Mercurial > hg > truffle
annotate src/share/vm/c1/c1_LinearScan.cpp @ 973:ad6585fd4087
6830542: Performance: JVM_DefineClass already verified.
Reviewed-by: kamg, phh
author | acorn |
---|---|
date | Fri, 04 Sep 2009 12:53:02 -0400 |
parents | bd02caa94611 |
children | 323bd24c6520 |
rev | line source |
---|---|
0 | 1 /* |
844 | 2 * Copyright 2005-2009 Sun Microsystems, Inc. All Rights Reserved. |
0 | 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 #include "incls/_precompiled.incl" | |
26 #include "incls/_c1_LinearScan.cpp.incl" | |
27 | |
28 | |
29 #ifndef PRODUCT | |
30 | |
31 static LinearScanStatistic _stat_before_alloc; | |
32 static LinearScanStatistic _stat_after_asign; | |
33 static LinearScanStatistic _stat_final; | |
34 | |
35 static LinearScanTimers _total_timer; | |
36 | |
37 // helper macro for short definition of timer | |
38 #define TIME_LINEAR_SCAN(timer_name) TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose); | |
39 | |
40 // helper macro for short definition of trace-output inside code | |
41 #define TRACE_LINEAR_SCAN(level, code) \ | |
42 if (TraceLinearScanLevel >= level) { \ | |
43 code; \ | |
44 } | |
45 | |
46 #else | |
47 | |
48 #define TIME_LINEAR_SCAN(timer_name) | |
49 #define TRACE_LINEAR_SCAN(level, code) | |
50 | |
51 #endif | |
52 | |
53 // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words | |
54 #ifdef _LP64 | |
55 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 0, 1, -1}; | |
56 #else | |
57 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 1, 1, 0, 1, -1}; | |
58 #endif | |
59 | |
60 | |
61 // Implementation of LinearScan | |
62 | |
63 LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map) | |
64 : _compilation(ir->compilation()) | |
65 , _ir(ir) | |
66 , _gen(gen) | |
67 , _frame_map(frame_map) | |
68 , _num_virtual_regs(gen->max_virtual_register_number()) | |
69 , _has_fpu_registers(false) | |
70 , _num_calls(-1) | |
71 , _max_spills(0) | |
72 , _unused_spill_slot(-1) | |
73 , _intervals(0) // initialized later with correct length | |
74 , _new_intervals_from_allocation(new IntervalList()) | |
75 , _sorted_intervals(NULL) | |
76 , _lir_ops(0) // initialized later with correct length | |
77 , _block_of_op(0) // initialized later with correct length | |
78 , _has_info(0) | |
79 , _has_call(0) | |
80 , _scope_value_cache(0) // initialized later with correct length | |
81 , _interval_in_loop(0, 0) // initialized later with correct length | |
82 , _cached_blocks(*ir->linear_scan_order()) | |
304 | 83 #ifdef X86 |
0 | 84 , _fpu_stack_allocator(NULL) |
85 #endif | |
86 { | |
87 // note: to use more than on instance of LinearScan at a time this function call has to | |
88 // be moved somewhere outside of this constructor: | |
89 Interval::initialize(); | |
90 | |
91 assert(this->ir() != NULL, "check if valid"); | |
92 assert(this->compilation() != NULL, "check if valid"); | |
93 assert(this->gen() != NULL, "check if valid"); | |
94 assert(this->frame_map() != NULL, "check if valid"); | |
95 } | |
96 | |
97 | |
98 // ********** functions for converting LIR-Operands to register numbers | |
99 // | |
100 // Emulate a flat register file comprising physical integer registers, | |
101 // physical floating-point registers and virtual registers, in that order. | |
102 // Virtual registers already have appropriate numbers, since V0 is | |
103 // the number of physical registers. | |
104 // Returns -1 for hi word if opr is a single word operand. | |
105 // | |
106 // Note: the inverse operation (calculating an operand for register numbers) | |
107 // is done in calc_operand_for_interval() | |
108 | |
109 int LinearScan::reg_num(LIR_Opr opr) { | |
110 assert(opr->is_register(), "should not call this otherwise"); | |
111 | |
112 if (opr->is_virtual_register()) { | |
113 assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number"); | |
114 return opr->vreg_number(); | |
115 } else if (opr->is_single_cpu()) { | |
116 return opr->cpu_regnr(); | |
117 } else if (opr->is_double_cpu()) { | |
118 return opr->cpu_regnrLo(); | |
304 | 119 #ifdef X86 |
0 | 120 } else if (opr->is_single_xmm()) { |
121 return opr->fpu_regnr() + pd_first_xmm_reg; | |
122 } else if (opr->is_double_xmm()) { | |
123 return opr->fpu_regnrLo() + pd_first_xmm_reg; | |
124 #endif | |
125 } else if (opr->is_single_fpu()) { | |
126 return opr->fpu_regnr() + pd_first_fpu_reg; | |
127 } else if (opr->is_double_fpu()) { | |
128 return opr->fpu_regnrLo() + pd_first_fpu_reg; | |
129 } else { | |
130 ShouldNotReachHere(); | |
304 | 131 return -1; |
0 | 132 } |
133 } | |
134 | |
135 int LinearScan::reg_numHi(LIR_Opr opr) { | |
136 assert(opr->is_register(), "should not call this otherwise"); | |
137 | |
138 if (opr->is_virtual_register()) { | |
139 return -1; | |
140 } else if (opr->is_single_cpu()) { | |
141 return -1; | |
142 } else if (opr->is_double_cpu()) { | |
143 return opr->cpu_regnrHi(); | |
304 | 144 #ifdef X86 |
0 | 145 } else if (opr->is_single_xmm()) { |
146 return -1; | |
147 } else if (opr->is_double_xmm()) { | |
148 return -1; | |
149 #endif | |
150 } else if (opr->is_single_fpu()) { | |
151 return -1; | |
152 } else if (opr->is_double_fpu()) { | |
153 return opr->fpu_regnrHi() + pd_first_fpu_reg; | |
154 } else { | |
155 ShouldNotReachHere(); | |
304 | 156 return -1; |
0 | 157 } |
158 } | |
159 | |
160 | |
161 // ********** functions for classification of intervals | |
162 | |
163 bool LinearScan::is_precolored_interval(const Interval* i) { | |
164 return i->reg_num() < LinearScan::nof_regs; | |
165 } | |
166 | |
167 bool LinearScan::is_virtual_interval(const Interval* i) { | |
168 return i->reg_num() >= LIR_OprDesc::vreg_base; | |
169 } | |
170 | |
171 bool LinearScan::is_precolored_cpu_interval(const Interval* i) { | |
172 return i->reg_num() < LinearScan::nof_cpu_regs; | |
173 } | |
174 | |
175 bool LinearScan::is_virtual_cpu_interval(const Interval* i) { | |
176 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE); | |
177 } | |
178 | |
179 bool LinearScan::is_precolored_fpu_interval(const Interval* i) { | |
180 return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs; | |
181 } | |
182 | |
183 bool LinearScan::is_virtual_fpu_interval(const Interval* i) { | |
184 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE); | |
185 } | |
186 | |
187 bool LinearScan::is_in_fpu_register(const Interval* i) { | |
188 // fixed intervals not needed for FPU stack allocation | |
189 return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg; | |
190 } | |
191 | |
192 bool LinearScan::is_oop_interval(const Interval* i) { | |
193 // fixed intervals never contain oops | |
194 return i->reg_num() >= nof_regs && i->type() == T_OBJECT; | |
195 } | |
196 | |
197 | |
198 // ********** General helper functions | |
199 | |
200 // compute next unused stack index that can be used for spilling | |
201 int LinearScan::allocate_spill_slot(bool double_word) { | |
202 int spill_slot; | |
203 if (double_word) { | |
204 if ((_max_spills & 1) == 1) { | |
205 // alignment of double-word values | |
206 // the hole because of the alignment is filled with the next single-word value | |
207 assert(_unused_spill_slot == -1, "wasting a spill slot"); | |
208 _unused_spill_slot = _max_spills; | |
209 _max_spills++; | |
210 } | |
211 spill_slot = _max_spills; | |
212 _max_spills += 2; | |
213 | |
214 } else if (_unused_spill_slot != -1) { | |
215 // re-use hole that was the result of a previous double-word alignment | |
216 spill_slot = _unused_spill_slot; | |
217 _unused_spill_slot = -1; | |
218 | |
219 } else { | |
220 spill_slot = _max_spills; | |
221 _max_spills++; | |
222 } | |
223 | |
224 int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount(); | |
225 | |
226 // the class OopMapValue uses only 11 bits for storing the name of the | |
227 // oop location. So a stack slot bigger than 2^11 leads to an overflow | |
228 // that is not reported in product builds. Prevent this by checking the | |
229 // spill slot here (altough this value and the later used location name | |
230 // are slightly different) | |
231 if (result > 2000) { | |
232 bailout("too many stack slots used"); | |
233 } | |
234 | |
235 return result; | |
236 } | |
237 | |
238 void LinearScan::assign_spill_slot(Interval* it) { | |
239 // assign the canonical spill slot of the parent (if a part of the interval | |
240 // is already spilled) or allocate a new spill slot | |
241 if (it->canonical_spill_slot() >= 0) { | |
242 it->assign_reg(it->canonical_spill_slot()); | |
243 } else { | |
244 int spill = allocate_spill_slot(type2spill_size[it->type()] == 2); | |
245 it->set_canonical_spill_slot(spill); | |
246 it->assign_reg(spill); | |
247 } | |
248 } | |
249 | |
250 void LinearScan::propagate_spill_slots() { | |
251 if (!frame_map()->finalize_frame(max_spills())) { | |
252 bailout("frame too large"); | |
253 } | |
254 } | |
255 | |
256 // create a new interval with a predefined reg_num | |
257 // (only used for parent intervals that are created during the building phase) | |
258 Interval* LinearScan::create_interval(int reg_num) { | |
259 assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval"); | |
260 | |
261 Interval* interval = new Interval(reg_num); | |
262 _intervals.at_put(reg_num, interval); | |
263 | |
264 // assign register number for precolored intervals | |
265 if (reg_num < LIR_OprDesc::vreg_base) { | |
266 interval->assign_reg(reg_num); | |
267 } | |
268 return interval; | |
269 } | |
270 | |
271 // assign a new reg_num to the interval and append it to the list of intervals | |
272 // (only used for child intervals that are created during register allocation) | |
273 void LinearScan::append_interval(Interval* it) { | |
274 it->set_reg_num(_intervals.length()); | |
275 _intervals.append(it); | |
276 _new_intervals_from_allocation->append(it); | |
277 } | |
278 | |
279 // copy the vreg-flags if an interval is split | |
280 void LinearScan::copy_register_flags(Interval* from, Interval* to) { | |
281 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) { | |
282 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg); | |
283 } | |
284 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) { | |
285 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved); | |
286 } | |
287 | |
288 // Note: do not copy the must_start_in_memory flag because it is not necessary for child | |
289 // intervals (only the very beginning of the interval must be in memory) | |
290 } | |
291 | |
292 | |
293 // ********** spill move optimization | |
294 // eliminate moves from register to stack if stack slot is known to be correct | |
295 | |
296 // called during building of intervals | |
297 void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) { | |
298 assert(interval->is_split_parent(), "can only be called for split parents"); | |
299 | |
300 switch (interval->spill_state()) { | |
301 case noDefinitionFound: | |
302 assert(interval->spill_definition_pos() == -1, "must no be set before"); | |
303 interval->set_spill_definition_pos(def_pos); | |
304 interval->set_spill_state(oneDefinitionFound); | |
305 break; | |
306 | |
307 case oneDefinitionFound: | |
308 assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created"); | |
309 if (def_pos < interval->spill_definition_pos() - 2) { | |
310 // second definition found, so no spill optimization possible for this interval | |
311 interval->set_spill_state(noOptimization); | |
312 } else { | |
313 // two consecutive definitions (because of two-operand LIR form) | |
314 assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal"); | |
315 } | |
316 break; | |
317 | |
318 case noOptimization: | |
319 // nothing to do | |
320 break; | |
321 | |
322 default: | |
323 assert(false, "other states not allowed at this time"); | |
324 } | |
325 } | |
326 | |
327 // called during register allocation | |
328 void LinearScan::change_spill_state(Interval* interval, int spill_pos) { | |
329 switch (interval->spill_state()) { | |
330 case oneDefinitionFound: { | |
331 int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth(); | |
332 int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth(); | |
333 | |
334 if (def_loop_depth < spill_loop_depth) { | |
335 // the loop depth of the spilling position is higher then the loop depth | |
336 // at the definition of the interval -> move write to memory out of loop | |
337 // by storing at definitin of the interval | |
338 interval->set_spill_state(storeAtDefinition); | |
339 } else { | |
340 // the interval is currently spilled only once, so for now there is no | |
341 // reason to store the interval at the definition | |
342 interval->set_spill_state(oneMoveInserted); | |
343 } | |
344 break; | |
345 } | |
346 | |
347 case oneMoveInserted: { | |
348 // the interval is spilled more then once, so it is better to store it to | |
349 // memory at the definition | |
350 interval->set_spill_state(storeAtDefinition); | |
351 break; | |
352 } | |
353 | |
354 case storeAtDefinition: | |
355 case startInMemory: | |
356 case noOptimization: | |
357 case noDefinitionFound: | |
358 // nothing to do | |
359 break; | |
360 | |
361 default: | |
362 assert(false, "other states not allowed at this time"); | |
363 } | |
364 } | |
365 | |
366 | |
367 bool LinearScan::must_store_at_definition(const Interval* i) { | |
368 return i->is_split_parent() && i->spill_state() == storeAtDefinition; | |
369 } | |
370 | |
371 // called once before asignment of register numbers | |
372 void LinearScan::eliminate_spill_moves() { | |
373 TIME_LINEAR_SCAN(timer_eliminate_spill_moves); | |
374 TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves")); | |
375 | |
376 // collect all intervals that must be stored after their definion. | |
377 // the list is sorted by Interval::spill_definition_pos | |
378 Interval* interval; | |
379 Interval* temp_list; | |
380 create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL); | |
381 | |
382 #ifdef ASSERT | |
383 Interval* prev = NULL; | |
384 Interval* temp = interval; | |
385 while (temp != Interval::end()) { | |
386 assert(temp->spill_definition_pos() > 0, "invalid spill definition pos"); | |
387 if (prev != NULL) { | |
388 assert(temp->from() >= prev->from(), "intervals not sorted"); | |
389 assert(temp->spill_definition_pos() >= prev->spill_definition_pos(), "when intervals are sorted by from, then they must also be sorted by spill_definition_pos"); | |
390 } | |
391 | |
392 assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned"); | |
393 assert(temp->spill_definition_pos() >= temp->from(), "invalid order"); | |
394 assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized"); | |
395 | |
396 TRACE_LINEAR_SCAN(4, tty->print_cr("interval %d (from %d to %d) must be stored at %d", temp->reg_num(), temp->from(), temp->to(), temp->spill_definition_pos())); | |
397 | |
398 temp = temp->next(); | |
399 } | |
400 #endif | |
401 | |
402 LIR_InsertionBuffer insertion_buffer; | |
403 int num_blocks = block_count(); | |
404 for (int i = 0; i < num_blocks; i++) { | |
405 BlockBegin* block = block_at(i); | |
406 LIR_OpList* instructions = block->lir()->instructions_list(); | |
407 int num_inst = instructions->length(); | |
408 bool has_new = false; | |
409 | |
410 // iterate all instructions of the block. skip the first because it is always a label | |
411 for (int j = 1; j < num_inst; j++) { | |
412 LIR_Op* op = instructions->at(j); | |
413 int op_id = op->id(); | |
414 | |
415 if (op_id == -1) { | |
416 // remove move from register to stack if the stack slot is guaranteed to be correct. | |
417 // only moves that have been inserted by LinearScan can be removed. | |
418 assert(op->code() == lir_move, "only moves can have a op_id of -1"); | |
419 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); | |
420 assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers"); | |
421 | |
422 LIR_Op1* op1 = (LIR_Op1*)op; | |
423 Interval* interval = interval_at(op1->result_opr()->vreg_number()); | |
424 | |
425 if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) { | |
426 // move target is a stack slot that is always correct, so eliminate instruction | |
427 TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number())); | |
428 instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num | |
429 } | |
430 | |
431 } else { | |
432 // insert move from register to stack just after the beginning of the interval | |
433 assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order"); | |
434 assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval"); | |
435 | |
436 while (interval != Interval::end() && interval->spill_definition_pos() == op_id) { | |
437 if (!has_new) { | |
438 // prepare insertion buffer (appended when all instructions of the block are processed) | |
439 insertion_buffer.init(block->lir()); | |
440 has_new = true; | |
441 } | |
442 | |
443 LIR_Opr from_opr = operand_for_interval(interval); | |
444 LIR_Opr to_opr = canonical_spill_opr(interval); | |
445 assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register"); | |
446 assert(to_opr->is_stack(), "to operand must be a stack slot"); | |
447 | |
448 insertion_buffer.move(j, from_opr, to_opr); | |
449 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting move after definition of interval %d to stack slot %d at op_id %d", interval->reg_num(), interval->canonical_spill_slot() - LinearScan::nof_regs, op_id)); | |
450 | |
451 interval = interval->next(); | |
452 } | |
453 } | |
454 } // end of instruction iteration | |
455 | |
456 if (has_new) { | |
457 block->lir()->append(&insertion_buffer); | |
458 } | |
459 } // end of block iteration | |
460 | |
461 assert(interval == Interval::end(), "missed an interval"); | |
462 } | |
463 | |
464 | |
465 // ********** Phase 1: number all instructions in all blocks | |
466 // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan. | |
467 | |
468 void LinearScan::number_instructions() { | |
469 { | |
470 // dummy-timer to measure the cost of the timer itself | |
471 // (this time is then subtracted from all other timers to get the real value) | |
472 TIME_LINEAR_SCAN(timer_do_nothing); | |
473 } | |
474 TIME_LINEAR_SCAN(timer_number_instructions); | |
475 | |
476 // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node. | |
477 int num_blocks = block_count(); | |
478 int num_instructions = 0; | |
479 int i; | |
480 for (i = 0; i < num_blocks; i++) { | |
481 num_instructions += block_at(i)->lir()->instructions_list()->length(); | |
482 } | |
483 | |
484 // initialize with correct length | |
485 _lir_ops = LIR_OpArray(num_instructions); | |
486 _block_of_op = BlockBeginArray(num_instructions); | |
487 | |
488 int op_id = 0; | |
489 int idx = 0; | |
490 | |
491 for (i = 0; i < num_blocks; i++) { | |
492 BlockBegin* block = block_at(i); | |
493 block->set_first_lir_instruction_id(op_id); | |
494 LIR_OpList* instructions = block->lir()->instructions_list(); | |
495 | |
496 int num_inst = instructions->length(); | |
497 for (int j = 0; j < num_inst; j++) { | |
498 LIR_Op* op = instructions->at(j); | |
499 op->set_id(op_id); | |
500 | |
501 _lir_ops.at_put(idx, op); | |
502 _block_of_op.at_put(idx, block); | |
503 assert(lir_op_with_id(op_id) == op, "must match"); | |
504 | |
505 idx++; | |
506 op_id += 2; // numbering of lir_ops by two | |
507 } | |
508 block->set_last_lir_instruction_id(op_id - 2); | |
509 } | |
510 assert(idx == num_instructions, "must match"); | |
511 assert(idx * 2 == op_id, "must match"); | |
512 | |
513 _has_call = BitMap(num_instructions); _has_call.clear(); | |
514 _has_info = BitMap(num_instructions); _has_info.clear(); | |
515 } | |
516 | |
517 | |
518 // ********** Phase 2: compute local live sets separately for each block | |
519 // (sets live_gen and live_kill for each block) | |
520 | |
521 void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) { | |
522 LIR_Opr opr = value->operand(); | |
523 Constant* con = value->as_Constant(); | |
524 | |
525 // check some asumptions about debug information | |
526 assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type"); | |
527 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands"); | |
528 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); | |
529 | |
530 if ((con == NULL || con->is_pinned()) && opr->is_register()) { | |
531 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
532 int reg = opr->vreg_number(); | |
533 if (!live_kill.at(reg)) { | |
534 live_gen.set_bit(reg); | |
535 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for value %c%d, LIR op_id %d, register number %d", value->type()->tchar(), value->id(), op->id(), reg)); | |
536 } | |
537 } | |
538 } | |
539 | |
540 | |
541 void LinearScan::compute_local_live_sets() { | |
542 TIME_LINEAR_SCAN(timer_compute_local_live_sets); | |
543 | |
544 int num_blocks = block_count(); | |
545 int live_size = live_set_size(); | |
546 bool local_has_fpu_registers = false; | |
547 int local_num_calls = 0; | |
548 LIR_OpVisitState visitor; | |
549 | |
550 BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops()); | |
551 local_interval_in_loop.clear(); | |
552 | |
553 // iterate all blocks | |
554 for (int i = 0; i < num_blocks; i++) { | |
555 BlockBegin* block = block_at(i); | |
556 | |
557 BitMap live_gen(live_size); live_gen.clear(); | |
558 BitMap live_kill(live_size); live_kill.clear(); | |
559 | |
560 if (block->is_set(BlockBegin::exception_entry_flag)) { | |
561 // Phi functions at the begin of an exception handler are | |
562 // implicitly defined (= killed) at the beginning of the block. | |
563 for_each_phi_fun(block, phi, | |
564 live_kill.set_bit(phi->operand()->vreg_number()) | |
565 ); | |
566 } | |
567 | |
568 LIR_OpList* instructions = block->lir()->instructions_list(); | |
569 int num_inst = instructions->length(); | |
570 | |
571 // iterate all instructions of the block. skip the first because it is always a label | |
572 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); | |
573 for (int j = 1; j < num_inst; j++) { | |
574 LIR_Op* op = instructions->at(j); | |
575 | |
576 // visit operation to collect all operands | |
577 visitor.visit(op); | |
578 | |
579 if (visitor.has_call()) { | |
580 _has_call.set_bit(op->id() >> 1); | |
581 local_num_calls++; | |
582 } | |
583 if (visitor.info_count() > 0) { | |
584 _has_info.set_bit(op->id() >> 1); | |
585 } | |
586 | |
587 // iterate input operands of instruction | |
588 int k, n, reg; | |
589 n = visitor.opr_count(LIR_OpVisitState::inputMode); | |
590 for (k = 0; k < n; k++) { | |
591 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); | |
592 assert(opr->is_register(), "visitor should only return register operands"); | |
593 | |
594 if (opr->is_virtual_register()) { | |
595 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
596 reg = opr->vreg_number(); | |
597 if (!live_kill.at(reg)) { | |
598 live_gen.set_bit(reg); | |
599 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d", reg, op->id())); | |
600 } | |
601 if (block->loop_index() >= 0) { | |
602 local_interval_in_loop.set_bit(reg, block->loop_index()); | |
603 } | |
604 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); | |
605 } | |
606 | |
607 #ifdef ASSERT | |
608 // fixed intervals are never live at block boundaries, so | |
609 // they need not be processed in live sets. | |
610 // this is checked by these assertions to be sure about it. | |
611 // the entry block may have incoming values in registers, which is ok. | |
612 if (!opr->is_virtual_register() && block != ir()->start()) { | |
613 reg = reg_num(opr); | |
614 if (is_processed_reg_num(reg)) { | |
615 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); | |
616 } | |
617 reg = reg_numHi(opr); | |
618 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { | |
619 assert(live_kill.at(reg), "using fixed register that is not defined in this block"); | |
620 } | |
621 } | |
622 #endif | |
623 } | |
624 | |
625 // Add uses of live locals from interpreter's point of view for proper debug information generation | |
626 n = visitor.info_count(); | |
627 for (k = 0; k < n; k++) { | |
628 CodeEmitInfo* info = visitor.info_at(k); | |
629 ValueStack* stack = info->stack(); | |
630 for_each_state_value(stack, value, | |
631 set_live_gen_kill(value, op, live_gen, live_kill) | |
632 ); | |
633 } | |
634 | |
635 // iterate temp operands of instruction | |
636 n = visitor.opr_count(LIR_OpVisitState::tempMode); | |
637 for (k = 0; k < n; k++) { | |
638 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); | |
639 assert(opr->is_register(), "visitor should only return register operands"); | |
640 | |
641 if (opr->is_virtual_register()) { | |
642 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
643 reg = opr->vreg_number(); | |
644 live_kill.set_bit(reg); | |
645 if (block->loop_index() >= 0) { | |
646 local_interval_in_loop.set_bit(reg, block->loop_index()); | |
647 } | |
648 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); | |
649 } | |
650 | |
651 #ifdef ASSERT | |
652 // fixed intervals are never live at block boundaries, so | |
653 // they need not be processed in live sets | |
654 // process them only in debug mode so that this can be checked | |
655 if (!opr->is_virtual_register()) { | |
656 reg = reg_num(opr); | |
657 if (is_processed_reg_num(reg)) { | |
658 live_kill.set_bit(reg_num(opr)); | |
659 } | |
660 reg = reg_numHi(opr); | |
661 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { | |
662 live_kill.set_bit(reg); | |
663 } | |
664 } | |
665 #endif | |
666 } | |
667 | |
668 // iterate output operands of instruction | |
669 n = visitor.opr_count(LIR_OpVisitState::outputMode); | |
670 for (k = 0; k < n; k++) { | |
671 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); | |
672 assert(opr->is_register(), "visitor should only return register operands"); | |
673 | |
674 if (opr->is_virtual_register()) { | |
675 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
676 reg = opr->vreg_number(); | |
677 live_kill.set_bit(reg); | |
678 if (block->loop_index() >= 0) { | |
679 local_interval_in_loop.set_bit(reg, block->loop_index()); | |
680 } | |
681 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); | |
682 } | |
683 | |
684 #ifdef ASSERT | |
685 // fixed intervals are never live at block boundaries, so | |
686 // they need not be processed in live sets | |
687 // process them only in debug mode so that this can be checked | |
688 if (!opr->is_virtual_register()) { | |
689 reg = reg_num(opr); | |
690 if (is_processed_reg_num(reg)) { | |
691 live_kill.set_bit(reg_num(opr)); | |
692 } | |
693 reg = reg_numHi(opr); | |
694 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { | |
695 live_kill.set_bit(reg); | |
696 } | |
697 } | |
698 #endif | |
699 } | |
700 } // end of instruction iteration | |
701 | |
702 block->set_live_gen (live_gen); | |
703 block->set_live_kill(live_kill); | |
704 block->set_live_in (BitMap(live_size)); block->live_in().clear(); | |
705 block->set_live_out (BitMap(live_size)); block->live_out().clear(); | |
706 | |
707 TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d ", block->block_id()); print_bitmap(block->live_gen())); | |
708 TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill())); | |
709 } // end of block iteration | |
710 | |
711 // propagate local calculated information into LinearScan object | |
712 _has_fpu_registers = local_has_fpu_registers; | |
713 compilation()->set_has_fpu_code(local_has_fpu_registers); | |
714 | |
715 _num_calls = local_num_calls; | |
716 _interval_in_loop = local_interval_in_loop; | |
717 } | |
718 | |
719 | |
720 // ********** Phase 3: perform a backward dataflow analysis to compute global live sets | |
721 // (sets live_in and live_out for each block) | |
722 | |
723 void LinearScan::compute_global_live_sets() { | |
724 TIME_LINEAR_SCAN(timer_compute_global_live_sets); | |
725 | |
726 int num_blocks = block_count(); | |
727 bool change_occurred; | |
728 bool change_occurred_in_block; | |
729 int iteration_count = 0; | |
730 BitMap live_out(live_set_size()); live_out.clear(); // scratch set for calculations | |
731 | |
732 // Perform a backward dataflow analysis to compute live_out and live_in for each block. | |
733 // The loop is executed until a fixpoint is reached (no changes in an iteration) | |
734 // Exception handlers must be processed because not all live values are | |
735 // present in the state array, e.g. because of global value numbering | |
736 do { | |
737 change_occurred = false; | |
738 | |
739 // iterate all blocks in reverse order | |
740 for (int i = num_blocks - 1; i >= 0; i--) { | |
741 BlockBegin* block = block_at(i); | |
742 | |
743 change_occurred_in_block = false; | |
744 | |
745 // live_out(block) is the union of live_in(sux), for successors sux of block | |
746 int n = block->number_of_sux(); | |
747 int e = block->number_of_exception_handlers(); | |
748 if (n + e > 0) { | |
749 // block has successors | |
750 if (n > 0) { | |
751 live_out.set_from(block->sux_at(0)->live_in()); | |
752 for (int j = 1; j < n; j++) { | |
753 live_out.set_union(block->sux_at(j)->live_in()); | |
754 } | |
755 } else { | |
756 live_out.clear(); | |
757 } | |
758 for (int j = 0; j < e; j++) { | |
759 live_out.set_union(block->exception_handler_at(j)->live_in()); | |
760 } | |
761 | |
762 if (!block->live_out().is_same(live_out)) { | |
763 // A change occurred. Swap the old and new live out sets to avoid copying. | |
764 BitMap temp = block->live_out(); | |
765 block->set_live_out(live_out); | |
766 live_out = temp; | |
767 | |
768 change_occurred = true; | |
769 change_occurred_in_block = true; | |
770 } | |
771 } | |
772 | |
773 if (iteration_count == 0 || change_occurred_in_block) { | |
774 // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block)) | |
775 // note: live_in has to be computed only in first iteration or if live_out has changed! | |
776 BitMap live_in = block->live_in(); | |
777 live_in.set_from(block->live_out()); | |
778 live_in.set_difference(block->live_kill()); | |
779 live_in.set_union(block->live_gen()); | |
780 } | |
781 | |
782 #ifndef PRODUCT | |
783 if (TraceLinearScanLevel >= 4) { | |
784 char c = ' '; | |
785 if (iteration_count == 0 || change_occurred_in_block) { | |
786 c = '*'; | |
787 } | |
788 tty->print("(%d) live_in%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in()); | |
789 tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out()); | |
790 } | |
791 #endif | |
792 } | |
793 iteration_count++; | |
794 | |
795 if (change_occurred && iteration_count > 50) { | |
796 BAILOUT("too many iterations in compute_global_live_sets"); | |
797 } | |
798 } while (change_occurred); | |
799 | |
800 | |
801 #ifdef ASSERT | |
802 // check that fixed intervals are not live at block boundaries | |
803 // (live set must be empty at fixed intervals) | |
804 for (int i = 0; i < num_blocks; i++) { | |
805 BlockBegin* block = block_at(i); | |
806 for (int j = 0; j < LIR_OprDesc::vreg_base; j++) { | |
807 assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty"); | |
808 assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty"); | |
809 assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty"); | |
810 } | |
811 } | |
812 #endif | |
813 | |
814 // check that the live_in set of the first block is empty | |
815 BitMap live_in_args(ir()->start()->live_in().size()); | |
816 live_in_args.clear(); | |
817 if (!ir()->start()->live_in().is_same(live_in_args)) { | |
818 #ifdef ASSERT | |
819 tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)"); | |
820 tty->print_cr("affected registers:"); | |
821 print_bitmap(ir()->start()->live_in()); | |
822 | |
823 // print some additional information to simplify debugging | |
824 for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) { | |
825 if (ir()->start()->live_in().at(i)) { | |
826 Instruction* instr = gen()->instruction_for_vreg(i); | |
827 tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id()); | |
828 | |
829 for (int j = 0; j < num_blocks; j++) { | |
830 BlockBegin* block = block_at(j); | |
831 if (block->live_gen().at(i)) { | |
832 tty->print_cr(" used in block B%d", block->block_id()); | |
833 } | |
834 if (block->live_kill().at(i)) { | |
835 tty->print_cr(" defined in block B%d", block->block_id()); | |
836 } | |
837 } | |
838 } | |
839 } | |
840 | |
841 #endif | |
842 // when this fails, virtual registers are used before they are defined. | |
843 assert(false, "live_in set of first block must be empty"); | |
844 // bailout of if this occurs in product mode. | |
845 bailout("live_in set of first block not empty"); | |
846 } | |
847 } | |
848 | |
849 | |
850 // ********** Phase 4: build intervals | |
851 // (fills the list _intervals) | |
852 | |
853 void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) { | |
854 assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type"); | |
855 LIR_Opr opr = value->operand(); | |
856 Constant* con = value->as_Constant(); | |
857 | |
858 if ((con == NULL || con->is_pinned()) && opr->is_register()) { | |
859 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
860 add_use(opr, from, to, use_kind); | |
861 } | |
862 } | |
863 | |
864 | |
865 void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) { | |
866 TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind)); | |
867 assert(opr->is_register(), "should not be called otherwise"); | |
868 | |
869 if (opr->is_virtual_register()) { | |
870 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
871 add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register()); | |
872 | |
873 } else { | |
874 int reg = reg_num(opr); | |
875 if (is_processed_reg_num(reg)) { | |
876 add_def(reg, def_pos, use_kind, opr->type_register()); | |
877 } | |
878 reg = reg_numHi(opr); | |
879 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { | |
880 add_def(reg, def_pos, use_kind, opr->type_register()); | |
881 } | |
882 } | |
883 } | |
884 | |
885 void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) { | |
886 TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind)); | |
887 assert(opr->is_register(), "should not be called otherwise"); | |
888 | |
889 if (opr->is_virtual_register()) { | |
890 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
891 add_use(opr->vreg_number(), from, to, use_kind, opr->type_register()); | |
892 | |
893 } else { | |
894 int reg = reg_num(opr); | |
895 if (is_processed_reg_num(reg)) { | |
896 add_use(reg, from, to, use_kind, opr->type_register()); | |
897 } | |
898 reg = reg_numHi(opr); | |
899 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { | |
900 add_use(reg, from, to, use_kind, opr->type_register()); | |
901 } | |
902 } | |
903 } | |
904 | |
905 void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) { | |
906 TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind)); | |
907 assert(opr->is_register(), "should not be called otherwise"); | |
908 | |
909 if (opr->is_virtual_register()) { | |
910 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below"); | |
911 add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register()); | |
912 | |
913 } else { | |
914 int reg = reg_num(opr); | |
915 if (is_processed_reg_num(reg)) { | |
916 add_temp(reg, temp_pos, use_kind, opr->type_register()); | |
917 } | |
918 reg = reg_numHi(opr); | |
919 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { | |
920 add_temp(reg, temp_pos, use_kind, opr->type_register()); | |
921 } | |
922 } | |
923 } | |
924 | |
925 | |
926 void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) { | |
927 Interval* interval = interval_at(reg_num); | |
928 if (interval != NULL) { | |
929 assert(interval->reg_num() == reg_num, "wrong interval"); | |
930 | |
931 if (type != T_ILLEGAL) { | |
932 interval->set_type(type); | |
933 } | |
934 | |
935 Range* r = interval->first(); | |
936 if (r->from() <= def_pos) { | |
937 // Update the starting point (when a range is first created for a use, its | |
938 // start is the beginning of the current block until a def is encountered.) | |
939 r->set_from(def_pos); | |
940 interval->add_use_pos(def_pos, use_kind); | |
941 | |
942 } else { | |
943 // Dead value - make vacuous interval | |
944 // also add use_kind for dead intervals | |
945 interval->add_range(def_pos, def_pos + 1); | |
946 interval->add_use_pos(def_pos, use_kind); | |
947 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos)); | |
948 } | |
949 | |
950 } else { | |
951 // Dead value - make vacuous interval | |
952 // also add use_kind for dead intervals | |
953 interval = create_interval(reg_num); | |
954 if (type != T_ILLEGAL) { | |
955 interval->set_type(type); | |
956 } | |
957 | |
958 interval->add_range(def_pos, def_pos + 1); | |
959 interval->add_use_pos(def_pos, use_kind); | |
960 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos)); | |
961 } | |
962 | |
963 change_spill_definition_pos(interval, def_pos); | |
964 if (use_kind == noUse && interval->spill_state() <= startInMemory) { | |
965 // detection of method-parameters and roundfp-results | |
966 // TODO: move this directly to position where use-kind is computed | |
967 interval->set_spill_state(startInMemory); | |
968 } | |
969 } | |
970 | |
971 void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) { | |
972 Interval* interval = interval_at(reg_num); | |
973 if (interval == NULL) { | |
974 interval = create_interval(reg_num); | |
975 } | |
976 assert(interval->reg_num() == reg_num, "wrong interval"); | |
977 | |
978 if (type != T_ILLEGAL) { | |
979 interval->set_type(type); | |
980 } | |
981 | |
982 interval->add_range(from, to); | |
983 interval->add_use_pos(to, use_kind); | |
984 } | |
985 | |
986 void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) { | |
987 Interval* interval = interval_at(reg_num); | |
988 if (interval == NULL) { | |
989 interval = create_interval(reg_num); | |
990 } | |
991 assert(interval->reg_num() == reg_num, "wrong interval"); | |
992 | |
993 if (type != T_ILLEGAL) { | |
994 interval->set_type(type); | |
995 } | |
996 | |
997 interval->add_range(temp_pos, temp_pos + 1); | |
998 interval->add_use_pos(temp_pos, use_kind); | |
999 } | |
1000 | |
1001 | |
1002 // the results of this functions are used for optimizing spilling and reloading | |
1003 // if the functions return shouldHaveRegister and the interval is spilled, | |
1004 // it is not reloaded to a register. | |
1005 IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) { | |
1006 if (op->code() == lir_move) { | |
1007 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); | |
1008 LIR_Op1* move = (LIR_Op1*)op; | |
1009 LIR_Opr res = move->result_opr(); | |
1010 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); | |
1011 | |
1012 if (result_in_memory) { | |
1013 // Begin of an interval with must_start_in_memory set. | |
1014 // This interval will always get a stack slot first, so return noUse. | |
1015 return noUse; | |
1016 | |
1017 } else if (move->in_opr()->is_stack()) { | |
1018 // method argument (condition must be equal to handle_method_arguments) | |
1019 return noUse; | |
1020 | |
1021 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { | |
1022 // Move from register to register | |
1023 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { | |
1024 // special handling of phi-function moves inside osr-entry blocks | |
1025 // input operand must have a register instead of output operand (leads to better register allocation) | |
1026 return shouldHaveRegister; | |
1027 } | |
1028 } | |
1029 } | |
1030 | |
1031 if (opr->is_virtual() && | |
1032 gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) { | |
1033 // result is a stack-slot, so prevent immediate reloading | |
1034 return noUse; | |
1035 } | |
1036 | |
1037 // all other operands require a register | |
1038 return mustHaveRegister; | |
1039 } | |
1040 | |
1041 IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) { | |
1042 if (op->code() == lir_move) { | |
1043 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1"); | |
1044 LIR_Op1* move = (LIR_Op1*)op; | |
1045 LIR_Opr res = move->result_opr(); | |
1046 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); | |
1047 | |
1048 if (result_in_memory) { | |
1049 // Move to an interval with must_start_in_memory set. | |
1050 // To avoid moves from stack to stack (not allowed) force the input operand to a register | |
1051 return mustHaveRegister; | |
1052 | |
1053 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { | |
1054 // Move from register to register | |
1055 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { | |
1056 // special handling of phi-function moves inside osr-entry blocks | |
1057 // input operand must have a register instead of output operand (leads to better register allocation) | |
1058 return mustHaveRegister; | |
1059 } | |
1060 | |
1061 // The input operand is not forced to a register (moves from stack to register are allowed), | |
1062 // but it is faster if the input operand is in a register | |
1063 return shouldHaveRegister; | |
1064 } | |
1065 } | |
1066 | |
1067 | |
304 | 1068 #ifdef X86 |
0 | 1069 if (op->code() == lir_cmove) { |
1070 // conditional moves can handle stack operands | |
1071 assert(op->result_opr()->is_register(), "result must always be in a register"); | |
1072 return shouldHaveRegister; | |
1073 } | |
1074 | |
1075 // optimizations for second input operand of arithmehtic operations on Intel | |
1076 // this operand is allowed to be on the stack in some cases | |
1077 BasicType opr_type = opr->type_register(); | |
1078 if (opr_type == T_FLOAT || opr_type == T_DOUBLE) { | |
1079 if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2) { | |
1080 // SSE float instruction (T_DOUBLE only supported with SSE2) | |
1081 switch (op->code()) { | |
1082 case lir_cmp: | |
1083 case lir_add: | |
1084 case lir_sub: | |
1085 case lir_mul: | |
1086 case lir_div: | |
1087 { | |
1088 assert(op->as_Op2() != NULL, "must be LIR_Op2"); | |
1089 LIR_Op2* op2 = (LIR_Op2*)op; | |
1090 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { | |
1091 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); | |
1092 return shouldHaveRegister; | |
1093 } | |
1094 } | |
1095 } | |
1096 } else { | |
1097 // FPU stack float instruction | |
1098 switch (op->code()) { | |
1099 case lir_add: | |
1100 case lir_sub: | |
1101 case lir_mul: | |
1102 case lir_div: | |
1103 { | |
1104 assert(op->as_Op2() != NULL, "must be LIR_Op2"); | |
1105 LIR_Op2* op2 = (LIR_Op2*)op; | |
1106 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { | |
1107 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); | |
1108 return shouldHaveRegister; | |
1109 } | |
1110 } | |
1111 } | |
1112 } | |
1113 | |
1114 } else if (opr_type != T_LONG) { | |
1115 // integer instruction (note: long operands must always be in register) | |
1116 switch (op->code()) { | |
1117 case lir_cmp: | |
1118 case lir_add: | |
1119 case lir_sub: | |
1120 case lir_logic_and: | |
1121 case lir_logic_or: | |
1122 case lir_logic_xor: | |
1123 { | |
1124 assert(op->as_Op2() != NULL, "must be LIR_Op2"); | |
1125 LIR_Op2* op2 = (LIR_Op2*)op; | |
1126 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { | |
1127 assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register"); | |
1128 return shouldHaveRegister; | |
1129 } | |
1130 } | |
1131 } | |
1132 } | |
304 | 1133 #endif // X86 |
0 | 1134 |
1135 // all other operands require a register | |
1136 return mustHaveRegister; | |
1137 } | |
1138 | |
1139 | |
1140 void LinearScan::handle_method_arguments(LIR_Op* op) { | |
1141 // special handling for method arguments (moves from stack to virtual register): | |
1142 // the interval gets no register assigned, but the stack slot. | |
1143 // it is split before the first use by the register allocator. | |
1144 | |
1145 if (op->code() == lir_move) { | |
1146 assert(op->as_Op1() != NULL, "must be LIR_Op1"); | |
1147 LIR_Op1* move = (LIR_Op1*)op; | |
1148 | |
1149 if (move->in_opr()->is_stack()) { | |
1150 #ifdef ASSERT | |
1151 int arg_size = compilation()->method()->arg_size(); | |
1152 LIR_Opr o = move->in_opr(); | |
1153 if (o->is_single_stack()) { | |
1154 assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range"); | |
1155 } else if (o->is_double_stack()) { | |
1156 assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range"); | |
1157 } else { | |
1158 ShouldNotReachHere(); | |
1159 } | |
1160 | |
1161 assert(move->id() > 0, "invalid id"); | |
1162 assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block"); | |
1163 assert(move->result_opr()->is_virtual(), "result of move must be a virtual register"); | |
1164 | |
1165 TRACE_LINEAR_SCAN(4, tty->print_cr("found move from stack slot %d to vreg %d", o->is_single_stack() ? o->single_stack_ix() : o->double_stack_ix(), reg_num(move->result_opr()))); | |
1166 #endif | |
1167 | |
1168 Interval* interval = interval_at(reg_num(move->result_opr())); | |
1169 | |
1170 int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix()); | |
1171 interval->set_canonical_spill_slot(stack_slot); | |
1172 interval->assign_reg(stack_slot); | |
1173 } | |
1174 } | |
1175 } | |
1176 | |
1177 void LinearScan::handle_doubleword_moves(LIR_Op* op) { | |
1178 // special handling for doubleword move from memory to register: | |
1179 // in this case the registers of the input address and the result | |
1180 // registers must not overlap -> add a temp range for the input registers | |
1181 if (op->code() == lir_move) { | |
1182 assert(op->as_Op1() != NULL, "must be LIR_Op1"); | |
1183 LIR_Op1* move = (LIR_Op1*)op; | |
1184 | |
1185 if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) { | |
1186 LIR_Address* address = move->in_opr()->as_address_ptr(); | |
1187 if (address != NULL) { | |
1188 if (address->base()->is_valid()) { | |
1189 add_temp(address->base(), op->id(), noUse); | |
1190 } | |
1191 if (address->index()->is_valid()) { | |
1192 add_temp(address->index(), op->id(), noUse); | |
1193 } | |
1194 } | |
1195 } | |
1196 } | |
1197 } | |
1198 | |
1199 void LinearScan::add_register_hints(LIR_Op* op) { | |
1200 switch (op->code()) { | |
1201 case lir_move: // fall through | |
1202 case lir_convert: { | |
1203 assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1"); | |
1204 LIR_Op1* move = (LIR_Op1*)op; | |
1205 | |
1206 LIR_Opr move_from = move->in_opr(); | |
1207 LIR_Opr move_to = move->result_opr(); | |
1208 | |
1209 if (move_to->is_register() && move_from->is_register()) { | |
1210 Interval* from = interval_at(reg_num(move_from)); | |
1211 Interval* to = interval_at(reg_num(move_to)); | |
1212 if (from != NULL && to != NULL) { | |
1213 to->set_register_hint(from); | |
1214 TRACE_LINEAR_SCAN(4, tty->print_cr("operation at op_id %d: added hint from interval %d to %d", move->id(), from->reg_num(), to->reg_num())); | |
1215 } | |
1216 } | |
1217 break; | |
1218 } | |
1219 case lir_cmove: { | |
1220 assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2"); | |
1221 LIR_Op2* cmove = (LIR_Op2*)op; | |
1222 | |
1223 LIR_Opr move_from = cmove->in_opr1(); | |
1224 LIR_Opr move_to = cmove->result_opr(); | |
1225 | |
1226 if (move_to->is_register() && move_from->is_register()) { | |
1227 Interval* from = interval_at(reg_num(move_from)); | |
1228 Interval* to = interval_at(reg_num(move_to)); | |
1229 if (from != NULL && to != NULL) { | |
1230 to->set_register_hint(from); | |
1231 TRACE_LINEAR_SCAN(4, tty->print_cr("operation at op_id %d: added hint from interval %d to %d", cmove->id(), from->reg_num(), to->reg_num())); | |
1232 } | |
1233 } | |
1234 break; | |
1235 } | |
1236 } | |
1237 } | |
1238 | |
1239 | |
1240 void LinearScan::build_intervals() { | |
1241 TIME_LINEAR_SCAN(timer_build_intervals); | |
1242 | |
1243 // initialize interval list with expected number of intervals | |
1244 // (32 is added to have some space for split children without having to resize the list) | |
1245 _intervals = IntervalList(num_virtual_regs() + 32); | |
1246 // initialize all slots that are used by build_intervals | |
1247 _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL); | |
1248 | |
1249 // create a list with all caller-save registers (cpu, fpu, xmm) | |
1250 // when an instruction is a call, a temp range is created for all these registers | |
1251 int num_caller_save_registers = 0; | |
1252 int caller_save_registers[LinearScan::nof_regs]; | |
1253 | |
1254 int i; | |
1255 for (i = 0; i < FrameMap::nof_caller_save_cpu_regs; i++) { | |
1256 LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i); | |
1257 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); | |
1258 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); | |
1259 caller_save_registers[num_caller_save_registers++] = reg_num(opr); | |
1260 } | |
1261 | |
1262 // temp ranges for fpu registers are only created when the method has | |
1263 // virtual fpu operands. Otherwise no allocation for fpu registers is | |
1264 // perfomed and so the temp ranges would be useless | |
1265 if (has_fpu_registers()) { | |
304 | 1266 #ifdef X86 |
0 | 1267 if (UseSSE < 2) { |
1268 #endif | |
1269 for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) { | |
1270 LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i); | |
1271 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); | |
1272 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); | |
1273 caller_save_registers[num_caller_save_registers++] = reg_num(opr); | |
1274 } | |
304 | 1275 #ifdef X86 |
0 | 1276 } |
1277 if (UseSSE > 0) { | |
1278 for (i = 0; i < FrameMap::nof_caller_save_xmm_regs; i++) { | |
1279 LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i); | |
1280 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands"); | |
1281 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register"); | |
1282 caller_save_registers[num_caller_save_registers++] = reg_num(opr); | |
1283 } | |
1284 } | |
1285 #endif | |
1286 } | |
1287 assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds"); | |
1288 | |
1289 | |
1290 LIR_OpVisitState visitor; | |
1291 | |
1292 // iterate all blocks in reverse order | |
1293 for (i = block_count() - 1; i >= 0; i--) { | |
1294 BlockBegin* block = block_at(i); | |
1295 LIR_OpList* instructions = block->lir()->instructions_list(); | |
1296 int block_from = block->first_lir_instruction_id(); | |
1297 int block_to = block->last_lir_instruction_id(); | |
1298 | |
1299 assert(block_from == instructions->at(0)->id(), "must be"); | |
1300 assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be"); | |
1301 | |
1302 // Update intervals for registers live at the end of this block; | |
1303 BitMap live = block->live_out(); | |
304 | 1304 int size = (int)live.size(); |
1305 for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) { | |
0 | 1306 assert(live.at(number), "should not stop here otherwise"); |
1307 assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds"); | |
1308 TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2)); | |
1309 | |
1310 add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL); | |
1311 | |
1312 // add special use positions for loop-end blocks when the | |
1313 // interval is used anywhere inside this loop. It's possible | |
1314 // that the block was part of a non-natural loop, so it might | |
1315 // have an invalid loop index. | |
1316 if (block->is_set(BlockBegin::linear_scan_loop_end_flag) && | |
1317 block->loop_index() != -1 && | |
1318 is_interval_in_loop(number, block->loop_index())) { | |
1319 interval_at(number)->add_use_pos(block_to + 1, loopEndMarker); | |
1320 } | |
1321 } | |
1322 | |
1323 // iterate all instructions of the block in reverse order. | |
1324 // skip the first instruction because it is always a label | |
1325 // definitions of intervals are processed before uses | |
1326 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label"); | |
1327 for (int j = instructions->length() - 1; j >= 1; j--) { | |
1328 LIR_Op* op = instructions->at(j); | |
1329 int op_id = op->id(); | |
1330 | |
1331 // visit operation to collect all operands | |
1332 visitor.visit(op); | |
1333 | |
1334 // add a temp range for each register if operation destroys caller-save registers | |
1335 if (visitor.has_call()) { | |
1336 for (int k = 0; k < num_caller_save_registers; k++) { | |
1337 add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL); | |
1338 } | |
1339 TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers")); | |
1340 } | |
1341 | |
1342 // Add any platform dependent temps | |
1343 pd_add_temps(op); | |
1344 | |
1345 // visit definitions (output and temp operands) | |
1346 int k, n; | |
1347 n = visitor.opr_count(LIR_OpVisitState::outputMode); | |
1348 for (k = 0; k < n; k++) { | |
1349 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); | |
1350 assert(opr->is_register(), "visitor should only return register operands"); | |
1351 add_def(opr, op_id, use_kind_of_output_operand(op, opr)); | |
1352 } | |
1353 | |
1354 n = visitor.opr_count(LIR_OpVisitState::tempMode); | |
1355 for (k = 0; k < n; k++) { | |
1356 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); | |
1357 assert(opr->is_register(), "visitor should only return register operands"); | |
1358 add_temp(opr, op_id, mustHaveRegister); | |
1359 } | |
1360 | |
1361 // visit uses (input operands) | |
1362 n = visitor.opr_count(LIR_OpVisitState::inputMode); | |
1363 for (k = 0; k < n; k++) { | |
1364 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); | |
1365 assert(opr->is_register(), "visitor should only return register operands"); | |
1366 add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr)); | |
1367 } | |
1368 | |
1369 // Add uses of live locals from interpreter's point of view for proper | |
1370 // debug information generation | |
1371 // Treat these operands as temp values (if the life range is extended | |
1372 // to a call site, the value would be in a register at the call otherwise) | |
1373 n = visitor.info_count(); | |
1374 for (k = 0; k < n; k++) { | |
1375 CodeEmitInfo* info = visitor.info_at(k); | |
1376 ValueStack* stack = info->stack(); | |
1377 for_each_state_value(stack, value, | |
1378 add_use(value, block_from, op_id + 1, noUse); | |
1379 ); | |
1380 } | |
1381 | |
1382 // special steps for some instructions (especially moves) | |
1383 handle_method_arguments(op); | |
1384 handle_doubleword_moves(op); | |
1385 add_register_hints(op); | |
1386 | |
1387 } // end of instruction iteration | |
1388 } // end of block iteration | |
1389 | |
1390 | |
1391 // add the range [0, 1[ to all fixed intervals | |
1392 // -> the register allocator need not handle unhandled fixed intervals | |
1393 for (int n = 0; n < LinearScan::nof_regs; n++) { | |
1394 Interval* interval = interval_at(n); | |
1395 if (interval != NULL) { | |
1396 interval->add_range(0, 1); | |
1397 } | |
1398 } | |
1399 } | |
1400 | |
1401 | |
1402 // ********** Phase 5: actual register allocation | |
1403 | |
1404 int LinearScan::interval_cmp(Interval** a, Interval** b) { | |
1405 if (*a != NULL) { | |
1406 if (*b != NULL) { | |
1407 return (*a)->from() - (*b)->from(); | |
1408 } else { | |
1409 return -1; | |
1410 } | |
1411 } else { | |
1412 if (*b != NULL) { | |
1413 return 1; | |
1414 } else { | |
1415 return 0; | |
1416 } | |
1417 } | |
1418 } | |
1419 | |
1420 #ifndef PRODUCT | |
1421 bool LinearScan::is_sorted(IntervalArray* intervals) { | |
1422 int from = -1; | |
1423 int i, j; | |
1424 for (i = 0; i < intervals->length(); i ++) { | |
1425 Interval* it = intervals->at(i); | |
1426 if (it != NULL) { | |
1427 if (from > it->from()) { | |
1428 assert(false, ""); | |
1429 return false; | |
1430 } | |
1431 from = it->from(); | |
1432 } | |
1433 } | |
1434 | |
1435 // check in both directions if sorted list and unsorted list contain same intervals | |
1436 for (i = 0; i < interval_count(); i++) { | |
1437 if (interval_at(i) != NULL) { | |
1438 int num_found = 0; | |
1439 for (j = 0; j < intervals->length(); j++) { | |
1440 if (interval_at(i) == intervals->at(j)) { | |
1441 num_found++; | |
1442 } | |
1443 } | |
1444 assert(num_found == 1, "lists do not contain same intervals"); | |
1445 } | |
1446 } | |
1447 for (j = 0; j < intervals->length(); j++) { | |
1448 int num_found = 0; | |
1449 for (i = 0; i < interval_count(); i++) { | |
1450 if (interval_at(i) == intervals->at(j)) { | |
1451 num_found++; | |
1452 } | |
1453 } | |
1454 assert(num_found == 1, "lists do not contain same intervals"); | |
1455 } | |
1456 | |
1457 return true; | |
1458 } | |
1459 #endif | |
1460 | |
1461 void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) { | |
1462 if (*prev != NULL) { | |
1463 (*prev)->set_next(interval); | |
1464 } else { | |
1465 *first = interval; | |
1466 } | |
1467 *prev = interval; | |
1468 } | |
1469 | |
1470 void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) { | |
1471 assert(is_sorted(_sorted_intervals), "interval list is not sorted"); | |
1472 | |
1473 *list1 = *list2 = Interval::end(); | |
1474 | |
1475 Interval* list1_prev = NULL; | |
1476 Interval* list2_prev = NULL; | |
1477 Interval* v; | |
1478 | |
1479 const int n = _sorted_intervals->length(); | |
1480 for (int i = 0; i < n; i++) { | |
1481 v = _sorted_intervals->at(i); | |
1482 if (v == NULL) continue; | |
1483 | |
1484 if (is_list1(v)) { | |
1485 add_to_list(list1, &list1_prev, v); | |
1486 } else if (is_list2 == NULL || is_list2(v)) { | |
1487 add_to_list(list2, &list2_prev, v); | |
1488 } | |
1489 } | |
1490 | |
1491 if (list1_prev != NULL) list1_prev->set_next(Interval::end()); | |
1492 if (list2_prev != NULL) list2_prev->set_next(Interval::end()); | |
1493 | |
1494 assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel"); | |
1495 assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel"); | |
1496 } | |
1497 | |
1498 | |
1499 void LinearScan::sort_intervals_before_allocation() { | |
1500 TIME_LINEAR_SCAN(timer_sort_intervals_before); | |
1501 | |
1502 IntervalList* unsorted_list = &_intervals; | |
1503 int unsorted_len = unsorted_list->length(); | |
1504 int sorted_len = 0; | |
1505 int unsorted_idx; | |
1506 int sorted_idx = 0; | |
1507 int sorted_from_max = -1; | |
1508 | |
1509 // calc number of items for sorted list (sorted list must not contain NULL values) | |
1510 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { | |
1511 if (unsorted_list->at(unsorted_idx) != NULL) { | |
1512 sorted_len++; | |
1513 } | |
1514 } | |
1515 IntervalArray* sorted_list = new IntervalArray(sorted_len); | |
1516 | |
1517 // special sorting algorithm: the original interval-list is almost sorted, | |
1518 // only some intervals are swapped. So this is much faster than a complete QuickSort | |
1519 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { | |
1520 Interval* cur_interval = unsorted_list->at(unsorted_idx); | |
1521 | |
1522 if (cur_interval != NULL) { | |
1523 int cur_from = cur_interval->from(); | |
1524 | |
1525 if (sorted_from_max <= cur_from) { | |
1526 sorted_list->at_put(sorted_idx++, cur_interval); | |
1527 sorted_from_max = cur_interval->from(); | |
1528 } else { | |
1529 // the asumption that the intervals are already sorted failed, | |
1530 // so this interval must be sorted in manually | |
1531 int j; | |
1532 for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) { | |
1533 sorted_list->at_put(j + 1, sorted_list->at(j)); | |
1534 } | |
1535 sorted_list->at_put(j + 1, cur_interval); | |
1536 sorted_idx++; | |
1537 } | |
1538 } | |
1539 } | |
1540 _sorted_intervals = sorted_list; | |
1541 } | |
1542 | |
1543 void LinearScan::sort_intervals_after_allocation() { | |
1544 TIME_LINEAR_SCAN(timer_sort_intervals_after); | |
1545 | |
1546 IntervalArray* old_list = _sorted_intervals; | |
1547 IntervalList* new_list = _new_intervals_from_allocation; | |
1548 int old_len = old_list->length(); | |
1549 int new_len = new_list->length(); | |
1550 | |
1551 if (new_len == 0) { | |
1552 // no intervals have been added during allocation, so sorted list is already up to date | |
1553 return; | |
1554 } | |
1555 | |
1556 // conventional sort-algorithm for new intervals | |
1557 new_list->sort(interval_cmp); | |
1558 | |
1559 // merge old and new list (both already sorted) into one combined list | |
1560 IntervalArray* combined_list = new IntervalArray(old_len + new_len); | |
1561 int old_idx = 0; | |
1562 int new_idx = 0; | |
1563 | |
1564 while (old_idx + new_idx < old_len + new_len) { | |
1565 if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) { | |
1566 combined_list->at_put(old_idx + new_idx, old_list->at(old_idx)); | |
1567 old_idx++; | |
1568 } else { | |
1569 combined_list->at_put(old_idx + new_idx, new_list->at(new_idx)); | |
1570 new_idx++; | |
1571 } | |
1572 } | |
1573 | |
1574 _sorted_intervals = combined_list; | |
1575 } | |
1576 | |
1577 | |
1578 void LinearScan::allocate_registers() { | |
1579 TIME_LINEAR_SCAN(timer_allocate_registers); | |
1580 | |
1581 Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals; | |
1582 Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals; | |
1583 | |
1584 create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals, is_precolored_cpu_interval, is_virtual_cpu_interval); | |
1585 if (has_fpu_registers()) { | |
1586 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); | |
1587 #ifdef ASSERT | |
1588 } else { | |
1589 // fpu register allocation is omitted because no virtual fpu registers are present | |
1590 // just check this again... | |
1591 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, is_precolored_fpu_interval, is_virtual_fpu_interval); | |
1592 assert(not_precolored_fpu_intervals == Interval::end(), "missed an uncolored fpu interval"); | |
1593 #endif | |
1594 } | |
1595 | |
1596 // allocate cpu registers | |
1597 LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals); | |
1598 cpu_lsw.walk(); | |
1599 cpu_lsw.finish_allocation(); | |
1600 | |
1601 if (has_fpu_registers()) { | |
1602 // allocate fpu registers | |
1603 LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals); | |
1604 fpu_lsw.walk(); | |
1605 fpu_lsw.finish_allocation(); | |
1606 } | |
1607 } | |
1608 | |
1609 | |
1610 // ********** Phase 6: resolve data flow | |
1611 // (insert moves at edges between blocks if intervals have been split) | |
1612 | |
1613 // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode | |
1614 // instead of returning NULL | |
1615 Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) { | |
1616 Interval* result = interval->split_child_at_op_id(op_id, mode); | |
1617 if (result != NULL) { | |
1618 return result; | |
1619 } | |
1620 | |
1621 assert(false, "must find an interval, but do a clean bailout in product mode"); | |
1622 result = new Interval(LIR_OprDesc::vreg_base); | |
1623 result->assign_reg(0); | |
1624 result->set_type(T_INT); | |
1625 BAILOUT_("LinearScan: interval is NULL", result); | |
1626 } | |
1627 | |
1628 | |
1629 Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) { | |
1630 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); | |
1631 assert(interval_at(reg_num) != NULL, "no interval found"); | |
1632 | |
1633 return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode); | |
1634 } | |
1635 | |
1636 Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) { | |
1637 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); | |
1638 assert(interval_at(reg_num) != NULL, "no interval found"); | |
1639 | |
1640 return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode); | |
1641 } | |
1642 | |
1643 Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) { | |
1644 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds"); | |
1645 assert(interval_at(reg_num) != NULL, "no interval found"); | |
1646 | |
1647 return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode); | |
1648 } | |
1649 | |
1650 | |
1651 void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { | |
1652 DEBUG_ONLY(move_resolver.check_empty()); | |
1653 | |
1654 const int num_regs = num_virtual_regs(); | |
1655 const int size = live_set_size(); | |
1656 const BitMap live_at_edge = to_block->live_in(); | |
1657 | |
1658 // visit all registers where the live_at_edge bit is set | |
304 | 1659 for (int r = (int)live_at_edge.get_next_one_offset(0, size); r < size; r = (int)live_at_edge.get_next_one_offset(r + 1, size)) { |
0 | 1660 assert(r < num_regs, "live information set for not exisiting interval"); |
1661 assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge"); | |
1662 | |
1663 Interval* from_interval = interval_at_block_end(from_block, r); | |
1664 Interval* to_interval = interval_at_block_begin(to_block, r); | |
1665 | |
1666 if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) { | |
1667 // need to insert move instruction | |
1668 move_resolver.add_mapping(from_interval, to_interval); | |
1669 } | |
1670 } | |
1671 } | |
1672 | |
1673 | |
1674 void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { | |
1675 if (from_block->number_of_sux() <= 1) { | |
1676 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id())); | |
1677 | |
1678 LIR_OpList* instructions = from_block->lir()->instructions_list(); | |
1679 LIR_OpBranch* branch = instructions->last()->as_OpBranch(); | |
1680 if (branch != NULL) { | |
1681 // insert moves before branch | |
1682 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); | |
1683 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2); | |
1684 } else { | |
1685 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1); | |
1686 } | |
1687 | |
1688 } else { | |
1689 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id())); | |
1690 #ifdef ASSERT | |
1691 assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label"); | |
1692 | |
1693 // because the number of predecessor edges matches the number of | |
1694 // successor edges, blocks which are reached by switch statements | |
1695 // may have be more than one predecessor but it will be guaranteed | |
1696 // that all predecessors will be the same. | |
1697 for (int i = 0; i < to_block->number_of_preds(); i++) { | |
1698 assert(from_block == to_block->pred_at(i), "all critical edges must be broken"); | |
1699 } | |
1700 #endif | |
1701 | |
1702 move_resolver.set_insert_position(to_block->lir(), 0); | |
1703 } | |
1704 } | |
1705 | |
1706 | |
1707 // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split | |
1708 void LinearScan::resolve_data_flow() { | |
1709 TIME_LINEAR_SCAN(timer_resolve_data_flow); | |
1710 | |
1711 int num_blocks = block_count(); | |
1712 MoveResolver move_resolver(this); | |
1713 BitMap block_completed(num_blocks); block_completed.clear(); | |
1714 BitMap already_resolved(num_blocks); already_resolved.clear(); | |
1715 | |
1716 int i; | |
1717 for (i = 0; i < num_blocks; i++) { | |
1718 BlockBegin* block = block_at(i); | |
1719 | |
1720 // check if block has only one predecessor and only one successor | |
1721 if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) { | |
1722 LIR_OpList* instructions = block->lir()->instructions_list(); | |
1723 assert(instructions->at(0)->code() == lir_label, "block must start with label"); | |
1724 assert(instructions->last()->code() == lir_branch, "block with successors must end with branch"); | |
1725 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch"); | |
1726 | |
1727 // check if block is empty (only label and branch) | |
1728 if (instructions->length() == 2) { | |
1729 BlockBegin* pred = block->pred_at(0); | |
1730 BlockBegin* sux = block->sux_at(0); | |
1731 | |
1732 // prevent optimization of two consecutive blocks | |
1733 if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) { | |
1734 TRACE_LINEAR_SCAN(3, tty->print_cr("**** optimizing empty block B%d (pred: B%d, sux: B%d)", block->block_id(), pred->block_id(), sux->block_id())); | |
1735 block_completed.set_bit(block->linear_scan_number()); | |
1736 | |
1737 // directly resolve between pred and sux (without looking at the empty block between) | |
1738 resolve_collect_mappings(pred, sux, move_resolver); | |
1739 if (move_resolver.has_mappings()) { | |
1740 move_resolver.set_insert_position(block->lir(), 0); | |
1741 move_resolver.resolve_and_append_moves(); | |
1742 } | |
1743 } | |
1744 } | |
1745 } | |
1746 } | |
1747 | |
1748 | |
1749 for (i = 0; i < num_blocks; i++) { | |
1750 if (!block_completed.at(i)) { | |
1751 BlockBegin* from_block = block_at(i); | |
1752 already_resolved.set_from(block_completed); | |
1753 | |
1754 int num_sux = from_block->number_of_sux(); | |
1755 for (int s = 0; s < num_sux; s++) { | |
1756 BlockBegin* to_block = from_block->sux_at(s); | |
1757 | |
1758 // check for duplicate edges between the same blocks (can happen with switch blocks) | |
1759 if (!already_resolved.at(to_block->linear_scan_number())) { | |
1760 TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id())); | |
1761 already_resolved.set_bit(to_block->linear_scan_number()); | |
1762 | |
1763 // collect all intervals that have been split between from_block and to_block | |
1764 resolve_collect_mappings(from_block, to_block, move_resolver); | |
1765 if (move_resolver.has_mappings()) { | |
1766 resolve_find_insert_pos(from_block, to_block, move_resolver); | |
1767 move_resolver.resolve_and_append_moves(); | |
1768 } | |
1769 } | |
1770 } | |
1771 } | |
1772 } | |
1773 } | |
1774 | |
1775 | |
1776 void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) { | |
1777 if (interval_at(reg_num) == NULL) { | |
1778 // if a phi function is never used, no interval is created -> ignore this | |
1779 return; | |
1780 } | |
1781 | |
1782 Interval* interval = interval_at_block_begin(block, reg_num); | |
1783 int reg = interval->assigned_reg(); | |
1784 int regHi = interval->assigned_regHi(); | |
1785 | |
1786 if ((reg < nof_regs && interval->always_in_memory()) || | |
1787 (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) { | |
1788 // the interval is split to get a short range that is located on the stack | |
1789 // in the following two cases: | |
1790 // * the interval started in memory (e.g. method parameter), but is currently in a register | |
1791 // this is an optimization for exception handling that reduces the number of moves that | |
1792 // are necessary for resolving the states when an exception uses this exception handler | |
1793 // * the interval would be on the fpu stack at the begin of the exception handler | |
1794 // this is not allowed because of the complicated fpu stack handling on Intel | |
1795 | |
1796 // range that will be spilled to memory | |
1797 int from_op_id = block->first_lir_instruction_id(); | |
1798 int to_op_id = from_op_id + 1; // short live range of length 1 | |
1799 assert(interval->from() <= from_op_id && interval->to() >= to_op_id, | |
1800 "no split allowed between exception entry and first instruction"); | |
1801 | |
1802 if (interval->from() != from_op_id) { | |
1803 // the part before from_op_id is unchanged | |
1804 interval = interval->split(from_op_id); | |
1805 interval->assign_reg(reg, regHi); | |
1806 append_interval(interval); | |
1807 } | |
1808 assert(interval->from() == from_op_id, "must be true now"); | |
1809 | |
1810 Interval* spilled_part = interval; | |
1811 if (interval->to() != to_op_id) { | |
1812 // the part after to_op_id is unchanged | |
1813 spilled_part = interval->split_from_start(to_op_id); | |
1814 append_interval(spilled_part); | |
1815 move_resolver.add_mapping(spilled_part, interval); | |
1816 } | |
1817 assign_spill_slot(spilled_part); | |
1818 | |
1819 assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking"); | |
1820 } | |
1821 } | |
1822 | |
1823 void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) { | |
1824 assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise"); | |
1825 DEBUG_ONLY(move_resolver.check_empty()); | |
1826 | |
1827 // visit all registers where the live_in bit is set | |
1828 int size = live_set_size(); | |
304 | 1829 for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { |
0 | 1830 resolve_exception_entry(block, r, move_resolver); |
1831 } | |
1832 | |
1833 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately | |
1834 for_each_phi_fun(block, phi, | |
1835 resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver) | |
1836 ); | |
1837 | |
1838 if (move_resolver.has_mappings()) { | |
1839 // insert moves after first instruction | |
1840 move_resolver.set_insert_position(block->lir(), 1); | |
1841 move_resolver.resolve_and_append_moves(); | |
1842 } | |
1843 } | |
1844 | |
1845 | |
1846 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) { | |
1847 if (interval_at(reg_num) == NULL) { | |
1848 // if a phi function is never used, no interval is created -> ignore this | |
1849 return; | |
1850 } | |
1851 | |
1852 // the computation of to_interval is equal to resolve_collect_mappings, | |
1853 // but from_interval is more complicated because of phi functions | |
1854 BlockBegin* to_block = handler->entry_block(); | |
1855 Interval* to_interval = interval_at_block_begin(to_block, reg_num); | |
1856 | |
1857 if (phi != NULL) { | |
1858 // phi function of the exception entry block | |
1859 // no moves are created for this phi function in the LIR_Generator, so the | |
1860 // interval at the throwing instruction must be searched using the operands | |
1861 // of the phi function | |
1862 Value from_value = phi->operand_at(handler->phi_operand()); | |
1863 | |
1864 // with phi functions it can happen that the same from_value is used in | |
1865 // multiple mappings, so notify move-resolver that this is allowed | |
1866 move_resolver.set_multiple_reads_allowed(); | |
1867 | |
1868 Constant* con = from_value->as_Constant(); | |
1869 if (con != NULL && !con->is_pinned()) { | |
1870 // unpinned constants may have no register, so add mapping from constant to interval | |
1871 move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval); | |
1872 } else { | |
1873 // search split child at the throwing op_id | |
1874 Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id); | |
1875 move_resolver.add_mapping(from_interval, to_interval); | |
1876 } | |
1877 | |
1878 } else { | |
1879 // no phi function, so use reg_num also for from_interval | |
1880 // search split child at the throwing op_id | |
1881 Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id); | |
1882 if (from_interval != to_interval) { | |
1883 // optimization to reduce number of moves: when to_interval is on stack and | |
1884 // the stack slot is known to be always correct, then no move is necessary | |
1885 if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) { | |
1886 move_resolver.add_mapping(from_interval, to_interval); | |
1887 } | |
1888 } | |
1889 } | |
1890 } | |
1891 | |
1892 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) { | |
1893 TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id)); | |
1894 | |
1895 DEBUG_ONLY(move_resolver.check_empty()); | |
1896 assert(handler->lir_op_id() == -1, "already processed this xhandler"); | |
1897 DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id)); | |
1898 assert(handler->entry_code() == NULL, "code already present"); | |
1899 | |
1900 // visit all registers where the live_in bit is set | |
1901 BlockBegin* block = handler->entry_block(); | |
1902 int size = live_set_size(); | |
304 | 1903 for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { |
0 | 1904 resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver); |
1905 } | |
1906 | |
1907 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately | |
1908 for_each_phi_fun(block, phi, | |
1909 resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver) | |
1910 ); | |
1911 | |
1912 if (move_resolver.has_mappings()) { | |
1913 LIR_List* entry_code = new LIR_List(compilation()); | |
1914 move_resolver.set_insert_position(entry_code, 0); | |
1915 move_resolver.resolve_and_append_moves(); | |
1916 | |
1917 entry_code->jump(handler->entry_block()); | |
1918 handler->set_entry_code(entry_code); | |
1919 } | |
1920 } | |
1921 | |
1922 | |
1923 void LinearScan::resolve_exception_handlers() { | |
1924 MoveResolver move_resolver(this); | |
1925 LIR_OpVisitState visitor; | |
1926 int num_blocks = block_count(); | |
1927 | |
1928 int i; | |
1929 for (i = 0; i < num_blocks; i++) { | |
1930 BlockBegin* block = block_at(i); | |
1931 if (block->is_set(BlockBegin::exception_entry_flag)) { | |
1932 resolve_exception_entry(block, move_resolver); | |
1933 } | |
1934 } | |
1935 | |
1936 for (i = 0; i < num_blocks; i++) { | |
1937 BlockBegin* block = block_at(i); | |
1938 LIR_List* ops = block->lir(); | |
1939 int num_ops = ops->length(); | |
1940 | |
1941 // iterate all instructions of the block. skip the first because it is always a label | |
1942 assert(visitor.no_operands(ops->at(0)), "first operation must always be a label"); | |
1943 for (int j = 1; j < num_ops; j++) { | |
1944 LIR_Op* op = ops->at(j); | |
1945 int op_id = op->id(); | |
1946 | |
1947 if (op_id != -1 && has_info(op_id)) { | |
1948 // visit operation to collect all operands | |
1949 visitor.visit(op); | |
1950 assert(visitor.info_count() > 0, "should not visit otherwise"); | |
1951 | |
1952 XHandlers* xhandlers = visitor.all_xhandler(); | |
1953 int n = xhandlers->length(); | |
1954 for (int k = 0; k < n; k++) { | |
1955 resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver); | |
1956 } | |
1957 | |
1958 #ifdef ASSERT | |
1959 } else { | |
1960 visitor.visit(op); | |
1961 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); | |
1962 #endif | |
1963 } | |
1964 } | |
1965 } | |
1966 } | |
1967 | |
1968 | |
1969 // ********** Phase 7: assign register numbers back to LIR | |
1970 // (includes computation of debug information and oop maps) | |
1971 | |
1972 VMReg LinearScan::vm_reg_for_interval(Interval* interval) { | |
1973 VMReg reg = interval->cached_vm_reg(); | |
1974 if (!reg->is_valid() ) { | |
1975 reg = vm_reg_for_operand(operand_for_interval(interval)); | |
1976 interval->set_cached_vm_reg(reg); | |
1977 } | |
1978 assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value"); | |
1979 return reg; | |
1980 } | |
1981 | |
1982 VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) { | |
1983 assert(opr->is_oop(), "currently only implemented for oop operands"); | |
1984 return frame_map()->regname(opr); | |
1985 } | |
1986 | |
1987 | |
1988 LIR_Opr LinearScan::operand_for_interval(Interval* interval) { | |
1989 LIR_Opr opr = interval->cached_opr(); | |
1990 if (opr->is_illegal()) { | |
1991 opr = calc_operand_for_interval(interval); | |
1992 interval->set_cached_opr(opr); | |
1993 } | |
1994 | |
1995 assert(opr == calc_operand_for_interval(interval), "wrong cached value"); | |
1996 return opr; | |
1997 } | |
1998 | |
1999 LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) { | |
2000 int assigned_reg = interval->assigned_reg(); | |
2001 BasicType type = interval->type(); | |
2002 | |
2003 if (assigned_reg >= nof_regs) { | |
2004 // stack slot | |
2005 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); | |
2006 return LIR_OprFact::stack(assigned_reg - nof_regs, type); | |
2007 | |
2008 } else { | |
2009 // register | |
2010 switch (type) { | |
2011 case T_OBJECT: { | |
2012 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); | |
2013 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); | |
2014 return LIR_OprFact::single_cpu_oop(assigned_reg); | |
2015 } | |
2016 | |
2017 case T_INT: { | |
2018 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); | |
2019 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); | |
2020 return LIR_OprFact::single_cpu(assigned_reg); | |
2021 } | |
2022 | |
2023 case T_LONG: { | |
2024 int assigned_regHi = interval->assigned_regHi(); | |
2025 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register"); | |
2026 assert(num_physical_regs(T_LONG) == 1 || | |
2027 (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register"); | |
2028 | |
2029 assert(assigned_reg != assigned_regHi, "invalid allocation"); | |
2030 assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi, | |
2031 "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)"); | |
2032 assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match"); | |
2033 if (requires_adjacent_regs(T_LONG)) { | |
2034 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even"); | |
2035 } | |
2036 | |
2037 #ifdef _LP64 | |
2038 return LIR_OprFact::double_cpu(assigned_reg, assigned_reg); | |
2039 #else | |
304 | 2040 #ifdef SPARC |
0 | 2041 return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg); |
2042 #else | |
2043 return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi); | |
304 | 2044 #endif // SPARC |
2045 #endif // LP64 | |
0 | 2046 } |
2047 | |
2048 case T_FLOAT: { | |
304 | 2049 #ifdef X86 |
0 | 2050 if (UseSSE >= 1) { |
2051 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register"); | |
2052 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); | |
2053 return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg); | |
2054 } | |
2055 #endif | |
2056 | |
2057 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); | |
2058 assert(interval->assigned_regHi() == any_reg, "must not have hi register"); | |
2059 return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg); | |
2060 } | |
2061 | |
2062 case T_DOUBLE: { | |
304 | 2063 #ifdef X86 |
0 | 2064 if (UseSSE >= 2) { |
2065 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= pd_last_xmm_reg, "no xmm register"); | |
2066 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)"); | |
2067 return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg); | |
2068 } | |
2069 #endif | |
2070 | |
2071 #ifdef SPARC | |
2072 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); | |
2073 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register"); | |
2074 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even"); | |
2075 LIR_Opr result = LIR_OprFact::double_fpu(interval->assigned_regHi() - pd_first_fpu_reg, assigned_reg - pd_first_fpu_reg); | |
2076 #else | |
2077 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register"); | |
2078 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)"); | |
2079 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg); | |
2080 #endif | |
2081 return result; | |
2082 } | |
2083 | |
2084 default: { | |
2085 ShouldNotReachHere(); | |
2086 return LIR_OprFact::illegalOpr; | |
2087 } | |
2088 } | |
2089 } | |
2090 } | |
2091 | |
2092 LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) { | |
2093 assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set"); | |
2094 return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type()); | |
2095 } | |
2096 | |
2097 LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) { | |
2098 assert(opr->is_virtual(), "should not call this otherwise"); | |
2099 | |
2100 Interval* interval = interval_at(opr->vreg_number()); | |
2101 assert(interval != NULL, "interval must exist"); | |
2102 | |
2103 if (op_id != -1) { | |
2104 #ifdef ASSERT | |
2105 BlockBegin* block = block_of_op_with_id(op_id); | |
2106 if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) { | |
2107 // check if spill moves could have been appended at the end of this block, but | |
2108 // before the branch instruction. So the split child information for this branch would | |
2109 // be incorrect. | |
2110 LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch(); | |
2111 if (branch != NULL) { | |
2112 if (block->live_out().at(opr->vreg_number())) { | |
2113 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump"); | |
2114 assert(false, "can't get split child for the last branch of a block because the information would be incorrect (moves are inserted before the branch in resolve_data_flow)"); | |
2115 } | |
2116 } | |
2117 } | |
2118 #endif | |
2119 | |
2120 // operands are not changed when an interval is split during allocation, | |
2121 // so search the right interval here | |
2122 interval = split_child_at_op_id(interval, op_id, mode); | |
2123 } | |
2124 | |
2125 LIR_Opr res = operand_for_interval(interval); | |
2126 | |
304 | 2127 #ifdef X86 |
0 | 2128 // new semantic for is_last_use: not only set on definite end of interval, |
2129 // but also before hole | |
2130 // This may still miss some cases (e.g. for dead values), but it is not necessary that the | |
2131 // last use information is completely correct | |
2132 // information is only needed for fpu stack allocation | |
2133 if (res->is_fpu_register()) { | |
2134 if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) { | |
2135 assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow"); | |
2136 res = res->make_last_use(); | |
2137 } | |
2138 } | |
2139 #endif | |
2140 | |
2141 assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation"); | |
2142 | |
2143 return res; | |
2144 } | |
2145 | |
2146 | |
2147 #ifdef ASSERT | |
2148 // some methods used to check correctness of debug information | |
2149 | |
2150 void assert_no_register_values(GrowableArray<ScopeValue*>* values) { | |
2151 if (values == NULL) { | |
2152 return; | |
2153 } | |
2154 | |
2155 for (int i = 0; i < values->length(); i++) { | |
2156 ScopeValue* value = values->at(i); | |
2157 | |
2158 if (value->is_location()) { | |
2159 Location location = ((LocationValue*)value)->location(); | |
2160 assert(location.where() == Location::on_stack, "value is in register"); | |
2161 } | |
2162 } | |
2163 } | |
2164 | |
2165 void assert_no_register_values(GrowableArray<MonitorValue*>* values) { | |
2166 if (values == NULL) { | |
2167 return; | |
2168 } | |
2169 | |
2170 for (int i = 0; i < values->length(); i++) { | |
2171 MonitorValue* value = values->at(i); | |
2172 | |
2173 if (value->owner()->is_location()) { | |
2174 Location location = ((LocationValue*)value->owner())->location(); | |
2175 assert(location.where() == Location::on_stack, "owner is in register"); | |
2176 } | |
2177 assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register"); | |
2178 } | |
2179 } | |
2180 | |
2181 void assert_equal(Location l1, Location l2) { | |
2182 assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), ""); | |
2183 } | |
2184 | |
2185 void assert_equal(ScopeValue* v1, ScopeValue* v2) { | |
2186 if (v1->is_location()) { | |
2187 assert(v2->is_location(), ""); | |
2188 assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location()); | |
2189 } else if (v1->is_constant_int()) { | |
2190 assert(v2->is_constant_int(), ""); | |
2191 assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), ""); | |
2192 } else if (v1->is_constant_double()) { | |
2193 assert(v2->is_constant_double(), ""); | |
2194 assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), ""); | |
2195 } else if (v1->is_constant_long()) { | |
2196 assert(v2->is_constant_long(), ""); | |
2197 assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), ""); | |
2198 } else if (v1->is_constant_oop()) { | |
2199 assert(v2->is_constant_oop(), ""); | |
2200 assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), ""); | |
2201 } else { | |
2202 ShouldNotReachHere(); | |
2203 } | |
2204 } | |
2205 | |
2206 void assert_equal(MonitorValue* m1, MonitorValue* m2) { | |
2207 assert_equal(m1->owner(), m2->owner()); | |
2208 assert_equal(m1->basic_lock(), m2->basic_lock()); | |
2209 } | |
2210 | |
2211 void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) { | |
2212 assert(d1->scope() == d2->scope(), "not equal"); | |
2213 assert(d1->bci() == d2->bci(), "not equal"); | |
2214 | |
2215 if (d1->locals() != NULL) { | |
2216 assert(d1->locals() != NULL && d2->locals() != NULL, "not equal"); | |
2217 assert(d1->locals()->length() == d2->locals()->length(), "not equal"); | |
2218 for (int i = 0; i < d1->locals()->length(); i++) { | |
2219 assert_equal(d1->locals()->at(i), d2->locals()->at(i)); | |
2220 } | |
2221 } else { | |
2222 assert(d1->locals() == NULL && d2->locals() == NULL, "not equal"); | |
2223 } | |
2224 | |
2225 if (d1->expressions() != NULL) { | |
2226 assert(d1->expressions() != NULL && d2->expressions() != NULL, "not equal"); | |
2227 assert(d1->expressions()->length() == d2->expressions()->length(), "not equal"); | |
2228 for (int i = 0; i < d1->expressions()->length(); i++) { | |
2229 assert_equal(d1->expressions()->at(i), d2->expressions()->at(i)); | |
2230 } | |
2231 } else { | |
2232 assert(d1->expressions() == NULL && d2->expressions() == NULL, "not equal"); | |
2233 } | |
2234 | |
2235 if (d1->monitors() != NULL) { | |
2236 assert(d1->monitors() != NULL && d2->monitors() != NULL, "not equal"); | |
2237 assert(d1->monitors()->length() == d2->monitors()->length(), "not equal"); | |
2238 for (int i = 0; i < d1->monitors()->length(); i++) { | |
2239 assert_equal(d1->monitors()->at(i), d2->monitors()->at(i)); | |
2240 } | |
2241 } else { | |
2242 assert(d1->monitors() == NULL && d2->monitors() == NULL, "not equal"); | |
2243 } | |
2244 | |
2245 if (d1->caller() != NULL) { | |
2246 assert(d1->caller() != NULL && d2->caller() != NULL, "not equal"); | |
2247 assert_equal(d1->caller(), d2->caller()); | |
2248 } else { | |
2249 assert(d1->caller() == NULL && d2->caller() == NULL, "not equal"); | |
2250 } | |
2251 } | |
2252 | |
2253 void check_stack_depth(CodeEmitInfo* info, int stack_end) { | |
2254 if (info->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) { | |
2255 Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->bci()); | |
2256 switch (code) { | |
2257 case Bytecodes::_ifnull : // fall through | |
2258 case Bytecodes::_ifnonnull : // fall through | |
2259 case Bytecodes::_ifeq : // fall through | |
2260 case Bytecodes::_ifne : // fall through | |
2261 case Bytecodes::_iflt : // fall through | |
2262 case Bytecodes::_ifge : // fall through | |
2263 case Bytecodes::_ifgt : // fall through | |
2264 case Bytecodes::_ifle : // fall through | |
2265 case Bytecodes::_if_icmpeq : // fall through | |
2266 case Bytecodes::_if_icmpne : // fall through | |
2267 case Bytecodes::_if_icmplt : // fall through | |
2268 case Bytecodes::_if_icmpge : // fall through | |
2269 case Bytecodes::_if_icmpgt : // fall through | |
2270 case Bytecodes::_if_icmple : // fall through | |
2271 case Bytecodes::_if_acmpeq : // fall through | |
2272 case Bytecodes::_if_acmpne : | |
2273 assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode"); | |
2274 break; | |
2275 } | |
2276 } | |
2277 } | |
2278 | |
2279 #endif // ASSERT | |
2280 | |
2281 | |
2282 IntervalWalker* LinearScan::init_compute_oop_maps() { | |
2283 // setup lists of potential oops for walking | |
2284 Interval* oop_intervals; | |
2285 Interval* non_oop_intervals; | |
2286 | |
2287 create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, NULL); | |
2288 | |
2289 // intervals that have no oops inside need not to be processed | |
2290 // to ensure a walking until the last instruction id, add a dummy interval | |
2291 // with a high operation id | |
2292 non_oop_intervals = new Interval(any_reg); | |
2293 non_oop_intervals->add_range(max_jint - 2, max_jint - 1); | |
2294 | |
2295 return new IntervalWalker(this, oop_intervals, non_oop_intervals); | |
2296 } | |
2297 | |
2298 | |
2299 OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) { | |
2300 TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d", op->id())); | |
2301 | |
2302 // walk before the current operation -> intervals that start at | |
2303 // the operation (= output operands of the operation) are not | |
2304 // included in the oop map | |
2305 iw->walk_before(op->id()); | |
2306 | |
2307 int frame_size = frame_map()->framesize(); | |
2308 int arg_count = frame_map()->oop_map_arg_count(); | |
2309 OopMap* map = new OopMap(frame_size, arg_count); | |
2310 | |
2311 // Check if this is a patch site. | |
2312 bool is_patch_info = false; | |
2313 if (op->code() == lir_move) { | |
2314 assert(!is_call_site, "move must not be a call site"); | |
2315 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); | |
2316 LIR_Op1* move = (LIR_Op1*)op; | |
2317 | |
2318 is_patch_info = move->patch_code() != lir_patch_none; | |
2319 } | |
2320 | |
2321 // Iterate through active intervals | |
2322 for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) { | |
2323 int assigned_reg = interval->assigned_reg(); | |
2324 | |
2325 assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise"); | |
2326 assert(interval->assigned_regHi() == any_reg, "oop must be single word"); | |
2327 assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found"); | |
2328 | |
2329 // Check if this range covers the instruction. Intervals that | |
2330 // start or end at the current operation are not included in the | |
2331 // oop map, except in the case of patching moves. For patching | |
2332 // moves, any intervals which end at this instruction are included | |
2333 // in the oop map since we may safepoint while doing the patch | |
2334 // before we've consumed the inputs. | |
2335 if (is_patch_info || op->id() < interval->current_to()) { | |
2336 | |
2337 // caller-save registers must not be included into oop-maps at calls | |
2338 assert(!is_call_site || assigned_reg >= nof_regs || !is_caller_save(assigned_reg), "interval is in a caller-save register at a call -> register will be overwritten"); | |
2339 | |
2340 VMReg name = vm_reg_for_interval(interval); | |
2341 map->set_oop(name); | |
2342 | |
2343 // Spill optimization: when the stack value is guaranteed to be always correct, | |
2344 // then it must be added to the oop map even if the interval is currently in a register | |
2345 if (interval->always_in_memory() && | |
2346 op->id() > interval->spill_definition_pos() && | |
2347 interval->assigned_reg() != interval->canonical_spill_slot()) { | |
2348 assert(interval->spill_definition_pos() > 0, "position not set correctly"); | |
2349 assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned"); | |
2350 assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice"); | |
2351 | |
2352 map->set_oop(frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs)); | |
2353 } | |
2354 } | |
2355 } | |
2356 | |
2357 // add oops from lock stack | |
2358 assert(info->stack() != NULL, "CodeEmitInfo must always have a stack"); | |
2359 int locks_count = info->stack()->locks_size(); | |
2360 for (int i = 0; i < locks_count; i++) { | |
2361 map->set_oop(frame_map()->monitor_object_regname(i)); | |
2362 } | |
2363 | |
2364 return map; | |
2365 } | |
2366 | |
2367 | |
2368 void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) { | |
2369 assert(visitor.info_count() > 0, "no oop map needed"); | |
2370 | |
2371 // compute oop_map only for first CodeEmitInfo | |
2372 // because it is (in most cases) equal for all other infos of the same operation | |
2373 CodeEmitInfo* first_info = visitor.info_at(0); | |
2374 OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call()); | |
2375 | |
2376 for (int i = 0; i < visitor.info_count(); i++) { | |
2377 CodeEmitInfo* info = visitor.info_at(i); | |
2378 OopMap* oop_map = first_oop_map; | |
2379 | |
2380 if (info->stack()->locks_size() != first_info->stack()->locks_size()) { | |
2381 // this info has a different number of locks then the precomputed oop map | |
2382 // (possible for lock and unlock instructions) -> compute oop map with | |
2383 // correct lock information | |
2384 oop_map = compute_oop_map(iw, op, info, visitor.has_call()); | |
2385 } | |
2386 | |
2387 if (info->_oop_map == NULL) { | |
2388 info->_oop_map = oop_map; | |
2389 } else { | |
2390 // a CodeEmitInfo can not be shared between different LIR-instructions | |
2391 // because interval splitting can occur anywhere between two instructions | |
2392 // and so the oop maps must be different | |
2393 // -> check if the already set oop_map is exactly the one calculated for this operation | |
2394 assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions"); | |
2395 } | |
2396 } | |
2397 } | |
2398 | |
2399 | |
2400 // frequently used constants | |
2401 ConstantOopWriteValue LinearScan::_oop_null_scope_value = ConstantOopWriteValue(NULL); | |
2402 ConstantIntValue LinearScan::_int_m1_scope_value = ConstantIntValue(-1); | |
2403 ConstantIntValue LinearScan::_int_0_scope_value = ConstantIntValue(0); | |
2404 ConstantIntValue LinearScan::_int_1_scope_value = ConstantIntValue(1); | |
2405 ConstantIntValue LinearScan::_int_2_scope_value = ConstantIntValue(2); | |
2406 LocationValue _illegal_value = LocationValue(Location()); | |
2407 | |
2408 void LinearScan::init_compute_debug_info() { | |
2409 // cache for frequently used scope values | |
2410 // (cpu registers and stack slots) | |
2411 _scope_value_cache = ScopeValueArray((LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2, NULL); | |
2412 } | |
2413 | |
2414 MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) { | |
2415 Location loc; | |
2416 if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) { | |
2417 bailout("too large frame"); | |
2418 } | |
2419 ScopeValue* object_scope_value = new LocationValue(loc); | |
2420 | |
2421 if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) { | |
2422 bailout("too large frame"); | |
2423 } | |
2424 return new MonitorValue(object_scope_value, loc); | |
2425 } | |
2426 | |
2427 LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) { | |
2428 Location loc; | |
2429 if (!frame_map()->locations_for_slot(name, loc_type, &loc)) { | |
2430 bailout("too large frame"); | |
2431 } | |
2432 return new LocationValue(loc); | |
2433 } | |
2434 | |
2435 | |
2436 int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { | |
2437 assert(opr->is_constant(), "should not be called otherwise"); | |
2438 | |
2439 LIR_Const* c = opr->as_constant_ptr(); | |
2440 BasicType t = c->type(); | |
2441 switch (t) { | |
2442 case T_OBJECT: { | |
2443 jobject value = c->as_jobject(); | |
2444 if (value == NULL) { | |
2445 scope_values->append(&_oop_null_scope_value); | |
2446 } else { | |
2447 scope_values->append(new ConstantOopWriteValue(c->as_jobject())); | |
2448 } | |
2449 return 1; | |
2450 } | |
2451 | |
2452 case T_INT: // fall through | |
2453 case T_FLOAT: { | |
2454 int value = c->as_jint_bits(); | |
2455 switch (value) { | |
2456 case -1: scope_values->append(&_int_m1_scope_value); break; | |
2457 case 0: scope_values->append(&_int_0_scope_value); break; | |
2458 case 1: scope_values->append(&_int_1_scope_value); break; | |
2459 case 2: scope_values->append(&_int_2_scope_value); break; | |
2460 default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break; | |
2461 } | |
2462 return 1; | |
2463 } | |
2464 | |
2465 case T_LONG: // fall through | |
2466 case T_DOUBLE: { | |
2467 if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) { | |
2468 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); | |
2469 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); | |
2470 } else { | |
2471 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); | |
2472 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); | |
2473 } | |
2474 | |
2475 return 2; | |
2476 } | |
2477 | |
2478 default: | |
2479 ShouldNotReachHere(); | |
304 | 2480 return -1; |
0 | 2481 } |
2482 } | |
2483 | |
2484 int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { | |
2485 if (opr->is_single_stack()) { | |
2486 int stack_idx = opr->single_stack_ix(); | |
2487 bool is_oop = opr->is_oop_register(); | |
2488 int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0); | |
2489 | |
2490 ScopeValue* sv = _scope_value_cache.at(cache_idx); | |
2491 if (sv == NULL) { | |
2492 Location::Type loc_type = is_oop ? Location::oop : Location::normal; | |
2493 sv = location_for_name(stack_idx, loc_type); | |
2494 _scope_value_cache.at_put(cache_idx, sv); | |
2495 } | |
2496 | |
2497 // check if cached value is correct | |
2498 DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal))); | |
2499 | |
2500 scope_values->append(sv); | |
2501 return 1; | |
2502 | |
2503 } else if (opr->is_single_cpu()) { | |
2504 bool is_oop = opr->is_oop_register(); | |
2505 int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0); | |
2506 | |
2507 ScopeValue* sv = _scope_value_cache.at(cache_idx); | |
2508 if (sv == NULL) { | |
2509 Location::Type loc_type = is_oop ? Location::oop : Location::normal; | |
2510 VMReg rname = frame_map()->regname(opr); | |
2511 sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); | |
2512 _scope_value_cache.at_put(cache_idx, sv); | |
2513 } | |
2514 | |
2515 // check if cached value is correct | |
2516 DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : Location::normal, frame_map()->regname(opr))))); | |
2517 | |
2518 scope_values->append(sv); | |
2519 return 1; | |
2520 | |
304 | 2521 #ifdef X86 |
0 | 2522 } else if (opr->is_single_xmm()) { |
2523 VMReg rname = opr->as_xmm_float_reg()->as_VMReg(); | |
2524 LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname)); | |
2525 | |
2526 scope_values->append(sv); | |
2527 return 1; | |
2528 #endif | |
2529 | |
2530 } else if (opr->is_single_fpu()) { | |
304 | 2531 #ifdef X86 |
0 | 2532 // the exact location of fpu stack values is only known |
2533 // during fpu stack allocation, so the stack allocator object | |
2534 // must be present | |
2535 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); | |
2536 assert(_fpu_stack_allocator != NULL, "must be present"); | |
2537 opr = _fpu_stack_allocator->to_fpu_stack(opr); | |
2538 #endif | |
2539 | |
2540 Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal; | |
2541 VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr()); | |
2542 LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); | |
2543 | |
2544 scope_values->append(sv); | |
2545 return 1; | |
2546 | |
2547 } else { | |
2548 // double-size operands | |
2549 | |
2550 ScopeValue* first; | |
2551 ScopeValue* second; | |
2552 | |
2553 if (opr->is_double_stack()) { | |
304 | 2554 #ifdef _LP64 |
2555 Location loc1; | |
2556 Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl; | |
2557 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) { | |
2558 bailout("too large frame"); | |
2559 } | |
2560 // Does this reverse on x86 vs. sparc? | |
2561 first = new LocationValue(loc1); | |
2562 second = &_int_0_scope_value; | |
2563 #else | |
0 | 2564 Location loc1, loc2; |
2565 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) { | |
2566 bailout("too large frame"); | |
2567 } | |
2568 first = new LocationValue(loc1); | |
2569 second = new LocationValue(loc2); | |
304 | 2570 #endif // _LP64 |
0 | 2571 |
2572 } else if (opr->is_double_cpu()) { | |
2573 #ifdef _LP64 | |
2574 VMReg rname_first = opr->as_register_lo()->as_VMReg(); | |
2575 first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first)); | |
2576 second = &_int_0_scope_value; | |
2577 #else | |
2578 VMReg rname_first = opr->as_register_lo()->as_VMReg(); | |
2579 VMReg rname_second = opr->as_register_hi()->as_VMReg(); | |
2580 | |
2581 if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) { | |
2582 // lo/hi and swapped relative to first and second, so swap them | |
2583 VMReg tmp = rname_first; | |
2584 rname_first = rname_second; | |
2585 rname_second = tmp; | |
2586 } | |
2587 | |
2588 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); | |
2589 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); | |
304 | 2590 #endif //_LP64 |
2591 | |
2592 | |
2593 #ifdef X86 | |
0 | 2594 } else if (opr->is_double_xmm()) { |
2595 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation"); | |
2596 VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg(); | |
2597 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); | |
2598 // %%% This is probably a waste but we'll keep things as they were for now | |
2599 if (true) { | |
2600 VMReg rname_second = rname_first->next(); | |
2601 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); | |
2602 } | |
2603 #endif | |
2604 | |
2605 } else if (opr->is_double_fpu()) { | |
2606 // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of | |
304 | 2607 // the double as float registers in the native ordering. On X86, |
0 | 2608 // fpu_regnrLo is a FPU stack slot whose VMReg represents |
2609 // the low-order word of the double and fpu_regnrLo + 1 is the | |
2610 // name for the other half. *first and *second must represent the | |
2611 // least and most significant words, respectively. | |
2612 | |
304 | 2613 #ifdef X86 |
0 | 2614 // the exact location of fpu stack values is only known |
2615 // during fpu stack allocation, so the stack allocator object | |
2616 // must be present | |
2617 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)"); | |
2618 assert(_fpu_stack_allocator != NULL, "must be present"); | |
2619 opr = _fpu_stack_allocator->to_fpu_stack(opr); | |
2620 | |
2621 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)"); | |
2622 #endif | |
2623 #ifdef SPARC | |
2624 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)"); | |
2625 #endif | |
2626 | |
2627 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi()); | |
2628 | |
2629 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); | |
2630 // %%% This is probably a waste but we'll keep things as they were for now | |
2631 if (true) { | |
2632 VMReg rname_second = rname_first->next(); | |
2633 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); | |
2634 } | |
2635 | |
2636 } else { | |
2637 ShouldNotReachHere(); | |
2638 first = NULL; | |
2639 second = NULL; | |
2640 } | |
2641 | |
2642 assert(first != NULL && second != NULL, "must be set"); | |
2643 // The convention the interpreter uses is that the second local | |
2644 // holds the first raw word of the native double representation. | |
2645 // This is actually reasonable, since locals and stack arrays | |
2646 // grow downwards in all implementations. | |
2647 // (If, on some machine, the interpreter's Java locals or stack | |
2648 // were to grow upwards, the embedded doubles would be word-swapped.) | |
2649 scope_values->append(second); | |
2650 scope_values->append(first); | |
2651 return 2; | |
2652 } | |
2653 } | |
2654 | |
2655 | |
2656 int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) { | |
2657 if (value != NULL) { | |
2658 LIR_Opr opr = value->operand(); | |
2659 Constant* con = value->as_Constant(); | |
2660 | |
2661 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands (or illegal if constant is optimized away)"); | |
2662 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands"); | |
2663 | |
2664 if (con != NULL && !con->is_pinned() && !opr->is_constant()) { | |
2665 // Unpinned constants may have a virtual operand for a part of the lifetime | |
2666 // or may be illegal when it was optimized away, | |
2667 // so always use a constant operand | |
2668 opr = LIR_OprFact::value_type(con->type()); | |
2669 } | |
2670 assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here"); | |
2671 | |
2672 if (opr->is_virtual()) { | |
2673 LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode; | |
2674 | |
2675 BlockBegin* block = block_of_op_with_id(op_id); | |
2676 if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) { | |
2677 // generating debug information for the last instruction of a block. | |
2678 // if this instruction is a branch, spill moves are inserted before this branch | |
2679 // and so the wrong operand would be returned (spill moves at block boundaries are not | |
2680 // considered in the live ranges of intervals) | |
2681 // Solution: use the first op_id of the branch target block instead. | |
2682 if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) { | |
2683 if (block->live_out().at(opr->vreg_number())) { | |
2684 op_id = block->sux_at(0)->first_lir_instruction_id(); | |
2685 mode = LIR_OpVisitState::outputMode; | |
2686 } | |
2687 } | |
2688 } | |
2689 | |
2690 // Get current location of operand | |
2691 // The operand must be live because debug information is considered when building the intervals | |
2692 // if the interval is not live, color_lir_opr will cause an assertion failure | |
2693 opr = color_lir_opr(opr, op_id, mode); | |
2694 assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls"); | |
2695 | |
2696 // Append to ScopeValue array | |
2697 return append_scope_value_for_operand(opr, scope_values); | |
2698 | |
2699 } else { | |
2700 assert(value->as_Constant() != NULL, "all other instructions have only virtual operands"); | |
2701 assert(opr->is_constant(), "operand must be constant"); | |
2702 | |
2703 return append_scope_value_for_constant(opr, scope_values); | |
2704 } | |
2705 } else { | |
2706 // append a dummy value because real value not needed | |
2707 scope_values->append(&_illegal_value); | |
2708 return 1; | |
2709 } | |
2710 } | |
2711 | |
2712 | |
2713 IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state, int cur_bci, int stack_end, int locks_end) { | |
2714 IRScopeDebugInfo* caller_debug_info = NULL; | |
2715 int stack_begin, locks_begin; | |
2716 | |
2717 ValueStack* caller_state = cur_scope->caller_state(); | |
2718 if (caller_state != NULL) { | |
2719 // process recursively to compute outermost scope first | |
2720 stack_begin = caller_state->stack_size(); | |
2721 locks_begin = caller_state->locks_size(); | |
2722 caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state, cur_scope->caller_bci(), stack_begin, locks_begin); | |
2723 } else { | |
2724 stack_begin = 0; | |
2725 locks_begin = 0; | |
2726 } | |
2727 | |
2728 // initialize these to null. | |
2729 // If we don't need deopt info or there are no locals, expressions or monitors, | |
2730 // then these get recorded as no information and avoids the allocation of 0 length arrays. | |
2731 GrowableArray<ScopeValue*>* locals = NULL; | |
2732 GrowableArray<ScopeValue*>* expressions = NULL; | |
2733 GrowableArray<MonitorValue*>* monitors = NULL; | |
2734 | |
2735 // describe local variable values | |
2736 int nof_locals = cur_scope->method()->max_locals(); | |
2737 if (nof_locals > 0) { | |
2738 locals = new GrowableArray<ScopeValue*>(nof_locals); | |
2739 | |
2740 int pos = 0; | |
2741 while (pos < nof_locals) { | |
2742 assert(pos < cur_state->locals_size(), "why not?"); | |
2743 | |
2744 Value local = cur_state->local_at(pos); | |
2745 pos += append_scope_value(op_id, local, locals); | |
2746 | |
2747 assert(locals->length() == pos, "must match"); | |
2748 } | |
2749 assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals"); | |
2750 assert(locals->length() == cur_state->locals_size(), "wrong number of locals"); | |
2751 } | |
2752 | |
2753 | |
2754 // describe expression stack | |
2755 // | |
2756 // When we inline methods containing exception handlers, the | |
2757 // "lock_stacks" are changed to preserve expression stack values | |
2758 // in caller scopes when exception handlers are present. This | |
2759 // can cause callee stacks to be smaller than caller stacks. | |
2760 if (stack_end > innermost_state->stack_size()) { | |
2761 stack_end = innermost_state->stack_size(); | |
2762 } | |
2763 | |
2764 | |
2765 | |
2766 int nof_stack = stack_end - stack_begin; | |
2767 if (nof_stack > 0) { | |
2768 expressions = new GrowableArray<ScopeValue*>(nof_stack); | |
2769 | |
2770 int pos = stack_begin; | |
2771 while (pos < stack_end) { | |
2772 Value expression = innermost_state->stack_at_inc(pos); | |
2773 append_scope_value(op_id, expression, expressions); | |
2774 | |
2775 assert(expressions->length() + stack_begin == pos, "must match"); | |
2776 } | |
2777 } | |
2778 | |
2779 // describe monitors | |
2780 assert(locks_begin <= locks_end, "error in scope iteration"); | |
2781 int nof_locks = locks_end - locks_begin; | |
2782 if (nof_locks > 0) { | |
2783 monitors = new GrowableArray<MonitorValue*>(nof_locks); | |
2784 for (int i = locks_begin; i < locks_end; i++) { | |
2785 monitors->append(location_for_monitor_index(i)); | |
2786 } | |
2787 } | |
2788 | |
2789 return new IRScopeDebugInfo(cur_scope, cur_bci, locals, expressions, monitors, caller_debug_info); | |
2790 } | |
2791 | |
2792 | |
2793 void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) { | |
2794 if (!compilation()->needs_debug_information()) { | |
2795 return; | |
2796 } | |
2797 TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id)); | |
2798 | |
2799 IRScope* innermost_scope = info->scope(); | |
2800 ValueStack* innermost_state = info->stack(); | |
2801 | |
2802 assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?"); | |
2803 | |
2804 int stack_end = innermost_state->stack_size(); | |
2805 int locks_end = innermost_state->locks_size(); | |
2806 | |
2807 DEBUG_ONLY(check_stack_depth(info, stack_end)); | |
2808 | |
2809 if (info->_scope_debug_info == NULL) { | |
2810 // compute debug information | |
2811 info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state, info->bci(), stack_end, locks_end); | |
2812 } else { | |
2813 // debug information already set. Check that it is correct from the current point of view | |
2814 DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state, info->bci(), stack_end, locks_end))); | |
2815 } | |
2816 } | |
2817 | |
2818 | |
2819 void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) { | |
2820 LIR_OpVisitState visitor; | |
2821 int num_inst = instructions->length(); | |
2822 bool has_dead = false; | |
2823 | |
2824 for (int j = 0; j < num_inst; j++) { | |
2825 LIR_Op* op = instructions->at(j); | |
2826 if (op == NULL) { // this can happen when spill-moves are removed in eliminate_spill_moves | |
2827 has_dead = true; | |
2828 continue; | |
2829 } | |
2830 int op_id = op->id(); | |
2831 | |
2832 // visit instruction to get list of operands | |
2833 visitor.visit(op); | |
2834 | |
2835 // iterate all modes of the visitor and process all virtual operands | |
2836 for_each_visitor_mode(mode) { | |
2837 int n = visitor.opr_count(mode); | |
2838 for (int k = 0; k < n; k++) { | |
2839 LIR_Opr opr = visitor.opr_at(mode, k); | |
2840 if (opr->is_virtual_register()) { | |
2841 visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode)); | |
2842 } | |
2843 } | |
2844 } | |
2845 | |
2846 if (visitor.info_count() > 0) { | |
2847 // exception handling | |
2848 if (compilation()->has_exception_handlers()) { | |
2849 XHandlers* xhandlers = visitor.all_xhandler(); | |
2850 int n = xhandlers->length(); | |
2851 for (int k = 0; k < n; k++) { | |
2852 XHandler* handler = xhandlers->handler_at(k); | |
2853 if (handler->entry_code() != NULL) { | |
2854 assign_reg_num(handler->entry_code()->instructions_list(), NULL); | |
2855 } | |
2856 } | |
2857 } else { | |
2858 assert(visitor.all_xhandler()->length() == 0, "missed exception handler"); | |
2859 } | |
2860 | |
2861 // compute oop map | |
2862 assert(iw != NULL, "needed for compute_oop_map"); | |
2863 compute_oop_map(iw, visitor, op); | |
2864 | |
2865 // compute debug information | |
2866 if (!use_fpu_stack_allocation()) { | |
2867 // compute debug information if fpu stack allocation is not needed. | |
2868 // when fpu stack allocation is needed, the debug information can not | |
2869 // be computed here because the exact location of fpu operands is not known | |
2870 // -> debug information is created inside the fpu stack allocator | |
2871 int n = visitor.info_count(); | |
2872 for (int k = 0; k < n; k++) { | |
2873 compute_debug_info(visitor.info_at(k), op_id); | |
2874 } | |
2875 } | |
2876 } | |
2877 | |
2878 #ifdef ASSERT | |
2879 // make sure we haven't made the op invalid. | |
2880 op->verify(); | |
2881 #endif | |
2882 | |
2883 // remove useless moves | |
2884 if (op->code() == lir_move) { | |
2885 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); | |
2886 LIR_Op1* move = (LIR_Op1*)op; | |
2887 LIR_Opr src = move->in_opr(); | |
2888 LIR_Opr dst = move->result_opr(); | |
2889 if (dst == src || | |
2890 !dst->is_pointer() && !src->is_pointer() && | |
2891 src->is_same_register(dst)) { | |
2892 instructions->at_put(j, NULL); | |
2893 has_dead = true; | |
2894 } | |
2895 } | |
2896 } | |
2897 | |
2898 if (has_dead) { | |
2899 // iterate all instructions of the block and remove all null-values. | |
2900 int insert_point = 0; | |
2901 for (int j = 0; j < num_inst; j++) { | |
2902 LIR_Op* op = instructions->at(j); | |
2903 if (op != NULL) { | |
2904 if (insert_point != j) { | |
2905 instructions->at_put(insert_point, op); | |
2906 } | |
2907 insert_point++; | |
2908 } | |
2909 } | |
2910 instructions->truncate(insert_point); | |
2911 } | |
2912 } | |
2913 | |
2914 void LinearScan::assign_reg_num() { | |
2915 TIME_LINEAR_SCAN(timer_assign_reg_num); | |
2916 | |
2917 init_compute_debug_info(); | |
2918 IntervalWalker* iw = init_compute_oop_maps(); | |
2919 | |
2920 int num_blocks = block_count(); | |
2921 for (int i = 0; i < num_blocks; i++) { | |
2922 BlockBegin* block = block_at(i); | |
2923 assign_reg_num(block->lir()->instructions_list(), iw); | |
2924 } | |
2925 } | |
2926 | |
2927 | |
2928 void LinearScan::do_linear_scan() { | |
2929 NOT_PRODUCT(_total_timer.begin_method()); | |
2930 | |
2931 number_instructions(); | |
2932 | |
2933 NOT_PRODUCT(print_lir(1, "Before Register Allocation")); | |
2934 | |
2935 compute_local_live_sets(); | |
2936 compute_global_live_sets(); | |
2937 CHECK_BAILOUT(); | |
2938 | |
2939 build_intervals(); | |
2940 CHECK_BAILOUT(); | |
2941 sort_intervals_before_allocation(); | |
2942 | |
2943 NOT_PRODUCT(print_intervals("Before Register Allocation")); | |
2944 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc)); | |
2945 | |
2946 allocate_registers(); | |
2947 CHECK_BAILOUT(); | |
2948 | |
2949 resolve_data_flow(); | |
2950 if (compilation()->has_exception_handlers()) { | |
2951 resolve_exception_handlers(); | |
2952 } | |
2953 // fill in number of spill slots into frame_map | |
2954 propagate_spill_slots(); | |
2955 CHECK_BAILOUT(); | |
2956 | |
2957 NOT_PRODUCT(print_intervals("After Register Allocation")); | |
2958 NOT_PRODUCT(print_lir(2, "LIR after register allocation:")); | |
2959 | |
2960 sort_intervals_after_allocation(); | |
722
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
2961 |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
2962 DEBUG_ONLY(verify()); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
2963 |
0 | 2964 eliminate_spill_moves(); |
2965 assign_reg_num(); | |
2966 CHECK_BAILOUT(); | |
2967 | |
2968 NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:")); | |
2969 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign)); | |
2970 | |
2971 { TIME_LINEAR_SCAN(timer_allocate_fpu_stack); | |
2972 | |
2973 if (use_fpu_stack_allocation()) { | |
2974 allocate_fpu_stack(); // Only has effect on Intel | |
2975 NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:")); | |
2976 } | |
2977 } | |
2978 | |
2979 { TIME_LINEAR_SCAN(timer_optimize_lir); | |
2980 | |
2981 EdgeMoveOptimizer::optimize(ir()->code()); | |
2982 ControlFlowOptimizer::optimize(ir()->code()); | |
2983 // check that cfg is still correct after optimizations | |
2984 ir()->verify(); | |
2985 } | |
2986 | |
2987 NOT_PRODUCT(print_lir(1, "Before Code Generation", false)); | |
2988 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final)); | |
2989 NOT_PRODUCT(_total_timer.end_method(this)); | |
2990 } | |
2991 | |
2992 | |
2993 // ********** Printing functions | |
2994 | |
2995 #ifndef PRODUCT | |
2996 | |
2997 void LinearScan::print_timers(double total) { | |
2998 _total_timer.print(total); | |
2999 } | |
3000 | |
3001 void LinearScan::print_statistics() { | |
3002 _stat_before_alloc.print("before allocation"); | |
3003 _stat_after_asign.print("after assignment of register"); | |
3004 _stat_final.print("after optimization"); | |
3005 } | |
3006 | |
3007 void LinearScan::print_bitmap(BitMap& b) { | |
3008 for (unsigned int i = 0; i < b.size(); i++) { | |
3009 if (b.at(i)) tty->print("%d ", i); | |
3010 } | |
3011 tty->cr(); | |
3012 } | |
3013 | |
3014 void LinearScan::print_intervals(const char* label) { | |
3015 if (TraceLinearScanLevel >= 1) { | |
3016 int i; | |
3017 tty->cr(); | |
3018 tty->print_cr("%s", label); | |
3019 | |
3020 for (i = 0; i < interval_count(); i++) { | |
3021 Interval* interval = interval_at(i); | |
3022 if (interval != NULL) { | |
3023 interval->print(); | |
3024 } | |
3025 } | |
3026 | |
3027 tty->cr(); | |
3028 tty->print_cr("--- Basic Blocks ---"); | |
3029 for (i = 0; i < block_count(); i++) { | |
3030 BlockBegin* block = block_at(i); | |
3031 tty->print("B%d [%d, %d, %d, %d] ", block->block_id(), block->first_lir_instruction_id(), block->last_lir_instruction_id(), block->loop_index(), block->loop_depth()); | |
3032 } | |
3033 tty->cr(); | |
3034 tty->cr(); | |
3035 } | |
3036 | |
3037 if (PrintCFGToFile) { | |
3038 CFGPrinter::print_intervals(&_intervals, label); | |
3039 } | |
3040 } | |
3041 | |
3042 void LinearScan::print_lir(int level, const char* label, bool hir_valid) { | |
3043 if (TraceLinearScanLevel >= level) { | |
3044 tty->cr(); | |
3045 tty->print_cr("%s", label); | |
3046 print_LIR(ir()->linear_scan_order()); | |
3047 tty->cr(); | |
3048 } | |
3049 | |
3050 if (level == 1 && PrintCFGToFile) { | |
3051 CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true); | |
3052 } | |
3053 } | |
3054 | |
3055 #endif //PRODUCT | |
3056 | |
3057 | |
3058 // ********** verification functions for allocation | |
3059 // (check that all intervals have a correct register and that no registers are overwritten) | |
3060 #ifdef ASSERT | |
3061 | |
3062 void LinearScan::verify() { | |
3063 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************")); | |
3064 verify_intervals(); | |
3065 | |
3066 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************")); | |
3067 verify_no_oops_in_fixed_intervals(); | |
3068 | |
3069 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries")); | |
3070 verify_constants(); | |
3071 | |
3072 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************")); | |
3073 verify_registers(); | |
3074 | |
3075 TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************")); | |
3076 } | |
3077 | |
3078 void LinearScan::verify_intervals() { | |
3079 int len = interval_count(); | |
3080 bool has_error = false; | |
3081 | |
3082 for (int i = 0; i < len; i++) { | |
3083 Interval* i1 = interval_at(i); | |
3084 if (i1 == NULL) continue; | |
3085 | |
3086 i1->check_split_children(); | |
3087 | |
3088 if (i1->reg_num() != i) { | |
3089 tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr(); | |
3090 has_error = true; | |
3091 } | |
3092 | |
3093 if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) { | |
3094 tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr(); | |
3095 has_error = true; | |
3096 } | |
3097 | |
3098 if (i1->assigned_reg() == any_reg) { | |
3099 tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr(); | |
3100 has_error = true; | |
3101 } | |
3102 | |
3103 if (i1->assigned_reg() == i1->assigned_regHi()) { | |
3104 tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr(); | |
3105 has_error = true; | |
3106 } | |
3107 | |
3108 if (!is_processed_reg_num(i1->assigned_reg())) { | |
3109 tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr(); | |
3110 has_error = true; | |
3111 } | |
3112 | |
3113 if (i1->first() == Range::end()) { | |
3114 tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr(); | |
3115 has_error = true; | |
3116 } | |
3117 | |
3118 for (Range* r = i1->first(); r != Range::end(); r = r->next()) { | |
3119 if (r->from() >= r->to()) { | |
3120 tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr(); | |
3121 has_error = true; | |
3122 } | |
3123 } | |
3124 | |
3125 for (int j = i + 1; j < len; j++) { | |
3126 Interval* i2 = interval_at(j); | |
3127 if (i2 == NULL) continue; | |
3128 | |
3129 // special intervals that are created in MoveResolver | |
3130 // -> ignore them because the range information has no meaning there | |
3131 if (i1->from() == 1 && i1->to() == 2) continue; | |
3132 if (i2->from() == 1 && i2->to() == 2) continue; | |
3133 | |
3134 int r1 = i1->assigned_reg(); | |
3135 int r1Hi = i1->assigned_regHi(); | |
3136 int r2 = i2->assigned_reg(); | |
3137 int r2Hi = i2->assigned_regHi(); | |
3138 if (i1->intersects(i2) && (r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi)))) { | |
3139 tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num()); | |
3140 i1->print(); tty->cr(); | |
3141 i2->print(); tty->cr(); | |
3142 has_error = true; | |
3143 } | |
3144 } | |
3145 } | |
3146 | |
3147 assert(has_error == false, "register allocation invalid"); | |
3148 } | |
3149 | |
3150 | |
3151 void LinearScan::verify_no_oops_in_fixed_intervals() { | |
722
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3152 Interval* fixed_intervals; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3153 Interval* other_intervals; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3154 create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3155 |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3156 // to ensure a walking until the last instruction id, add a dummy interval |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3157 // with a high operation id |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3158 other_intervals = new Interval(any_reg); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3159 other_intervals->add_range(max_jint - 2, max_jint - 1); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3160 IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3161 |
0 | 3162 LIR_OpVisitState visitor; |
3163 for (int i = 0; i < block_count(); i++) { | |
3164 BlockBegin* block = block_at(i); | |
3165 | |
3166 LIR_OpList* instructions = block->lir()->instructions_list(); | |
3167 | |
3168 for (int j = 0; j < instructions->length(); j++) { | |
3169 LIR_Op* op = instructions->at(j); | |
3170 int op_id = op->id(); | |
3171 | |
3172 visitor.visit(op); | |
3173 | |
722
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3174 if (visitor.info_count() > 0) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3175 iw->walk_before(op->id()); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3176 bool check_live = true; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3177 if (op->code() == lir_move) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3178 LIR_Op1* move = (LIR_Op1*)op; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3179 check_live = (move->patch_code() == lir_patch_none); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3180 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3181 LIR_OpBranch* branch = op->as_OpBranch(); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3182 if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3183 // Don't bother checking the stub in this case since the |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3184 // exception stub will never return to normal control flow. |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3185 check_live = false; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3186 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3187 |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3188 // Make sure none of the fixed registers is live across an |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3189 // oopmap since we can't handle that correctly. |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3190 if (check_live) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3191 for (Interval* interval = iw->active_first(fixedKind); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3192 interval != Interval::end(); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3193 interval = interval->next()) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3194 if (interval->current_to() > op->id() + 1) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3195 // This interval is live out of this op so make sure |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3196 // that this interval represents some value that's |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3197 // referenced by this op either as an input or output. |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3198 bool ok = false; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3199 for_each_visitor_mode(mode) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3200 int n = visitor.opr_count(mode); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3201 for (int k = 0; k < n; k++) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3202 LIR_Opr opr = visitor.opr_at(mode, k); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3203 if (opr->is_fixed_cpu()) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3204 if (interval_at(reg_num(opr)) == interval) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3205 ok = true; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3206 break; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3207 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3208 int hi = reg_numHi(opr); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3209 if (hi != -1 && interval_at(hi) == interval) { |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3210 ok = true; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3211 break; |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3212 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3213 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3214 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3215 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3216 assert(ok, "fixed intervals should never be live across an oopmap point"); |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3217 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3218 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3219 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3220 } |
a134d9824964
6828024: verification of fixed interval usage is too weak
never
parents:
337
diff
changeset
|
3221 |
0 | 3222 // oop-maps at calls do not contain registers, so check is not needed |
3223 if (!visitor.has_call()) { | |
3224 | |
3225 for_each_visitor_mode(mode) { | |
3226 int n = visitor.opr_count(mode); | |
3227 for (int k = 0; k < n; k++) { | |
3228 LIR_Opr opr = visitor.opr_at(mode, k); | |
3229 | |
3230 if (opr->is_fixed_cpu() && opr->is_oop()) { | |
3231 // operand is a non-virtual cpu register and contains an oop | |
3232 TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr()); | |
3233 | |
3234 Interval* interval = interval_at(reg_num(opr)); | |
3235 assert(interval != NULL, "no interval"); | |
3236 | |
3237 if (mode == LIR_OpVisitState::inputMode) { | |
3238 if (interval->to() >= op_id + 1) { | |
3239 assert(interval->to() < op_id + 2 || | |
3240 interval->has_hole_between(op_id, op_id + 2), | |
3241 "oop input operand live after instruction"); | |
3242 } | |
3243 } else if (mode == LIR_OpVisitState::outputMode) { | |
3244 if (interval->from() <= op_id - 1) { | |
3245 assert(interval->has_hole_between(op_id - 1, op_id), | |
3246 "oop input operand live after instruction"); | |
3247 } | |
3248 } | |
3249 } | |
3250 } | |
3251 } | |
3252 } | |
3253 } | |
3254 } | |
3255 } | |
3256 | |
3257 | |
3258 void LinearScan::verify_constants() { | |
3259 int num_regs = num_virtual_regs(); | |
3260 int size = live_set_size(); | |
3261 int num_blocks = block_count(); | |
3262 | |
3263 for (int i = 0; i < num_blocks; i++) { | |
3264 BlockBegin* block = block_at(i); | |
3265 BitMap live_at_edge = block->live_in(); | |
3266 | |
3267 // visit all registers where the live_at_edge bit is set | |
304 | 3268 for (int r = (int)live_at_edge.get_next_one_offset(0, size); r < size; r = (int)live_at_edge.get_next_one_offset(r + 1, size)) { |
0 | 3269 TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id())); |
3270 | |
3271 Value value = gen()->instruction_for_vreg(r); | |
3272 | |
3273 assert(value != NULL, "all intervals live across block boundaries must have Value"); | |
3274 assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand"); | |
3275 assert(value->operand()->vreg_number() == r, "register number must match"); | |
3276 // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries"); | |
3277 } | |
3278 } | |
3279 } | |
3280 | |
3281 | |
3282 class RegisterVerifier: public StackObj { | |
3283 private: | |
3284 LinearScan* _allocator; | |
3285 BlockList _work_list; // all blocks that must be processed | |
3286 IntervalsList _saved_states; // saved information of previous check | |
3287 | |
3288 // simplified access to methods of LinearScan | |
3289 Compilation* compilation() const { return _allocator->compilation(); } | |
3290 Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); } | |
3291 int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); } | |
3292 | |
3293 // currently, only registers are processed | |
3294 int state_size() { return LinearScan::nof_regs; } | |
3295 | |
3296 // accessors | |
3297 IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); } | |
3298 void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); } | |
3299 void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); } | |
3300 | |
3301 // helper functions | |
3302 IntervalList* copy(IntervalList* input_state); | |
3303 void state_put(IntervalList* input_state, int reg, Interval* interval); | |
3304 bool check_state(IntervalList* input_state, int reg, Interval* interval); | |
3305 | |
3306 void process_block(BlockBegin* block); | |
3307 void process_xhandler(XHandler* xhandler, IntervalList* input_state); | |
3308 void process_successor(BlockBegin* block, IntervalList* input_state); | |
3309 void process_operations(LIR_List* ops, IntervalList* input_state); | |
3310 | |
3311 public: | |
3312 RegisterVerifier(LinearScan* allocator) | |
3313 : _allocator(allocator) | |
3314 , _work_list(16) | |
3315 , _saved_states(BlockBegin::number_of_blocks(), NULL) | |
3316 { } | |
3317 | |
3318 void verify(BlockBegin* start); | |
3319 }; | |
3320 | |
3321 | |
3322 // entry function from LinearScan that starts the verification | |
3323 void LinearScan::verify_registers() { | |
3324 RegisterVerifier verifier(this); | |
3325 verifier.verify(block_at(0)); | |
3326 } | |
3327 | |
3328 | |
3329 void RegisterVerifier::verify(BlockBegin* start) { | |
3330 // setup input registers (method arguments) for first block | |
3331 IntervalList* input_state = new IntervalList(state_size(), NULL); | |
3332 CallingConvention* args = compilation()->frame_map()->incoming_arguments(); | |
3333 for (int n = 0; n < args->length(); n++) { | |
3334 LIR_Opr opr = args->at(n); | |
3335 if (opr->is_register()) { | |
3336 Interval* interval = interval_at(reg_num(opr)); | |
3337 | |
3338 if (interval->assigned_reg() < state_size()) { | |
3339 input_state->at_put(interval->assigned_reg(), interval); | |
3340 } | |
3341 if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) { | |
3342 input_state->at_put(interval->assigned_regHi(), interval); | |
3343 } | |
3344 } | |
3345 } | |
3346 | |
3347 set_state_for_block(start, input_state); | |
3348 add_to_work_list(start); | |
3349 | |
3350 // main loop for verification | |
3351 do { | |
3352 BlockBegin* block = _work_list.at(0); | |
3353 _work_list.remove_at(0); | |
3354 | |
3355 process_block(block); | |
3356 } while (!_work_list.is_empty()); | |
3357 } | |
3358 | |
3359 void RegisterVerifier::process_block(BlockBegin* block) { | |
3360 TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id())); | |
3361 | |
3362 // must copy state because it is modified | |
3363 IntervalList* input_state = copy(state_for_block(block)); | |
3364 | |
3365 if (TraceLinearScanLevel >= 4) { | |
3366 tty->print_cr("Input-State of intervals:"); | |
3367 tty->print(" "); | |
3368 for (int i = 0; i < state_size(); i++) { | |
3369 if (input_state->at(i) != NULL) { | |
3370 tty->print(" %4d", input_state->at(i)->reg_num()); | |
3371 } else { | |
3372 tty->print(" __"); | |
3373 } | |
3374 } | |
3375 tty->cr(); | |
3376 tty->cr(); | |
3377 } | |
3378 | |
3379 // process all operations of the block | |
3380 process_operations(block->lir(), input_state); | |
3381 | |
3382 // iterate all successors | |
3383 for (int i = 0; i < block->number_of_sux(); i++) { | |
3384 process_successor(block->sux_at(i), input_state); | |
3385 } | |
3386 } | |
3387 | |
3388 void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) { | |
3389 TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id())); | |
3390 | |
3391 // must copy state because it is modified | |
3392 input_state = copy(input_state); | |
3393 | |
3394 if (xhandler->entry_code() != NULL) { | |
3395 process_operations(xhandler->entry_code(), input_state); | |
3396 } | |
3397 process_successor(xhandler->entry_block(), input_state); | |
3398 } | |
3399 | |
3400 void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) { | |
3401 IntervalList* saved_state = state_for_block(block); | |
3402 | |
3403 if (saved_state != NULL) { | |
3404 // this block was already processed before. | |
3405 // check if new input_state is consistent with saved_state | |
3406 | |
3407 bool saved_state_correct = true; | |
3408 for (int i = 0; i < state_size(); i++) { | |
3409 if (input_state->at(i) != saved_state->at(i)) { | |
3410 // current input_state and previous saved_state assume a different | |
3411 // interval in this register -> assume that this register is invalid | |
3412 if (saved_state->at(i) != NULL) { | |
3413 // invalidate old calculation only if it assumed that | |
3414 // register was valid. when the register was already invalid, | |
3415 // then the old calculation was correct. | |
3416 saved_state_correct = false; | |
3417 saved_state->at_put(i, NULL); | |
3418 | |
3419 TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i)); | |
3420 } | |
3421 } | |
3422 } | |
3423 | |
3424 if (saved_state_correct) { | |
3425 // already processed block with correct input_state | |
3426 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id())); | |
3427 } else { | |
3428 // must re-visit this block | |
3429 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id())); | |
3430 add_to_work_list(block); | |
3431 } | |
3432 | |
3433 } else { | |
3434 // block was not processed before, so set initial input_state | |
3435 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id())); | |
3436 | |
3437 set_state_for_block(block, copy(input_state)); | |
3438 add_to_work_list(block); | |
3439 } | |
3440 } | |
3441 | |
3442 | |
3443 IntervalList* RegisterVerifier::copy(IntervalList* input_state) { | |
3444 IntervalList* copy_state = new IntervalList(input_state->length()); | |
3445 copy_state->push_all(input_state); | |
3446 return copy_state; | |
3447 } | |
3448 | |
3449 void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) { | |
3450 if (reg != LinearScan::any_reg && reg < state_size()) { | |
3451 if (interval != NULL) { | |
3452 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d", reg, interval->reg_num())); | |
3453 } else if (input_state->at(reg) != NULL) { | |
3454 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = NULL", reg)); | |
3455 } | |
3456 | |
3457 input_state->at_put(reg, interval); | |
3458 } | |
3459 } | |
3460 | |
3461 bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) { | |
3462 if (reg != LinearScan::any_reg && reg < state_size()) { | |
3463 if (input_state->at(reg) != interval) { | |
3464 tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num()); | |
3465 return true; | |
3466 } | |
3467 } | |
3468 return false; | |
3469 } | |
3470 | |
3471 void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) { | |
3472 // visit all instructions of the block | |
3473 LIR_OpVisitState visitor; | |
3474 bool has_error = false; | |
3475 | |
3476 for (int i = 0; i < ops->length(); i++) { | |
3477 LIR_Op* op = ops->at(i); | |
3478 visitor.visit(op); | |
3479 | |
3480 TRACE_LINEAR_SCAN(4, op->print_on(tty)); | |
3481 | |
3482 // check if input operands are correct | |
3483 int j; | |
3484 int n = visitor.opr_count(LIR_OpVisitState::inputMode); | |
3485 for (j = 0; j < n; j++) { | |
3486 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j); | |
3487 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { | |
3488 Interval* interval = interval_at(reg_num(opr)); | |
3489 if (op->id() != -1) { | |
3490 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode); | |
3491 } | |
3492 | |
3493 has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent()); | |
3494 has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent()); | |
3495 | |
3496 // When an operand is marked with is_last_use, then the fpu stack allocator | |
3497 // removes the register from the fpu stack -> the register contains no value | |
3498 if (opr->is_last_use()) { | |
3499 state_put(input_state, interval->assigned_reg(), NULL); | |
3500 state_put(input_state, interval->assigned_regHi(), NULL); | |
3501 } | |
3502 } | |
3503 } | |
3504 | |
3505 // invalidate all caller save registers at calls | |
3506 if (visitor.has_call()) { | |
3507 for (j = 0; j < FrameMap::nof_caller_save_cpu_regs; j++) { | |
3508 state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL); | |
3509 } | |
3510 for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) { | |
3511 state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL); | |
3512 } | |
3513 | |
304 | 3514 #ifdef X86 |
0 | 3515 for (j = 0; j < FrameMap::nof_caller_save_xmm_regs; j++) { |
3516 state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL); | |
3517 } | |
3518 #endif | |
3519 } | |
3520 | |
3521 // process xhandler before output and temp operands | |
3522 XHandlers* xhandlers = visitor.all_xhandler(); | |
3523 n = xhandlers->length(); | |
3524 for (int k = 0; k < n; k++) { | |
3525 process_xhandler(xhandlers->handler_at(k), input_state); | |
3526 } | |
3527 | |
3528 // set temp operands (some operations use temp operands also as output operands, so can't set them NULL) | |
3529 n = visitor.opr_count(LIR_OpVisitState::tempMode); | |
3530 for (j = 0; j < n; j++) { | |
3531 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j); | |
3532 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { | |
3533 Interval* interval = interval_at(reg_num(opr)); | |
3534 if (op->id() != -1) { | |
3535 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode); | |
3536 } | |
3537 | |
3538 state_put(input_state, interval->assigned_reg(), interval->split_parent()); | |
3539 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); | |
3540 } | |
3541 } | |
3542 | |
3543 // set output operands | |
3544 n = visitor.opr_count(LIR_OpVisitState::outputMode); | |
3545 for (j = 0; j < n; j++) { | |
3546 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j); | |
3547 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { | |
3548 Interval* interval = interval_at(reg_num(opr)); | |
3549 if (op->id() != -1) { | |
3550 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode); | |
3551 } | |
3552 | |
3553 state_put(input_state, interval->assigned_reg(), interval->split_parent()); | |
3554 state_put(input_state, interval->assigned_regHi(), interval->split_parent()); | |
3555 } | |
3556 } | |
3557 } | |
3558 assert(has_error == false, "Error in register allocation"); | |
3559 } | |
3560 | |
3561 #endif // ASSERT | |
3562 | |
3563 | |
3564 | |
3565 // **** Implementation of MoveResolver ****************************** | |
3566 | |
3567 MoveResolver::MoveResolver(LinearScan* allocator) : | |
3568 _allocator(allocator), | |
3569 _multiple_reads_allowed(false), | |
3570 _mapping_from(8), | |
3571 _mapping_from_opr(8), | |
3572 _mapping_to(8), | |
3573 _insert_list(NULL), | |
3574 _insert_idx(-1), | |
3575 _insertion_buffer() | |
3576 { | |
3577 for (int i = 0; i < LinearScan::nof_regs; i++) { | |
3578 _register_blocked[i] = 0; | |
3579 } | |
3580 DEBUG_ONLY(check_empty()); | |
3581 } | |
3582 | |
3583 | |
3584 #ifdef ASSERT | |
3585 | |
3586 void MoveResolver::check_empty() { | |
3587 assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing"); | |
3588 for (int i = 0; i < LinearScan::nof_regs; i++) { | |
3589 assert(register_blocked(i) == 0, "register map must be empty before and after processing"); | |
3590 } | |
3591 assert(_multiple_reads_allowed == false, "must have default value"); | |
3592 } | |
3593 | |
3594 void MoveResolver::verify_before_resolve() { | |
3595 assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal"); | |
3596 assert(_mapping_from.length() == _mapping_to.length(), "length must be equal"); | |
3597 assert(_insert_list != NULL && _insert_idx != -1, "insert position not set"); | |
3598 | |
3599 int i, j; | |
3600 if (!_multiple_reads_allowed) { | |
3601 for (i = 0; i < _mapping_from.length(); i++) { | |
3602 for (j = i + 1; j < _mapping_from.length(); j++) { | |
3603 assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice"); | |
3604 } | |
3605 } | |
3606 } | |
3607 | |
3608 for (i = 0; i < _mapping_to.length(); i++) { | |
3609 for (j = i + 1; j < _mapping_to.length(); j++) { | |
3610 assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice"); | |
3611 } | |
3612 } | |
3613 | |
3614 | |
3615 BitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills()); | |
3616 used_regs.clear(); | |
3617 if (!_multiple_reads_allowed) { | |
3618 for (i = 0; i < _mapping_from.length(); i++) { | |
3619 Interval* it = _mapping_from.at(i); | |
3620 if (it != NULL) { | |
3621 assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice"); | |
3622 used_regs.set_bit(it->assigned_reg()); | |
3623 | |
3624 if (it->assigned_regHi() != LinearScan::any_reg) { | |
3625 assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice"); | |
3626 used_regs.set_bit(it->assigned_regHi()); | |
3627 } | |
3628 } | |
3629 } | |
3630 } | |
3631 | |
3632 used_regs.clear(); | |
3633 for (i = 0; i < _mapping_to.length(); i++) { | |
3634 Interval* it = _mapping_to.at(i); | |
3635 assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice"); | |
3636 used_regs.set_bit(it->assigned_reg()); | |
3637 | |
3638 if (it->assigned_regHi() != LinearScan::any_reg) { | |
3639 assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice"); | |
3640 used_regs.set_bit(it->assigned_regHi()); | |
3641 } | |
3642 } | |
3643 | |
3644 used_regs.clear(); | |
3645 for (i = 0; i < _mapping_from.length(); i++) { | |
3646 Interval* it = _mapping_from.at(i); | |
3647 if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) { | |
3648 used_regs.set_bit(it->assigned_reg()); | |
3649 } | |
3650 } | |
3651 for (i = 0; i < _mapping_to.length(); i++) { | |
3652 Interval* it = _mapping_to.at(i); | |
3653 assert(!used_regs.at(it->assigned_reg()) || it->assigned_reg() == _mapping_from.at(i)->assigned_reg(), "stack slots used in _mapping_from must be disjoint to _mapping_to"); | |
3654 } | |
3655 } | |
3656 | |
3657 #endif // ASSERT | |
3658 | |
3659 | |
3660 // mark assigned_reg and assigned_regHi of the interval as blocked | |
3661 void MoveResolver::block_registers(Interval* it) { | |
3662 int reg = it->assigned_reg(); | |
3663 if (reg < LinearScan::nof_regs) { | |
3664 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); | |
3665 set_register_blocked(reg, 1); | |
3666 } | |
3667 reg = it->assigned_regHi(); | |
3668 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { | |
3669 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used"); | |
3670 set_register_blocked(reg, 1); | |
3671 } | |
3672 } | |
3673 | |
3674 // mark assigned_reg and assigned_regHi of the interval as unblocked | |
3675 void MoveResolver::unblock_registers(Interval* it) { | |
3676 int reg = it->assigned_reg(); | |
3677 if (reg < LinearScan::nof_regs) { | |
3678 assert(register_blocked(reg) > 0, "register already marked as unused"); | |
3679 set_register_blocked(reg, -1); | |
3680 } | |
3681 reg = it->assigned_regHi(); | |
3682 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { | |
3683 assert(register_blocked(reg) > 0, "register already marked as unused"); | |
3684 set_register_blocked(reg, -1); | |
3685 } | |
3686 } | |
3687 | |
3688 // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from) | |
3689 bool MoveResolver::save_to_process_move(Interval* from, Interval* to) { | |
3690 int from_reg = -1; | |
3691 int from_regHi = -1; | |
3692 if (from != NULL) { | |
3693 from_reg = from->assigned_reg(); | |
3694 from_regHi = from->assigned_regHi(); | |
3695 } | |
3696 | |
3697 int reg = to->assigned_reg(); | |
3698 if (reg < LinearScan::nof_regs) { | |
3699 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { | |
3700 return false; | |
3701 } | |
3702 } | |
3703 reg = to->assigned_regHi(); | |
3704 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { | |
3705 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { | |
3706 return false; | |
3707 } | |
3708 } | |
3709 | |
3710 return true; | |
3711 } | |
3712 | |
3713 | |
3714 void MoveResolver::create_insertion_buffer(LIR_List* list) { | |
3715 assert(!_insertion_buffer.initialized(), "overwriting existing buffer"); | |
3716 _insertion_buffer.init(list); | |
3717 } | |
3718 | |
3719 void MoveResolver::append_insertion_buffer() { | |
3720 if (_insertion_buffer.initialized()) { | |
3721 _insertion_buffer.lir_list()->append(&_insertion_buffer); | |
3722 } | |
3723 assert(!_insertion_buffer.initialized(), "must be uninitialized now"); | |
3724 | |
3725 _insert_list = NULL; | |
3726 _insert_idx = -1; | |
3727 } | |
3728 | |
3729 void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) { | |
3730 assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal"); | |
3731 assert(from_interval->type() == to_interval->type(), "move between different types"); | |
3732 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); | |
3733 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); | |
3734 | |
3735 LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type()); | |
3736 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); | |
3737 | |
3738 if (!_multiple_reads_allowed) { | |
3739 // the last_use flag is an optimization for FPU stack allocation. When the same | |
3740 // input interval is used in more than one move, then it is too difficult to determine | |
3741 // if this move is really the last use. | |
3742 from_opr = from_opr->make_last_use(); | |
3743 } | |
3744 _insertion_buffer.move(_insert_idx, from_opr, to_opr); | |
3745 | |
3746 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: inserted move from register %d (%d, %d) to %d (%d, %d)", from_interval->reg_num(), from_interval->assigned_reg(), from_interval->assigned_regHi(), to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); | |
3747 } | |
3748 | |
3749 void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) { | |
3750 assert(from_opr->type() == to_interval->type(), "move between different types"); | |
3751 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first"); | |
3752 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer"); | |
3753 | |
3754 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); | |
3755 _insertion_buffer.move(_insert_idx, from_opr, to_opr); | |
3756 | |
3757 TRACE_LINEAR_SCAN(4, tty->print("MoveResolver: inserted move from constant "); from_opr->print(); tty->print_cr(" to %d (%d, %d)", to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); | |
3758 } | |
3759 | |
3760 | |
3761 void MoveResolver::resolve_mappings() { | |
3762 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: resolving mappings for Block B%d, index %d", _insert_list->block() != NULL ? _insert_list->block()->block_id() : -1, _insert_idx)); | |
3763 DEBUG_ONLY(verify_before_resolve()); | |
3764 | |
3765 // Block all registers that are used as input operands of a move. | |
3766 // When a register is blocked, no move to this register is emitted. | |
3767 // This is necessary for detecting cycles in moves. | |
3768 int i; | |
3769 for (i = _mapping_from.length() - 1; i >= 0; i--) { | |
3770 Interval* from_interval = _mapping_from.at(i); | |
3771 if (from_interval != NULL) { | |
3772 block_registers(from_interval); | |
3773 } | |
3774 } | |
3775 | |
3776 int spill_candidate = -1; | |
3777 while (_mapping_from.length() > 0) { | |
3778 bool processed_interval = false; | |
3779 | |
3780 for (i = _mapping_from.length() - 1; i >= 0; i--) { | |
3781 Interval* from_interval = _mapping_from.at(i); | |
3782 Interval* to_interval = _mapping_to.at(i); | |
3783 | |
3784 if (save_to_process_move(from_interval, to_interval)) { | |
3785 // this inverval can be processed because target is free | |
3786 if (from_interval != NULL) { | |
3787 insert_move(from_interval, to_interval); | |
3788 unblock_registers(from_interval); | |
3789 } else { | |
3790 insert_move(_mapping_from_opr.at(i), to_interval); | |
3791 } | |
3792 _mapping_from.remove_at(i); | |
3793 _mapping_from_opr.remove_at(i); | |
3794 _mapping_to.remove_at(i); | |
3795 | |
3796 processed_interval = true; | |
3797 } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) { | |
3798 // this interval cannot be processed now because target is not free | |
3799 // it starts in a register, so it is a possible candidate for spilling | |
3800 spill_candidate = i; | |
3801 } | |
3802 } | |
3803 | |
3804 if (!processed_interval) { | |
3805 // no move could be processed because there is a cycle in the move list | |
3806 // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory | |
3807 assert(spill_candidate != -1, "no interval in register for spilling found"); | |
3808 | |
3809 // create a new spill interval and assign a stack slot to it | |
3810 Interval* from_interval = _mapping_from.at(spill_candidate); | |
3811 Interval* spill_interval = new Interval(-1); | |
3812 spill_interval->set_type(from_interval->type()); | |
3813 | |
3814 // add a dummy range because real position is difficult to calculate | |
3815 // Note: this range is a special case when the integrity of the allocation is checked | |
3816 spill_interval->add_range(1, 2); | |
3817 | |
3818 // do not allocate a new spill slot for temporary interval, but | |
3819 // use spill slot assigned to from_interval. Otherwise moves from | |
3820 // one stack slot to another can happen (not allowed by LIR_Assembler | |
3821 int spill_slot = from_interval->canonical_spill_slot(); | |
3822 if (spill_slot < 0) { | |
3823 spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2); | |
3824 from_interval->set_canonical_spill_slot(spill_slot); | |
3825 } | |
3826 spill_interval->assign_reg(spill_slot); | |
3827 allocator()->append_interval(spill_interval); | |
3828 | |
3829 TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num())); | |
3830 | |
3831 // insert a move from register to stack and update the mapping | |
3832 insert_move(from_interval, spill_interval); | |
3833 _mapping_from.at_put(spill_candidate, spill_interval); | |
3834 unblock_registers(from_interval); | |
3835 } | |
3836 } | |
3837 | |
3838 // reset to default value | |
3839 _multiple_reads_allowed = false; | |
3840 | |
3841 // check that all intervals have been processed | |
3842 DEBUG_ONLY(check_empty()); | |
3843 } | |
3844 | |
3845 | |
3846 void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) { | |
3847 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: setting insert position to Block B%d, index %d", insert_list->block() != NULL ? insert_list->block()->block_id() : -1, insert_idx)); | |
3848 assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set"); | |
3849 | |
3850 create_insertion_buffer(insert_list); | |
3851 _insert_list = insert_list; | |
3852 _insert_idx = insert_idx; | |
3853 } | |
3854 | |
3855 void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) { | |
3856 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: moving insert position to Block B%d, index %d", insert_list->block() != NULL ? insert_list->block()->block_id() : -1, insert_idx)); | |
3857 | |
3858 if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) { | |
3859 // insert position changed -> resolve current mappings | |
3860 resolve_mappings(); | |
3861 } | |
3862 | |
3863 if (insert_list != _insert_list) { | |
3864 // block changed -> append insertion_buffer because it is | |
3865 // bound to a specific block and create a new insertion_buffer | |
3866 append_insertion_buffer(); | |
3867 create_insertion_buffer(insert_list); | |
3868 } | |
3869 | |
3870 _insert_list = insert_list; | |
3871 _insert_idx = insert_idx; | |
3872 } | |
3873 | |
3874 void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) { | |
3875 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: adding mapping from %d (%d, %d) to %d (%d, %d)", from_interval->reg_num(), from_interval->assigned_reg(), from_interval->assigned_regHi(), to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); | |
3876 | |
3877 _mapping_from.append(from_interval); | |
3878 _mapping_from_opr.append(LIR_OprFact::illegalOpr); | |
3879 _mapping_to.append(to_interval); | |
3880 } | |
3881 | |
3882 | |
3883 void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) { | |
3884 TRACE_LINEAR_SCAN(4, tty->print("MoveResolver: adding mapping from "); from_opr->print(); tty->print_cr(" to %d (%d, %d)", to_interval->reg_num(), to_interval->assigned_reg(), to_interval->assigned_regHi())); | |
3885 assert(from_opr->is_constant(), "only for constants"); | |
3886 | |
3887 _mapping_from.append(NULL); | |
3888 _mapping_from_opr.append(from_opr); | |
3889 _mapping_to.append(to_interval); | |
3890 } | |
3891 | |
3892 void MoveResolver::resolve_and_append_moves() { | |
3893 if (has_mappings()) { | |
3894 resolve_mappings(); | |
3895 } | |
3896 append_insertion_buffer(); | |
3897 } | |
3898 | |
3899 | |
3900 | |
3901 // **** Implementation of Range ************************************* | |
3902 | |
3903 Range::Range(int from, int to, Range* next) : | |
3904 _from(from), | |
3905 _to(to), | |
3906 _next(next) | |
3907 { | |
3908 } | |
3909 | |
3910 // initialize sentinel | |
3911 Range* Range::_end = NULL; | |
3912 void Range::initialize() { | |
3913 _end = new Range(max_jint, max_jint, NULL); | |
3914 } | |
3915 | |
3916 int Range::intersects_at(Range* r2) const { | |
3917 const Range* r1 = this; | |
3918 | |
3919 assert(r1 != NULL && r2 != NULL, "null ranges not allowed"); | |
3920 assert(r1 != _end && r2 != _end, "empty ranges not allowed"); | |
3921 | |
3922 do { | |
3923 if (r1->from() < r2->from()) { | |
3924 if (r1->to() <= r2->from()) { | |
3925 r1 = r1->next(); if (r1 == _end) return -1; | |
3926 } else { | |
3927 return r2->from(); | |
3928 } | |
3929 } else if (r2->from() < r1->from()) { | |
3930 if (r2->to() <= r1->from()) { | |
3931 r2 = r2->next(); if (r2 == _end) return -1; | |
3932 } else { | |
3933 return r1->from(); | |
3934 } | |
3935 } else { // r1->from() == r2->from() | |
3936 if (r1->from() == r1->to()) { | |
3937 r1 = r1->next(); if (r1 == _end) return -1; | |
3938 } else if (r2->from() == r2->to()) { | |
3939 r2 = r2->next(); if (r2 == _end) return -1; | |
3940 } else { | |
3941 return r1->from(); | |
3942 } | |
3943 } | |
3944 } while (true); | |
3945 } | |
3946 | |
3947 #ifndef PRODUCT | |
3948 void Range::print(outputStream* out) const { | |
3949 out->print("[%d, %d[ ", _from, _to); | |
3950 } | |
3951 #endif | |
3952 | |
3953 | |
3954 | |
3955 // **** Implementation of Interval ********************************** | |
3956 | |
3957 // initialize sentinel | |
3958 Interval* Interval::_end = NULL; | |
3959 void Interval::initialize() { | |
3960 Range::initialize(); | |
3961 _end = new Interval(-1); | |
3962 } | |
3963 | |
3964 Interval::Interval(int reg_num) : | |
3965 _reg_num(reg_num), | |
3966 _type(T_ILLEGAL), | |
3967 _first(Range::end()), | |
3968 _use_pos_and_kinds(12), | |
3969 _current(Range::end()), | |
3970 _next(_end), | |
3971 _state(invalidState), | |
3972 _assigned_reg(LinearScan::any_reg), | |
3973 _assigned_regHi(LinearScan::any_reg), | |
3974 _cached_to(-1), | |
3975 _cached_opr(LIR_OprFact::illegalOpr), | |
3976 _cached_vm_reg(VMRegImpl::Bad()), | |
3977 _split_children(0), | |
3978 _canonical_spill_slot(-1), | |
3979 _insert_move_when_activated(false), | |
3980 _register_hint(NULL), | |
3981 _spill_state(noDefinitionFound), | |
3982 _spill_definition_pos(-1) | |
3983 { | |
3984 _split_parent = this; | |
3985 _current_split_child = this; | |
3986 } | |
3987 | |
3988 int Interval::calc_to() { | |
3989 assert(_first != Range::end(), "interval has no range"); | |
3990 | |
3991 Range* r = _first; | |
3992 while (r->next() != Range::end()) { | |
3993 r = r->next(); | |
3994 } | |
3995 return r->to(); | |
3996 } | |
3997 | |
3998 | |
3999 #ifdef ASSERT | |
4000 // consistency check of split-children | |
4001 void Interval::check_split_children() { | |
4002 if (_split_children.length() > 0) { | |
4003 assert(is_split_parent(), "only split parents can have children"); | |
4004 | |
4005 for (int i = 0; i < _split_children.length(); i++) { | |
4006 Interval* i1 = _split_children.at(i); | |
4007 | |
4008 assert(i1->split_parent() == this, "not a split child of this interval"); | |
4009 assert(i1->type() == type(), "must be equal for all split children"); | |
4010 assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children"); | |
4011 | |
4012 for (int j = i + 1; j < _split_children.length(); j++) { | |
4013 Interval* i2 = _split_children.at(j); | |
4014 | |
4015 assert(i1->reg_num() != i2->reg_num(), "same register number"); | |
4016 | |
4017 if (i1->from() < i2->from()) { | |
4018 assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping"); | |
4019 } else { | |
4020 assert(i2->from() < i1->from(), "intervals start at same op_id"); | |
4021 assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping"); | |
4022 } | |
4023 } | |
4024 } | |
4025 } | |
4026 } | |
4027 #endif // ASSERT | |
4028 | |
4029 Interval* Interval::register_hint(bool search_split_child) const { | |
4030 if (!search_split_child) { | |
4031 return _register_hint; | |
4032 } | |
4033 | |
4034 if (_register_hint != NULL) { | |
4035 assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers"); | |
4036 | |
4037 if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) { | |
4038 return _register_hint; | |
4039 | |
4040 } else if (_register_hint->_split_children.length() > 0) { | |
4041 // search the first split child that has a register assigned | |
4042 int len = _register_hint->_split_children.length(); | |
4043 for (int i = 0; i < len; i++) { | |
4044 Interval* cur = _register_hint->_split_children.at(i); | |
4045 | |
4046 if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) { | |
4047 return cur; | |
4048 } | |
4049 } | |
4050 } | |
4051 } | |
4052 | |
4053 // no hint interval found that has a register assigned | |
4054 return NULL; | |
4055 } | |
4056 | |
4057 | |
4058 Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) { | |
4059 assert(is_split_parent(), "can only be called for split parents"); | |
4060 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); | |
4061 | |
4062 Interval* result; | |
4063 if (_split_children.length() == 0) { | |
4064 result = this; | |
4065 } else { | |
4066 result = NULL; | |
4067 int len = _split_children.length(); | |
4068 | |
4069 // in outputMode, the end of the interval (op_id == cur->to()) is not valid | |
4070 int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1); | |
4071 | |
4072 int i; | |
4073 for (i = 0; i < len; i++) { | |
4074 Interval* cur = _split_children.at(i); | |
4075 if (cur->from() <= op_id && op_id < cur->to() + to_offset) { | |
4076 if (i > 0) { | |
4077 // exchange current split child to start of list (faster access for next call) | |
4078 _split_children.at_put(i, _split_children.at(0)); | |
4079 _split_children.at_put(0, cur); | |
4080 } | |
4081 | |
4082 // interval found | |
4083 result = cur; | |
4084 break; | |
4085 } | |
4086 } | |
4087 | |
4088 #ifdef ASSERT | |
4089 for (i = 0; i < len; i++) { | |
4090 Interval* tmp = _split_children.at(i); | |
4091 if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) { | |
4092 tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num()); | |
4093 result->print(); | |
4094 tmp->print(); | |
4095 assert(false, "two valid result intervals found"); | |
4096 } | |
4097 } | |
4098 #endif | |
4099 } | |
4100 | |
4101 assert(result != NULL, "no matching interval found"); | |
4102 assert(result->covers(op_id, mode), "op_id not covered by interval"); | |
4103 | |
4104 return result; | |
4105 } | |
4106 | |
4107 | |
4108 // returns the last split child that ends before the given op_id | |
4109 Interval* Interval::split_child_before_op_id(int op_id) { | |
4110 assert(op_id >= 0, "invalid op_id"); | |
4111 | |
4112 Interval* parent = split_parent(); | |
4113 Interval* result = NULL; | |
4114 | |
4115 int len = parent->_split_children.length(); | |
4116 assert(len > 0, "no split children available"); | |
4117 | |
4118 for (int i = len - 1; i >= 0; i--) { | |
4119 Interval* cur = parent->_split_children.at(i); | |
4120 if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) { | |
4121 result = cur; | |
4122 } | |
4123 } | |
4124 | |
4125 assert(result != NULL, "no split child found"); | |
4126 return result; | |
4127 } | |
4128 | |
4129 | |
4130 // checks if op_id is covered by any split child | |
4131 bool Interval::split_child_covers(int op_id, LIR_OpVisitState::OprMode mode) { | |
4132 assert(is_split_parent(), "can only be called for split parents"); | |
4133 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)"); | |
4134 | |
4135 if (_split_children.length() == 0) { | |
4136 // simple case if interval was not split | |
4137 return covers(op_id, mode); | |
4138 | |
4139 } else { | |
4140 // extended case: check all split children | |
4141 int len = _split_children.length(); | |
4142 for (int i = 0; i < len; i++) { | |
4143 Interval* cur = _split_children.at(i); | |
4144 if (cur->covers(op_id, mode)) { | |
4145 return true; | |
4146 } | |
4147 } | |
4148 return false; | |
4149 } | |
4150 } | |
4151 | |
4152 | |
4153 // Note: use positions are sorted descending -> first use has highest index | |
4154 int Interval::first_usage(IntervalUseKind min_use_kind) const { | |
4155 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); | |
4156 | |
4157 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { | |
4158 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { | |
4159 return _use_pos_and_kinds.at(i); | |
4160 } | |
4161 } | |
4162 return max_jint; | |
4163 } | |
4164 | |
4165 int Interval::next_usage(IntervalUseKind min_use_kind, int from) const { | |
4166 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); | |
4167 | |
4168 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { | |
4169 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) { | |
4170 return _use_pos_and_kinds.at(i); | |
4171 } | |
4172 } | |
4173 return max_jint; | |
4174 } | |
4175 | |
4176 int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const { | |
4177 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); | |
4178 | |
4179 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { | |
4180 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) { | |
4181 return _use_pos_and_kinds.at(i); | |
4182 } | |
4183 } | |
4184 return max_jint; | |
4185 } | |
4186 | |
4187 int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const { | |
4188 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals"); | |
4189 | |
4190 int prev = 0; | |
4191 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { | |
4192 if (_use_pos_and_kinds.at(i) > from) { | |
4193 return prev; | |
4194 } | |
4195 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { | |
4196 prev = _use_pos_and_kinds.at(i); | |
4197 } | |
4198 } | |
4199 return prev; | |
4200 } | |
4201 | |
4202 void Interval::add_use_pos(int pos, IntervalUseKind use_kind) { | |
4203 assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range"); | |
4204 | |
4205 // do not add use positions for precolored intervals because | |
4206 // they are never used | |
4207 if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) { | |
4208 #ifdef ASSERT | |
4209 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); | |
4210 for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) { | |
4211 assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position"); | |
4212 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); | |
4213 if (i > 0) { | |
4214 assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending"); | |
4215 } | |
4216 } | |
4217 #endif | |
4218 | |
4219 // Note: add_use is called in descending order, so list gets sorted | |
4220 // automatically by just appending new use positions | |
4221 int len = _use_pos_and_kinds.length(); | |
4222 if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) { | |
4223 _use_pos_and_kinds.append(pos); | |
4224 _use_pos_and_kinds.append(use_kind); | |
4225 } else if (_use_pos_and_kinds.at(len - 1) < use_kind) { | |
4226 assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly"); | |
4227 _use_pos_and_kinds.at_put(len - 1, use_kind); | |
4228 } | |
4229 } | |
4230 } | |
4231 | |
4232 void Interval::add_range(int from, int to) { | |
4233 assert(from < to, "invalid range"); | |
4234 assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval"); | |
4235 assert(from <= first()->to(), "not inserting at begin of interval"); | |
4236 | |
4237 if (first()->from() <= to) { | |
4238 // join intersecting ranges | |
4239 first()->set_from(MIN2(from, first()->from())); | |
4240 first()->set_to (MAX2(to, first()->to())); | |
4241 } else { | |
4242 // insert new range | |
4243 _first = new Range(from, to, first()); | |
4244 } | |
4245 } | |
4246 | |
4247 Interval* Interval::new_split_child() { | |
4248 // allocate new interval | |
4249 Interval* result = new Interval(-1); | |
4250 result->set_type(type()); | |
4251 | |
4252 Interval* parent = split_parent(); | |
4253 result->_split_parent = parent; | |
4254 result->set_register_hint(parent); | |
4255 | |
4256 // insert new interval in children-list of parent | |
4257 if (parent->_split_children.length() == 0) { | |
4258 assert(is_split_parent(), "list must be initialized at first split"); | |
4259 | |
4260 parent->_split_children = IntervalList(4); | |
4261 parent->_split_children.append(this); | |
4262 } | |
4263 parent->_split_children.append(result); | |
4264 | |
4265 return result; | |
4266 } | |
4267 | |
4268 // split this interval at the specified position and return | |
4269 // the remainder as a new interval. | |
4270 // | |
4271 // when an interval is split, a bi-directional link is established between the original interval | |
4272 // (the split parent) and the intervals that are split off this interval (the split children) | |
4273 // When a split child is split again, the new created interval is also a direct child | |
4274 // of the original parent (there is no tree of split children stored, but a flat list) | |
4275 // All split children are spilled to the same stack slot (stored in _canonical_spill_slot) | |
4276 // | |
4277 // Note: The new interval has no valid reg_num | |
4278 Interval* Interval::split(int split_pos) { | |
4279 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); | |
4280 | |
4281 // allocate new interval | |
4282 Interval* result = new_split_child(); | |
4283 | |
4284 // split the ranges | |
4285 Range* prev = NULL; | |
4286 Range* cur = _first; | |
4287 while (cur != Range::end() && cur->to() <= split_pos) { | |
4288 prev = cur; | |
4289 cur = cur->next(); | |
4290 } | |
4291 assert(cur != Range::end(), "split interval after end of last range"); | |
4292 | |
4293 if (cur->from() < split_pos) { | |
4294 result->_first = new Range(split_pos, cur->to(), cur->next()); | |
4295 cur->set_to(split_pos); | |
4296 cur->set_next(Range::end()); | |
4297 | |
4298 } else { | |
4299 assert(prev != NULL, "split before start of first range"); | |
4300 result->_first = cur; | |
4301 prev->set_next(Range::end()); | |
4302 } | |
4303 result->_current = result->_first; | |
4304 _cached_to = -1; // clear cached value | |
4305 | |
4306 // split list of use positions | |
4307 int total_len = _use_pos_and_kinds.length(); | |
4308 int start_idx = total_len - 2; | |
4309 while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) { | |
4310 start_idx -= 2; | |
4311 } | |
4312 | |
4313 intStack new_use_pos_and_kinds(total_len - start_idx); | |
4314 int i; | |
4315 for (i = start_idx + 2; i < total_len; i++) { | |
4316 new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i)); | |
4317 } | |
4318 | |
4319 _use_pos_and_kinds.truncate(start_idx + 2); | |
4320 result->_use_pos_and_kinds = _use_pos_and_kinds; | |
4321 _use_pos_and_kinds = new_use_pos_and_kinds; | |
4322 | |
4323 #ifdef ASSERT | |
4324 assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); | |
4325 assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos"); | |
4326 assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries"); | |
4327 | |
4328 for (i = 0; i < _use_pos_and_kinds.length(); i += 2) { | |
4329 assert(_use_pos_and_kinds.at(i) < split_pos, "must be"); | |
4330 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); | |
4331 } | |
4332 for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) { | |
4333 assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be"); | |
4334 assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); | |
4335 } | |
4336 #endif | |
4337 | |
4338 return result; | |
4339 } | |
4340 | |
4341 // split this interval at the specified position and return | |
4342 // the head as a new interval (the original interval is the tail) | |
4343 // | |
4344 // Currently, only the first range can be split, and the new interval | |
4345 // must not have split positions | |
4346 Interval* Interval::split_from_start(int split_pos) { | |
4347 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals"); | |
4348 assert(split_pos > from() && split_pos < to(), "can only split inside interval"); | |
4349 assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range"); | |
4350 assert(first_usage(noUse) > split_pos, "can not split when use positions are present"); | |
4351 | |
4352 // allocate new interval | |
4353 Interval* result = new_split_child(); | |
4354 | |
4355 // the new created interval has only one range (checked by assertion above), | |
4356 // so the splitting of the ranges is very simple | |
4357 result->add_range(_first->from(), split_pos); | |
4358 | |
4359 if (split_pos == _first->to()) { | |
4360 assert(_first->next() != Range::end(), "must not be at end"); | |
4361 _first = _first->next(); | |
4362 } else { | |
4363 _first->set_from(split_pos); | |
4364 } | |
4365 | |
4366 return result; | |
4367 } | |
4368 | |
4369 | |
4370 // returns true if the op_id is inside the interval | |
4371 bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const { | |
4372 Range* cur = _first; | |
4373 | |
4374 while (cur != Range::end() && cur->to() < op_id) { | |
4375 cur = cur->next(); | |
4376 } | |
4377 if (cur != Range::end()) { | |
4378 assert(cur->to() != cur->next()->from(), "ranges not separated"); | |
4379 | |
4380 if (mode == LIR_OpVisitState::outputMode) { | |
4381 return cur->from() <= op_id && op_id < cur->to(); | |
4382 } else { | |
4383 return cur->from() <= op_id && op_id <= cur->to(); | |
4384 } | |
4385 } | |
4386 return false; | |
4387 } | |
4388 | |
4389 // returns true if the interval has any hole between hole_from and hole_to | |
4390 // (even if the hole has only the length 1) | |
4391 bool Interval::has_hole_between(int hole_from, int hole_to) { | |
4392 assert(hole_from < hole_to, "check"); | |
4393 assert(from() <= hole_from && hole_to <= to(), "index out of interval"); | |
4394 | |
4395 Range* cur = _first; | |
4396 while (cur != Range::end()) { | |
4397 assert(cur->to() < cur->next()->from(), "no space between ranges"); | |
4398 | |
4399 // hole-range starts before this range -> hole | |
4400 if (hole_from < cur->from()) { | |
4401 return true; | |
4402 | |
4403 // hole-range completely inside this range -> no hole | |
4404 } else if (hole_to <= cur->to()) { | |
4405 return false; | |
4406 | |
4407 // overlapping of hole-range with this range -> hole | |
4408 } else if (hole_from <= cur->to()) { | |
4409 return true; | |
4410 } | |
4411 | |
4412 cur = cur->next(); | |
4413 } | |
4414 | |
4415 return false; | |
4416 } | |
4417 | |
4418 | |
4419 #ifndef PRODUCT | |
4420 void Interval::print(outputStream* out) const { | |
4421 const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" }; | |
4422 const char* UseKind2Name[] = { "N", "L", "S", "M" }; | |
4423 | |
4424 const char* type_name; | |
4425 LIR_Opr opr = LIR_OprFact::illegal(); | |
4426 if (reg_num() < LIR_OprDesc::vreg_base) { | |
4427 type_name = "fixed"; | |
4428 // need a temporary operand for fixed intervals because type() cannot be called | |
4429 if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) { | |
4430 opr = LIR_OprFact::single_cpu(assigned_reg()); | |
4431 } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) { | |
4432 opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg); | |
304 | 4433 #ifdef X86 |
0 | 4434 } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= pd_last_xmm_reg) { |
4435 opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg); | |
4436 #endif | |
4437 } else { | |
4438 ShouldNotReachHere(); | |
4439 } | |
4440 } else { | |
4441 type_name = type2name(type()); | |
4442 if (assigned_reg() != -1) { | |
4443 opr = LinearScan::calc_operand_for_interval(this); | |
4444 } | |
4445 } | |
4446 | |
4447 out->print("%d %s ", reg_num(), type_name); | |
4448 if (opr->is_valid()) { | |
4449 out->print("\""); | |
4450 opr->print(out); | |
4451 out->print("\" "); | |
4452 } | |
4453 out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1)); | |
4454 | |
4455 // print ranges | |
4456 Range* cur = _first; | |
4457 while (cur != Range::end()) { | |
4458 cur->print(out); | |
4459 cur = cur->next(); | |
4460 assert(cur != NULL, "range list not closed with range sentinel"); | |
4461 } | |
4462 | |
4463 // print use positions | |
4464 int prev = 0; | |
4465 assert(_use_pos_and_kinds.length() % 2 == 0, "must be"); | |
4466 for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { | |
4467 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind"); | |
4468 assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted"); | |
4469 | |
4470 out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]); | |
4471 prev = _use_pos_and_kinds.at(i); | |
4472 } | |
4473 | |
4474 out->print(" \"%s\"", SpillState2Name[spill_state()]); | |
4475 out->cr(); | |
4476 } | |
4477 #endif | |
4478 | |
4479 | |
4480 | |
4481 // **** Implementation of IntervalWalker **************************** | |
4482 | |
4483 IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) | |
4484 : _compilation(allocator->compilation()) | |
4485 , _allocator(allocator) | |
4486 { | |
4487 _unhandled_first[fixedKind] = unhandled_fixed_first; | |
4488 _unhandled_first[anyKind] = unhandled_any_first; | |
4489 _active_first[fixedKind] = Interval::end(); | |
4490 _inactive_first[fixedKind] = Interval::end(); | |
4491 _active_first[anyKind] = Interval::end(); | |
4492 _inactive_first[anyKind] = Interval::end(); | |
4493 _current_position = -1; | |
4494 _current = NULL; | |
4495 next_interval(); | |
4496 } | |
4497 | |
4498 | |
4499 // append interval at top of list | |
4500 void IntervalWalker::append_unsorted(Interval** list, Interval* interval) { | |
4501 interval->set_next(*list); *list = interval; | |
4502 } | |
4503 | |
4504 | |
4505 // append interval in order of current range from() | |
4506 void IntervalWalker::append_sorted(Interval** list, Interval* interval) { | |
4507 Interval* prev = NULL; | |
4508 Interval* cur = *list; | |
4509 while (cur->current_from() < interval->current_from()) { | |
4510 prev = cur; cur = cur->next(); | |
4511 } | |
4512 if (prev == NULL) { | |
4513 *list = interval; | |
4514 } else { | |
4515 prev->set_next(interval); | |
4516 } | |
4517 interval->set_next(cur); | |
4518 } | |
4519 | |
4520 void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) { | |
4521 assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position"); | |
4522 | |
4523 Interval* prev = NULL; | |
4524 Interval* cur = *list; | |
4525 while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) { | |
4526 prev = cur; cur = cur->next(); | |
4527 } | |
4528 if (prev == NULL) { | |
4529 *list = interval; | |
4530 } else { | |
4531 prev->set_next(interval); | |
4532 } | |
4533 interval->set_next(cur); | |
4534 } | |
4535 | |
4536 | |
4537 inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) { | |
4538 while (*list != Interval::end() && *list != i) { | |
4539 list = (*list)->next_addr(); | |
4540 } | |
4541 if (*list != Interval::end()) { | |
4542 assert(*list == i, "check"); | |
4543 *list = (*list)->next(); | |
4544 return true; | |
4545 } else { | |
4546 return false; | |
4547 } | |
4548 } | |
4549 | |
4550 void IntervalWalker::remove_from_list(Interval* i) { | |
4551 bool deleted; | |
4552 | |
4553 if (i->state() == activeState) { | |
4554 deleted = remove_from_list(active_first_addr(anyKind), i); | |
4555 } else { | |
4556 assert(i->state() == inactiveState, "invalid state"); | |
4557 deleted = remove_from_list(inactive_first_addr(anyKind), i); | |
4558 } | |
4559 | |
4560 assert(deleted, "interval has not been found in list"); | |
4561 } | |
4562 | |
4563 | |
4564 void IntervalWalker::walk_to(IntervalState state, int from) { | |
4565 assert (state == activeState || state == inactiveState, "wrong state"); | |
4566 for_each_interval_kind(kind) { | |
4567 Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind); | |
4568 Interval* next = *prev; | |
4569 while (next->current_from() <= from) { | |
4570 Interval* cur = next; | |
4571 next = cur->next(); | |
4572 | |
4573 bool range_has_changed = false; | |
4574 while (cur->current_to() <= from) { | |
4575 cur->next_range(); | |
4576 range_has_changed = true; | |
4577 } | |
4578 | |
4579 // also handle move from inactive list to active list | |
4580 range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from); | |
4581 | |
4582 if (range_has_changed) { | |
4583 // remove cur from list | |
4584 *prev = next; | |
4585 if (cur->current_at_end()) { | |
4586 // move to handled state (not maintained as a list) | |
4587 cur->set_state(handledState); | |
4588 interval_moved(cur, kind, state, handledState); | |
4589 } else if (cur->current_from() <= from){ | |
4590 // sort into active list | |
4591 append_sorted(active_first_addr(kind), cur); | |
4592 cur->set_state(activeState); | |
4593 if (*prev == cur) { | |
4594 assert(state == activeState, "check"); | |
4595 prev = cur->next_addr(); | |
4596 } | |
4597 interval_moved(cur, kind, state, activeState); | |
4598 } else { | |
4599 // sort into inactive list | |
4600 append_sorted(inactive_first_addr(kind), cur); | |
4601 cur->set_state(inactiveState); | |
4602 if (*prev == cur) { | |
4603 assert(state == inactiveState, "check"); | |
4604 prev = cur->next_addr(); | |
4605 } | |
4606 interval_moved(cur, kind, state, inactiveState); | |
4607 } | |
4608 } else { | |
4609 prev = cur->next_addr(); | |
4610 continue; | |
4611 } | |
4612 } | |
4613 } | |
4614 } | |
4615 | |
4616 | |
4617 void IntervalWalker::next_interval() { | |
4618 IntervalKind kind; | |
4619 Interval* any = _unhandled_first[anyKind]; | |
4620 Interval* fixed = _unhandled_first[fixedKind]; | |
4621 | |
4622 if (any != Interval::end()) { | |
4623 // intervals may start at same position -> prefer fixed interval | |
4624 kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind; | |
4625 | |
4626 assert (kind == fixedKind && fixed->from() <= any->from() || | |
4627 kind == anyKind && any->from() <= fixed->from(), "wrong interval!!!"); | |
4628 assert(any == Interval::end() || fixed == Interval::end() || any->from() != fixed->from() || kind == fixedKind, "if fixed and any-Interval start at same position, fixed must be processed first"); | |
4629 | |
4630 } else if (fixed != Interval::end()) { | |
4631 kind = fixedKind; | |
4632 } else { | |
4633 _current = NULL; return; | |
4634 } | |
4635 _current_kind = kind; | |
4636 _current = _unhandled_first[kind]; | |
4637 _unhandled_first[kind] = _current->next(); | |
4638 _current->set_next(Interval::end()); | |
4639 _current->rewind_range(); | |
4640 } | |
4641 | |
4642 | |
4643 void IntervalWalker::walk_to(int lir_op_id) { | |
4644 assert(_current_position <= lir_op_id, "can not walk backwards"); | |
4645 while (current() != NULL) { | |
4646 bool is_active = current()->from() <= lir_op_id; | |
4647 int id = is_active ? current()->from() : lir_op_id; | |
4648 | |
4649 TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); }) | |
4650 | |
4651 // set _current_position prior to call of walk_to | |
4652 _current_position = id; | |
4653 | |
4654 // call walk_to even if _current_position == id | |
4655 walk_to(activeState, id); | |
4656 walk_to(inactiveState, id); | |
4657 | |
4658 if (is_active) { | |
4659 current()->set_state(activeState); | |
4660 if (activate_current()) { | |
4661 append_sorted(active_first_addr(current_kind()), current()); | |
4662 interval_moved(current(), current_kind(), unhandledState, activeState); | |
4663 } | |
4664 | |
4665 next_interval(); | |
4666 } else { | |
4667 return; | |
4668 } | |
4669 } | |
4670 } | |
4671 | |
4672 void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) { | |
4673 #ifndef PRODUCT | |
4674 if (TraceLinearScanLevel >= 4) { | |
4675 #define print_state(state) \ | |
4676 switch(state) {\ | |
4677 case unhandledState: tty->print("unhandled"); break;\ | |
4678 case activeState: tty->print("active"); break;\ | |
4679 case inactiveState: tty->print("inactive"); break;\ | |
4680 case handledState: tty->print("handled"); break;\ | |
4681 default: ShouldNotReachHere(); \ | |
4682 } | |
4683 | |
4684 print_state(from); tty->print(" to "); print_state(to); | |
4685 tty->fill_to(23); | |
4686 interval->print(); | |
4687 | |
4688 #undef print_state | |
4689 } | |
4690 #endif | |
4691 } | |
4692 | |
4693 | |
4694 | |
4695 // **** Implementation of LinearScanWalker ************************** | |
4696 | |
4697 LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) | |
4698 : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first) | |
4699 , _move_resolver(allocator) | |
4700 { | |
4701 for (int i = 0; i < LinearScan::nof_regs; i++) { | |
4702 _spill_intervals[i] = new IntervalList(2); | |
4703 } | |
4704 } | |
4705 | |
4706 | |
4707 inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) { | |
4708 for (int i = _first_reg; i <= _last_reg; i++) { | |
4709 _use_pos[i] = max_jint; | |
4710 | |
4711 if (!only_process_use_pos) { | |
4712 _block_pos[i] = max_jint; | |
4713 _spill_intervals[i]->clear(); | |
4714 } | |
4715 } | |
4716 } | |
4717 | |
4718 inline void LinearScanWalker::exclude_from_use(int reg) { | |
4719 assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)"); | |
4720 if (reg >= _first_reg && reg <= _last_reg) { | |
4721 _use_pos[reg] = 0; | |
4722 } | |
4723 } | |
4724 inline void LinearScanWalker::exclude_from_use(Interval* i) { | |
4725 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); | |
4726 | |
4727 exclude_from_use(i->assigned_reg()); | |
4728 exclude_from_use(i->assigned_regHi()); | |
4729 } | |
4730 | |
4731 inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) { | |
4732 assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0"); | |
4733 | |
4734 if (reg >= _first_reg && reg <= _last_reg) { | |
4735 if (_use_pos[reg] > use_pos) { | |
4736 _use_pos[reg] = use_pos; | |
4737 } | |
4738 if (!only_process_use_pos) { | |
4739 _spill_intervals[reg]->append(i); | |
4740 } | |
4741 } | |
4742 } | |
4743 inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) { | |
4744 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); | |
4745 if (use_pos != -1) { | |
4746 set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos); | |
4747 set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos); | |
4748 } | |
4749 } | |
4750 | |
4751 inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) { | |
4752 if (reg >= _first_reg && reg <= _last_reg) { | |
4753 if (_block_pos[reg] > block_pos) { | |
4754 _block_pos[reg] = block_pos; | |
4755 } | |
4756 if (_use_pos[reg] > block_pos) { | |
4757 _use_pos[reg] = block_pos; | |
4758 } | |
4759 } | |
4760 } | |
4761 inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) { | |
4762 assert(i->assigned_reg() != any_reg, "interval has no register assigned"); | |
4763 if (block_pos != -1) { | |
4764 set_block_pos(i->assigned_reg(), i, block_pos); | |
4765 set_block_pos(i->assigned_regHi(), i, block_pos); | |
4766 } | |
4767 } | |
4768 | |
4769 | |
4770 void LinearScanWalker::free_exclude_active_fixed() { | |
4771 Interval* list = active_first(fixedKind); | |
4772 while (list != Interval::end()) { | |
4773 assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned"); | |
4774 exclude_from_use(list); | |
4775 list = list->next(); | |
4776 } | |
4777 } | |
4778 | |
4779 void LinearScanWalker::free_exclude_active_any() { | |
4780 Interval* list = active_first(anyKind); | |
4781 while (list != Interval::end()) { | |
4782 exclude_from_use(list); | |
4783 list = list->next(); | |
4784 } | |
4785 } | |
4786 | |
4787 void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) { | |
4788 Interval* list = inactive_first(fixedKind); | |
4789 while (list != Interval::end()) { | |
4790 if (cur->to() <= list->current_from()) { | |
4791 assert(list->current_intersects_at(cur) == -1, "must not intersect"); | |
4792 set_use_pos(list, list->current_from(), true); | |
4793 } else { | |
4794 set_use_pos(list, list->current_intersects_at(cur), true); | |
4795 } | |
4796 list = list->next(); | |
4797 } | |
4798 } | |
4799 | |
4800 void LinearScanWalker::free_collect_inactive_any(Interval* cur) { | |
4801 Interval* list = inactive_first(anyKind); | |
4802 while (list != Interval::end()) { | |
4803 set_use_pos(list, list->current_intersects_at(cur), true); | |
4804 list = list->next(); | |
4805 } | |
4806 } | |
4807 | |
4808 void LinearScanWalker::free_collect_unhandled(IntervalKind kind, Interval* cur) { | |
4809 Interval* list = unhandled_first(kind); | |
4810 while (list != Interval::end()) { | |
4811 set_use_pos(list, list->intersects_at(cur), true); | |
4812 if (kind == fixedKind && cur->to() <= list->from()) { | |
4813 set_use_pos(list, list->from(), true); | |
4814 } | |
4815 list = list->next(); | |
4816 } | |
4817 } | |
4818 | |
4819 void LinearScanWalker::spill_exclude_active_fixed() { | |
4820 Interval* list = active_first(fixedKind); | |
4821 while (list != Interval::end()) { | |
4822 exclude_from_use(list); | |
4823 list = list->next(); | |
4824 } | |
4825 } | |
4826 | |
4827 void LinearScanWalker::spill_block_unhandled_fixed(Interval* cur) { | |
4828 Interval* list = unhandled_first(fixedKind); | |
4829 while (list != Interval::end()) { | |
4830 set_block_pos(list, list->intersects_at(cur)); | |
4831 list = list->next(); | |
4832 } | |
4833 } | |
4834 | |
4835 void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) { | |
4836 Interval* list = inactive_first(fixedKind); | |
4837 while (list != Interval::end()) { | |
4838 if (cur->to() > list->current_from()) { | |
4839 set_block_pos(list, list->current_intersects_at(cur)); | |
4840 } else { | |
4841 assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect"); | |
4842 } | |
4843 | |
4844 list = list->next(); | |
4845 } | |
4846 } | |
4847 | |
4848 void LinearScanWalker::spill_collect_active_any() { | |
4849 Interval* list = active_first(anyKind); | |
4850 while (list != Interval::end()) { | |
4851 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); | |
4852 list = list->next(); | |
4853 } | |
4854 } | |
4855 | |
4856 void LinearScanWalker::spill_collect_inactive_any(Interval* cur) { | |
4857 Interval* list = inactive_first(anyKind); | |
4858 while (list != Interval::end()) { | |
4859 if (list->current_intersects(cur)) { | |
4860 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); | |
4861 } | |
4862 list = list->next(); | |
4863 } | |
4864 } | |
4865 | |
4866 | |
4867 void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) { | |
4868 // output all moves here. When source and target are equal, the move is | |
4869 // optimized away later in assign_reg_nums | |
4870 | |
4871 op_id = (op_id + 1) & ~1; | |
4872 BlockBegin* op_block = allocator()->block_of_op_with_id(op_id); | |
4873 assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary"); | |
4874 | |
4875 // calculate index of instruction inside instruction list of current block | |
4876 // the minimal index (for a block with no spill moves) can be calculated because the | |
4877 // numbering of instructions is known. | |
4878 // When the block already contains spill moves, the index must be increased until the | |
4879 // correct index is reached. | |
4880 LIR_OpList* list = op_block->lir()->instructions_list(); | |
4881 int index = (op_id - list->at(0)->id()) / 2; | |
4882 assert(list->at(index)->id() <= op_id, "error in calculation"); | |
4883 | |
4884 while (list->at(index)->id() != op_id) { | |
4885 index++; | |
4886 assert(0 <= index && index < list->length(), "index out of bounds"); | |
4887 } | |
4888 assert(1 <= index && index < list->length(), "index out of bounds"); | |
4889 assert(list->at(index)->id() == op_id, "error in calculation"); | |
4890 | |
4891 // insert new instruction before instruction at position index | |
4892 _move_resolver.move_insert_position(op_block->lir(), index - 1); | |
4893 _move_resolver.add_mapping(src_it, dst_it); | |
4894 } | |
4895 | |
4896 | |
4897 int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) { | |
4898 int from_block_nr = min_block->linear_scan_number(); | |
4899 int to_block_nr = max_block->linear_scan_number(); | |
4900 | |
4901 assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range"); | |
4902 assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range"); | |
4903 assert(from_block_nr < to_block_nr, "must cross block boundary"); | |
4904 | |
4905 // Try to split at end of max_block. If this would be after | |
4906 // max_split_pos, then use the begin of max_block | |
4907 int optimal_split_pos = max_block->last_lir_instruction_id() + 2; | |
4908 if (optimal_split_pos > max_split_pos) { | |
4909 optimal_split_pos = max_block->first_lir_instruction_id(); | |
4910 } | |
4911 | |
4912 int min_loop_depth = max_block->loop_depth(); | |
4913 for (int i = to_block_nr - 1; i >= from_block_nr; i--) { | |
4914 BlockBegin* cur = block_at(i); | |
4915 | |
4916 if (cur->loop_depth() < min_loop_depth) { | |
4917 // block with lower loop-depth found -> split at the end of this block | |
4918 min_loop_depth = cur->loop_depth(); | |
4919 optimal_split_pos = cur->last_lir_instruction_id() + 2; | |
4920 } | |
4921 } | |
4922 assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary"); | |
4923 | |
4924 return optimal_split_pos; | |
4925 } | |
4926 | |
4927 | |
4928 int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) { | |
4929 int optimal_split_pos = -1; | |
4930 if (min_split_pos == max_split_pos) { | |
4931 // trivial case, no optimization of split position possible | |
4932 TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible")); | |
4933 optimal_split_pos = min_split_pos; | |
4934 | |
4935 } else { | |
4936 assert(min_split_pos < max_split_pos, "must be true then"); | |
4937 assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise"); | |
4938 | |
4939 // reason for using min_split_pos - 1: when the minimal split pos is exactly at the | |
4940 // beginning of a block, then min_split_pos is also a possible split position. | |
4941 // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos | |
4942 BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1); | |
4943 | |
4944 // reason for using max_split_pos - 1: otherwise there would be an assertion failure | |
4945 // when an interval ends at the end of the last block of the method | |
4946 // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no | |
4947 // block at this op_id) | |
4948 BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1); | |
4949 | |
4950 assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order"); | |
4951 if (min_block == max_block) { | |
4952 // split position cannot be moved to block boundary, so split as late as possible | |
4953 TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block")); | |
4954 optimal_split_pos = max_split_pos; | |
4955 | |
4956 } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) { | |
4957 // Do not move split position if the interval has a hole before max_split_pos. | |
4958 // Intervals resulting from Phi-Functions have more than one definition (marked | |
4959 // as mustHaveRegister) with a hole before each definition. When the register is needed | |
4960 // for the second definition, an earlier reloading is unnecessary. | |
4961 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos")); | |
4962 optimal_split_pos = max_split_pos; | |
4963 | |
4964 } else { | |
4965 // seach optimal block boundary between min_split_pos and max_split_pos | |
4966 TRACE_LINEAR_SCAN(4, tty->print_cr(" moving split pos to optimal block boundary between block B%d and B%d", min_block->block_id(), max_block->block_id())); | |
4967 | |
4968 if (do_loop_optimization) { | |
4969 // Loop optimization: if a loop-end marker is found between min- and max-position, | |
4970 // then split before this loop | |
4971 int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2); | |
4972 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d", loop_end_pos)); | |
4973 | |
4974 assert(loop_end_pos > min_split_pos, "invalid order"); | |
4975 if (loop_end_pos < max_split_pos) { | |
4976 // loop-end marker found between min- and max-position | |
4977 // if it is not the end marker for the same loop as the min-position, then move | |
4978 // the max-position to this loop block. | |
4979 // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading | |
4980 // of the interval (normally, only mustHaveRegister causes a reloading) | |
4981 BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos); | |
4982 | |
4983 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval is used in loop that ends in block B%d, so trying to move max_block back from B%d to B%d", loop_block->block_id(), max_block->block_id(), loop_block->block_id())); | |
4984 assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between"); | |
4985 | |
4986 optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2); | |
4987 if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) { | |
4988 optimal_split_pos = -1; | |
4989 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary")); | |
4990 } else { | |
4991 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful")); | |
4992 } | |
4993 } | |
4994 } | |
4995 | |
4996 if (optimal_split_pos == -1) { | |
4997 // not calculated by loop optimization | |
4998 optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos); | |
4999 } | |
5000 } | |
5001 } | |
5002 TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d", optimal_split_pos)); | |
5003 | |
5004 return optimal_split_pos; | |
5005 } | |
5006 | |
5007 | |
5008 /* | |
5009 split an interval at the optimal position between min_split_pos and | |
5010 max_split_pos in two parts: | |
5011 1) the left part has already a location assigned | |
5012 2) the right part is sorted into to the unhandled-list | |
5013 */ | |
5014 void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) { | |
5015 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: "); it->print()); | |
5016 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); | |
5017 | |
5018 assert(it->from() < min_split_pos, "cannot split at start of interval"); | |
5019 assert(current_position() < min_split_pos, "cannot split before current position"); | |
5020 assert(min_split_pos <= max_split_pos, "invalid order"); | |
5021 assert(max_split_pos <= it->to(), "cannot split after end of interval"); | |
5022 | |
5023 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true); | |
5024 | |
5025 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); | |
5026 assert(optimal_split_pos <= it->to(), "cannot split after end of interval"); | |
5027 assert(optimal_split_pos > it->from(), "cannot split at start of interval"); | |
5028 | |
5029 if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) { | |
5030 // the split position would be just before the end of the interval | |
5031 // -> no split at all necessary | |
5032 TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval")); | |
5033 return; | |
5034 } | |
5035 | |
5036 // must calculate this before the actual split is performed and before split position is moved to odd op_id | |
5037 bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos); | |
5038 | |
5039 if (!allocator()->is_block_begin(optimal_split_pos)) { | |
5040 // move position before actual instruction (odd op_id) | |
5041 optimal_split_pos = (optimal_split_pos - 1) | 1; | |
5042 } | |
5043 | |
5044 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); | |
5045 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); | |
5046 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); | |
5047 | |
5048 Interval* split_part = it->split(optimal_split_pos); | |
5049 | |
5050 allocator()->append_interval(split_part); | |
5051 allocator()->copy_register_flags(it, split_part); | |
5052 split_part->set_insert_move_when_activated(move_necessary); | |
5053 append_to_unhandled(unhandled_first_addr(anyKind), split_part); | |
5054 | |
5055 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)", move_necessary)); | |
5056 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); | |
5057 TRACE_LINEAR_SCAN(2, tty->print (" "); split_part->print()); | |
5058 } | |
5059 | |
5060 /* | |
5061 split an interval at the optimal position between min_split_pos and | |
5062 max_split_pos in two parts: | |
5063 1) the left part has already a location assigned | |
5064 2) the right part is always on the stack and therefore ignored in further processing | |
5065 */ | |
5066 void LinearScanWalker::split_for_spilling(Interval* it) { | |
5067 // calculate allowed range of splitting position | |
5068 int max_split_pos = current_position(); | |
5069 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from()); | |
5070 | |
5071 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: "); it->print()); | |
5072 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos)); | |
5073 | |
5074 assert(it->state() == activeState, "why spill interval that is not active?"); | |
5075 assert(it->from() <= min_split_pos, "cannot split before start of interval"); | |
5076 assert(min_split_pos <= max_split_pos, "invalid order"); | |
5077 assert(max_split_pos < it->to(), "cannot split at end end of interval"); | |
5078 assert(current_position() < it->to(), "interval must not end before current position"); | |
5079 | |
5080 if (min_split_pos == it->from()) { | |
5081 // the whole interval is never used, so spill it entirely to memory | |
5082 TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval")); | |
5083 assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position"); | |
5084 | |
5085 allocator()->assign_spill_slot(it); | |
5086 allocator()->change_spill_state(it, min_split_pos); | |
5087 | |
5088 // Also kick parent intervals out of register to memory when they have no use | |
5089 // position. This avoids short interval in register surrounded by intervals in | |
5090 // memory -> avoid useless moves from memory to register and back | |
5091 Interval* parent = it; | |
5092 while (parent != NULL && parent->is_split_child()) { | |
5093 parent = parent->split_child_before_op_id(parent->from()); | |
5094 | |
5095 if (parent->assigned_reg() < LinearScan::nof_regs) { | |
5096 if (parent->first_usage(shouldHaveRegister) == max_jint) { | |
5097 // parent is never used, so kick it out of its assigned register | |
5098 TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used", parent->reg_num())); | |
5099 allocator()->assign_spill_slot(parent); | |
5100 } else { | |
5101 // do not go further back because the register is actually used by the interval | |
5102 parent = NULL; | |
5103 } | |
5104 } | |
5105 } | |
5106 | |
5107 } else { | |
5108 // search optimal split pos, split interval and spill only the right hand part | |
5109 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false); | |
5110 | |
5111 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range"); | |
5112 assert(optimal_split_pos < it->to(), "cannot split at end of interval"); | |
5113 assert(optimal_split_pos >= it->from(), "cannot split before start of interval"); | |
5114 | |
5115 if (!allocator()->is_block_begin(optimal_split_pos)) { | |
5116 // move position before actual instruction (odd op_id) | |
5117 optimal_split_pos = (optimal_split_pos - 1) | 1; | |
5118 } | |
5119 | |
5120 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos)); | |
5121 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary"); | |
5122 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary"); | |
5123 | |
5124 Interval* spilled_part = it->split(optimal_split_pos); | |
5125 allocator()->append_interval(spilled_part); | |
5126 allocator()->assign_spill_slot(spilled_part); | |
5127 allocator()->change_spill_state(spilled_part, optimal_split_pos); | |
5128 | |
5129 if (!allocator()->is_block_begin(optimal_split_pos)) { | |
5130 TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num())); | |
5131 insert_move(optimal_split_pos, it, spilled_part); | |
5132 } | |
5133 | |
5134 // the current_split_child is needed later when moves are inserted for reloading | |
5135 assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child"); | |
5136 spilled_part->make_current_split_child(); | |
5137 | |
5138 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts")); | |
5139 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print()); | |
5140 TRACE_LINEAR_SCAN(2, tty->print (" "); spilled_part->print()); | |
5141 } | |
5142 } | |
5143 | |
5144 | |
5145 void LinearScanWalker::split_stack_interval(Interval* it) { | |
5146 int min_split_pos = current_position() + 1; | |
5147 int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to()); | |
5148 | |
5149 split_before_usage(it, min_split_pos, max_split_pos); | |
5150 } | |
5151 | |
5152 void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) { | |
5153 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1); | |
5154 int max_split_pos = register_available_until; | |
5155 | |
5156 split_before_usage(it, min_split_pos, max_split_pos); | |
5157 } | |
5158 | |
5159 void LinearScanWalker::split_and_spill_interval(Interval* it) { | |
5160 assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed"); | |
5161 | |
5162 int current_pos = current_position(); | |
5163 if (it->state() == inactiveState) { | |
5164 // the interval is currently inactive, so no spill slot is needed for now. | |
5165 // when the split part is activated, the interval has a new chance to get a register, | |
5166 // so in the best case no stack slot is necessary | |
5167 assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise"); | |
5168 split_before_usage(it, current_pos + 1, current_pos + 1); | |
5169 | |
5170 } else { | |
5171 // search the position where the interval must have a register and split | |
5172 // at the optimal position before. | |
5173 // The new created part is added to the unhandled list and will get a register | |
5174 // when it is activated | |
5175 int min_split_pos = current_pos + 1; | |
5176 int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to()); | |
5177 | |
5178 split_before_usage(it, min_split_pos, max_split_pos); | |
5179 | |
5180 assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register"); | |
5181 split_for_spilling(it); | |
5182 } | |
5183 } | |
5184 | |
5185 | |
5186 int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { | |
5187 int min_full_reg = any_reg; | |
5188 int max_partial_reg = any_reg; | |
5189 | |
5190 for (int i = _first_reg; i <= _last_reg; i++) { | |
5191 if (i == ignore_reg) { | |
5192 // this register must be ignored | |
5193 | |
5194 } else if (_use_pos[i] >= interval_to) { | |
5195 // this register is free for the full interval | |
5196 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { | |
5197 min_full_reg = i; | |
5198 } | |
5199 } else if (_use_pos[i] > reg_needed_until) { | |
5200 // this register is at least free until reg_needed_until | |
5201 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { | |
5202 max_partial_reg = i; | |
5203 } | |
5204 } | |
5205 } | |
5206 | |
5207 if (min_full_reg != any_reg) { | |
5208 return min_full_reg; | |
5209 } else if (max_partial_reg != any_reg) { | |
5210 *need_split = true; | |
5211 return max_partial_reg; | |
5212 } else { | |
5213 return any_reg; | |
5214 } | |
5215 } | |
5216 | |
5217 int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { | |
5218 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); | |
5219 | |
5220 int min_full_reg = any_reg; | |
5221 int max_partial_reg = any_reg; | |
5222 | |
5223 for (int i = _first_reg; i < _last_reg; i+=2) { | |
5224 if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) { | |
5225 // this register is free for the full interval | |
5226 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { | |
5227 min_full_reg = i; | |
5228 } | |
5229 } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { | |
5230 // this register is at least free until reg_needed_until | |
5231 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { | |
5232 max_partial_reg = i; | |
5233 } | |
5234 } | |
5235 } | |
5236 | |
5237 if (min_full_reg != any_reg) { | |
5238 return min_full_reg; | |
5239 } else if (max_partial_reg != any_reg) { | |
5240 *need_split = true; | |
5241 return max_partial_reg; | |
5242 } else { | |
5243 return any_reg; | |
5244 } | |
5245 } | |
5246 | |
5247 | |
5248 bool LinearScanWalker::alloc_free_reg(Interval* cur) { | |
5249 TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print()); | |
5250 | |
5251 init_use_lists(true); | |
5252 free_exclude_active_fixed(); | |
5253 free_exclude_active_any(); | |
5254 free_collect_inactive_fixed(cur); | |
5255 free_collect_inactive_any(cur); | |
5256 // free_collect_unhandled(fixedKind, cur); | |
5257 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); | |
5258 | |
5259 // _use_pos contains the start of the next interval that has this register assigned | |
5260 // (either as a fixed register or a normal allocated register in the past) | |
5261 // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely | |
5262 TRACE_LINEAR_SCAN(4, tty->print_cr(" state of registers:")); | |
5263 TRACE_LINEAR_SCAN(4, for (int i = _first_reg; i <= _last_reg; i++) tty->print_cr(" reg %d: use_pos: %d", i, _use_pos[i])); | |
5264 | |
5265 int hint_reg, hint_regHi; | |
5266 Interval* register_hint = cur->register_hint(); | |
5267 if (register_hint != NULL) { | |
5268 hint_reg = register_hint->assigned_reg(); | |
5269 hint_regHi = register_hint->assigned_regHi(); | |
5270 | |
5271 if (allocator()->is_precolored_cpu_interval(register_hint)) { | |
5272 assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals"); | |
5273 hint_regHi = hint_reg + 1; // connect e.g. eax-edx | |
5274 } | |
5275 TRACE_LINEAR_SCAN(4, tty->print(" hint registers %d, %d from interval ", hint_reg, hint_regHi); register_hint->print()); | |
5276 | |
5277 } else { | |
5278 hint_reg = any_reg; | |
5279 hint_regHi = any_reg; | |
5280 } | |
5281 assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal"); | |
5282 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval"); | |
5283 | |
5284 // the register must be free at least until this position | |
5285 int reg_needed_until = cur->from() + 1; | |
5286 int interval_to = cur->to(); | |
5287 | |
5288 bool need_split = false; | |
5289 int split_pos = -1; | |
5290 int reg = any_reg; | |
5291 int regHi = any_reg; | |
5292 | |
5293 if (_adjacent_regs) { | |
5294 reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split); | |
5295 regHi = reg + 1; | |
5296 if (reg == any_reg) { | |
5297 return false; | |
5298 } | |
5299 split_pos = MIN2(_use_pos[reg], _use_pos[regHi]); | |
5300 | |
5301 } else { | |
5302 reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split); | |
5303 if (reg == any_reg) { | |
5304 return false; | |
5305 } | |
5306 split_pos = _use_pos[reg]; | |
5307 | |
5308 if (_num_phys_regs == 2) { | |
5309 regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split); | |
5310 | |
5311 if (_use_pos[reg] < interval_to && regHi == any_reg) { | |
5312 // do not split interval if only one register can be assigned until the split pos | |
5313 // (when one register is found for the whole interval, split&spill is only | |
5314 // performed for the hi register) | |
5315 return false; | |
5316 | |
5317 } else if (regHi != any_reg) { | |
5318 split_pos = MIN2(split_pos, _use_pos[regHi]); | |
5319 | |
5320 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax | |
5321 if (reg > regHi) { | |
5322 int temp = reg; | |
5323 reg = regHi; | |
5324 regHi = temp; | |
5325 } | |
5326 } | |
5327 } | |
5328 } | |
5329 | |
5330 cur->assign_reg(reg, regHi); | |
5331 TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d", reg, regHi)); | |
5332 | |
5333 assert(split_pos > 0, "invalid split_pos"); | |
5334 if (need_split) { | |
5335 // register not available for full interval, so split it | |
5336 split_when_partial_register_available(cur, split_pos); | |
5337 } | |
5338 | |
5339 // only return true if interval is completely assigned | |
5340 return _num_phys_regs == 1 || regHi != any_reg; | |
5341 } | |
5342 | |
5343 | |
5344 int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { | |
5345 int max_reg = any_reg; | |
5346 | |
5347 for (int i = _first_reg; i <= _last_reg; i++) { | |
5348 if (i == ignore_reg) { | |
5349 // this register must be ignored | |
5350 | |
5351 } else if (_use_pos[i] > reg_needed_until) { | |
5352 if (max_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_reg] && max_reg != hint_reg)) { | |
5353 max_reg = i; | |
5354 } | |
5355 } | |
5356 } | |
5357 | |
5358 if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) { | |
5359 *need_split = true; | |
5360 } | |
5361 | |
5362 return max_reg; | |
5363 } | |
5364 | |
5365 int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { | |
5366 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm"); | |
5367 | |
5368 int max_reg = any_reg; | |
5369 | |
5370 for (int i = _first_reg; i < _last_reg; i+=2) { | |
5371 if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { | |
5372 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) { | |
5373 max_reg = i; | |
5374 } | |
5375 } | |
5376 } | |
5377 | |
5378 if (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to) { | |
5379 *need_split = true; | |
5380 } | |
5381 | |
5382 return max_reg; | |
5383 } | |
5384 | |
5385 void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) { | |
5386 assert(reg != any_reg, "no register assigned"); | |
5387 | |
5388 for (int i = 0; i < _spill_intervals[reg]->length(); i++) { | |
5389 Interval* it = _spill_intervals[reg]->at(i); | |
5390 remove_from_list(it); | |
5391 split_and_spill_interval(it); | |
5392 } | |
5393 | |
5394 if (regHi != any_reg) { | |
5395 IntervalList* processed = _spill_intervals[reg]; | |
5396 for (int i = 0; i < _spill_intervals[regHi]->length(); i++) { | |
5397 Interval* it = _spill_intervals[regHi]->at(i); | |
5398 if (processed->index_of(it) == -1) { | |
5399 remove_from_list(it); | |
5400 split_and_spill_interval(it); | |
5401 } | |
5402 } | |
5403 } | |
5404 } | |
5405 | |
5406 | |
5407 // Split an Interval and spill it to memory so that cur can be placed in a register | |
5408 void LinearScanWalker::alloc_locked_reg(Interval* cur) { | |
5409 TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print()); | |
5410 | |
5411 // collect current usage of registers | |
5412 init_use_lists(false); | |
5413 spill_exclude_active_fixed(); | |
5414 // spill_block_unhandled_fixed(cur); | |
5415 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0"); | |
5416 spill_block_inactive_fixed(cur); | |
5417 spill_collect_active_any(); | |
5418 spill_collect_inactive_any(cur); | |
5419 | |
5420 #ifndef PRODUCT | |
5421 if (TraceLinearScanLevel >= 4) { | |
5422 tty->print_cr(" state of registers:"); | |
5423 for (int i = _first_reg; i <= _last_reg; i++) { | |
5424 tty->print(" reg %d: use_pos: %d, block_pos: %d, intervals: ", i, _use_pos[i], _block_pos[i]); | |
5425 for (int j = 0; j < _spill_intervals[i]->length(); j++) { | |
5426 tty->print("%d ", _spill_intervals[i]->at(j)->reg_num()); | |
5427 } | |
5428 tty->cr(); | |
5429 } | |
5430 } | |
5431 #endif | |
5432 | |
5433 // the register must be free at least until this position | |
5434 int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1); | |
5435 int interval_to = cur->to(); | |
5436 assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use"); | |
5437 | |
5438 int split_pos = 0; | |
5439 int use_pos = 0; | |
5440 bool need_split = false; | |
5441 int reg, regHi; | |
5442 | |
5443 if (_adjacent_regs) { | |
5444 reg = find_locked_double_reg(reg_needed_until, interval_to, any_reg, &need_split); | |
5445 regHi = reg + 1; | |
5446 | |
5447 if (reg != any_reg) { | |
5448 use_pos = MIN2(_use_pos[reg], _use_pos[regHi]); | |
5449 split_pos = MIN2(_block_pos[reg], _block_pos[regHi]); | |
5450 } | |
5451 } else { | |
5452 reg = find_locked_reg(reg_needed_until, interval_to, any_reg, cur->assigned_reg(), &need_split); | |
5453 regHi = any_reg; | |
5454 | |
5455 if (reg != any_reg) { | |
5456 use_pos = _use_pos[reg]; | |
5457 split_pos = _block_pos[reg]; | |
5458 | |
5459 if (_num_phys_regs == 2) { | |
5460 if (cur->assigned_reg() != any_reg) { | |
5461 regHi = reg; | |
5462 reg = cur->assigned_reg(); | |
5463 } else { | |
5464 regHi = find_locked_reg(reg_needed_until, interval_to, any_reg, reg, &need_split); | |
5465 if (regHi != any_reg) { | |
5466 use_pos = MIN2(use_pos, _use_pos[regHi]); | |
5467 split_pos = MIN2(split_pos, _block_pos[regHi]); | |
5468 } | |
5469 } | |
5470 | |
5471 if (regHi != any_reg && reg > regHi) { | |
5472 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax | |
5473 int temp = reg; | |
5474 reg = regHi; | |
5475 regHi = temp; | |
5476 } | |
5477 } | |
5478 } | |
5479 } | |
5480 | |
5481 if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) { | |
5482 // the first use of cur is later than the spilling position -> spill cur | |
5483 TRACE_LINEAR_SCAN(4, tty->print_cr("able to spill current interval. first_usage(register): %d, use_pos: %d", cur->first_usage(mustHaveRegister), use_pos)); | |
5484 | |
5485 if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) { | |
5486 assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)"); | |
5487 // assign a reasonable register and do a bailout in product mode to avoid errors | |
5488 allocator()->assign_spill_slot(cur); | |
5489 BAILOUT("LinearScan: no register found"); | |
5490 } | |
5491 | |
5492 split_and_spill_interval(cur); | |
5493 } else { | |
5494 TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d", reg, regHi)); | |
5495 assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found"); | |
5496 assert(split_pos > 0, "invalid split_pos"); | |
5497 assert(need_split == false || split_pos > cur->from(), "splitting interval at from"); | |
5498 | |
5499 cur->assign_reg(reg, regHi); | |
5500 if (need_split) { | |
5501 // register not available for full interval, so split it | |
5502 split_when_partial_register_available(cur, split_pos); | |
5503 } | |
5504 | |
5505 // perform splitting and spilling for all affected intervalls | |
5506 split_and_spill_intersecting_intervals(reg, regHi); | |
5507 } | |
5508 } | |
5509 | |
5510 bool LinearScanWalker::no_allocation_possible(Interval* cur) { | |
304 | 5511 #ifdef X86 |
0 | 5512 // fast calculation of intervals that can never get a register because the |
5513 // the next instruction is a call that blocks all registers | |
5514 // Note: this does not work if callee-saved registers are available (e.g. on Sparc) | |
5515 | |
5516 // check if this interval is the result of a split operation | |
5517 // (an interval got a register until this position) | |
5518 int pos = cur->from(); | |
5519 if ((pos & 1) == 1) { | |
5520 // the current instruction is a call that blocks all registers | |
5521 if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) { | |
5522 TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call")); | |
5523 | |
5524 // safety check that there is really no register available | |
5525 assert(alloc_free_reg(cur) == false, "found a register for this interval"); | |
5526 return true; | |
5527 } | |
5528 | |
5529 } | |
5530 #endif | |
5531 return false; | |
5532 } | |
5533 | |
5534 void LinearScanWalker::init_vars_for_alloc(Interval* cur) { | |
5535 BasicType type = cur->type(); | |
5536 _num_phys_regs = LinearScan::num_physical_regs(type); | |
5537 _adjacent_regs = LinearScan::requires_adjacent_regs(type); | |
5538 | |
5539 if (pd_init_regs_for_alloc(cur)) { | |
5540 // the appropriate register range was selected. | |
5541 } else if (type == T_FLOAT || type == T_DOUBLE) { | |
5542 _first_reg = pd_first_fpu_reg; | |
5543 _last_reg = pd_last_fpu_reg; | |
5544 } else { | |
5545 _first_reg = pd_first_cpu_reg; | |
5546 _last_reg = pd_last_cpu_reg; | |
5547 } | |
5548 | |
5549 assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range"); | |
5550 assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range"); | |
5551 } | |
5552 | |
5553 | |
5554 bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) { | |
5555 if (op->code() != lir_move) { | |
5556 return false; | |
5557 } | |
5558 assert(op->as_Op1() != NULL, "move must be LIR_Op1"); | |
5559 | |
5560 LIR_Opr in = ((LIR_Op1*)op)->in_opr(); | |
5561 LIR_Opr res = ((LIR_Op1*)op)->result_opr(); | |
5562 return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num(); | |
5563 } | |
5564 | |
5565 // optimization (especially for phi functions of nested loops): | |
5566 // assign same spill slot to non-intersecting intervals | |
5567 void LinearScanWalker::combine_spilled_intervals(Interval* cur) { | |
5568 if (cur->is_split_child()) { | |
5569 // optimization is only suitable for split parents | |
5570 return; | |
5571 } | |
5572 | |
5573 Interval* register_hint = cur->register_hint(false); | |
5574 if (register_hint == NULL) { | |
5575 // cur is not the target of a move, otherwise register_hint would be set | |
5576 return; | |
5577 } | |
5578 assert(register_hint->is_split_parent(), "register hint must be split parent"); | |
5579 | |
5580 if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) { | |
5581 // combining the stack slots for intervals where spill move optimization is applied | |
5582 // is not benefitial and would cause problems | |
5583 return; | |
5584 } | |
5585 | |
5586 int begin_pos = cur->from(); | |
5587 int end_pos = cur->to(); | |
5588 if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) { | |
5589 // safety check that lir_op_with_id is allowed | |
5590 return; | |
5591 } | |
5592 | |
5593 if (!is_move(allocator()->lir_op_with_id(begin_pos), register_hint, cur) || !is_move(allocator()->lir_op_with_id(end_pos), cur, register_hint)) { | |
5594 // cur and register_hint are not connected with two moves | |
5595 return; | |
5596 } | |
5597 | |
5598 Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode); | |
5599 Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode); | |
5600 if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) { | |
5601 // register_hint must be split, otherwise the re-writing of use positions does not work | |
5602 return; | |
5603 } | |
5604 | |
5605 assert(begin_hint->assigned_reg() != any_reg, "must have register assigned"); | |
5606 assert(end_hint->assigned_reg() == any_reg, "must not have register assigned"); | |
5607 assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move"); | |
5608 assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move"); | |
5609 | |
5610 if (begin_hint->assigned_reg() < LinearScan::nof_regs) { | |
5611 // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur | |
5612 return; | |
5613 } | |
5614 assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled"); | |
5615 | |
5616 // modify intervals such that cur gets the same stack slot as register_hint | |
5617 // delete use positions to prevent the intervals to get a register at beginning | |
5618 cur->set_canonical_spill_slot(register_hint->canonical_spill_slot()); | |
5619 cur->remove_first_use_pos(); | |
5620 end_hint->remove_first_use_pos(); | |
5621 } | |
5622 | |
5623 | |
5624 // allocate a physical register or memory location to an interval | |
5625 bool LinearScanWalker::activate_current() { | |
5626 Interval* cur = current(); | |
5627 bool result = true; | |
5628 | |
5629 TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval "); cur->print()); | |
5630 TRACE_LINEAR_SCAN(4, tty->print_cr(" split_parent: %d, insert_move_when_activated: %d", cur->split_parent()->reg_num(), cur->insert_move_when_activated())); | |
5631 | |
5632 if (cur->assigned_reg() >= LinearScan::nof_regs) { | |
5633 // activating an interval that has a stack slot assigned -> split it at first use position | |
5634 // used for method parameters | |
5635 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use")); | |
5636 | |
5637 split_stack_interval(cur); | |
5638 result = false; | |
5639 | |
5640 } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) { | |
5641 // activating an interval that must start in a stack slot, but may get a register later | |
5642 // used for lir_roundfp: rounding is done by store to stack and reload later | |
5643 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use")); | |
5644 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned"); | |
5645 | |
5646 allocator()->assign_spill_slot(cur); | |
5647 split_stack_interval(cur); | |
5648 result = false; | |
5649 | |
5650 } else if (cur->assigned_reg() == any_reg) { | |
5651 // interval has not assigned register -> normal allocation | |
5652 // (this is the normal case for most intervals) | |
5653 TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register")); | |
5654 | |
5655 // assign same spill slot to non-intersecting intervals | |
5656 combine_spilled_intervals(cur); | |
5657 | |
5658 init_vars_for_alloc(cur); | |
5659 if (no_allocation_possible(cur) || !alloc_free_reg(cur)) { | |
5660 // no empty register available. | |
5661 // split and spill another interval so that this interval gets a register | |
5662 alloc_locked_reg(cur); | |
5663 } | |
5664 | |
5665 // spilled intervals need not be move to active-list | |
5666 if (cur->assigned_reg() >= LinearScan::nof_regs) { | |
5667 result = false; | |
5668 } | |
5669 } | |
5670 | |
5671 // load spilled values that become active from stack slot to register | |
5672 if (cur->insert_move_when_activated()) { | |
5673 assert(cur->is_split_child(), "must be"); | |
5674 assert(cur->current_split_child() != NULL, "must be"); | |
5675 assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval"); | |
5676 TRACE_LINEAR_SCAN(4, tty->print_cr("Inserting move from interval %d to %d because insert_move_when_activated is set", cur->current_split_child()->reg_num(), cur->reg_num())); | |
5677 | |
5678 insert_move(cur->from(), cur->current_split_child(), cur); | |
5679 } | |
5680 cur->make_current_split_child(); | |
5681 | |
5682 return result; // true = interval is moved to active list | |
5683 } | |
5684 | |
5685 | |
5686 // Implementation of EdgeMoveOptimizer | |
5687 | |
5688 EdgeMoveOptimizer::EdgeMoveOptimizer() : | |
5689 _edge_instructions(4), | |
5690 _edge_instructions_idx(4) | |
5691 { | |
5692 } | |
5693 | |
5694 void EdgeMoveOptimizer::optimize(BlockList* code) { | |
5695 EdgeMoveOptimizer optimizer = EdgeMoveOptimizer(); | |
5696 | |
5697 // ignore the first block in the list (index 0 is not processed) | |
5698 for (int i = code->length() - 1; i >= 1; i--) { | |
5699 BlockBegin* block = code->at(i); | |
5700 | |
5701 if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) { | |
5702 optimizer.optimize_moves_at_block_end(block); | |
5703 } | |
5704 if (block->number_of_sux() == 2) { | |
5705 optimizer.optimize_moves_at_block_begin(block); | |
5706 } | |
5707 } | |
5708 } | |
5709 | |
5710 | |
5711 // clear all internal data structures | |
5712 void EdgeMoveOptimizer::init_instructions() { | |
5713 _edge_instructions.clear(); | |
5714 _edge_instructions_idx.clear(); | |
5715 } | |
5716 | |
5717 // append a lir-instruction-list and the index of the current operation in to the list | |
5718 void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) { | |
5719 _edge_instructions.append(instructions); | |
5720 _edge_instructions_idx.append(instructions_idx); | |
5721 } | |
5722 | |
5723 // return the current operation of the given edge (predecessor or successor) | |
5724 LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) { | |
5725 LIR_OpList* instructions = _edge_instructions.at(edge); | |
5726 int idx = _edge_instructions_idx.at(edge); | |
5727 | |
5728 if (idx < instructions->length()) { | |
5729 return instructions->at(idx); | |
5730 } else { | |
5731 return NULL; | |
5732 } | |
5733 } | |
5734 | |
5735 // removes the current operation of the given edge (predecessor or successor) | |
5736 void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) { | |
5737 LIR_OpList* instructions = _edge_instructions.at(edge); | |
5738 int idx = _edge_instructions_idx.at(edge); | |
5739 instructions->remove_at(idx); | |
5740 | |
5741 if (decrement_index) { | |
5742 _edge_instructions_idx.at_put(edge, idx - 1); | |
5743 } | |
5744 } | |
5745 | |
5746 | |
5747 bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) { | |
5748 if (op1 == NULL || op2 == NULL) { | |
5749 // at least one block is already empty -> no optimization possible | |
5750 return true; | |
5751 } | |
5752 | |
5753 if (op1->code() == lir_move && op2->code() == lir_move) { | |
5754 assert(op1->as_Op1() != NULL, "move must be LIR_Op1"); | |
5755 assert(op2->as_Op1() != NULL, "move must be LIR_Op1"); | |
5756 LIR_Op1* move1 = (LIR_Op1*)op1; | |
5757 LIR_Op1* move2 = (LIR_Op1*)op2; | |
5758 if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) { | |
5759 // these moves are exactly equal and can be optimized | |
5760 return false; | |
5761 } | |
5762 | |
5763 } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) { | |
5764 assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1"); | |
5765 assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1"); | |
5766 LIR_Op1* fxch1 = (LIR_Op1*)op1; | |
5767 LIR_Op1* fxch2 = (LIR_Op1*)op2; | |
5768 if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) { | |
5769 // equal FPU stack operations can be optimized | |
5770 return false; | |
5771 } | |
5772 | |
5773 } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) { | |
5774 // equal FPU stack operations can be optimized | |
5775 return false; | |
5776 } | |
5777 | |
5778 // no optimization possible | |
5779 return true; | |
5780 } | |
5781 | |
5782 void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) { | |
5783 TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id())); | |
5784 | |
5785 if (block->is_predecessor(block)) { | |
5786 // currently we can't handle this correctly. | |
5787 return; | |
5788 } | |
5789 | |
5790 init_instructions(); | |
5791 int num_preds = block->number_of_preds(); | |
5792 assert(num_preds > 1, "do not call otherwise"); | |
5793 assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); | |
5794 | |
5795 // setup a list with the lir-instructions of all predecessors | |
5796 int i; | |
5797 for (i = 0; i < num_preds; i++) { | |
5798 BlockBegin* pred = block->pred_at(i); | |
5799 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); | |
5800 | |
5801 if (pred->number_of_sux() != 1) { | |
5802 // this can happen with switch-statements where multiple edges are between | |
5803 // the same blocks. | |
5804 return; | |
5805 } | |
5806 | |
5807 assert(pred->number_of_sux() == 1, "can handle only one successor"); | |
5808 assert(pred->sux_at(0) == block, "invalid control flow"); | |
5809 assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch"); | |
5810 assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); | |
5811 assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); | |
5812 | |
5813 if (pred_instructions->last()->info() != NULL) { | |
5814 // can not optimize instructions when debug info is needed | |
5815 return; | |
5816 } | |
5817 | |
5818 // ignore the unconditional branch at the end of the block | |
5819 append_instructions(pred_instructions, pred_instructions->length() - 2); | |
5820 } | |
5821 | |
5822 | |
5823 // process lir-instructions while all predecessors end with the same instruction | |
5824 while (true) { | |
5825 LIR_Op* op = instruction_at(0); | |
5826 for (i = 1; i < num_preds; i++) { | |
5827 if (operations_different(op, instruction_at(i))) { | |
5828 // these instructions are different and cannot be optimized -> | |
5829 // no further optimization possible | |
5830 return; | |
5831 } | |
5832 } | |
5833 | |
5834 TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print()); | |
5835 | |
5836 // insert the instruction at the beginning of the current block | |
5837 block->lir()->insert_before(1, op); | |
5838 | |
5839 // delete the instruction at the end of all predecessors | |
5840 for (i = 0; i < num_preds; i++) { | |
5841 remove_cur_instruction(i, true); | |
5842 } | |
5843 } | |
5844 } | |
5845 | |
5846 | |
5847 void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) { | |
5848 TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id())); | |
5849 | |
5850 init_instructions(); | |
5851 int num_sux = block->number_of_sux(); | |
5852 | |
5853 LIR_OpList* cur_instructions = block->lir()->instructions_list(); | |
5854 | |
5855 assert(num_sux == 2, "method should not be called otherwise"); | |
5856 assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch"); | |
5857 assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); | |
5858 assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch"); | |
5859 | |
5860 if (cur_instructions->last()->info() != NULL) { | |
5861 // can no optimize instructions when debug info is needed | |
5862 return; | |
5863 } | |
5864 | |
5865 LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2); | |
5866 if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) { | |
5867 // not a valid case for optimization | |
5868 // currently, only blocks that end with two branches (conditional branch followed | |
5869 // by unconditional branch) are optimized | |
5870 return; | |
5871 } | |
5872 | |
5873 // now it is guaranteed that the block ends with two branch instructions. | |
5874 // the instructions are inserted at the end of the block before these two branches | |
5875 int insert_idx = cur_instructions->length() - 2; | |
5876 | |
5877 int i; | |
5878 #ifdef ASSERT | |
5879 for (i = insert_idx - 1; i >= 0; i--) { | |
5880 LIR_Op* op = cur_instructions->at(i); | |
5881 if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) { | |
5882 assert(false, "block with two successors can have only two branch instructions"); | |
5883 } | |
5884 } | |
5885 #endif | |
5886 | |
5887 // setup a list with the lir-instructions of all successors | |
5888 for (i = 0; i < num_sux; i++) { | |
5889 BlockBegin* sux = block->sux_at(i); | |
5890 LIR_OpList* sux_instructions = sux->lir()->instructions_list(); | |
5891 | |
5892 assert(sux_instructions->at(0)->code() == lir_label, "block must start with label"); | |
5893 | |
5894 if (sux->number_of_preds() != 1) { | |
5895 // this can happen with switch-statements where multiple edges are between | |
5896 // the same blocks. | |
5897 return; | |
5898 } | |
5899 assert(sux->pred_at(0) == block, "invalid control flow"); | |
5900 assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed"); | |
5901 | |
5902 // ignore the label at the beginning of the block | |
5903 append_instructions(sux_instructions, 1); | |
5904 } | |
5905 | |
5906 // process lir-instructions while all successors begin with the same instruction | |
5907 while (true) { | |
5908 LIR_Op* op = instruction_at(0); | |
5909 for (i = 1; i < num_sux; i++) { | |
5910 if (operations_different(op, instruction_at(i))) { | |
5911 // these instructions are different and cannot be optimized -> | |
5912 // no further optimization possible | |
5913 return; | |
5914 } | |
5915 } | |
5916 | |
5917 TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print()); | |
5918 | |
5919 // insert instruction at end of current block | |
5920 block->lir()->insert_before(insert_idx, op); | |
5921 insert_idx++; | |
5922 | |
5923 // delete the instructions at the beginning of all successors | |
5924 for (i = 0; i < num_sux; i++) { | |
5925 remove_cur_instruction(i, false); | |
5926 } | |
5927 } | |
5928 } | |
5929 | |
5930 | |
5931 // Implementation of ControlFlowOptimizer | |
5932 | |
5933 ControlFlowOptimizer::ControlFlowOptimizer() : | |
5934 _original_preds(4) | |
5935 { | |
5936 } | |
5937 | |
5938 void ControlFlowOptimizer::optimize(BlockList* code) { | |
5939 ControlFlowOptimizer optimizer = ControlFlowOptimizer(); | |
5940 | |
5941 // push the OSR entry block to the end so that we're not jumping over it. | |
5942 BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry(); | |
5943 if (osr_entry) { | |
5944 int index = osr_entry->linear_scan_number(); | |
5945 assert(code->at(index) == osr_entry, "wrong index"); | |
5946 code->remove_at(index); | |
5947 code->append(osr_entry); | |
5948 } | |
5949 | |
5950 optimizer.reorder_short_loops(code); | |
5951 optimizer.delete_empty_blocks(code); | |
5952 optimizer.delete_unnecessary_jumps(code); | |
5953 optimizer.delete_jumps_to_return(code); | |
5954 } | |
5955 | |
5956 void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) { | |
5957 int i = header_idx + 1; | |
5958 int max_end = MIN2(header_idx + ShortLoopSize, code->length()); | |
5959 while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) { | |
5960 i++; | |
5961 } | |
5962 | |
5963 if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) { | |
5964 int end_idx = i - 1; | |
5965 BlockBegin* end_block = code->at(end_idx); | |
5966 | |
5967 if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) { | |
5968 // short loop from header_idx to end_idx found -> reorder blocks such that | |
5969 // the header_block is the last block instead of the first block of the loop | |
5970 TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d", | |
5971 end_idx - header_idx + 1, | |
5972 header_block->block_id(), end_block->block_id())); | |
5973 | |
5974 for (int j = header_idx; j < end_idx; j++) { | |
5975 code->at_put(j, code->at(j + 1)); | |
5976 } | |
5977 code->at_put(end_idx, header_block); | |
5978 | |
5979 // correct the flags so that any loop alignment occurs in the right place. | |
5980 assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target"); | |
5981 code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag); | |
5982 code->at(header_idx)->set(BlockBegin::backward_branch_target_flag); | |
5983 } | |
5984 } | |
5985 } | |
5986 | |
5987 void ControlFlowOptimizer::reorder_short_loops(BlockList* code) { | |
5988 for (int i = code->length() - 1; i >= 0; i--) { | |
5989 BlockBegin* block = code->at(i); | |
5990 | |
5991 if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) { | |
5992 reorder_short_loop(code, block, i); | |
5993 } | |
5994 } | |
5995 | |
5996 DEBUG_ONLY(verify(code)); | |
5997 } | |
5998 | |
5999 // only blocks with exactly one successor can be deleted. Such blocks | |
6000 // must always end with an unconditional branch to this successor | |
6001 bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) { | |
6002 if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) { | |
6003 return false; | |
6004 } | |
6005 | |
6006 LIR_OpList* instructions = block->lir()->instructions_list(); | |
6007 | |
6008 assert(instructions->length() >= 2, "block must have label and branch"); | |
6009 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); | |
6010 assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch"); | |
6011 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional"); | |
6012 assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor"); | |
6013 | |
6014 // block must have exactly one successor | |
6015 | |
6016 if (instructions->length() == 2 && instructions->last()->info() == NULL) { | |
6017 return true; | |
6018 } | |
6019 return false; | |
6020 } | |
6021 | |
6022 // substitute branch targets in all branch-instructions of this blocks | |
6023 void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) { | |
6024 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting empty block: substituting from B%d to B%d inside B%d", target_from->block_id(), target_to->block_id(), block->block_id())); | |
6025 | |
6026 LIR_OpList* instructions = block->lir()->instructions_list(); | |
6027 | |
6028 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label"); | |
6029 for (int i = instructions->length() - 1; i >= 1; i--) { | |
6030 LIR_Op* op = instructions->at(i); | |
6031 | |
6032 if (op->code() == lir_branch || op->code() == lir_cond_float_branch) { | |
6033 assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); | |
6034 LIR_OpBranch* branch = (LIR_OpBranch*)op; | |
6035 | |
6036 if (branch->block() == target_from) { | |
6037 branch->change_block(target_to); | |
6038 } | |
6039 if (branch->ublock() == target_from) { | |
6040 branch->change_ublock(target_to); | |
6041 } | |
6042 } | |
6043 } | |
6044 } | |
6045 | |
6046 void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) { | |
6047 int old_pos = 0; | |
6048 int new_pos = 0; | |
6049 int num_blocks = code->length(); | |
6050 | |
6051 while (old_pos < num_blocks) { | |
6052 BlockBegin* block = code->at(old_pos); | |
6053 | |
6054 if (can_delete_block(block)) { | |
6055 BlockBegin* new_target = block->sux_at(0); | |
6056 | |
6057 // propagate backward branch target flag for correct code alignment | |
6058 if (block->is_set(BlockBegin::backward_branch_target_flag)) { | |
6059 new_target->set(BlockBegin::backward_branch_target_flag); | |
6060 } | |
6061 | |
6062 // collect a list with all predecessors that contains each predecessor only once | |
6063 // the predecessors of cur are changed during the substitution, so a copy of the | |
6064 // predecessor list is necessary | |
6065 int j; | |
6066 _original_preds.clear(); | |
6067 for (j = block->number_of_preds() - 1; j >= 0; j--) { | |
6068 BlockBegin* pred = block->pred_at(j); | |
6069 if (_original_preds.index_of(pred) == -1) { | |
6070 _original_preds.append(pred); | |
6071 } | |
6072 } | |
6073 | |
6074 for (j = _original_preds.length() - 1; j >= 0; j--) { | |
6075 BlockBegin* pred = _original_preds.at(j); | |
6076 substitute_branch_target(pred, block, new_target); | |
6077 pred->substitute_sux(block, new_target); | |
6078 } | |
6079 } else { | |
6080 // adjust position of this block in the block list if blocks before | |
6081 // have been deleted | |
6082 if (new_pos != old_pos) { | |
6083 code->at_put(new_pos, code->at(old_pos)); | |
6084 } | |
6085 new_pos++; | |
6086 } | |
6087 old_pos++; | |
6088 } | |
6089 code->truncate(new_pos); | |
6090 | |
6091 DEBUG_ONLY(verify(code)); | |
6092 } | |
6093 | |
6094 void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) { | |
6095 // skip the last block because there a branch is always necessary | |
6096 for (int i = code->length() - 2; i >= 0; i--) { | |
6097 BlockBegin* block = code->at(i); | |
6098 LIR_OpList* instructions = block->lir()->instructions_list(); | |
6099 | |
6100 LIR_Op* last_op = instructions->last(); | |
6101 if (last_op->code() == lir_branch) { | |
6102 assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); | |
6103 LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op; | |
6104 | |
6105 assert(last_branch->block() != NULL, "last branch must always have a block as target"); | |
6106 assert(last_branch->label() == last_branch->block()->label(), "must be equal"); | |
6107 | |
6108 if (last_branch->info() == NULL) { | |
6109 if (last_branch->block() == code->at(i + 1)) { | |
6110 | |
6111 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id())); | |
6112 | |
6113 // delete last branch instruction | |
6114 instructions->truncate(instructions->length() - 1); | |
6115 | |
6116 } else { | |
6117 LIR_Op* prev_op = instructions->at(instructions->length() - 2); | |
6118 if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) { | |
6119 assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch"); | |
6120 LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op; | |
6121 | |
6122 if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) { | |
6123 | |
6124 TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d", block->block_id())); | |
6125 | |
6126 // eliminate a conditional branch to the immediate successor | |
6127 prev_branch->change_block(last_branch->block()); | |
6128 prev_branch->negate_cond(); | |
6129 instructions->truncate(instructions->length() - 1); | |
6130 } | |
6131 } | |
6132 } | |
6133 } | |
6134 } | |
6135 } | |
6136 | |
6137 DEBUG_ONLY(verify(code)); | |
6138 } | |
6139 | |
6140 void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) { | |
6141 #ifdef ASSERT | |
6142 BitMap return_converted(BlockBegin::number_of_blocks()); | |
6143 return_converted.clear(); | |
6144 #endif | |
6145 | |
6146 for (int i = code->length() - 1; i >= 0; i--) { | |
6147 BlockBegin* block = code->at(i); | |
6148 LIR_OpList* cur_instructions = block->lir()->instructions_list(); | |
6149 LIR_Op* cur_last_op = cur_instructions->last(); | |
6150 | |
6151 assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label"); | |
6152 if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) { | |
6153 // the block contains only a label and a return | |
6154 // if a predecessor ends with an unconditional jump to this block, then the jump | |
6155 // can be replaced with a return instruction | |
6156 // | |
6157 // Note: the original block with only a return statement cannot be deleted completely | |
6158 // because the predecessors might have other (conditional) jumps to this block | |
6159 // -> this may lead to unnecesary return instructions in the final code | |
6160 | |
6161 assert(cur_last_op->info() == NULL, "return instructions do not have debug information"); | |
6162 assert(block->number_of_sux() == 0 || | |
6163 (return_converted.at(block->block_id()) && block->number_of_sux() == 1), | |
6164 "blocks that end with return must not have successors"); | |
6165 | |
6166 assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1"); | |
6167 LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr(); | |
6168 | |
6169 for (int j = block->number_of_preds() - 1; j >= 0; j--) { | |
6170 BlockBegin* pred = block->pred_at(j); | |
6171 LIR_OpList* pred_instructions = pred->lir()->instructions_list(); | |
6172 LIR_Op* pred_last_op = pred_instructions->last(); | |
6173 | |
6174 if (pred_last_op->code() == lir_branch) { | |
6175 assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch"); | |
6176 LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op; | |
6177 | |
6178 if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) { | |
6179 // replace the jump to a return with a direct return | |
6180 // Note: currently the edge between the blocks is not deleted | |
6181 pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr)); | |
6182 #ifdef ASSERT | |
6183 return_converted.set_bit(pred->block_id()); | |
6184 #endif | |
6185 } | |
6186 } | |
6187 } | |
6188 } | |
6189 } | |
6190 } | |
6191 | |
6192 | |
6193 #ifdef ASSERT | |
6194 void ControlFlowOptimizer::verify(BlockList* code) { | |
6195 for (int i = 0; i < code->length(); i++) { | |
6196 BlockBegin* block = code->at(i); | |
6197 LIR_OpList* instructions = block->lir()->instructions_list(); | |
6198 | |
6199 int j; | |
6200 for (j = 0; j < instructions->length(); j++) { | |
6201 LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch(); | |
6202 | |
6203 if (op_branch != NULL) { | |
6204 assert(op_branch->block() == NULL || code->index_of(op_branch->block()) != -1, "branch target not valid"); | |
6205 assert(op_branch->ublock() == NULL || code->index_of(op_branch->ublock()) != -1, "branch target not valid"); | |
6206 } | |
6207 } | |
6208 | |
6209 for (j = 0; j < block->number_of_sux() - 1; j++) { | |
6210 BlockBegin* sux = block->sux_at(j); | |
6211 assert(code->index_of(sux) != -1, "successor not valid"); | |
6212 } | |
6213 | |
6214 for (j = 0; j < block->number_of_preds() - 1; j++) { | |
6215 BlockBegin* pred = block->pred_at(j); | |
6216 assert(code->index_of(pred) != -1, "successor not valid"); | |
6217 } | |
6218 } | |
6219 } | |
6220 #endif | |
6221 | |
6222 | |
6223 #ifndef PRODUCT | |
6224 | |
6225 // Implementation of LinearStatistic | |
6226 | |
6227 const char* LinearScanStatistic::counter_name(int counter_idx) { | |
6228 switch (counter_idx) { | |
6229 case counter_method: return "compiled methods"; | |
6230 case counter_fpu_method: return "methods using fpu"; | |
6231 case counter_loop_method: return "methods with loops"; | |
6232 case counter_exception_method:return "methods with xhandler"; | |
6233 | |
6234 case counter_loop: return "loops"; | |
6235 case counter_block: return "blocks"; | |
6236 case counter_loop_block: return "blocks inside loop"; | |
6237 case counter_exception_block: return "exception handler entries"; | |
6238 case counter_interval: return "intervals"; | |
6239 case counter_fixed_interval: return "fixed intervals"; | |
6240 case counter_range: return "ranges"; | |
6241 case counter_fixed_range: return "fixed ranges"; | |
6242 case counter_use_pos: return "use positions"; | |
6243 case counter_fixed_use_pos: return "fixed use positions"; | |
6244 case counter_spill_slots: return "spill slots"; | |
6245 | |
6246 // counter for classes of lir instructions | |
6247 case counter_instruction: return "total instructions"; | |
6248 case counter_label: return "labels"; | |
6249 case counter_entry: return "method entries"; | |
6250 case counter_return: return "method returns"; | |
6251 case counter_call: return "method calls"; | |
6252 case counter_move: return "moves"; | |
6253 case counter_cmp: return "compare"; | |
6254 case counter_cond_branch: return "conditional branches"; | |
6255 case counter_uncond_branch: return "unconditional branches"; | |
6256 case counter_stub_branch: return "branches to stub"; | |
6257 case counter_alu: return "artithmetic + logic"; | |
6258 case counter_alloc: return "allocations"; | |
6259 case counter_sync: return "synchronisation"; | |
6260 case counter_throw: return "throw"; | |
6261 case counter_unwind: return "unwind"; | |
6262 case counter_typecheck: return "type+null-checks"; | |
6263 case counter_fpu_stack: return "fpu-stack"; | |
6264 case counter_misc_inst: return "other instructions"; | |
6265 case counter_other_inst: return "misc. instructions"; | |
6266 | |
6267 // counter for different types of moves | |
6268 case counter_move_total: return "total moves"; | |
6269 case counter_move_reg_reg: return "register->register"; | |
6270 case counter_move_reg_stack: return "register->stack"; | |
6271 case counter_move_stack_reg: return "stack->register"; | |
6272 case counter_move_stack_stack:return "stack->stack"; | |
6273 case counter_move_reg_mem: return "register->memory"; | |
6274 case counter_move_mem_reg: return "memory->register"; | |
6275 case counter_move_const_any: return "constant->any"; | |
6276 | |
6277 case blank_line_1: return ""; | |
6278 case blank_line_2: return ""; | |
6279 | |
6280 default: ShouldNotReachHere(); return ""; | |
6281 } | |
6282 } | |
6283 | |
6284 LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) { | |
6285 if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) { | |
6286 return counter_method; | |
6287 } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) { | |
6288 return counter_block; | |
6289 } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) { | |
6290 return counter_instruction; | |
6291 } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) { | |
6292 return counter_move_total; | |
6293 } | |
6294 return invalid_counter; | |
6295 } | |
6296 | |
6297 LinearScanStatistic::LinearScanStatistic() { | |
6298 for (int i = 0; i < number_of_counters; i++) { | |
6299 _counters_sum[i] = 0; | |
6300 _counters_max[i] = -1; | |
6301 } | |
6302 | |
6303 } | |
6304 | |
6305 // add the method-local numbers to the total sum | |
6306 void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) { | |
6307 for (int i = 0; i < number_of_counters; i++) { | |
6308 _counters_sum[i] += method_statistic._counters_sum[i]; | |
6309 _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]); | |
6310 } | |
6311 } | |
6312 | |
6313 void LinearScanStatistic::print(const char* title) { | |
6314 if (CountLinearScan || TraceLinearScanLevel > 0) { | |
6315 tty->cr(); | |
6316 tty->print_cr("***** LinearScan statistic - %s *****", title); | |
6317 | |
6318 for (int i = 0; i < number_of_counters; i++) { | |
6319 if (_counters_sum[i] > 0 || _counters_max[i] >= 0) { | |
6320 tty->print("%25s: %8d", counter_name(i), _counters_sum[i]); | |
6321 | |
6322 if (base_counter(i) != invalid_counter) { | |
6323 tty->print(" (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[base_counter(i)]); | |
6324 } else { | |
6325 tty->print(" "); | |
6326 } | |
6327 | |
6328 if (_counters_max[i] >= 0) { | |
6329 tty->print("%8d", _counters_max[i]); | |
6330 } | |
6331 } | |
6332 tty->cr(); | |
6333 } | |
6334 } | |
6335 } | |
6336 | |
6337 void LinearScanStatistic::collect(LinearScan* allocator) { | |
6338 inc_counter(counter_method); | |
6339 if (allocator->has_fpu_registers()) { | |
6340 inc_counter(counter_fpu_method); | |
6341 } | |
6342 if (allocator->num_loops() > 0) { | |
6343 inc_counter(counter_loop_method); | |
6344 } | |
6345 inc_counter(counter_loop, allocator->num_loops()); | |
6346 inc_counter(counter_spill_slots, allocator->max_spills()); | |
6347 | |
6348 int i; | |
6349 for (i = 0; i < allocator->interval_count(); i++) { | |
6350 Interval* cur = allocator->interval_at(i); | |
6351 | |
6352 if (cur != NULL) { | |
6353 inc_counter(counter_interval); | |
6354 inc_counter(counter_use_pos, cur->num_use_positions()); | |
6355 if (LinearScan::is_precolored_interval(cur)) { | |
6356 inc_counter(counter_fixed_interval); | |
6357 inc_counter(counter_fixed_use_pos, cur->num_use_positions()); | |
6358 } | |
6359 | |
6360 Range* range = cur->first(); | |
6361 while (range != Range::end()) { | |
6362 inc_counter(counter_range); | |
6363 if (LinearScan::is_precolored_interval(cur)) { | |
6364 inc_counter(counter_fixed_range); | |
6365 } | |
6366 range = range->next(); | |
6367 } | |
6368 } | |
6369 } | |
6370 | |
6371 bool has_xhandlers = false; | |
6372 // Note: only count blocks that are in code-emit order | |
6373 for (i = 0; i < allocator->ir()->code()->length(); i++) { | |
6374 BlockBegin* cur = allocator->ir()->code()->at(i); | |
6375 | |
6376 inc_counter(counter_block); | |
6377 if (cur->loop_depth() > 0) { | |
6378 inc_counter(counter_loop_block); | |
6379 } | |
6380 if (cur->is_set(BlockBegin::exception_entry_flag)) { | |
6381 inc_counter(counter_exception_block); | |
6382 has_xhandlers = true; | |
6383 } | |
6384 | |
6385 LIR_OpList* instructions = cur->lir()->instructions_list(); | |
6386 for (int j = 0; j < instructions->length(); j++) { | |
6387 LIR_Op* op = instructions->at(j); | |
6388 | |
6389 inc_counter(counter_instruction); | |
6390 | |
6391 switch (op->code()) { | |
6392 case lir_label: inc_counter(counter_label); break; | |
6393 case lir_std_entry: | |
6394 case lir_osr_entry: inc_counter(counter_entry); break; | |
6395 case lir_return: inc_counter(counter_return); break; | |
6396 | |
6397 case lir_rtcall: | |
6398 case lir_static_call: | |
6399 case lir_optvirtual_call: | |
6400 case lir_virtual_call: inc_counter(counter_call); break; | |
6401 | |
6402 case lir_move: { | |
6403 inc_counter(counter_move); | |
6404 inc_counter(counter_move_total); | |
6405 | |
6406 LIR_Opr in = op->as_Op1()->in_opr(); | |
6407 LIR_Opr res = op->as_Op1()->result_opr(); | |
6408 if (in->is_register()) { | |
6409 if (res->is_register()) { | |
6410 inc_counter(counter_move_reg_reg); | |
6411 } else if (res->is_stack()) { | |
6412 inc_counter(counter_move_reg_stack); | |
6413 } else if (res->is_address()) { | |
6414 inc_counter(counter_move_reg_mem); | |
6415 } else { | |
6416 ShouldNotReachHere(); | |
6417 } | |
6418 } else if (in->is_stack()) { | |
6419 if (res->is_register()) { | |
6420 inc_counter(counter_move_stack_reg); | |
6421 } else { | |
6422 inc_counter(counter_move_stack_stack); | |
6423 } | |
6424 } else if (in->is_address()) { | |
6425 assert(res->is_register(), "must be"); | |
6426 inc_counter(counter_move_mem_reg); | |
6427 } else if (in->is_constant()) { | |
6428 inc_counter(counter_move_const_any); | |
6429 } else { | |
6430 ShouldNotReachHere(); | |
6431 } | |
6432 break; | |
6433 } | |
6434 | |
6435 case lir_cmp: inc_counter(counter_cmp); break; | |
6436 | |
6437 case lir_branch: | |
6438 case lir_cond_float_branch: { | |
6439 LIR_OpBranch* branch = op->as_OpBranch(); | |
6440 if (branch->block() == NULL) { | |
6441 inc_counter(counter_stub_branch); | |
6442 } else if (branch->cond() == lir_cond_always) { | |
6443 inc_counter(counter_uncond_branch); | |
6444 } else { | |
6445 inc_counter(counter_cond_branch); | |
6446 } | |
6447 break; | |
6448 } | |
6449 | |
6450 case lir_neg: | |
6451 case lir_add: | |
6452 case lir_sub: | |
6453 case lir_mul: | |
6454 case lir_mul_strictfp: | |
6455 case lir_div: | |
6456 case lir_div_strictfp: | |
6457 case lir_rem: | |
6458 case lir_sqrt: | |
6459 case lir_sin: | |
6460 case lir_cos: | |
6461 case lir_abs: | |
6462 case lir_log10: | |
6463 case lir_log: | |
6464 case lir_logic_and: | |
6465 case lir_logic_or: | |
6466 case lir_logic_xor: | |
6467 case lir_shl: | |
6468 case lir_shr: | |
6469 case lir_ushr: inc_counter(counter_alu); break; | |
6470 | |
6471 case lir_alloc_object: | |
6472 case lir_alloc_array: inc_counter(counter_alloc); break; | |
6473 | |
6474 case lir_monaddr: | |
6475 case lir_lock: | |
6476 case lir_unlock: inc_counter(counter_sync); break; | |
6477 | |
6478 case lir_throw: inc_counter(counter_throw); break; | |
6479 | |
6480 case lir_unwind: inc_counter(counter_unwind); break; | |
6481 | |
6482 case lir_null_check: | |
6483 case lir_leal: | |
6484 case lir_instanceof: | |
6485 case lir_checkcast: | |
6486 case lir_store_check: inc_counter(counter_typecheck); break; | |
6487 | |
6488 case lir_fpop_raw: | |
6489 case lir_fxch: | |
6490 case lir_fld: inc_counter(counter_fpu_stack); break; | |
6491 | |
6492 case lir_nop: | |
6493 case lir_push: | |
6494 case lir_pop: | |
6495 case lir_convert: | |
6496 case lir_roundfp: | |
6497 case lir_cmove: inc_counter(counter_misc_inst); break; | |
6498 | |
6499 default: inc_counter(counter_other_inst); break; | |
6500 } | |
6501 } | |
6502 } | |
6503 | |
6504 if (has_xhandlers) { | |
6505 inc_counter(counter_exception_method); | |
6506 } | |
6507 } | |
6508 | |
6509 void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) { | |
6510 if (CountLinearScan || TraceLinearScanLevel > 0) { | |
6511 | |
6512 LinearScanStatistic local_statistic = LinearScanStatistic(); | |
6513 | |
6514 local_statistic.collect(allocator); | |
6515 global_statistic.sum_up(local_statistic); | |
6516 | |
6517 if (TraceLinearScanLevel > 2) { | |
6518 local_statistic.print("current local statistic"); | |
6519 } | |
6520 } | |
6521 } | |
6522 | |
6523 | |
6524 // Implementation of LinearTimers | |
6525 | |
6526 LinearScanTimers::LinearScanTimers() { | |
6527 for (int i = 0; i < number_of_timers; i++) { | |
6528 timer(i)->reset(); | |
6529 } | |
6530 } | |
6531 | |
6532 const char* LinearScanTimers::timer_name(int idx) { | |
6533 switch (idx) { | |
6534 case timer_do_nothing: return "Nothing (Time Check)"; | |
6535 case timer_number_instructions: return "Number Instructions"; | |
6536 case timer_compute_local_live_sets: return "Local Live Sets"; | |
6537 case timer_compute_global_live_sets: return "Global Live Sets"; | |
6538 case timer_build_intervals: return "Build Intervals"; | |
6539 case timer_sort_intervals_before: return "Sort Intervals Before"; | |
6540 case timer_allocate_registers: return "Allocate Registers"; | |
6541 case timer_resolve_data_flow: return "Resolve Data Flow"; | |
6542 case timer_sort_intervals_after: return "Sort Intervals After"; | |
6543 case timer_eliminate_spill_moves: return "Spill optimization"; | |
6544 case timer_assign_reg_num: return "Assign Reg Num"; | |
6545 case timer_allocate_fpu_stack: return "Allocate FPU Stack"; | |
6546 case timer_optimize_lir: return "Optimize LIR"; | |
6547 default: ShouldNotReachHere(); return ""; | |
6548 } | |
6549 } | |
6550 | |
6551 void LinearScanTimers::begin_method() { | |
6552 if (TimeEachLinearScan) { | |
6553 // reset all timers to measure only current method | |
6554 for (int i = 0; i < number_of_timers; i++) { | |
6555 timer(i)->reset(); | |
6556 } | |
6557 } | |
6558 } | |
6559 | |
6560 void LinearScanTimers::end_method(LinearScan* allocator) { | |
6561 if (TimeEachLinearScan) { | |
6562 | |
6563 double c = timer(timer_do_nothing)->seconds(); | |
6564 double total = 0; | |
6565 for (int i = 1; i < number_of_timers; i++) { | |
6566 total += timer(i)->seconds() - c; | |
6567 } | |
6568 | |
6569 if (total >= 0.0005) { | |
6570 // print all information in one line for automatic processing | |
6571 tty->print("@"); allocator->compilation()->method()->print_name(); | |
6572 | |
6573 tty->print("@ %d ", allocator->compilation()->method()->code_size()); | |
6574 tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2); | |
6575 tty->print("@ %d ", allocator->block_count()); | |
6576 tty->print("@ %d ", allocator->num_virtual_regs()); | |
6577 tty->print("@ %d ", allocator->interval_count()); | |
6578 tty->print("@ %d ", allocator->_num_calls); | |
6579 tty->print("@ %d ", allocator->num_loops()); | |
6580 | |
6581 tty->print("@ %6.6f ", total); | |
6582 for (int i = 1; i < number_of_timers; i++) { | |
6583 tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100); | |
6584 } | |
6585 tty->cr(); | |
6586 } | |
6587 } | |
6588 } | |
6589 | |
6590 void LinearScanTimers::print(double total_time) { | |
6591 if (TimeLinearScan) { | |
6592 // correction value: sum of dummy-timer that only measures the time that | |
6593 // is necesary to start and stop itself | |
6594 double c = timer(timer_do_nothing)->seconds(); | |
6595 | |
6596 for (int i = 0; i < number_of_timers; i++) { | |
6597 double t = timer(i)->seconds(); | |
6598 tty->print_cr(" %25s: %6.3f s (%4.1f%%) corrected: %6.3f s (%4.1f%%)", timer_name(i), t, (t / total_time) * 100.0, t - c, (t - c) / (total_time - 2 * number_of_timers * c) * 100); | |
6599 } | |
6600 } | |
6601 } | |
6602 | |
6603 #endif // #ifndef PRODUCT |