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