init_spill_mask( Node *ret );
177 // Convert machine register number to register mask
178 static uint mreg2regmask_max;
179 static RegMask mreg2regmask[];
180 static RegMask STACK_ONLY_mask;
181 static RegMask caller_save_regmask;
182 static RegMask caller_save_regmask_exclude_soe;
183 static RegMask mh_caller_save_regmask;
184 static RegMask mh_caller_save_regmask_exclude_soe;
185
186 MachNode* mach_null() const { return _mach_null; }
187
188 bool is_shared( Node *n ) { return _shared.test(n->_idx) != 0; }
189 void set_shared( Node *n ) { _shared.set(n->_idx); }
190 bool is_visited( Node *n ) { return _visited.test(n->_idx) != 0; }
191 void set_visited( Node *n ) { _visited.set(n->_idx); }
192 bool is_dontcare( Node *n ) { return _dontcare.test(n->_idx) != 0; }
193 void set_dontcare( Node *n ) { _dontcare.set(n->_idx); }
194
195 // Mode bit to tell DFA and expand rules whether we are running after
196 // (or during) register selection. Usually, the matcher runs before,
197 // but it will also get called to generate post-allocation spill code.
198 // In this situation, it is a deadly error to attempt to allocate more
199 // temporary registers.
200 bool _allocation_started;
201
202 // Machine register names
203 static const char *regName[];
204 // Machine register encodings
205 static const unsigned char _regEncode[];
206 // Machine Node names
207 const char **_ruleName;
208 // Rules that are cheaper to rematerialize than to spill
209 static const uint _begin_rematerialize;
210 static const uint _end_rematerialize;
211
212 // An array of chars, from 0 to _last_Mach_Reg.
213 // No Save = 'N' (for register windows)
214 // Save on Entry = 'E'
215 // Save on Call = 'C'
216 // Always Save = 'A' (same as SOE + SOC)
217 const char *_register_save_policy;
218 const char *_c_reg_save_policy;
219 // Convert a machine register to a machine register type, so-as to
220 // properly match spill code.
221 const int *_register_save_type;
222 // Maps from machine register to boolean; true if machine register can
223 // be holding a call argument in some signature.
224 static bool can_be_java_arg( int reg );
225 // Maps from machine register to boolean; true if machine register holds
226 // a spillable argument.
227 static bool is_spillable_arg( int reg );
228
229 // List of IfFalse or IfTrue Nodes that indicate a taken null test.
230 // List is valid in the post-matching space.
231 Node_List _null_check_tests;
232 void collect_null_checks( Node *proj, Node *orig_proj );
233 void validate_null_checks( );
234
235 Matcher();
236
237 // Get a projection node at position pos
238 Node* get_projection(uint pos) {
239 return _projection_list[pos];
240 }
241
242 // Push a projection node onto the projection list
243 void push_projection(Node* node) {
244 _projection_list.push(node);
245 }
246
247 Node* pop_projection() {
248 return _projection_list.pop();
249 }
250
251 // Number of nodes in the projection list
252 uint number_of_projections() const {
253 return _projection_list.size();
254 }
255
256 // Select instructions for entire method
257 void match();
258
259 // Helper for match
260 OptoReg::Name warp_incoming_stk_arg( VMReg reg );
261
262 // Transform, then walk. Does implicit DCE while walking.
263 // Name changed from "transform" to avoid it being virtual.
264 Node *xform( Node *old_space_node, int Nodes );
265
266 // Match a single Ideal Node - turn it into a 1-Node tree; Label & Reduce.
267 MachNode *match_tree( const Node *n );
268 MachNode *match_sfpt( SafePointNode *sfpt );
269 // Helper for match_sfpt
270 OptoReg::Name warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call );
271
272 // Initialize first stack mask and related masks.
273 void init_first_stack_mask();
274
275 // If we should save-on-entry this register
276 bool is_save_on_entry( int reg );
277
278 // Fixup the save-on-entry registers
279 void Fixup_Save_On_Entry( );
280
281 // --- Frame handling ---
282
283 // Register number of the stack slot corresponding to the incoming SP.
284 // Per the Big Picture in the AD file, it is:
285 // SharedInfo::stack0 + locks + in_preserve_stack_slots + pad2.
286 OptoReg::Name _old_SP;
287
288 // Register number of the stack slot corresponding to the highest incoming
289 // argument on the stack. Per the Big Picture in the AD file, it is:
290 // _old_SP + out_preserve_stack_slots + incoming argument size.
291 OptoReg::Name _in_arg_limit;
292
293 // Register number of the stack slot corresponding to the new SP.
294 // Per the Big Picture in the AD file, it is:
295 // _in_arg_limit + pad0
296 OptoReg::Name _new_SP;
297
298 // Register number of the stack slot corresponding to the highest outgoing
299 // argument on the stack. Per the Big Picture in the AD file, it is:
300 // _new_SP + max outgoing arguments of all calls
301 OptoReg::Name _out_arg_limit;
302
303 OptoRegPair *_parm_regs; // Array of machine registers per argument
304 RegMask *_calling_convention_mask; // Array of RegMasks per argument
305
306 // Does matcher have a match rule for this ideal node?
307 static const bool has_match_rule(int opcode);
308 static const bool _hasMatchRule[_last_opcode];
309
310 // Does matcher have a match rule for this ideal node and is the
311 // predicate (if there is one) true?
312 // NOTE: If this function is used more commonly in the future, ADLC
313 // should generate this one.
314 static const bool match_rule_supported(int opcode);
315
316 // identify extra cases that we might want to provide match rules for
317 // e.g. Op_ vector nodes and other intrinsics while guarding with vlen
318 static const bool match_rule_supported_vector(int opcode, int vlen, BasicType bt);
319
320 // Some microarchitectures have mask registers used on vectors
321 static const bool has_predicated_vectors(void);
322
323 // Some uarchs have different sized float register resources
324 static const int float_pressure(int default_pressure_threshold);
325
326 // Used to determine if we have fast l2f conversion
327 // USII has it, USIII doesn't
328 static const bool convL2FSupported(void);
329
330 // Vector width in bytes
331 static const int vector_width_in_bytes(BasicType bt);
332
333 // Limits on vector size (number of elements).
334 static const int max_vector_size(const BasicType bt);
335 static const int min_vector_size(const BasicType bt);
336 static const bool vector_size_supported(const BasicType bt, int size) {
337 return (Matcher::max_vector_size(bt) >= size &&
338 Matcher::min_vector_size(bt) <= size);
339 }
340
341 // Vector ideal reg
342 static const uint vector_ideal_reg(int len);
343
344 // CPU supports misaligned vectors store/load.
345 static const bool misaligned_vectors_ok();
346
347 // Should original key array reference be passed to AES stubs
348 static const bool pass_original_key_for_aes();
349
350 // Used to determine a "low complexity" 64-bit constant. (Zero is simple.)
351 // The standard of comparison is one (StoreL ConL) vs. two (StoreI ConI).
352 // Depends on the details of 64-bit constant generation on the CPU.
353 static const bool isSimpleConstant64(jlong con);
354
355 // These calls are all generated by the ADLC
356
357 // TRUE - grows up, FALSE - grows down (Intel)
358 virtual bool stack_direction() const;
359
360 // Java-Java calling convention
361 // (what you use when Java calls Java)
362
363 // Alignment of stack in bytes, standard Intel word alignment is 4.
364 // Sparc probably wants at least double-word (8).
365 static uint stack_alignment_in_bytes();
366 // Alignment of stack, measured in stack slots.
367 // The size of stack slots is defined by VMRegImpl::stack_slot_size.
368 static uint stack_alignment_in_slots() {
369 return stack_alignment_in_bytes() / (VMRegImpl::stack_slot_size);
370 }
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 calling_convention( BasicType *, VMRegPair *, uint len, bool is_outgoing );
375
376 // Convert a sig into a calling convention register layout
377 // and find interesting things about it.
378 static OptoReg::Name find_receiver( bool is_outgoing );
379 // Return address register. On Intel it is a stack-slot. On PowerPC
380 // it is the Link register. On Sparc it is r31?
381 virtual OptoReg::Name return_addr() const;
382 RegMask _return_addr_mask;
383 // Return value register. On Intel it is EAX. On Sparc i0/o0.
384 static OptoRegPair return_value(uint ideal_reg, bool is_outgoing);
385 static OptoRegPair c_return_value(uint ideal_reg, bool is_outgoing);
386 RegMask _return_value_mask;
387 // Inline Cache Register
388 static OptoReg::Name inline_cache_reg();
389 static int inline_cache_reg_encode();
390
391 // Register for DIVI projection of divmodI
392 static RegMask divI_proj_mask();
393 // Register for MODI projection of divmodI
394 static RegMask modI_proj_mask();
395
396 // Register for DIVL projection of divmodL
397 static RegMask divL_proj_mask();
398 // Register for MODL projection of divmodL
399 static RegMask modL_proj_mask();
400
401 // Use hardware DIV instruction when it is faster than
402 // a code which use multiply for division by constant.
403 static bool use_asm_for_ldiv_by_con( jlong divisor );
404
405 static const RegMask method_handle_invoke_SP_save_mask();
406
407 // Java-Interpreter calling convention
408 // (what you use when calling between compiled-Java and Interpreted-Java
409
410 // Number of callee-save + always-save registers
411 // Ignores frame pointer and "special" registers
412 static int number_of_saved_registers();
413
414 // The Method-klass-holder may be passed in the inline_cache_reg
415 // and then expanded into the inline_cache_reg and a method_ptr register
416
417 static OptoReg::Name interpreter_method_oop_reg();
418 static int interpreter_method_oop_reg_encode();
419
420 static OptoReg::Name compiler_method_oop_reg();
421 static const RegMask &compiler_method_oop_reg_mask();
422 static int compiler_method_oop_reg_encode();
423
424 // Interpreter's Frame Pointer Register
425 static OptoReg::Name interpreter_frame_pointer_reg();
426
427 // Java-Native calling convention
428 // (what you use when intercalling between Java and C++ code)
429
430 // Array mapping arguments to registers. Argument 0 is usually the 'this'
431 // pointer. Registers can include stack-slots and regular registers.
432 static void c_calling_convention( BasicType*, VMRegPair *, uint );
433 // Frame pointer. The frame pointer is kept at the base of the stack
434 // and so is probably the stack pointer for most machines. On Intel
435 // it is ESP. On the PowerPC it is R1. On Sparc it is SP.
436 OptoReg::Name c_frame_pointer() const;
437 static RegMask c_frame_ptr_mask;
438
439 // !!!!! Special stuff for building ScopeDescs
440 virtual int regnum_to_fpu_offset(int regnum);
441
442 // Is this branch offset small enough to be addressed by a short branch?
443 bool is_short_branch_offset(int rule, int br_size, int offset);
444
445 // Optional scaling for the parameter to the ClearArray/CopyArray node.
446 static const bool init_array_count_is_in_bytes;
447
448 // Some hardware needs 2 CMOV's for longs.
449 static const int long_cmove_cost();
450
451 // Some hardware have expensive CMOV for float and double.
452 static const int float_cmove_cost();
453
454 // Should the input 'm' of node 'n' be cloned during matching?
455 // Reports back whether the node was cloned or not.
456 bool clone_node(Node* n, Node* m, Matcher::MStack& mstack);
457 bool pd_clone_node(Node* n, Node* m, Matcher::MStack& mstack);
458
459 // Should the Matcher clone shifts on addressing modes, expecting them to
460 // be subsumed into complex addressing expressions or compute them into
461 // registers? True for Intel but false for most RISCs
462 bool pd_clone_address_expressions(AddPNode* m, MStack& mstack, VectorSet& address_visited);
463 // Clone base + offset address expression
464 bool clone_base_plus_offset_address(AddPNode* m, MStack& mstack, VectorSet& address_visited);
465
466 static bool narrow_oop_use_complex_address();
467 static bool narrow_klass_use_complex_address();
468
469 static bool const_oop_prefer_decode();
470 static bool const_klass_prefer_decode();
471
472 // Generate implicit null check for narrow oops if it can fold
473 // into address expression (x64).
474 //
475 // [R12 + narrow_oop_reg<<3 + offset] // fold into address expression
476 // NullCheck narrow_oop_reg
477 //
478 // When narrow oops can't fold into address expression (Sparc) and
479 // base is not null use decode_not_null and normal implicit null check.
480 // Note, decode_not_null node can be used here since it is referenced
481 // only on non null path but it requires special handling, see
482 // collect_null_checks():
483 //
484 // decode_not_null narrow_oop_reg, oop_reg // 'shift' and 'add base'
485 // [oop_reg + offset]
486 // NullCheck oop_reg
487 //
488 // With Zero base and when narrow oops can not fold into address
489 // expression use normal implicit null check since only shift
490 // is needed to decode narrow oop.
491 //
492 // decode narrow_oop_reg, oop_reg // only 'shift'
493 // [oop_reg + offset]
494 // NullCheck oop_reg
495 //
496 static bool gen_narrow_oop_implicit_null_checks();
497
498 // Is it better to copy float constants, or load them directly from memory?
499 // Intel can load a float constant from a direct address, requiring no
500 // extra registers. Most RISCs will have to materialize an address into a
501 // register first, so they may as well materialize the constant immediately.
502 static const bool rematerialize_float_constants;
503
504 // If CPU can load and store mis-aligned doubles directly then no fixup is
505 // needed. Else we split the double into 2 integer pieces and move it
506 // piece-by-piece. Only happens when passing doubles into C code or when
507 // calling i2c adapters as the Java calling convention forces doubles to be
508 // aligned.
509 static const bool misaligned_doubles_ok;
510
511 // Does the CPU require postalloc expand (see block.cpp for description of
512 // postalloc expand)?
513 static const bool require_postalloc_expand;
514
515 // Does the platform support generic vector operands?
516 // Requires cleanup after selection phase.
517 static const bool supports_generic_vector_operands;
518
519 private:
520 void do_postselect_cleanup();
521
522 void specialize_generic_vector_operands();
523 void specialize_mach_node(MachNode* m);
524 void specialize_temp_node(MachTempNode* tmp, MachNode* use, uint idx);
525 MachOper* specialize_vector_operand(MachNode* m, uint opnd_idx);
526
527 static MachOper* pd_specialize_generic_vector_operand(MachOper* generic_opnd, uint ideal_reg, bool is_temp);
528 static bool is_generic_reg2reg_move(MachNode* m);
529 static bool is_generic_vector(MachOper* opnd);
530
531 const RegMask* regmask_for_ideal_register(uint ideal_reg, Node* ret);
532
533 // Graph verification code
534 DEBUG_ONLY( bool verify_after_postselect_cleanup(); )
535
536 public:
537 // Perform a platform dependent implicit null fixup. This is needed
538 // on windows95 to take care of some unusual register constraints.
539 void pd_implicit_null_fixup(MachNode *load, uint idx);
540
541 // Advertise here if the CPU requires explicit rounding operations to implement strictfp mode.
542 static const bool strict_fp_requires_explicit_rounding;
543
544 // Are floats conerted to double when stored to stack during deoptimization?
545 static bool float_in_double();
546 // Do ints take an entire long register or just half?
547 static const bool int_in_long;
548
549 // Do the processor's shift instructions only use the low 5/6 bits
550 // of the count for 32/64 bit ints? If not we need to do the masking
551 // ourselves.
552 static const bool need_masked_shift_count;
553
554 // Whether code generation need accurate ConvI2L types.
555 static const bool convi2l_type_required;
556
557 // This routine is run whenever a graph fails to match.
558 // If it returns, the compiler should bailout to interpreter without error.
559 // In non-product mode, SoftMatchFailure is false to detect non-canonical
560 // graphs. Print a message and exit.
561 static void soft_match_failure() {
562 if( SoftMatchFailure ) return;
563 else { fatal("SoftMatchFailure is not allowed except in product"); }
564 }
565
566 // Check for a following volatile memory barrier without an
567 // intervening load and thus we don't need a barrier here. We
568 // retain the Node to act as a compiler ordering barrier.
569 static bool post_store_load_barrier(const Node* mb);
570
571 // Does n lead to an uncommon trap that can cause deoptimization?
572 static bool branches_to_uncommon_trap(const Node *n);
573
574 #ifdef ASSERT
575 void dump_old2new_map(); // machine-independent to machine-dependent
576
577 Node* find_old_node(Node* new_node) {
578 return _new2old_map[new_node->_idx];
579 }
580 #endif
581 };
582
583 #endif // SHARE_OPTO_MATCHER_HPP