0
|
1 /*
|
|
2 * Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
|
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or
|
|
21 * have any questions.
|
|
22 *
|
|
23 */
|
|
24
|
|
25 class Compile;
|
|
26 class Node;
|
|
27 class MachNode;
|
|
28 class MachTypeNode;
|
|
29 class MachOper;
|
|
30
|
|
31 //---------------------------Matcher-------------------------------------------
|
|
32 class Matcher : public PhaseTransform {
|
|
33 friend class VMStructs;
|
|
34 // Private arena of State objects
|
|
35 ResourceArea _states_arena;
|
|
36
|
|
37 VectorSet _visited; // Visit bits
|
|
38
|
|
39 // Used to control the Label pass
|
|
40 VectorSet _shared; // Shared Ideal Node
|
|
41 VectorSet _dontcare; // Nothing the matcher cares about
|
|
42
|
|
43 // Private methods which perform the actual matching and reduction
|
|
44 // Walks the label tree, generating machine nodes
|
|
45 MachNode *ReduceInst( State *s, int rule, Node *&mem);
|
|
46 void ReduceInst_Chain_Rule( State *s, int rule, Node *&mem, MachNode *mach);
|
|
47 uint ReduceInst_Interior(State *s, int rule, Node *&mem, MachNode *mach, uint num_opnds);
|
|
48 void ReduceOper( State *s, int newrule, Node *&mem, MachNode *mach );
|
|
49
|
|
50 // If this node already matched using "rule", return the MachNode for it.
|
|
51 MachNode* find_shared_constant(Node* con, uint rule);
|
|
52
|
|
53 // Convert a dense opcode number to an expanded rule number
|
|
54 const int *_reduceOp;
|
|
55 const int *_leftOp;
|
|
56 const int *_rightOp;
|
|
57
|
|
58 // Map dense opcode number to info on when rule is swallowed constant.
|
|
59 const bool *_swallowed;
|
|
60
|
|
61 // Map dense rule number to determine if this is an instruction chain rule
|
|
62 const uint _begin_inst_chain_rule;
|
|
63 const uint _end_inst_chain_rule;
|
|
64
|
|
65 // We want to clone constants and possible CmpI-variants.
|
|
66 // If we do not clone CmpI, then we can have many instances of
|
|
67 // condition codes alive at once. This is OK on some chips and
|
|
68 // bad on others. Hence the machine-dependent table lookup.
|
|
69 const char *_must_clone;
|
|
70
|
|
71 // Find shared Nodes, or Nodes that otherwise are Matcher roots
|
|
72 void find_shared( Node *n );
|
|
73
|
|
74 // Debug and profile information for nodes in old space:
|
|
75 GrowableArray<Node_Notes*>* _old_node_note_array;
|
|
76
|
|
77 // Node labeling iterator for instruction selection
|
|
78 Node *Label_Root( const Node *n, State *svec, Node *control, const Node *mem );
|
|
79
|
|
80 Node *transform( Node *dummy );
|
|
81
|
|
82 Node_List &_proj_list; // For Machine nodes killing many values
|
|
83
|
|
84 Node_Array _shared_constants;
|
|
85
|
|
86 debug_only(Node_Array _old2new_map;) // Map roots of ideal-trees to machine-roots
|
|
87
|
|
88 // Accessors for the inherited field PhaseTransform::_nodes:
|
|
89 void grow_new_node_array(uint idx_limit) {
|
|
90 _nodes.map(idx_limit-1, NULL);
|
|
91 }
|
|
92 bool has_new_node(const Node* n) const {
|
|
93 return _nodes.at(n->_idx) != NULL;
|
|
94 }
|
|
95 Node* new_node(const Node* n) const {
|
|
96 assert(has_new_node(n), "set before get");
|
|
97 return _nodes.at(n->_idx);
|
|
98 }
|
|
99 void set_new_node(const Node* n, Node *nn) {
|
|
100 assert(!has_new_node(n), "set only once");
|
|
101 _nodes.map(n->_idx, nn);
|
|
102 }
|
|
103
|
|
104 #ifdef ASSERT
|
|
105 // Make sure only new nodes are reachable from this node
|
|
106 void verify_new_nodes_only(Node* root);
|
|
107 #endif
|
|
108
|
|
109 public:
|
|
110 int LabelRootDepth;
|
|
111 static const int base2reg[]; // Map Types to machine register types
|
|
112 // Convert ideal machine register to a register mask for spill-loads
|
|
113 static const RegMask *idealreg2regmask[];
|
|
114 RegMask *idealreg2spillmask[_last_machine_leaf];
|
|
115 RegMask *idealreg2debugmask[_last_machine_leaf];
|
|
116 void init_spill_mask( Node *ret );
|
|
117 // Convert machine register number to register mask
|
|
118 static uint mreg2regmask_max;
|
|
119 static RegMask mreg2regmask[];
|
|
120 static RegMask STACK_ONLY_mask;
|
|
121
|
|
122 bool is_shared( Node *n ) { return _shared.test(n->_idx) != 0; }
|
|
123 void set_shared( Node *n ) { _shared.set(n->_idx); }
|
|
124 bool is_visited( Node *n ) { return _visited.test(n->_idx) != 0; }
|
|
125 void set_visited( Node *n ) { _visited.set(n->_idx); }
|
|
126 bool is_dontcare( Node *n ) { return _dontcare.test(n->_idx) != 0; }
|
|
127 void set_dontcare( Node *n ) { _dontcare.set(n->_idx); }
|
|
128
|
|
129 // Mode bit to tell DFA and expand rules whether we are running after
|
|
130 // (or during) register selection. Usually, the matcher runs before,
|
|
131 // but it will also get called to generate post-allocation spill code.
|
|
132 // In this situation, it is a deadly error to attempt to allocate more
|
|
133 // temporary registers.
|
|
134 bool _allocation_started;
|
|
135
|
|
136 // Machine register names
|
|
137 static const char *regName[];
|
|
138 // Machine register encodings
|
|
139 static const unsigned char _regEncode[];
|
|
140 // Machine Node names
|
|
141 const char **_ruleName;
|
|
142 // Rules that are cheaper to rematerialize than to spill
|
|
143 static const uint _begin_rematerialize;
|
|
144 static const uint _end_rematerialize;
|
|
145
|
|
146 // An array of chars, from 0 to _last_Mach_Reg.
|
|
147 // No Save = 'N' (for register windows)
|
|
148 // Save on Entry = 'E'
|
|
149 // Save on Call = 'C'
|
|
150 // Always Save = 'A' (same as SOE + SOC)
|
|
151 const char *_register_save_policy;
|
|
152 const char *_c_reg_save_policy;
|
|
153 // Convert a machine register to a machine register type, so-as to
|
|
154 // properly match spill code.
|
|
155 const int *_register_save_type;
|
|
156 // Maps from machine register to boolean; true if machine register can
|
|
157 // be holding a call argument in some signature.
|
|
158 static bool can_be_java_arg( int reg );
|
|
159 // Maps from machine register to boolean; true if machine register holds
|
|
160 // a spillable argument.
|
|
161 static bool is_spillable_arg( int reg );
|
|
162
|
|
163 // List of IfFalse or IfTrue Nodes that indicate a taken null test.
|
|
164 // List is valid in the post-matching space.
|
|
165 Node_List _null_check_tests;
|
|
166 void collect_null_checks( Node *proj );
|
|
167 void validate_null_checks( );
|
|
168
|
|
169 Matcher( Node_List &proj_list );
|
|
170
|
|
171 // Select instructions for entire method
|
|
172 void match( );
|
|
173 // Helper for match
|
|
174 OptoReg::Name warp_incoming_stk_arg( VMReg reg );
|
|
175
|
|
176 // Transform, then walk. Does implicit DCE while walking.
|
|
177 // Name changed from "transform" to avoid it being virtual.
|
|
178 Node *xform( Node *old_space_node, int Nodes );
|
|
179
|
|
180 // Match a single Ideal Node - turn it into a 1-Node tree; Label & Reduce.
|
|
181 MachNode *match_tree( const Node *n );
|
|
182 MachNode *match_sfpt( SafePointNode *sfpt );
|
|
183 // Helper for match_sfpt
|
|
184 OptoReg::Name warp_outgoing_stk_arg( VMReg reg, OptoReg::Name begin_out_arg_area, OptoReg::Name &out_arg_limit_per_call );
|
|
185
|
|
186 // Initialize first stack mask and related masks.
|
|
187 void init_first_stack_mask();
|
|
188
|
|
189 // If we should save-on-entry this register
|
|
190 bool is_save_on_entry( int reg );
|
|
191
|
|
192 // Fixup the save-on-entry registers
|
|
193 void Fixup_Save_On_Entry( );
|
|
194
|
|
195 // --- Frame handling ---
|
|
196
|
|
197 // Register number of the stack slot corresponding to the incoming SP.
|
|
198 // Per the Big Picture in the AD file, it is:
|
|
199 // SharedInfo::stack0 + locks + in_preserve_stack_slots + pad2.
|
|
200 OptoReg::Name _old_SP;
|
|
201
|
|
202 // Register number of the stack slot corresponding to the highest incoming
|
|
203 // argument on the stack. Per the Big Picture in the AD file, it is:
|
|
204 // _old_SP + out_preserve_stack_slots + incoming argument size.
|
|
205 OptoReg::Name _in_arg_limit;
|
|
206
|
|
207 // Register number of the stack slot corresponding to the new SP.
|
|
208 // Per the Big Picture in the AD file, it is:
|
|
209 // _in_arg_limit + pad0
|
|
210 OptoReg::Name _new_SP;
|
|
211
|
|
212 // Register number of the stack slot corresponding to the highest outgoing
|
|
213 // argument on the stack. Per the Big Picture in the AD file, it is:
|
|
214 // _new_SP + max outgoing arguments of all calls
|
|
215 OptoReg::Name _out_arg_limit;
|
|
216
|
|
217 OptoRegPair *_parm_regs; // Array of machine registers per argument
|
|
218 RegMask *_calling_convention_mask; // Array of RegMasks per argument
|
|
219
|
|
220 // Does matcher support this ideal node?
|
|
221 static const bool has_match_rule(int opcode);
|
|
222 static const bool _hasMatchRule[_last_opcode];
|
|
223
|
|
224 // Used to determine if we have fast l2f conversion
|
|
225 // USII has it, USIII doesn't
|
|
226 static const bool convL2FSupported(void);
|
|
227
|
|
228 // Vector width in bytes
|
|
229 static const uint vector_width_in_bytes(void);
|
|
230
|
|
231 // Vector ideal reg
|
|
232 static const uint vector_ideal_reg(void);
|
|
233
|
|
234 // Used to determine a "low complexity" 64-bit constant. (Zero is simple.)
|
|
235 // The standard of comparison is one (StoreL ConL) vs. two (StoreI ConI).
|
|
236 // Depends on the details of 64-bit constant generation on the CPU.
|
|
237 static const bool isSimpleConstant64(jlong con);
|
|
238
|
|
239 // These calls are all generated by the ADLC
|
|
240
|
|
241 // TRUE - grows up, FALSE - grows down (Intel)
|
|
242 virtual bool stack_direction() const;
|
|
243
|
|
244 // Java-Java calling convention
|
|
245 // (what you use when Java calls Java)
|
|
246
|
|
247 // Alignment of stack in bytes, standard Intel word alignment is 4.
|
|
248 // Sparc probably wants at least double-word (8).
|
|
249 static uint stack_alignment_in_bytes();
|
|
250 // Alignment of stack, measured in stack slots.
|
|
251 // The size of stack slots is defined by VMRegImpl::stack_slot_size.
|
|
252 static uint stack_alignment_in_slots() {
|
|
253 return stack_alignment_in_bytes() / (VMRegImpl::stack_slot_size);
|
|
254 }
|
|
255
|
|
256 // Array mapping arguments to registers. Argument 0 is usually the 'this'
|
|
257 // pointer. Registers can include stack-slots and regular registers.
|
|
258 static void calling_convention( BasicType *, VMRegPair *, uint len, bool is_outgoing );
|
|
259
|
|
260 // Convert a sig into a calling convention register layout
|
|
261 // and find interesting things about it.
|
|
262 static OptoReg::Name find_receiver( bool is_outgoing );
|
|
263 // Return address register. On Intel it is a stack-slot. On PowerPC
|
|
264 // it is the Link register. On Sparc it is r31?
|
|
265 virtual OptoReg::Name return_addr() const;
|
|
266 RegMask _return_addr_mask;
|
|
267 // Return value register. On Intel it is EAX. On Sparc i0/o0.
|
|
268 static OptoRegPair return_value(int ideal_reg, bool is_outgoing);
|
|
269 static OptoRegPair c_return_value(int ideal_reg, bool is_outgoing);
|
|
270 RegMask _return_value_mask;
|
|
271 // Inline Cache Register
|
|
272 static OptoReg::Name inline_cache_reg();
|
|
273 static const RegMask &inline_cache_reg_mask();
|
|
274 static int inline_cache_reg_encode();
|
|
275
|
|
276 // Register for DIVI projection of divmodI
|
|
277 static RegMask divI_proj_mask();
|
|
278 // Register for MODI projection of divmodI
|
|
279 static RegMask modI_proj_mask();
|
|
280
|
|
281 // Register for DIVL projection of divmodL
|
|
282 static RegMask divL_proj_mask();
|
|
283 // Register for MODL projection of divmodL
|
|
284 static RegMask modL_proj_mask();
|
|
285
|
|
286 // Java-Interpreter calling convention
|
|
287 // (what you use when calling between compiled-Java and Interpreted-Java
|
|
288
|
|
289 // Number of callee-save + always-save registers
|
|
290 // Ignores frame pointer and "special" registers
|
|
291 static int number_of_saved_registers();
|
|
292
|
|
293 // The Method-klass-holder may be passed in the inline_cache_reg
|
|
294 // and then expanded into the inline_cache_reg and a method_oop register
|
|
295
|
|
296 static OptoReg::Name interpreter_method_oop_reg();
|
|
297 static const RegMask &interpreter_method_oop_reg_mask();
|
|
298 static int interpreter_method_oop_reg_encode();
|
|
299
|
|
300 static OptoReg::Name compiler_method_oop_reg();
|
|
301 static const RegMask &compiler_method_oop_reg_mask();
|
|
302 static int compiler_method_oop_reg_encode();
|
|
303
|
|
304 // Interpreter's Frame Pointer Register
|
|
305 static OptoReg::Name interpreter_frame_pointer_reg();
|
|
306 static const RegMask &interpreter_frame_pointer_reg_mask();
|
|
307
|
|
308 // Java-Native calling convention
|
|
309 // (what you use when intercalling between Java and C++ code)
|
|
310
|
|
311 // Array mapping arguments to registers. Argument 0 is usually the 'this'
|
|
312 // pointer. Registers can include stack-slots and regular registers.
|
|
313 static void c_calling_convention( BasicType*, VMRegPair *, uint );
|
|
314 // Frame pointer. The frame pointer is kept at the base of the stack
|
|
315 // and so is probably the stack pointer for most machines. On Intel
|
|
316 // it is ESP. On the PowerPC it is R1. On Sparc it is SP.
|
|
317 OptoReg::Name c_frame_pointer() const;
|
|
318 static RegMask c_frame_ptr_mask;
|
|
319
|
|
320 // !!!!! Special stuff for building ScopeDescs
|
|
321 virtual int regnum_to_fpu_offset(int regnum);
|
|
322
|
|
323 // Is this branch offset small enough to be addressed by a short branch?
|
|
324 bool is_short_branch_offset(int offset);
|
|
325
|
|
326 // Optional scaling for the parameter to the ClearArray/CopyArray node.
|
|
327 static const bool init_array_count_is_in_bytes;
|
|
328
|
|
329 // Threshold small size (in bytes) for a ClearArray/CopyArray node.
|
|
330 // Anything this size or smaller may get converted to discrete scalar stores.
|
|
331 static const int init_array_short_size;
|
|
332
|
|
333 // Should the Matcher clone shifts on addressing modes, expecting them to
|
|
334 // be subsumed into complex addressing expressions or compute them into
|
|
335 // registers? True for Intel but false for most RISCs
|
|
336 static const bool clone_shift_expressions;
|
|
337
|
|
338 // Is it better to copy float constants, or load them directly from memory?
|
|
339 // Intel can load a float constant from a direct address, requiring no
|
|
340 // extra registers. Most RISCs will have to materialize an address into a
|
|
341 // register first, so they may as well materialize the constant immediately.
|
|
342 static const bool rematerialize_float_constants;
|
|
343
|
|
344 // If CPU can load and store mis-aligned doubles directly then no fixup is
|
|
345 // needed. Else we split the double into 2 integer pieces and move it
|
|
346 // piece-by-piece. Only happens when passing doubles into C code or when
|
|
347 // calling i2c adapters as the Java calling convention forces doubles to be
|
|
348 // aligned.
|
|
349 static const bool misaligned_doubles_ok;
|
|
350
|
|
351 // Perform a platform dependent implicit null fixup. This is needed
|
|
352 // on windows95 to take care of some unusual register constraints.
|
|
353 void pd_implicit_null_fixup(MachNode *load, uint idx);
|
|
354
|
|
355 // Advertise here if the CPU requires explicit rounding operations
|
|
356 // to implement the UseStrictFP mode.
|
|
357 static const bool strict_fp_requires_explicit_rounding;
|
|
358
|
|
359 // Do floats take an entire double register or just half?
|
|
360 static const bool float_in_double;
|
|
361 // Do ints take an entire long register or just half?
|
|
362 static const bool int_in_long;
|
|
363
|
|
364 // This routine is run whenever a graph fails to match.
|
|
365 // If it returns, the compiler should bailout to interpreter without error.
|
|
366 // In non-product mode, SoftMatchFailure is false to detect non-canonical
|
|
367 // graphs. Print a message and exit.
|
|
368 static void soft_match_failure() {
|
|
369 if( SoftMatchFailure ) return;
|
|
370 else { fatal("SoftMatchFailure is not allowed except in product"); }
|
|
371 }
|
|
372
|
|
373 // Used by the DFA in dfa_sparc.cpp. Check for a prior FastLock
|
|
374 // acting as an Acquire and thus we don't need an Acquire here. We
|
|
375 // retain the Node to act as a compiler ordering barrier.
|
|
376 static bool prior_fast_lock( const Node *acq );
|
|
377
|
|
378 // Used by the DFA in dfa_sparc.cpp. Check for a following
|
|
379 // FastUnLock acting as a Release and thus we don't need a Release
|
|
380 // here. We retain the Node to act as a compiler ordering barrier.
|
|
381 static bool post_fast_unlock( const Node *rel );
|
|
382
|
|
383 // Check for a following volatile memory barrier without an
|
|
384 // intervening load and thus we don't need a barrier here. We
|
|
385 // retain the Node to act as a compiler ordering barrier.
|
|
386 static bool post_store_load_barrier(const Node* mb);
|
|
387
|
|
388
|
|
389 #ifdef ASSERT
|
|
390 void dump_old2new_map(); // machine-independent to machine-dependent
|
|
391 #endif
|
|
392 };
|