1 | /* |
2 | * Copyright (c) 1997, 2019, 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_COMPILE_HPP |
26 | #define SHARE_OPTO_COMPILE_HPP |
27 | |
28 | #include "asm/codeBuffer.hpp" |
29 | #include "ci/compilerInterface.hpp" |
30 | #include "code/debugInfoRec.hpp" |
31 | #include "code/exceptionHandlerTable.hpp" |
32 | #include "compiler/compilerOracle.hpp" |
33 | #include "compiler/compileBroker.hpp" |
34 | #include "libadt/dict.hpp" |
35 | #include "libadt/vectset.hpp" |
36 | #include "jfr/jfrEvents.hpp" |
37 | #include "memory/resourceArea.hpp" |
38 | #include "oops/methodData.hpp" |
39 | #include "opto/idealGraphPrinter.hpp" |
40 | #include "opto/phasetype.hpp" |
41 | #include "opto/phase.hpp" |
42 | #include "opto/regmask.hpp" |
43 | #include "runtime/deoptimization.hpp" |
44 | #include "runtime/timerTrace.hpp" |
45 | #include "runtime/vmThread.hpp" |
46 | #include "utilities/ticks.hpp" |
47 | |
48 | class AddPNode; |
49 | class Block; |
50 | class Bundle; |
51 | class C2Compiler; |
52 | class CallGenerator; |
53 | class CloneMap; |
54 | class ConnectionGraph; |
55 | class IdealGraphPrinter; |
56 | class InlineTree; |
57 | class Int_Array; |
58 | class LoadBarrierNode; |
59 | class Matcher; |
60 | class MachConstantNode; |
61 | class MachConstantBaseNode; |
62 | class MachNode; |
63 | class MachOper; |
64 | class MachSafePointNode; |
65 | class Node; |
66 | class Node_Array; |
67 | class Node_Notes; |
68 | class NodeCloneInfo; |
69 | class OptoReg; |
70 | class PhaseCFG; |
71 | class PhaseGVN; |
72 | class PhaseIterGVN; |
73 | class PhaseRegAlloc; |
74 | class PhaseCCP; |
75 | class PhaseCCP_DCE; |
76 | class RootNode; |
77 | class relocInfo; |
78 | class Scope; |
79 | class StartNode; |
80 | class SafePointNode; |
81 | class JVMState; |
82 | class Type; |
83 | class TypeData; |
84 | class TypeInt; |
85 | class TypePtr; |
86 | class TypeOopPtr; |
87 | class TypeFunc; |
88 | class Unique_Node_List; |
89 | class nmethod; |
90 | class WarmCallInfo; |
91 | class Node_Stack; |
92 | struct Final_Reshape_Counts; |
93 | |
94 | enum LoopOptsMode { |
95 | LoopOptsDefault, |
96 | LoopOptsNone, |
97 | LoopOptsShenandoahExpand, |
98 | LoopOptsShenandoahPostExpand, |
99 | LoopOptsZBarrierInsertion, |
100 | LoopOptsSkipSplitIf, |
101 | LoopOptsVerify |
102 | }; |
103 | |
104 | typedef unsigned int node_idx_t; |
105 | class NodeCloneInfo { |
106 | private: |
107 | uint64_t _idx_clone_orig; |
108 | public: |
109 | |
110 | void set_idx(node_idx_t idx) { |
111 | _idx_clone_orig = (_idx_clone_orig & CONST64(0xFFFFFFFF00000000)) | idx; |
112 | } |
113 | node_idx_t idx() const { return (node_idx_t)(_idx_clone_orig & 0xFFFFFFFF); } |
114 | |
115 | void set_gen(int generation) { |
116 | uint64_t g = (uint64_t)generation << 32; |
117 | _idx_clone_orig = (_idx_clone_orig & 0xFFFFFFFF) | g; |
118 | } |
119 | int gen() const { return (int)(_idx_clone_orig >> 32); } |
120 | |
121 | void set(uint64_t x) { _idx_clone_orig = x; } |
122 | void set(node_idx_t x, int g) { set_idx(x); set_gen(g); } |
123 | uint64_t get() const { return _idx_clone_orig; } |
124 | |
125 | NodeCloneInfo(uint64_t idx_clone_orig) : _idx_clone_orig(idx_clone_orig) {} |
126 | NodeCloneInfo(node_idx_t x, int g) : _idx_clone_orig(0) { set(x, g); } |
127 | |
128 | void dump() const; |
129 | }; |
130 | |
131 | class CloneMap { |
132 | friend class Compile; |
133 | private: |
134 | bool _debug; |
135 | Dict* _dict; |
136 | int _clone_idx; // current cloning iteration/generation in loop unroll |
137 | public: |
138 | void* _2p(node_idx_t key) const { return (void*)(intptr_t)key; } // 2 conversion functions to make gcc happy |
139 | node_idx_t _2_node_idx_t(const void* k) const { return (node_idx_t)(intptr_t)k; } |
140 | Dict* dict() const { return _dict; } |
141 | void insert(node_idx_t key, uint64_t val) { assert(_dict->operator[](_2p(key)) == NULL, "key existed" ); _dict->Insert(_2p(key), (void*)val); } |
142 | void insert(node_idx_t key, NodeCloneInfo& ci) { insert(key, ci.get()); } |
143 | void remove(node_idx_t key) { _dict->Delete(_2p(key)); } |
144 | uint64_t value(node_idx_t key) const { return (uint64_t)_dict->operator[](_2p(key)); } |
145 | node_idx_t idx(node_idx_t key) const { return NodeCloneInfo(value(key)).idx(); } |
146 | int gen(node_idx_t key) const { return NodeCloneInfo(value(key)).gen(); } |
147 | int gen(const void* k) const { return gen(_2_node_idx_t(k)); } |
148 | int max_gen() const; |
149 | void clone(Node* old, Node* nnn, int gen); |
150 | void verify_insert_and_clone(Node* old, Node* nnn, int gen); |
151 | void dump(node_idx_t key) const; |
152 | |
153 | int clone_idx() const { return _clone_idx; } |
154 | void set_clone_idx(int x) { _clone_idx = x; } |
155 | bool is_debug() const { return _debug; } |
156 | void set_debug(bool debug) { _debug = debug; } |
157 | static const char* debug_option_name; |
158 | |
159 | bool same_idx(node_idx_t k1, node_idx_t k2) const { return idx(k1) == idx(k2); } |
160 | bool same_gen(node_idx_t k1, node_idx_t k2) const { return gen(k1) == gen(k2); } |
161 | }; |
162 | |
163 | //------------------------------Compile---------------------------------------- |
164 | // This class defines a top-level Compiler invocation. |
165 | |
166 | class Compile : public Phase { |
167 | friend class VMStructs; |
168 | |
169 | public: |
170 | // Fixed alias indexes. (See also MergeMemNode.) |
171 | enum { |
172 | AliasIdxTop = 1, // pseudo-index, aliases to nothing (used as sentinel value) |
173 | AliasIdxBot = 2, // pseudo-index, aliases to everything |
174 | AliasIdxRaw = 3 // hard-wired index for TypeRawPtr::BOTTOM |
175 | }; |
176 | |
177 | // Variant of TraceTime(NULL, &_t_accumulator, CITime); |
178 | // Integrated with logging. If logging is turned on, and CITimeVerbose is true, |
179 | // then brackets are put into the log, with time stamps and node counts. |
180 | // (The time collection itself is always conditionalized on CITime.) |
181 | class TracePhase : public TraceTime { |
182 | private: |
183 | Compile* C; |
184 | CompileLog* _log; |
185 | const char* _phase_name; |
186 | bool _dolog; |
187 | public: |
188 | TracePhase(const char* name, elapsedTimer* accumulator); |
189 | ~TracePhase(); |
190 | }; |
191 | |
192 | // Information per category of alias (memory slice) |
193 | class AliasType { |
194 | private: |
195 | friend class Compile; |
196 | |
197 | int _index; // unique index, used with MergeMemNode |
198 | const TypePtr* _adr_type; // normalized address type |
199 | ciField* _field; // relevant instance field, or null if none |
200 | const Type* _element; // relevant array element type, or null if none |
201 | bool _is_rewritable; // false if the memory is write-once only |
202 | int _general_index; // if this is type is an instance, the general |
203 | // type that this is an instance of |
204 | |
205 | void Init(int i, const TypePtr* at); |
206 | |
207 | public: |
208 | int index() const { return _index; } |
209 | const TypePtr* adr_type() const { return _adr_type; } |
210 | ciField* field() const { return _field; } |
211 | const Type* element() const { return _element; } |
212 | bool is_rewritable() const { return _is_rewritable; } |
213 | bool is_volatile() const { return (_field ? _field->is_volatile() : false); } |
214 | int general_index() const { return (_general_index != 0) ? _general_index : _index; } |
215 | |
216 | void set_rewritable(bool z) { _is_rewritable = z; } |
217 | void set_field(ciField* f) { |
218 | assert(!_field,"" ); |
219 | _field = f; |
220 | if (f->is_final() || f->is_stable()) { |
221 | // In the case of @Stable, multiple writes are possible but may be assumed to be no-ops. |
222 | _is_rewritable = false; |
223 | } |
224 | } |
225 | void set_element(const Type* e) { |
226 | assert(_element == NULL, "" ); |
227 | _element = e; |
228 | } |
229 | |
230 | BasicType basic_type() const; |
231 | |
232 | void print_on(outputStream* st) PRODUCT_RETURN; |
233 | }; |
234 | |
235 | enum { |
236 | logAliasCacheSize = 6, |
237 | AliasCacheSize = (1<<logAliasCacheSize) |
238 | }; |
239 | struct AliasCacheEntry { const TypePtr* _adr_type; int _index; }; // simple duple type |
240 | enum { |
241 | trapHistLength = MethodData::_trap_hist_limit |
242 | }; |
243 | |
244 | // Constant entry of the constant table. |
245 | class Constant { |
246 | private: |
247 | BasicType _type; |
248 | union { |
249 | jvalue _value; |
250 | Metadata* _metadata; |
251 | } _v; |
252 | int _offset; // offset of this constant (in bytes) relative to the constant table base. |
253 | float _freq; |
254 | bool _can_be_reused; // true (default) if the value can be shared with other users. |
255 | |
256 | public: |
257 | Constant() : _type(T_ILLEGAL), _offset(-1), _freq(0.0f), _can_be_reused(true) { _v._value.l = 0; } |
258 | Constant(BasicType type, jvalue value, float freq = 0.0f, bool can_be_reused = true) : |
259 | _type(type), |
260 | _offset(-1), |
261 | _freq(freq), |
262 | _can_be_reused(can_be_reused) |
263 | { |
264 | assert(type != T_METADATA, "wrong constructor" ); |
265 | _v._value = value; |
266 | } |
267 | Constant(Metadata* metadata, bool can_be_reused = true) : |
268 | _type(T_METADATA), |
269 | _offset(-1), |
270 | _freq(0.0f), |
271 | _can_be_reused(can_be_reused) |
272 | { |
273 | _v._metadata = metadata; |
274 | } |
275 | |
276 | bool operator==(const Constant& other); |
277 | |
278 | BasicType type() const { return _type; } |
279 | |
280 | jint get_jint() const { return _v._value.i; } |
281 | jlong get_jlong() const { return _v._value.j; } |
282 | jfloat get_jfloat() const { return _v._value.f; } |
283 | jdouble get_jdouble() const { return _v._value.d; } |
284 | jobject get_jobject() const { return _v._value.l; } |
285 | |
286 | Metadata* get_metadata() const { return _v._metadata; } |
287 | |
288 | int offset() const { return _offset; } |
289 | void set_offset(int offset) { _offset = offset; } |
290 | |
291 | float freq() const { return _freq; } |
292 | void inc_freq(float freq) { _freq += freq; } |
293 | |
294 | bool can_be_reused() const { return _can_be_reused; } |
295 | }; |
296 | |
297 | // Constant table. |
298 | class ConstantTable { |
299 | private: |
300 | GrowableArray<Constant> _constants; // Constants of this table. |
301 | int _size; // Size in bytes the emitted constant table takes (including padding). |
302 | int _table_base_offset; // Offset of the table base that gets added to the constant offsets. |
303 | int _nof_jump_tables; // Number of jump-tables in this constant table. |
304 | |
305 | static int qsort_comparator(Constant* a, Constant* b); |
306 | |
307 | // We use negative frequencies to keep the order of the |
308 | // jump-tables in which they were added. Otherwise we get into |
309 | // trouble with relocation. |
310 | float next_jump_table_freq() { return -1.0f * (++_nof_jump_tables); } |
311 | |
312 | public: |
313 | ConstantTable() : |
314 | _size(-1), |
315 | _table_base_offset(-1), // We can use -1 here since the constant table is always bigger than 2 bytes (-(size / 2), see MachConstantBaseNode::emit). |
316 | _nof_jump_tables(0) |
317 | {} |
318 | |
319 | int size() const { assert(_size != -1, "not calculated yet" ); return _size; } |
320 | |
321 | int calculate_table_base_offset() const; // AD specific |
322 | void set_table_base_offset(int x) { assert(_table_base_offset == -1 || x == _table_base_offset, "can't change" ); _table_base_offset = x; } |
323 | int table_base_offset() const { assert(_table_base_offset != -1, "not set yet" ); return _table_base_offset; } |
324 | |
325 | void emit(CodeBuffer& cb); |
326 | |
327 | // Returns the offset of the last entry (the top) of the constant table. |
328 | int top_offset() const { assert(_constants.top().offset() != -1, "not bound yet" ); return _constants.top().offset(); } |
329 | |
330 | void calculate_offsets_and_size(); |
331 | int find_offset(Constant& con) const; |
332 | |
333 | void add(Constant& con); |
334 | Constant add(MachConstantNode* n, BasicType type, jvalue value); |
335 | Constant add(Metadata* metadata); |
336 | Constant add(MachConstantNode* n, MachOper* oper); |
337 | Constant add(MachConstantNode* n, jint i) { |
338 | jvalue value; value.i = i; |
339 | return add(n, T_INT, value); |
340 | } |
341 | Constant add(MachConstantNode* n, jlong j) { |
342 | jvalue value; value.j = j; |
343 | return add(n, T_LONG, value); |
344 | } |
345 | Constant add(MachConstantNode* n, jfloat f) { |
346 | jvalue value; value.f = f; |
347 | return add(n, T_FLOAT, value); |
348 | } |
349 | Constant add(MachConstantNode* n, jdouble d) { |
350 | jvalue value; value.d = d; |
351 | return add(n, T_DOUBLE, value); |
352 | } |
353 | |
354 | // Jump-table |
355 | Constant add_jump_table(MachConstantNode* n); |
356 | void fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const; |
357 | }; |
358 | |
359 | private: |
360 | // Fixed parameters to this compilation. |
361 | const int _compile_id; |
362 | const bool _save_argument_registers; // save/restore arg regs for trampolines |
363 | const bool _subsume_loads; // Load can be matched as part of a larger op. |
364 | const bool _do_escape_analysis; // Do escape analysis. |
365 | const bool _eliminate_boxing; // Do boxing elimination. |
366 | ciMethod* _method; // The method being compiled. |
367 | int _entry_bci; // entry bci for osr methods. |
368 | const TypeFunc* _tf; // My kind of signature |
369 | InlineTree* _ilt; // Ditto (temporary). |
370 | address _stub_function; // VM entry for stub being compiled, or NULL |
371 | const char* _stub_name; // Name of stub or adapter being compiled, or NULL |
372 | address _stub_entry_point; // Compile code entry for generated stub, or NULL |
373 | |
374 | // Control of this compilation. |
375 | int _max_inline_size; // Max inline size for this compilation |
376 | int _freq_inline_size; // Max hot method inline size for this compilation |
377 | int _fixed_slots; // count of frame slots not allocated by the register |
378 | // allocator i.e. locks, original deopt pc, etc. |
379 | uintx _max_node_limit; // Max unique node count during a single compilation. |
380 | // For deopt |
381 | int _orig_pc_slot; |
382 | int _orig_pc_slot_offset_in_bytes; |
383 | |
384 | int _major_progress; // Count of something big happening |
385 | bool _inlining_progress; // progress doing incremental inlining? |
386 | bool _inlining_incrementally;// Are we doing incremental inlining (post parse) |
387 | bool _do_cleanup; // Cleanup is needed before proceeding with incremental inlining |
388 | bool _has_loops; // True if the method _may_ have some loops |
389 | bool _has_split_ifs; // True if the method _may_ have some split-if |
390 | bool _has_unsafe_access; // True if the method _may_ produce faults in unsafe loads or stores. |
391 | bool _has_stringbuilder; // True StringBuffers or StringBuilders are allocated |
392 | bool _has_boxed_value; // True if a boxed object is allocated |
393 | bool _has_reserved_stack_access; // True if the method or an inlined method is annotated with ReservedStackAccess |
394 | uint _max_vector_size; // Maximum size of generated vectors |
395 | bool _clear_upper_avx; // Clear upper bits of ymm registers using vzeroupper |
396 | uint _trap_hist[trapHistLength]; // Cumulative traps |
397 | bool _trap_can_recompile; // Have we emitted a recompiling trap? |
398 | uint _decompile_count; // Cumulative decompilation counts. |
399 | bool _do_inlining; // True if we intend to do inlining |
400 | bool _do_scheduling; // True if we intend to do scheduling |
401 | bool _do_freq_based_layout; // True if we intend to do frequency based block layout |
402 | bool _do_count_invocations; // True if we generate code to count invocations |
403 | bool _do_method_data_update; // True if we generate code to update MethodData*s |
404 | bool _do_vector_loop; // True if allowed to execute loop in parallel iterations |
405 | bool _use_cmove; // True if CMove should be used without profitability analysis |
406 | bool _age_code; // True if we need to profile code age (decrement the aging counter) |
407 | int _AliasLevel; // Locally-adjusted version of AliasLevel flag. |
408 | bool _print_assembly; // True if we should dump assembly code for this compilation |
409 | bool _print_inlining; // True if we should print inlining for this compilation |
410 | bool _print_intrinsics; // True if we should print intrinsics for this compilation |
411 | #ifndef PRODUCT |
412 | bool _trace_opto_output; |
413 | bool _parsed_irreducible_loop; // True if ciTypeFlow detected irreducible loops during parsing |
414 | #endif |
415 | bool _has_irreducible_loop; // Found irreducible loops |
416 | // JSR 292 |
417 | bool _has_method_handle_invokes; // True if this method has MethodHandle invokes. |
418 | RTMState _rtm_state; // State of Restricted Transactional Memory usage |
419 | int _loop_opts_cnt; // loop opts round |
420 | bool _clinit_barrier_on_entry; // True if clinit barrier is needed on nmethod entry |
421 | |
422 | // Compilation environment. |
423 | Arena _comp_arena; // Arena with lifetime equivalent to Compile |
424 | void* _barrier_set_state; // Potential GC barrier state for Compile |
425 | ciEnv* _env; // CI interface |
426 | DirectiveSet* _directive; // Compiler directive |
427 | CompileLog* _log; // from CompilerThread |
428 | const char* _failure_reason; // for record_failure/failing pattern |
429 | GrowableArray<CallGenerator*>* _intrinsics; // List of intrinsics. |
430 | GrowableArray<Node*>* _macro_nodes; // List of nodes which need to be expanded before matching. |
431 | GrowableArray<Node*>* _predicate_opaqs; // List of Opaque1 nodes for the loop predicates. |
432 | GrowableArray<Node*>* _expensive_nodes; // List of nodes that are expensive to compute and that we'd better not let the GVN freely common |
433 | GrowableArray<Node*>* _range_check_casts; // List of CastII nodes with a range check dependency |
434 | GrowableArray<Node*>* _opaque4_nodes; // List of Opaque4 nodes that have a default value |
435 | ConnectionGraph* _congraph; |
436 | #ifndef PRODUCT |
437 | IdealGraphPrinter* _printer; |
438 | #endif |
439 | |
440 | |
441 | // Node management |
442 | uint _unique; // Counter for unique Node indices |
443 | VectorSet _dead_node_list; // Set of dead nodes |
444 | uint _dead_node_count; // Number of dead nodes; VectorSet::Size() is O(N). |
445 | // So use this to keep count and make the call O(1). |
446 | DEBUG_ONLY( Unique_Node_List* _modified_nodes; ) // List of nodes which inputs were modified |
447 | |
448 | debug_only(static int _debug_idx;) // Monotonic counter (not reset), use -XX:BreakAtNode=<idx> |
449 | Arena _node_arena; // Arena for new-space Nodes |
450 | Arena _old_arena; // Arena for old-space Nodes, lifetime during xform |
451 | RootNode* _root; // Unique root of compilation, or NULL after bail-out. |
452 | Node* _top; // Unique top node. (Reset by various phases.) |
453 | |
454 | Node* _immutable_memory; // Initial memory state |
455 | |
456 | Node* _recent_alloc_obj; |
457 | Node* _recent_alloc_ctl; |
458 | |
459 | // Constant table |
460 | ConstantTable _constant_table; // The constant table for this compile. |
461 | MachConstantBaseNode* _mach_constant_base_node; // Constant table base node singleton. |
462 | |
463 | |
464 | // Blocked array of debugging and profiling information, |
465 | // tracked per node. |
466 | enum { _log2_node_notes_block_size = 8, |
467 | _node_notes_block_size = (1<<_log2_node_notes_block_size) |
468 | }; |
469 | GrowableArray<Node_Notes*>* _node_note_array; |
470 | Node_Notes* _default_node_notes; // default notes for new nodes |
471 | |
472 | // After parsing and every bulk phase we hang onto the Root instruction. |
473 | // The RootNode instruction is where the whole program begins. It produces |
474 | // the initial Control and BOTTOM for everybody else. |
475 | |
476 | // Type management |
477 | Arena _Compile_types; // Arena for all types |
478 | Arena* _type_arena; // Alias for _Compile_types except in Initialize_shared() |
479 | Dict* _type_dict; // Intern table |
480 | CloneMap _clone_map; // used for recording history of cloned nodes |
481 | void* _type_hwm; // Last allocation (see Type::operator new/delete) |
482 | size_t _type_last_size; // Last allocation size (see Type::operator new/delete) |
483 | ciMethod* _last_tf_m; // Cache for |
484 | const TypeFunc* _last_tf; // TypeFunc::make |
485 | AliasType** _alias_types; // List of alias types seen so far. |
486 | int _num_alias_types; // Logical length of _alias_types |
487 | int _max_alias_types; // Physical length of _alias_types |
488 | AliasCacheEntry _alias_cache[AliasCacheSize]; // Gets aliases w/o data structure walking |
489 | |
490 | // Parsing, optimization |
491 | PhaseGVN* _initial_gvn; // Results of parse-time PhaseGVN |
492 | Unique_Node_List* _for_igvn; // Initial work-list for next round of Iterative GVN |
493 | WarmCallInfo* _warm_calls; // Sorted work-list for heat-based inlining. |
494 | |
495 | GrowableArray<CallGenerator*> _late_inlines; // List of CallGenerators to be revisited after |
496 | // main parsing has finished. |
497 | GrowableArray<CallGenerator*> _string_late_inlines; // same but for string operations |
498 | |
499 | GrowableArray<CallGenerator*> _boxing_late_inlines; // same but for boxing operations |
500 | |
501 | int _late_inlines_pos; // Where in the queue should the next late inlining candidate go (emulate depth first inlining) |
502 | uint _number_of_mh_late_inlines; // number of method handle late inlining still pending |
503 | |
504 | |
505 | // Inlining may not happen in parse order which would make |
506 | // PrintInlining output confusing. Keep track of PrintInlining |
507 | // pieces in order. |
508 | class PrintInliningBuffer : public ResourceObj { |
509 | private: |
510 | CallGenerator* _cg; |
511 | stringStream* _ss; |
512 | |
513 | public: |
514 | PrintInliningBuffer() |
515 | : _cg(NULL) { _ss = new stringStream(); } |
516 | |
517 | stringStream* ss() const { return _ss; } |
518 | CallGenerator* cg() const { return _cg; } |
519 | void set_cg(CallGenerator* cg) { _cg = cg; } |
520 | }; |
521 | |
522 | stringStream* _print_inlining_stream; |
523 | GrowableArray<PrintInliningBuffer>* _print_inlining_list; |
524 | int _print_inlining_idx; |
525 | char* _print_inlining_output; |
526 | |
527 | // Only keep nodes in the expensive node list that need to be optimized |
528 | void cleanup_expensive_nodes(PhaseIterGVN &igvn); |
529 | // Use for sorting expensive nodes to bring similar nodes together |
530 | static int cmp_expensive_nodes(Node** n1, Node** n2); |
531 | // Expensive nodes list already sorted? |
532 | bool expensive_nodes_sorted() const; |
533 | // Remove the speculative part of types and clean up the graph |
534 | void remove_speculative_types(PhaseIterGVN &igvn); |
535 | |
536 | void* _replay_inline_data; // Pointer to data loaded from file |
537 | |
538 | void print_inlining_init(); |
539 | void print_inlining_reinit(); |
540 | void print_inlining_commit(); |
541 | void print_inlining_push(); |
542 | PrintInliningBuffer& print_inlining_current(); |
543 | |
544 | void log_late_inline_failure(CallGenerator* cg, const char* msg); |
545 | |
546 | public: |
547 | |
548 | void* barrier_set_state() const { return _barrier_set_state; } |
549 | |
550 | outputStream* print_inlining_stream() const { |
551 | assert(print_inlining() || print_intrinsics(), "PrintInlining off?" ); |
552 | return _print_inlining_stream; |
553 | } |
554 | |
555 | void print_inlining_update(CallGenerator* cg); |
556 | void print_inlining_update_delayed(CallGenerator* cg); |
557 | void print_inlining_move_to(CallGenerator* cg); |
558 | void print_inlining_assert_ready(); |
559 | void print_inlining_reset(); |
560 | |
561 | void print_inlining(ciMethod* method, int inline_level, int bci, const char* msg = NULL) { |
562 | stringStream ss; |
563 | CompileTask::print_inlining_inner(&ss, method, inline_level, bci, msg); |
564 | print_inlining_stream()->print("%s" , ss.as_string()); |
565 | } |
566 | |
567 | #ifndef PRODUCT |
568 | IdealGraphPrinter* printer() { return _printer; } |
569 | #endif |
570 | |
571 | void log_late_inline(CallGenerator* cg); |
572 | void log_inline_id(CallGenerator* cg); |
573 | void log_inline_failure(const char* msg); |
574 | |
575 | void* replay_inline_data() const { return _replay_inline_data; } |
576 | |
577 | // Dump inlining replay data to the stream. |
578 | void dump_inline_data(outputStream* out); |
579 | |
580 | private: |
581 | // Matching, CFG layout, allocation, code generation |
582 | PhaseCFG* _cfg; // Results of CFG finding |
583 | bool _select_24_bit_instr; // We selected an instruction with a 24-bit result |
584 | bool _in_24_bit_fp_mode; // We are emitting instructions with 24-bit results |
585 | int _java_calls; // Number of java calls in the method |
586 | int _inner_loops; // Number of inner loops in the method |
587 | Matcher* _matcher; // Engine to map ideal to machine instructions |
588 | PhaseRegAlloc* _regalloc; // Results of register allocation. |
589 | int _frame_slots; // Size of total frame in stack slots |
590 | CodeOffsets _code_offsets; // Offsets into the code for various interesting entries |
591 | RegMask _FIRST_STACK_mask; // All stack slots usable for spills (depends on frame layout) |
592 | Arena* _indexSet_arena; // control IndexSet allocation within PhaseChaitin |
593 | void* _indexSet_free_block_list; // free list of IndexSet bit blocks |
594 | int _interpreter_frame_size; |
595 | |
596 | uint _node_bundling_limit; |
597 | Bundle* _node_bundling_base; // Information for instruction bundling |
598 | |
599 | // Instruction bits passed off to the VM |
600 | int _method_size; // Size of nmethod code segment in bytes |
601 | CodeBuffer _code_buffer; // Where the code is assembled |
602 | int _first_block_size; // Size of unvalidated entry point code / OSR poison code |
603 | ExceptionHandlerTable _handler_table; // Table of native-code exception handlers |
604 | ImplicitExceptionTable _inc_table; // Table of implicit null checks in native code |
605 | OopMapSet* _oop_map_set; // Table of oop maps (one for each safepoint location) |
606 | static int _CompiledZap_count; // counter compared against CompileZap[First/Last] |
607 | BufferBlob* _scratch_buffer_blob; // For temporary code buffers. |
608 | relocInfo* _scratch_locs_memory; // For temporary code buffers. |
609 | int _scratch_const_size; // For temporary code buffers. |
610 | bool _in_scratch_emit_size; // true when in scratch_emit_size. |
611 | |
612 | void reshape_address(AddPNode* n); |
613 | |
614 | public: |
615 | // Accessors |
616 | |
617 | // The Compile instance currently active in this (compiler) thread. |
618 | static Compile* current() { |
619 | return (Compile*) ciEnv::current()->compiler_data(); |
620 | } |
621 | |
622 | // ID for this compilation. Useful for setting breakpoints in the debugger. |
623 | int compile_id() const { return _compile_id; } |
624 | DirectiveSet* directive() const { return _directive; } |
625 | |
626 | // Does this compilation allow instructions to subsume loads? User |
627 | // instructions that subsume a load may result in an unschedulable |
628 | // instruction sequence. |
629 | bool subsume_loads() const { return _subsume_loads; } |
630 | /** Do escape analysis. */ |
631 | bool do_escape_analysis() const { return _do_escape_analysis; } |
632 | /** Do boxing elimination. */ |
633 | bool eliminate_boxing() const { return _eliminate_boxing; } |
634 | /** Do aggressive boxing elimination. */ |
635 | bool aggressive_unboxing() const { return _eliminate_boxing && AggressiveUnboxing; } |
636 | bool save_argument_registers() const { return _save_argument_registers; } |
637 | |
638 | |
639 | // Other fixed compilation parameters. |
640 | ciMethod* method() const { return _method; } |
641 | int entry_bci() const { return _entry_bci; } |
642 | bool is_osr_compilation() const { return _entry_bci != InvocationEntryBci; } |
643 | bool is_method_compilation() const { return (_method != NULL && !_method->flags().is_native()); } |
644 | const TypeFunc* tf() const { assert(_tf!=NULL, "" ); return _tf; } |
645 | void init_tf(const TypeFunc* tf) { assert(_tf==NULL, "" ); _tf = tf; } |
646 | InlineTree* ilt() const { return _ilt; } |
647 | address stub_function() const { return _stub_function; } |
648 | const char* stub_name() const { return _stub_name; } |
649 | address stub_entry_point() const { return _stub_entry_point; } |
650 | |
651 | // Control of this compilation. |
652 | int fixed_slots() const { assert(_fixed_slots >= 0, "" ); return _fixed_slots; } |
653 | void set_fixed_slots(int n) { _fixed_slots = n; } |
654 | int major_progress() const { return _major_progress; } |
655 | void set_inlining_progress(bool z) { _inlining_progress = z; } |
656 | int inlining_progress() const { return _inlining_progress; } |
657 | void set_inlining_incrementally(bool z) { _inlining_incrementally = z; } |
658 | int inlining_incrementally() const { return _inlining_incrementally; } |
659 | void set_do_cleanup(bool z) { _do_cleanup = z; } |
660 | int do_cleanup() const { return _do_cleanup; } |
661 | void set_major_progress() { _major_progress++; } |
662 | void restore_major_progress(int progress) { _major_progress += progress; } |
663 | void clear_major_progress() { _major_progress = 0; } |
664 | int max_inline_size() const { return _max_inline_size; } |
665 | void set_freq_inline_size(int n) { _freq_inline_size = n; } |
666 | int freq_inline_size() const { return _freq_inline_size; } |
667 | void set_max_inline_size(int n) { _max_inline_size = n; } |
668 | bool has_loops() const { return _has_loops; } |
669 | void set_has_loops(bool z) { _has_loops = z; } |
670 | bool has_split_ifs() const { return _has_split_ifs; } |
671 | void set_has_split_ifs(bool z) { _has_split_ifs = z; } |
672 | bool has_unsafe_access() const { return _has_unsafe_access; } |
673 | void set_has_unsafe_access(bool z) { _has_unsafe_access = z; } |
674 | bool has_stringbuilder() const { return _has_stringbuilder; } |
675 | void set_has_stringbuilder(bool z) { _has_stringbuilder = z; } |
676 | bool has_boxed_value() const { return _has_boxed_value; } |
677 | void set_has_boxed_value(bool z) { _has_boxed_value = z; } |
678 | bool has_reserved_stack_access() const { return _has_reserved_stack_access; } |
679 | void set_has_reserved_stack_access(bool z) { _has_reserved_stack_access = z; } |
680 | uint max_vector_size() const { return _max_vector_size; } |
681 | void set_max_vector_size(uint s) { _max_vector_size = s; } |
682 | bool clear_upper_avx() const { return _clear_upper_avx; } |
683 | void set_clear_upper_avx(bool s) { _clear_upper_avx = s; } |
684 | void set_trap_count(uint r, uint c) { assert(r < trapHistLength, "oob" ); _trap_hist[r] = c; } |
685 | uint trap_count(uint r) const { assert(r < trapHistLength, "oob" ); return _trap_hist[r]; } |
686 | bool trap_can_recompile() const { return _trap_can_recompile; } |
687 | void set_trap_can_recompile(bool z) { _trap_can_recompile = z; } |
688 | uint decompile_count() const { return _decompile_count; } |
689 | void set_decompile_count(uint c) { _decompile_count = c; } |
690 | bool allow_range_check_smearing() const; |
691 | bool do_inlining() const { return _do_inlining; } |
692 | void set_do_inlining(bool z) { _do_inlining = z; } |
693 | bool do_scheduling() const { return _do_scheduling; } |
694 | void set_do_scheduling(bool z) { _do_scheduling = z; } |
695 | bool do_freq_based_layout() const{ return _do_freq_based_layout; } |
696 | void set_do_freq_based_layout(bool z){ _do_freq_based_layout = z; } |
697 | bool do_count_invocations() const{ return _do_count_invocations; } |
698 | void set_do_count_invocations(bool z){ _do_count_invocations = z; } |
699 | bool do_method_data_update() const { return _do_method_data_update; } |
700 | void set_do_method_data_update(bool z) { _do_method_data_update = z; } |
701 | bool do_vector_loop() const { return _do_vector_loop; } |
702 | void set_do_vector_loop(bool z) { _do_vector_loop = z; } |
703 | bool use_cmove() const { return _use_cmove; } |
704 | void set_use_cmove(bool z) { _use_cmove = z; } |
705 | bool age_code() const { return _age_code; } |
706 | void set_age_code(bool z) { _age_code = z; } |
707 | int AliasLevel() const { return _AliasLevel; } |
708 | bool print_assembly() const { return _print_assembly; } |
709 | void set_print_assembly(bool z) { _print_assembly = z; } |
710 | bool print_inlining() const { return _print_inlining; } |
711 | void set_print_inlining(bool z) { _print_inlining = z; } |
712 | bool print_intrinsics() const { return _print_intrinsics; } |
713 | void set_print_intrinsics(bool z) { _print_intrinsics = z; } |
714 | RTMState rtm_state() const { return _rtm_state; } |
715 | void set_rtm_state(RTMState s) { _rtm_state = s; } |
716 | bool use_rtm() const { return (_rtm_state & NoRTM) == 0; } |
717 | bool profile_rtm() const { return _rtm_state == ProfileRTM; } |
718 | uint max_node_limit() const { return (uint)_max_node_limit; } |
719 | void set_max_node_limit(uint n) { _max_node_limit = n; } |
720 | bool clinit_barrier_on_entry() { return _clinit_barrier_on_entry; } |
721 | void set_clinit_barrier_on_entry(bool z) { _clinit_barrier_on_entry = z; } |
722 | |
723 | // check the CompilerOracle for special behaviours for this compile |
724 | bool method_has_option(const char * option) { |
725 | return method() != NULL && method()->has_option(option); |
726 | } |
727 | |
728 | #ifndef PRODUCT |
729 | bool trace_opto_output() const { return _trace_opto_output; } |
730 | bool parsed_irreducible_loop() const { return _parsed_irreducible_loop; } |
731 | void set_parsed_irreducible_loop(bool z) { _parsed_irreducible_loop = z; } |
732 | int _in_dump_cnt; // Required for dumping ir nodes. |
733 | #endif |
734 | bool has_irreducible_loop() const { return _has_irreducible_loop; } |
735 | void set_has_irreducible_loop(bool z) { _has_irreducible_loop = z; } |
736 | |
737 | // JSR 292 |
738 | bool has_method_handle_invokes() const { return _has_method_handle_invokes; } |
739 | void set_has_method_handle_invokes(bool z) { _has_method_handle_invokes = z; } |
740 | |
741 | Ticks _latest_stage_start_counter; |
742 | |
743 | void begin_method() { |
744 | #ifndef PRODUCT |
745 | if (_printer && _printer->should_print(1)) { |
746 | _printer->begin_method(); |
747 | } |
748 | #endif |
749 | C->_latest_stage_start_counter.stamp(); |
750 | } |
751 | |
752 | bool should_print(int level = 1) { |
753 | #ifndef PRODUCT |
754 | return (_printer && _printer->should_print(level)); |
755 | #else |
756 | return false; |
757 | #endif |
758 | } |
759 | |
760 | void print_method(CompilerPhaseType cpt, int level = 1, int idx = 0) { |
761 | EventCompilerPhase event; |
762 | if (event.should_commit()) { |
763 | event.set_starttime(C->_latest_stage_start_counter); |
764 | event.set_phase((u1) cpt); |
765 | event.set_compileId(C->_compile_id); |
766 | event.set_phaseLevel(level); |
767 | event.commit(); |
768 | } |
769 | |
770 | #ifndef PRODUCT |
771 | if (should_print(level)) { |
772 | char output[1024]; |
773 | if (idx != 0) { |
774 | sprintf(output, "%s:%d" , CompilerPhaseTypeHelper::to_string(cpt), idx); |
775 | } else { |
776 | sprintf(output, "%s" , CompilerPhaseTypeHelper::to_string(cpt)); |
777 | } |
778 | _printer->print_method(output, level); |
779 | } |
780 | #endif |
781 | C->_latest_stage_start_counter.stamp(); |
782 | } |
783 | |
784 | void end_method(int level = 1) { |
785 | EventCompilerPhase event; |
786 | if (event.should_commit()) { |
787 | event.set_starttime(C->_latest_stage_start_counter); |
788 | event.set_phase((u1) PHASE_END); |
789 | event.set_compileId(C->_compile_id); |
790 | event.set_phaseLevel(level); |
791 | event.commit(); |
792 | } |
793 | #ifndef PRODUCT |
794 | if (_printer && _printer->should_print(level)) { |
795 | _printer->end_method(); |
796 | } |
797 | #endif |
798 | } |
799 | |
800 | int macro_count() const { return _macro_nodes->length(); } |
801 | int predicate_count() const { return _predicate_opaqs->length();} |
802 | int expensive_count() const { return _expensive_nodes->length(); } |
803 | Node* macro_node(int idx) const { return _macro_nodes->at(idx); } |
804 | Node* predicate_opaque1_node(int idx) const { return _predicate_opaqs->at(idx);} |
805 | Node* expensive_node(int idx) const { return _expensive_nodes->at(idx); } |
806 | ConnectionGraph* congraph() { return _congraph;} |
807 | void set_congraph(ConnectionGraph* congraph) { _congraph = congraph;} |
808 | void add_macro_node(Node * n) { |
809 | //assert(n->is_macro(), "must be a macro node"); |
810 | assert(!_macro_nodes->contains(n), "duplicate entry in expand list" ); |
811 | _macro_nodes->append(n); |
812 | } |
813 | void remove_macro_node(Node * n) { |
814 | // this function may be called twice for a node so check |
815 | // that the node is in the array before attempting to remove it |
816 | if (_macro_nodes->contains(n)) |
817 | _macro_nodes->remove(n); |
818 | // remove from _predicate_opaqs list also if it is there |
819 | if (predicate_count() > 0 && _predicate_opaqs->contains(n)){ |
820 | _predicate_opaqs->remove(n); |
821 | } |
822 | } |
823 | void add_expensive_node(Node * n); |
824 | void remove_expensive_node(Node * n) { |
825 | if (_expensive_nodes->contains(n)) { |
826 | _expensive_nodes->remove(n); |
827 | } |
828 | } |
829 | void add_predicate_opaq(Node * n) { |
830 | assert(!_predicate_opaqs->contains(n), "duplicate entry in predicate opaque1" ); |
831 | assert(_macro_nodes->contains(n), "should have already been in macro list" ); |
832 | _predicate_opaqs->append(n); |
833 | } |
834 | |
835 | // Range check dependent CastII nodes that can be removed after loop optimizations |
836 | void add_range_check_cast(Node* n); |
837 | void remove_range_check_cast(Node* n) { |
838 | if (_range_check_casts->contains(n)) { |
839 | _range_check_casts->remove(n); |
840 | } |
841 | } |
842 | Node* range_check_cast_node(int idx) const { return _range_check_casts->at(idx); } |
843 | int range_check_cast_count() const { return _range_check_casts->length(); } |
844 | // Remove all range check dependent CastIINodes. |
845 | void remove_range_check_casts(PhaseIterGVN &igvn); |
846 | |
847 | void add_opaque4_node(Node* n); |
848 | void remove_opaque4_node(Node* n) { |
849 | if (_opaque4_nodes->contains(n)) { |
850 | _opaque4_nodes->remove(n); |
851 | } |
852 | } |
853 | Node* opaque4_node(int idx) const { return _opaque4_nodes->at(idx); } |
854 | int opaque4_count() const { return _opaque4_nodes->length(); } |
855 | void remove_opaque4_nodes(PhaseIterGVN &igvn); |
856 | |
857 | // remove the opaque nodes that protect the predicates so that the unused checks and |
858 | // uncommon traps will be eliminated from the graph. |
859 | void cleanup_loop_predicates(PhaseIterGVN &igvn); |
860 | bool is_predicate_opaq(Node * n) { |
861 | return _predicate_opaqs->contains(n); |
862 | } |
863 | |
864 | // Are there candidate expensive nodes for optimization? |
865 | bool should_optimize_expensive_nodes(PhaseIterGVN &igvn); |
866 | // Check whether n1 and n2 are similar |
867 | static int cmp_expensive_nodes(Node* n1, Node* n2); |
868 | // Sort expensive nodes to locate similar expensive nodes |
869 | void sort_expensive_nodes(); |
870 | |
871 | // Compilation environment. |
872 | Arena* comp_arena() { return &_comp_arena; } |
873 | ciEnv* env() const { return _env; } |
874 | CompileLog* log() const { return _log; } |
875 | bool failing() const { return _env->failing() || _failure_reason != NULL; } |
876 | const char* failure_reason() const { return (_env->failing()) ? _env->failure_reason() : _failure_reason; } |
877 | |
878 | bool failure_reason_is(const char* r) const { |
879 | return (r == _failure_reason) || (r != NULL && _failure_reason != NULL && strcmp(r, _failure_reason) == 0); |
880 | } |
881 | |
882 | void record_failure(const char* reason); |
883 | void record_method_not_compilable(const char* reason) { |
884 | // Bailouts cover "all_tiers" when TieredCompilation is off. |
885 | env()->record_method_not_compilable(reason, !TieredCompilation); |
886 | // Record failure reason. |
887 | record_failure(reason); |
888 | } |
889 | bool check_node_count(uint margin, const char* reason) { |
890 | if (live_nodes() + margin > max_node_limit()) { |
891 | record_method_not_compilable(reason); |
892 | return true; |
893 | } else { |
894 | return false; |
895 | } |
896 | } |
897 | |
898 | // Node management |
899 | uint unique() const { return _unique; } |
900 | uint next_unique() { return _unique++; } |
901 | void set_unique(uint i) { _unique = i; } |
902 | static int debug_idx() { return debug_only(_debug_idx)+0; } |
903 | static void set_debug_idx(int i) { debug_only(_debug_idx = i); } |
904 | Arena* node_arena() { return &_node_arena; } |
905 | Arena* old_arena() { return &_old_arena; } |
906 | RootNode* root() const { return _root; } |
907 | void set_root(RootNode* r) { _root = r; } |
908 | StartNode* start() const; // (Derived from root.) |
909 | void init_start(StartNode* s); |
910 | Node* immutable_memory(); |
911 | |
912 | Node* recent_alloc_ctl() const { return _recent_alloc_ctl; } |
913 | Node* recent_alloc_obj() const { return _recent_alloc_obj; } |
914 | void set_recent_alloc(Node* ctl, Node* obj) { |
915 | _recent_alloc_ctl = ctl; |
916 | _recent_alloc_obj = obj; |
917 | } |
918 | void record_dead_node(uint idx) { if (_dead_node_list.test_set(idx)) return; |
919 | _dead_node_count++; |
920 | } |
921 | bool is_dead_node(uint idx) { return _dead_node_list.test(idx) != 0; } |
922 | uint dead_node_count() { return _dead_node_count; } |
923 | void reset_dead_node_list() { _dead_node_list.Reset(); |
924 | _dead_node_count = 0; |
925 | } |
926 | uint live_nodes() const { |
927 | int val = _unique - _dead_node_count; |
928 | assert (val >= 0, "number of tracked dead nodes %d more than created nodes %d" , _unique, _dead_node_count); |
929 | return (uint) val; |
930 | } |
931 | #ifdef ASSERT |
932 | uint count_live_nodes_by_graph_walk(); |
933 | void print_missing_nodes(); |
934 | #endif |
935 | |
936 | // Record modified nodes to check that they are put on IGVN worklist |
937 | void record_modified_node(Node* n) NOT_DEBUG_RETURN; |
938 | void remove_modified_node(Node* n) NOT_DEBUG_RETURN; |
939 | DEBUG_ONLY( Unique_Node_List* modified_nodes() const { return _modified_nodes; } ) |
940 | |
941 | // Constant table |
942 | ConstantTable& constant_table() { return _constant_table; } |
943 | |
944 | MachConstantBaseNode* mach_constant_base_node(); |
945 | bool has_mach_constant_base_node() const { return _mach_constant_base_node != NULL; } |
946 | // Generated by adlc, true if CallNode requires MachConstantBase. |
947 | bool needs_clone_jvms(); |
948 | |
949 | // Handy undefined Node |
950 | Node* top() const { return _top; } |
951 | |
952 | // these are used by guys who need to know about creation and transformation of top: |
953 | Node* cached_top_node() { return _top; } |
954 | void set_cached_top_node(Node* tn); |
955 | |
956 | GrowableArray<Node_Notes*>* node_note_array() const { return _node_note_array; } |
957 | void set_node_note_array(GrowableArray<Node_Notes*>* arr) { _node_note_array = arr; } |
958 | Node_Notes* default_node_notes() const { return _default_node_notes; } |
959 | void set_default_node_notes(Node_Notes* n) { _default_node_notes = n; } |
960 | |
961 | Node_Notes* node_notes_at(int idx) { |
962 | return locate_node_notes(_node_note_array, idx, false); |
963 | } |
964 | inline bool set_node_notes_at(int idx, Node_Notes* value); |
965 | |
966 | // Copy notes from source to dest, if they exist. |
967 | // Overwrite dest only if source provides something. |
968 | // Return true if information was moved. |
969 | bool copy_node_notes_to(Node* dest, Node* source); |
970 | |
971 | // Workhorse function to sort out the blocked Node_Notes array: |
972 | inline Node_Notes* locate_node_notes(GrowableArray<Node_Notes*>* arr, |
973 | int idx, bool can_grow = false); |
974 | |
975 | void grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by); |
976 | |
977 | // Type management |
978 | Arena* type_arena() { return _type_arena; } |
979 | Dict* type_dict() { return _type_dict; } |
980 | void* type_hwm() { return _type_hwm; } |
981 | size_t type_last_size() { return _type_last_size; } |
982 | int num_alias_types() { return _num_alias_types; } |
983 | |
984 | void init_type_arena() { _type_arena = &_Compile_types; } |
985 | void set_type_arena(Arena* a) { _type_arena = a; } |
986 | void set_type_dict(Dict* d) { _type_dict = d; } |
987 | void set_type_hwm(void* p) { _type_hwm = p; } |
988 | void set_type_last_size(size_t sz) { _type_last_size = sz; } |
989 | |
990 | const TypeFunc* last_tf(ciMethod* m) { |
991 | return (m == _last_tf_m) ? _last_tf : NULL; |
992 | } |
993 | void set_last_tf(ciMethod* m, const TypeFunc* tf) { |
994 | assert(m != NULL || tf == NULL, "" ); |
995 | _last_tf_m = m; |
996 | _last_tf = tf; |
997 | } |
998 | |
999 | AliasType* alias_type(int idx) { assert(idx < num_alias_types(), "oob" ); return _alias_types[idx]; } |
1000 | AliasType* alias_type(const TypePtr* adr_type, ciField* field = NULL) { return find_alias_type(adr_type, false, field); } |
1001 | bool have_alias_type(const TypePtr* adr_type); |
1002 | AliasType* alias_type(ciField* field); |
1003 | |
1004 | int get_alias_index(const TypePtr* at) { return alias_type(at)->index(); } |
1005 | const TypePtr* get_adr_type(uint aidx) { return alias_type(aidx)->adr_type(); } |
1006 | int get_general_index(uint aidx) { return alias_type(aidx)->general_index(); } |
1007 | |
1008 | // Building nodes |
1009 | void rethrow_exceptions(JVMState* jvms); |
1010 | void return_values(JVMState* jvms); |
1011 | JVMState* build_start_state(StartNode* start, const TypeFunc* tf); |
1012 | |
1013 | // Decide how to build a call. |
1014 | // The profile factor is a discount to apply to this site's interp. profile. |
1015 | CallGenerator* call_generator(ciMethod* call_method, int vtable_index, bool call_does_dispatch, |
1016 | JVMState* jvms, bool allow_inline, float profile_factor, ciKlass* speculative_receiver_type = NULL, |
1017 | bool allow_intrinsics = true, bool delayed_forbidden = false); |
1018 | bool should_delay_inlining(ciMethod* call_method, JVMState* jvms) { |
1019 | return should_delay_string_inlining(call_method, jvms) || |
1020 | should_delay_boxing_inlining(call_method, jvms); |
1021 | } |
1022 | bool should_delay_string_inlining(ciMethod* call_method, JVMState* jvms); |
1023 | bool should_delay_boxing_inlining(ciMethod* call_method, JVMState* jvms); |
1024 | |
1025 | // Helper functions to identify inlining potential at call-site |
1026 | ciMethod* optimize_virtual_call(ciMethod* caller, int bci, ciInstanceKlass* klass, |
1027 | ciKlass* holder, ciMethod* callee, |
1028 | const TypeOopPtr* receiver_type, bool is_virtual, |
1029 | bool &call_does_dispatch, int &vtable_index, |
1030 | bool check_access = true); |
1031 | ciMethod* optimize_inlining(ciMethod* caller, int bci, ciInstanceKlass* klass, |
1032 | ciMethod* callee, const TypeOopPtr* receiver_type, |
1033 | bool check_access = true); |
1034 | |
1035 | // Report if there were too many traps at a current method and bci. |
1036 | // Report if a trap was recorded, and/or PerMethodTrapLimit was exceeded. |
1037 | // If there is no MDO at all, report no trap unless told to assume it. |
1038 | bool too_many_traps(ciMethod* method, int bci, Deoptimization::DeoptReason reason); |
1039 | // This version, unspecific to a particular bci, asks if |
1040 | // PerMethodTrapLimit was exceeded for all inlined methods seen so far. |
1041 | bool too_many_traps(Deoptimization::DeoptReason reason, |
1042 | // Privately used parameter for logging: |
1043 | ciMethodData* logmd = NULL); |
1044 | // Report if there were too many recompiles at a method and bci. |
1045 | bool too_many_recompiles(ciMethod* method, int bci, Deoptimization::DeoptReason reason); |
1046 | // Report if there were too many traps or recompiles at a method and bci. |
1047 | bool too_many_traps_or_recompiles(ciMethod* method, int bci, Deoptimization::DeoptReason reason) { |
1048 | return too_many_traps(method, bci, reason) || |
1049 | too_many_recompiles(method, bci, reason); |
1050 | } |
1051 | // Return a bitset with the reasons where deoptimization is allowed, |
1052 | // i.e., where there were not too many uncommon traps. |
1053 | int _allowed_reasons; |
1054 | int allowed_deopt_reasons() { return _allowed_reasons; } |
1055 | void set_allowed_deopt_reasons(); |
1056 | |
1057 | // Parsing, optimization |
1058 | PhaseGVN* initial_gvn() { return _initial_gvn; } |
1059 | Unique_Node_List* for_igvn() { return _for_igvn; } |
1060 | inline void record_for_igvn(Node* n); // Body is after class Unique_Node_List. |
1061 | void set_initial_gvn(PhaseGVN *gvn) { _initial_gvn = gvn; } |
1062 | void set_for_igvn(Unique_Node_List *for_igvn) { _for_igvn = for_igvn; } |
1063 | |
1064 | // Replace n by nn using initial_gvn, calling hash_delete and |
1065 | // record_for_igvn as needed. |
1066 | void gvn_replace_by(Node* n, Node* nn); |
1067 | |
1068 | |
1069 | void identify_useful_nodes(Unique_Node_List &useful); |
1070 | void update_dead_node_list(Unique_Node_List &useful); |
1071 | void remove_useless_nodes (Unique_Node_List &useful); |
1072 | |
1073 | WarmCallInfo* warm_calls() const { return _warm_calls; } |
1074 | void set_warm_calls(WarmCallInfo* l) { _warm_calls = l; } |
1075 | WarmCallInfo* pop_warm_call(); |
1076 | |
1077 | // Record this CallGenerator for inlining at the end of parsing. |
1078 | void add_late_inline(CallGenerator* cg) { |
1079 | _late_inlines.insert_before(_late_inlines_pos, cg); |
1080 | _late_inlines_pos++; |
1081 | } |
1082 | |
1083 | void prepend_late_inline(CallGenerator* cg) { |
1084 | _late_inlines.insert_before(0, cg); |
1085 | } |
1086 | |
1087 | void add_string_late_inline(CallGenerator* cg) { |
1088 | _string_late_inlines.push(cg); |
1089 | } |
1090 | |
1091 | void add_boxing_late_inline(CallGenerator* cg) { |
1092 | _boxing_late_inlines.push(cg); |
1093 | } |
1094 | |
1095 | void remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful); |
1096 | |
1097 | void process_print_inlining(); |
1098 | void dump_print_inlining(); |
1099 | |
1100 | bool over_inlining_cutoff() const { |
1101 | if (!inlining_incrementally()) { |
1102 | return unique() > (uint)NodeCountInliningCutoff; |
1103 | } else { |
1104 | // Give some room for incremental inlining algorithm to "breathe" |
1105 | // and avoid thrashing when live node count is close to the limit. |
1106 | // Keep in mind that live_nodes() isn't accurate during inlining until |
1107 | // dead node elimination step happens (see Compile::inline_incrementally). |
1108 | return live_nodes() > (uint)LiveNodeCountInliningCutoff * 11 / 10; |
1109 | } |
1110 | } |
1111 | |
1112 | void inc_number_of_mh_late_inlines() { _number_of_mh_late_inlines++; } |
1113 | void dec_number_of_mh_late_inlines() { assert(_number_of_mh_late_inlines > 0, "_number_of_mh_late_inlines < 0 !" ); _number_of_mh_late_inlines--; } |
1114 | bool has_mh_late_inlines() const { return _number_of_mh_late_inlines > 0; } |
1115 | |
1116 | bool inline_incrementally_one(); |
1117 | void inline_incrementally_cleanup(PhaseIterGVN& igvn); |
1118 | void inline_incrementally(PhaseIterGVN& igvn); |
1119 | void inline_string_calls(bool parse_time); |
1120 | void inline_boxing_calls(PhaseIterGVN& igvn); |
1121 | bool optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode); |
1122 | void remove_root_to_sfpts_edges(PhaseIterGVN& igvn); |
1123 | |
1124 | // Matching, CFG layout, allocation, code generation |
1125 | PhaseCFG* cfg() { return _cfg; } |
1126 | bool select_24_bit_instr() const { return _select_24_bit_instr; } |
1127 | bool in_24_bit_fp_mode() const { return _in_24_bit_fp_mode; } |
1128 | bool has_java_calls() const { return _java_calls > 0; } |
1129 | int java_calls() const { return _java_calls; } |
1130 | int inner_loops() const { return _inner_loops; } |
1131 | Matcher* matcher() { return _matcher; } |
1132 | PhaseRegAlloc* regalloc() { return _regalloc; } |
1133 | int frame_slots() const { return _frame_slots; } |
1134 | int frame_size_in_words() const; // frame_slots in units of the polymorphic 'words' |
1135 | int frame_size_in_bytes() const { return _frame_slots << LogBytesPerInt; } |
1136 | RegMask& FIRST_STACK_mask() { return _FIRST_STACK_mask; } |
1137 | Arena* indexSet_arena() { return _indexSet_arena; } |
1138 | void* indexSet_free_block_list() { return _indexSet_free_block_list; } |
1139 | uint node_bundling_limit() { return _node_bundling_limit; } |
1140 | Bundle* node_bundling_base() { return _node_bundling_base; } |
1141 | void set_node_bundling_limit(uint n) { _node_bundling_limit = n; } |
1142 | void set_node_bundling_base(Bundle* b) { _node_bundling_base = b; } |
1143 | bool starts_bundle(const Node *n) const; |
1144 | bool need_stack_bang(int frame_size_in_bytes) const; |
1145 | bool need_register_stack_bang() const; |
1146 | |
1147 | void update_interpreter_frame_size(int size) { |
1148 | if (_interpreter_frame_size < size) { |
1149 | _interpreter_frame_size = size; |
1150 | } |
1151 | } |
1152 | int bang_size_in_bytes() const; |
1153 | |
1154 | void set_matcher(Matcher* m) { _matcher = m; } |
1155 | //void set_regalloc(PhaseRegAlloc* ra) { _regalloc = ra; } |
1156 | void set_indexSet_arena(Arena* a) { _indexSet_arena = a; } |
1157 | void set_indexSet_free_block_list(void* p) { _indexSet_free_block_list = p; } |
1158 | |
1159 | // Remember if this compilation changes hardware mode to 24-bit precision |
1160 | void set_24_bit_selection_and_mode(bool selection, bool mode) { |
1161 | _select_24_bit_instr = selection; |
1162 | _in_24_bit_fp_mode = mode; |
1163 | } |
1164 | |
1165 | void set_java_calls(int z) { _java_calls = z; } |
1166 | void set_inner_loops(int z) { _inner_loops = z; } |
1167 | |
1168 | // Instruction bits passed off to the VM |
1169 | int code_size() { return _method_size; } |
1170 | CodeBuffer* code_buffer() { return &_code_buffer; } |
1171 | int first_block_size() { return _first_block_size; } |
1172 | void set_frame_complete(int off) { if (!in_scratch_emit_size()) { _code_offsets.set_value(CodeOffsets::Frame_Complete, off); } } |
1173 | ExceptionHandlerTable* handler_table() { return &_handler_table; } |
1174 | ImplicitExceptionTable* inc_table() { return &_inc_table; } |
1175 | OopMapSet* oop_map_set() { return _oop_map_set; } |
1176 | DebugInformationRecorder* debug_info() { return env()->debug_info(); } |
1177 | Dependencies* dependencies() { return env()->dependencies(); } |
1178 | static int CompiledZap_count() { return _CompiledZap_count; } |
1179 | BufferBlob* scratch_buffer_blob() { return _scratch_buffer_blob; } |
1180 | void init_scratch_buffer_blob(int const_size); |
1181 | void clear_scratch_buffer_blob(); |
1182 | void set_scratch_buffer_blob(BufferBlob* b) { _scratch_buffer_blob = b; } |
1183 | relocInfo* scratch_locs_memory() { return _scratch_locs_memory; } |
1184 | void set_scratch_locs_memory(relocInfo* b) { _scratch_locs_memory = b; } |
1185 | |
1186 | // emit to scratch blob, report resulting size |
1187 | uint scratch_emit_size(const Node* n); |
1188 | void set_in_scratch_emit_size(bool x) { _in_scratch_emit_size = x; } |
1189 | bool in_scratch_emit_size() const { return _in_scratch_emit_size; } |
1190 | |
1191 | enum ScratchBufferBlob { |
1192 | #if defined(PPC64) |
1193 | MAX_inst_size = 2048, |
1194 | #else |
1195 | MAX_inst_size = 1024, |
1196 | #endif |
1197 | MAX_locs_size = 128, // number of relocInfo elements |
1198 | MAX_const_size = 128, |
1199 | MAX_stubs_size = 128 |
1200 | }; |
1201 | |
1202 | // Major entry point. Given a Scope, compile the associated method. |
1203 | // For normal compilations, entry_bci is InvocationEntryBci. For on stack |
1204 | // replacement, entry_bci indicates the bytecode for which to compile a |
1205 | // continuation. |
1206 | Compile(ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, |
1207 | int entry_bci, bool subsume_loads, bool do_escape_analysis, |
1208 | bool eliminate_boxing, DirectiveSet* directive); |
1209 | |
1210 | // Second major entry point. From the TypeFunc signature, generate code |
1211 | // to pass arguments from the Java calling convention to the C calling |
1212 | // convention. |
1213 | Compile(ciEnv* ci_env, const TypeFunc *(*gen)(), |
1214 | address stub_function, const char *stub_name, |
1215 | int is_fancy_jump, bool pass_tls, |
1216 | bool save_arg_registers, bool return_pc, DirectiveSet* directive); |
1217 | |
1218 | // From the TypeFunc signature, generate code to pass arguments |
1219 | // from Compiled calling convention to Interpreter's calling convention |
1220 | void Generate_Compiled_To_Interpreter_Graph(const TypeFunc *tf, address interpreter_entry); |
1221 | |
1222 | // From the TypeFunc signature, generate code to pass arguments |
1223 | // from Interpreter's calling convention to Compiler's calling convention |
1224 | void Generate_Interpreter_To_Compiled_Graph(const TypeFunc *tf); |
1225 | |
1226 | // Are we compiling a method? |
1227 | bool has_method() { return method() != NULL; } |
1228 | |
1229 | // Maybe print some information about this compile. |
1230 | void print_compile_messages(); |
1231 | |
1232 | // Final graph reshaping, a post-pass after the regular optimizer is done. |
1233 | bool final_graph_reshaping(); |
1234 | |
1235 | // returns true if adr is completely contained in the given alias category |
1236 | bool must_alias(const TypePtr* adr, int alias_idx); |
1237 | |
1238 | // returns true if adr overlaps with the given alias category |
1239 | bool can_alias(const TypePtr* adr, int alias_idx); |
1240 | |
1241 | // Driver for converting compiler's IR into machine code bits |
1242 | void Output(); |
1243 | |
1244 | // Accessors for node bundling info. |
1245 | Bundle* node_bundling(const Node *n); |
1246 | bool valid_bundle_info(const Node *n); |
1247 | |
1248 | // Schedule and Bundle the instructions |
1249 | void ScheduleAndBundle(); |
1250 | |
1251 | // Build OopMaps for each GC point |
1252 | void BuildOopMaps(); |
1253 | |
1254 | // Append debug info for the node "local" at safepoint node "sfpt" to the |
1255 | // "array", May also consult and add to "objs", which describes the |
1256 | // scalar-replaced objects. |
1257 | void FillLocArray( int idx, MachSafePointNode* sfpt, |
1258 | Node *local, GrowableArray<ScopeValue*> *array, |
1259 | GrowableArray<ScopeValue*> *objs ); |
1260 | |
1261 | // If "objs" contains an ObjectValue whose id is "id", returns it, else NULL. |
1262 | static ObjectValue* sv_for_node_id(GrowableArray<ScopeValue*> *objs, int id); |
1263 | // Requres that "objs" does not contains an ObjectValue whose id matches |
1264 | // that of "sv. Appends "sv". |
1265 | static void set_sv_for_object_node(GrowableArray<ScopeValue*> *objs, |
1266 | ObjectValue* sv ); |
1267 | |
1268 | // Process an OopMap Element while emitting nodes |
1269 | void Process_OopMap_Node(MachNode *mach, int code_offset); |
1270 | |
1271 | // Initialize code buffer |
1272 | CodeBuffer* init_buffer(uint* blk_starts); |
1273 | |
1274 | // Write out basic block data to code buffer |
1275 | void fill_buffer(CodeBuffer* cb, uint* blk_starts); |
1276 | |
1277 | // Determine which variable sized branches can be shortened |
1278 | void shorten_branches(uint* blk_starts, int& code_size, int& reloc_size, int& stub_size); |
1279 | |
1280 | // Compute the size of first NumberOfLoopInstrToAlign instructions |
1281 | // at the head of a loop. |
1282 | void compute_loop_first_inst_sizes(); |
1283 | |
1284 | // Compute the information for the exception tables |
1285 | void FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels); |
1286 | |
1287 | // Stack slots that may be unused by the calling convention but must |
1288 | // otherwise be preserved. On Intel this includes the return address. |
1289 | // On PowerPC it includes the 4 words holding the old TOC & LR glue. |
1290 | uint in_preserve_stack_slots(); |
1291 | |
1292 | // "Top of Stack" slots that may be unused by the calling convention but must |
1293 | // otherwise be preserved. |
1294 | // On Intel these are not necessary and the value can be zero. |
1295 | // On Sparc this describes the words reserved for storing a register window |
1296 | // when an interrupt occurs. |
1297 | static uint out_preserve_stack_slots(); |
1298 | |
1299 | // Number of outgoing stack slots killed above the out_preserve_stack_slots |
1300 | // for calls to C. Supports the var-args backing area for register parms. |
1301 | uint varargs_C_out_slots_killed() const; |
1302 | |
1303 | // Number of Stack Slots consumed by a synchronization entry |
1304 | int sync_stack_slots() const; |
1305 | |
1306 | // Compute the name of old_SP. See <arch>.ad for frame layout. |
1307 | OptoReg::Name compute_old_SP(); |
1308 | |
1309 | private: |
1310 | // Phase control: |
1311 | void Init(int aliaslevel); // Prepare for a single compilation |
1312 | int Inline_Warm(); // Find more inlining work. |
1313 | void Finish_Warm(); // Give up on further inlines. |
1314 | void Optimize(); // Given a graph, optimize it |
1315 | void Code_Gen(); // Generate code from a graph |
1316 | |
1317 | // Management of the AliasType table. |
1318 | void grow_alias_types(); |
1319 | AliasCacheEntry* probe_alias_cache(const TypePtr* adr_type); |
1320 | const TypePtr *flatten_alias_type(const TypePtr* adr_type) const; |
1321 | AliasType* find_alias_type(const TypePtr* adr_type, bool no_create, ciField* field); |
1322 | |
1323 | void verify_top(Node*) const PRODUCT_RETURN; |
1324 | |
1325 | // Intrinsic setup. |
1326 | void register_library_intrinsics(); // initializer |
1327 | CallGenerator* make_vm_intrinsic(ciMethod* m, bool is_virtual); // constructor |
1328 | int intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found); // helper |
1329 | CallGenerator* find_intrinsic(ciMethod* m, bool is_virtual); // query fn |
1330 | void register_intrinsic(CallGenerator* cg); // update fn |
1331 | |
1332 | #ifndef PRODUCT |
1333 | static juint _intrinsic_hist_count[vmIntrinsics::ID_LIMIT]; |
1334 | static jubyte _intrinsic_hist_flags[vmIntrinsics::ID_LIMIT]; |
1335 | #endif |
1336 | // Function calls made by the public function final_graph_reshaping. |
1337 | // No need to be made public as they are not called elsewhere. |
1338 | void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc); |
1339 | void final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop); |
1340 | void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ); |
1341 | void eliminate_redundant_card_marks(Node* n); |
1342 | |
1343 | public: |
1344 | |
1345 | // Note: Histogram array size is about 1 Kb. |
1346 | enum { // flag bits: |
1347 | _intrinsic_worked = 1, // succeeded at least once |
1348 | _intrinsic_failed = 2, // tried it but it failed |
1349 | _intrinsic_disabled = 4, // was requested but disabled (e.g., -XX:-InlineUnsafeOps) |
1350 | _intrinsic_virtual = 8, // was seen in the virtual form (rare) |
1351 | _intrinsic_both = 16 // was seen in the non-virtual form (usual) |
1352 | }; |
1353 | // Update histogram. Return boolean if this is a first-time occurrence. |
1354 | static bool gather_intrinsic_statistics(vmIntrinsics::ID id, |
1355 | bool is_virtual, int flags) PRODUCT_RETURN0; |
1356 | static void print_intrinsic_statistics() PRODUCT_RETURN; |
1357 | |
1358 | // Graph verification code |
1359 | // Walk the node list, verifying that there is a one-to-one |
1360 | // correspondence between Use-Def edges and Def-Use edges |
1361 | // The option no_dead_code enables stronger checks that the |
1362 | // graph is strongly connected from root in both directions. |
1363 | void verify_graph_edges(bool no_dead_code = false) PRODUCT_RETURN; |
1364 | |
1365 | // End-of-run dumps. |
1366 | static void print_statistics() PRODUCT_RETURN; |
1367 | |
1368 | // Dump formatted assembly |
1369 | #if defined(SUPPORT_OPTO_ASSEMBLY) |
1370 | void dump_asm_on(outputStream* ost, int* pcs, uint pc_limit); |
1371 | void dump_asm(int* pcs = NULL, uint pc_limit = 0) { dump_asm_on(tty, pcs, pc_limit); } |
1372 | #else |
1373 | void dump_asm_on(outputStream* ost, int* pcs, uint pc_limit) { return; } |
1374 | void dump_asm(int* pcs = NULL, uint pc_limit = 0) { return; } |
1375 | #endif |
1376 | void dump_pc(int *pcs, int pc_limit, Node *n); |
1377 | |
1378 | // Verify ADLC assumptions during startup |
1379 | static void adlc_verification() PRODUCT_RETURN; |
1380 | |
1381 | // Definitions of pd methods |
1382 | static void pd_compiler2_init(); |
1383 | |
1384 | // Static parse-time type checking logic for gen_subtype_check: |
1385 | enum { SSC_always_false, SSC_always_true, SSC_easy_test, SSC_full_test }; |
1386 | int static_subtype_check(ciKlass* superk, ciKlass* subk); |
1387 | |
1388 | static Node* conv_I2X_index(PhaseGVN* phase, Node* offset, const TypeInt* sizetype, |
1389 | // Optional control dependency (for example, on range check) |
1390 | Node* ctrl = NULL); |
1391 | |
1392 | // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) |
1393 | static Node* constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl); |
1394 | |
1395 | // Auxiliary method for randomized fuzzing/stressing |
1396 | static bool randomized_select(int count); |
1397 | |
1398 | // supporting clone_map |
1399 | CloneMap& clone_map(); |
1400 | void set_clone_map(Dict* d); |
1401 | |
1402 | bool needs_clinit_barrier(ciField* ik, ciMethod* accessing_method); |
1403 | bool needs_clinit_barrier(ciMethod* ik, ciMethod* accessing_method); |
1404 | bool needs_clinit_barrier(ciInstanceKlass* ik, ciMethod* accessing_method); |
1405 | }; |
1406 | |
1407 | #endif // SHARE_OPTO_COMPILE_HPP |
1408 | |