1 /*
2 * Copyright (c) 2005, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "c1/c1_CFGPrinter.hpp"
27 #include "c1/c1_CodeStubs.hpp"
28 #include "c1/c1_Compilation.hpp"
29 #include "c1/c1_FrameMap.hpp"
30 #include "c1/c1_IR.hpp"
31 #include "c1/c1_LIRGenerator.hpp"
32 #include "c1/c1_LinearScan.hpp"
33 #include "c1/c1_ValueStack.hpp"
34 #include "code/vmreg.inline.hpp"
35 #include "runtime/timerTrace.hpp"
36 #include "utilities/bitMap.inline.hpp"
37
38 #ifndef PRODUCT
39
40 static LinearScanStatistic _stat_before_alloc;
41 static LinearScanStatistic _stat_after_asign;
42 static LinearScanStatistic _stat_final;
43
44 static LinearScanTimers _total_timer;
45
46 // helper macro for short definition of timer
47 #define TIME_LINEAR_SCAN(timer_name) TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose);
48
49 // helper macro for short definition of trace-output inside code
50 #define TRACE_LINEAR_SCAN(level, code) \
51 if (TraceLinearScanLevel >= level) { \
52 code; \
53 }
54
55 #else
56
57 #define TIME_LINEAR_SCAN(timer_name)
58 #define TRACE_LINEAR_SCAN(level, code)
59
60 #endif
61
62 // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words
63 #ifdef _LP64
64 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 0, 2, 1, 2, 1, -1};
65 #else
66 static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 1, 1, 0, 1, -1, 1, 1, -1};
67 #endif
68
69
70 // Implementation of LinearScan
71
72 LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map)
73 : _compilation(ir->compilation())
74 , _ir(ir)
75 , _gen(gen)
76 , _frame_map(frame_map)
77 , _num_virtual_regs(gen->max_virtual_register_number())
78 , _has_fpu_registers(false)
79 , _num_calls(-1)
80 , _max_spills(0)
81 , _unused_spill_slot(-1)
82 , _intervals(0) // initialized later with correct length
83 , _new_intervals_from_allocation(new IntervalList())
84 , _sorted_intervals(NULL)
85 , _needs_full_resort(false)
86 , _lir_ops(0) // initialized later with correct length
87 , _block_of_op(0) // initialized later with correct length
88 , _has_info(0)
89 , _has_call(0)
90 , _scope_value_cache(0) // initialized later with correct length
91 , _interval_in_loop(0) // initialized later with correct length
92 , _cached_blocks(*ir->linear_scan_order())
93 #ifdef X86
94 , _fpu_stack_allocator(NULL)
95 #endif
96 {
97 assert(this->ir() != NULL, "check if valid");
98 assert(this->compilation() != NULL, "check if valid");
99 assert(this->gen() != NULL, "check if valid");
100 assert(this->frame_map() != NULL, "check if valid");
101 }
102
103
104 // ********** functions for converting LIR-Operands to register numbers
105 //
106 // Emulate a flat register file comprising physical integer registers,
107 // physical floating-point registers and virtual registers, in that order.
108 // Virtual registers already have appropriate numbers, since V0 is
109 // the number of physical registers.
110 // Returns -1 for hi word if opr is a single word operand.
111 //
112 // Note: the inverse operation (calculating an operand for register numbers)
113 // is done in calc_operand_for_interval()
114
115 int LinearScan::reg_num(LIR_Opr opr) {
116 assert(opr->is_register(), "should not call this otherwise");
117
118 if (opr->is_virtual_register()) {
119 assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number");
120 return opr->vreg_number();
121 } else if (opr->is_single_cpu()) {
122 return opr->cpu_regnr();
123 } else if (opr->is_double_cpu()) {
124 return opr->cpu_regnrLo();
125 #ifdef X86
126 } else if (opr->is_single_xmm()) {
127 return opr->fpu_regnr() + pd_first_xmm_reg;
128 } else if (opr->is_double_xmm()) {
129 return opr->fpu_regnrLo() + pd_first_xmm_reg;
130 #endif
131 } else if (opr->is_single_fpu()) {
132 return opr->fpu_regnr() + pd_first_fpu_reg;
133 } else if (opr->is_double_fpu()) {
134 return opr->fpu_regnrLo() + pd_first_fpu_reg;
135 } else {
136 ShouldNotReachHere();
137 return -1;
138 }
139 }
140
141 int LinearScan::reg_numHi(LIR_Opr opr) {
142 assert(opr->is_register(), "should not call this otherwise");
143
144 if (opr->is_virtual_register()) {
145 return -1;
146 } else if (opr->is_single_cpu()) {
147 return -1;
148 } else if (opr->is_double_cpu()) {
149 return opr->cpu_regnrHi();
150 #ifdef X86
151 } else if (opr->is_single_xmm()) {
152 return -1;
153 } else if (opr->is_double_xmm()) {
154 return -1;
155 #endif
156 } else if (opr->is_single_fpu()) {
157 return -1;
158 } else if (opr->is_double_fpu()) {
159 return opr->fpu_regnrHi() + pd_first_fpu_reg;
160 } else {
161 ShouldNotReachHere();
162 return -1;
163 }
164 }
165
166
167 // ********** functions for classification of intervals
168
169 bool LinearScan::is_precolored_interval(const Interval* i) {
170 return i->reg_num() < LinearScan::nof_regs;
171 }
172
173 bool LinearScan::is_virtual_interval(const Interval* i) {
174 return i->reg_num() >= LIR_OprDesc::vreg_base;
175 }
176
177 bool LinearScan::is_precolored_cpu_interval(const Interval* i) {
178 return i->reg_num() < LinearScan::nof_cpu_regs;
179 }
180
181 bool LinearScan::is_virtual_cpu_interval(const Interval* i) {
182 #if defined(__SOFTFP__) || defined(E500V2)
183 return i->reg_num() >= LIR_OprDesc::vreg_base;
184 #else
185 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE);
186 #endif // __SOFTFP__ or E500V2
187 }
188
189 bool LinearScan::is_precolored_fpu_interval(const Interval* i) {
190 return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs;
191 }
192
193 bool LinearScan::is_virtual_fpu_interval(const Interval* i) {
194 #if defined(__SOFTFP__) || defined(E500V2)
195 return false;
196 #else
197 return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE);
198 #endif // __SOFTFP__ or E500V2
199 }
200
201 bool LinearScan::is_in_fpu_register(const Interval* i) {
202 // fixed intervals not needed for FPU stack allocation
203 return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg;
204 }
205
206 bool LinearScan::is_oop_interval(const Interval* i) {
207 // fixed intervals never contain oops
208 return i->reg_num() >= nof_regs && i->type() == T_OBJECT;
209 }
210
211
212 // ********** General helper functions
213
214 // compute next unused stack index that can be used for spilling
215 int LinearScan::allocate_spill_slot(bool double_word) {
216 int spill_slot;
217 if (double_word) {
218 if ((_max_spills & 1) == 1) {
219 // alignment of double-word values
220 // the hole because of the alignment is filled with the next single-word value
221 assert(_unused_spill_slot == -1, "wasting a spill slot");
222 _unused_spill_slot = _max_spills;
223 _max_spills++;
224 }
225 spill_slot = _max_spills;
226 _max_spills += 2;
227
228 } else if (_unused_spill_slot != -1) {
229 // re-use hole that was the result of a previous double-word alignment
230 spill_slot = _unused_spill_slot;
231 _unused_spill_slot = -1;
232
233 } else {
234 spill_slot = _max_spills;
235 _max_spills++;
236 }
237
238 int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount();
239
240 // the class OopMapValue uses only 11 bits for storing the name of the
241 // oop location. So a stack slot bigger than 2^11 leads to an overflow
242 // that is not reported in product builds. Prevent this by checking the
243 // spill slot here (altough this value and the later used location name
244 // are slightly different)
245 if (result > 2000) {
246 bailout("too many stack slots used");
247 }
248
249 return result;
250 }
251
252 void LinearScan::assign_spill_slot(Interval* it) {
253 // assign the canonical spill slot of the parent (if a part of the interval
254 // is already spilled) or allocate a new spill slot
255 if (it->canonical_spill_slot() >= 0) {
256 it->assign_reg(it->canonical_spill_slot());
257 } else {
258 int spill = allocate_spill_slot(type2spill_size[it->type()] == 2);
259 it->set_canonical_spill_slot(spill);
260 it->assign_reg(spill);
261 }
262 }
263
264 void LinearScan::propagate_spill_slots() {
265 if (!frame_map()->finalize_frame(max_spills())) {
266 bailout("frame too large");
267 }
268 }
269
270 // create a new interval with a predefined reg_num
271 // (only used for parent intervals that are created during the building phase)
272 Interval* LinearScan::create_interval(int reg_num) {
273 assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval");
274
275 Interval* interval = new Interval(reg_num);
276 _intervals.at_put(reg_num, interval);
277
278 // assign register number for precolored intervals
279 if (reg_num < LIR_OprDesc::vreg_base) {
280 interval->assign_reg(reg_num);
281 }
282 return interval;
283 }
284
285 // assign a new reg_num to the interval and append it to the list of intervals
286 // (only used for child intervals that are created during register allocation)
287 void LinearScan::append_interval(Interval* it) {
288 it->set_reg_num(_intervals.length());
289 _intervals.append(it);
290 _new_intervals_from_allocation->append(it);
291 }
292
293 // copy the vreg-flags if an interval is split
294 void LinearScan::copy_register_flags(Interval* from, Interval* to) {
295 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) {
296 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg);
297 }
298 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) {
299 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved);
300 }
301
302 // Note: do not copy the must_start_in_memory flag because it is not necessary for child
303 // intervals (only the very beginning of the interval must be in memory)
304 }
305
306
307 // ********** spill move optimization
308 // eliminate moves from register to stack if stack slot is known to be correct
309
310 // called during building of intervals
311 void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) {
312 assert(interval->is_split_parent(), "can only be called for split parents");
313
314 switch (interval->spill_state()) {
315 case noDefinitionFound:
316 assert(interval->spill_definition_pos() == -1, "must no be set before");
317 interval->set_spill_definition_pos(def_pos);
318 interval->set_spill_state(oneDefinitionFound);
319 break;
320
321 case oneDefinitionFound:
322 assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created");
323 if (def_pos < interval->spill_definition_pos() - 2) {
324 // second definition found, so no spill optimization possible for this interval
325 interval->set_spill_state(noOptimization);
326 } else {
327 // two consecutive definitions (because of two-operand LIR form)
328 assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal");
329 }
330 break;
331
332 case noOptimization:
333 // nothing to do
334 break;
335
336 default:
337 assert(false, "other states not allowed at this time");
338 }
339 }
340
341 // called during register allocation
342 void LinearScan::change_spill_state(Interval* interval, int spill_pos) {
343 switch (interval->spill_state()) {
344 case oneDefinitionFound: {
345 int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth();
346 int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth();
347
348 if (def_loop_depth < spill_loop_depth) {
349 // the loop depth of the spilling position is higher then the loop depth
350 // at the definition of the interval -> move write to memory out of loop
351 // by storing at definitin of the interval
352 interval->set_spill_state(storeAtDefinition);
353 } else {
354 // the interval is currently spilled only once, so for now there is no
355 // reason to store the interval at the definition
356 interval->set_spill_state(oneMoveInserted);
357 }
358 break;
359 }
360
361 case oneMoveInserted: {
362 // the interval is spilled more then once, so it is better to store it to
363 // memory at the definition
364 interval->set_spill_state(storeAtDefinition);
365 break;
366 }
367
368 case storeAtDefinition:
369 case startInMemory:
370 case noOptimization:
371 case noDefinitionFound:
372 // nothing to do
373 break;
374
375 default:
376 assert(false, "other states not allowed at this time");
377 }
378 }
379
380
381 bool LinearScan::must_store_at_definition(const Interval* i) {
382 return i->is_split_parent() && i->spill_state() == storeAtDefinition;
383 }
384
385 // called once before asignment of register numbers
386 void LinearScan::eliminate_spill_moves() {
387 TIME_LINEAR_SCAN(timer_eliminate_spill_moves);
388 TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves"));
389
390 // collect all intervals that must be stored after their definion.
391 // the list is sorted by Interval::spill_definition_pos
392 Interval* interval;
393 Interval* temp_list;
394 create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL);
395
396 #ifdef ASSERT
397 Interval* prev = NULL;
398 Interval* temp = interval;
399 while (temp != Interval::end()) {
400 assert(temp->spill_definition_pos() > 0, "invalid spill definition pos");
401 if (prev != NULL) {
402 assert(temp->from() >= prev->from(), "intervals not sorted");
403 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");
404 }
405
406 assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned");
407 assert(temp->spill_definition_pos() >= temp->from(), "invalid order");
408 assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized");
409
410 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()));
411
412 temp = temp->next();
413 }
414 #endif
415
416 LIR_InsertionBuffer insertion_buffer;
417 int num_blocks = block_count();
418 for (int i = 0; i < num_blocks; i++) {
419 BlockBegin* block = block_at(i);
420 LIR_OpList* instructions = block->lir()->instructions_list();
421 int num_inst = instructions->length();
422 bool has_new = false;
423
424 // iterate all instructions of the block. skip the first because it is always a label
425 for (int j = 1; j < num_inst; j++) {
426 LIR_Op* op = instructions->at(j);
427 int op_id = op->id();
428
429 if (op_id == -1) {
430 // remove move from register to stack if the stack slot is guaranteed to be correct.
431 // only moves that have been inserted by LinearScan can be removed.
432 assert(op->code() == lir_move, "only moves can have a op_id of -1");
433 assert(op->as_Op1() != NULL, "move must be LIR_Op1");
434 assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers");
435
436 LIR_Op1* op1 = (LIR_Op1*)op;
437 Interval* interval = interval_at(op1->result_opr()->vreg_number());
438
439 if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) {
440 // move target is a stack slot that is always correct, so eliminate instruction
441 TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number()));
442 instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num
443 }
444
445 } else {
446 // insert move from register to stack just after the beginning of the interval
447 assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order");
448 assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval");
449
450 while (interval != Interval::end() && interval->spill_definition_pos() == op_id) {
451 if (!has_new) {
452 // prepare insertion buffer (appended when all instructions of the block are processed)
453 insertion_buffer.init(block->lir());
454 has_new = true;
455 }
456
457 LIR_Opr from_opr = operand_for_interval(interval);
458 LIR_Opr to_opr = canonical_spill_opr(interval);
459 assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register");
460 assert(to_opr->is_stack(), "to operand must be a stack slot");
461
462 insertion_buffer.move(j, from_opr, to_opr);
463 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));
464
465 interval = interval->next();
466 }
467 }
468 } // end of instruction iteration
469
470 if (has_new) {
471 block->lir()->append(&insertion_buffer);
472 }
473 } // end of block iteration
474
475 assert(interval == Interval::end(), "missed an interval");
476 }
477
478
479 // ********** Phase 1: number all instructions in all blocks
480 // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan.
481
482 void LinearScan::number_instructions() {
483 {
484 // dummy-timer to measure the cost of the timer itself
485 // (this time is then subtracted from all other timers to get the real value)
486 TIME_LINEAR_SCAN(timer_do_nothing);
487 }
488 TIME_LINEAR_SCAN(timer_number_instructions);
489
490 // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node.
491 int num_blocks = block_count();
492 int num_instructions = 0;
493 int i;
494 for (i = 0; i < num_blocks; i++) {
495 num_instructions += block_at(i)->lir()->instructions_list()->length();
496 }
497
498 // initialize with correct length
499 _lir_ops = LIR_OpArray(num_instructions, num_instructions, NULL);
500 _block_of_op = BlockBeginArray(num_instructions, num_instructions, NULL);
501
502 int op_id = 0;
503 int idx = 0;
504
505 for (i = 0; i < num_blocks; i++) {
506 BlockBegin* block = block_at(i);
507 block->set_first_lir_instruction_id(op_id);
508 LIR_OpList* instructions = block->lir()->instructions_list();
509
510 int num_inst = instructions->length();
511 for (int j = 0; j < num_inst; j++) {
512 LIR_Op* op = instructions->at(j);
513 op->set_id(op_id);
514
515 _lir_ops.at_put(idx, op);
516 _block_of_op.at_put(idx, block);
517 assert(lir_op_with_id(op_id) == op, "must match");
518
519 idx++;
520 op_id += 2; // numbering of lir_ops by two
521 }
522 block->set_last_lir_instruction_id(op_id - 2);
523 }
524 assert(idx == num_instructions, "must match");
525 assert(idx * 2 == op_id, "must match");
526
527 _has_call.initialize(num_instructions);
528 _has_info.initialize(num_instructions);
529 }
530
531
532 // ********** Phase 2: compute local live sets separately for each block
533 // (sets live_gen and live_kill for each block)
534
535 void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) {
536 LIR_Opr opr = value->operand();
537 Constant* con = value->as_Constant();
538
539 // check some asumptions about debug information
540 assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type");
541 assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands");
542 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands");
543
544 if ((con == NULL || con->is_pinned()) && opr->is_register()) {
545 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
546 int reg = opr->vreg_number();
547 if (!live_kill.at(reg)) {
548 live_gen.set_bit(reg);
549 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));
550 }
551 }
552 }
553
554
555 void LinearScan::compute_local_live_sets() {
556 TIME_LINEAR_SCAN(timer_compute_local_live_sets);
557
558 int num_blocks = block_count();
559 int live_size = live_set_size();
560 bool local_has_fpu_registers = false;
561 int local_num_calls = 0;
562 LIR_OpVisitState visitor;
563
564 BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops());
565
566 // iterate all blocks
567 for (int i = 0; i < num_blocks; i++) {
568 BlockBegin* block = block_at(i);
569
570 ResourceBitMap live_gen(live_size);
571 ResourceBitMap live_kill(live_size);
572
573 if (block->is_set(BlockBegin::exception_entry_flag)) {
574 // Phi functions at the begin of an exception handler are
575 // implicitly defined (= killed) at the beginning of the block.
576 for_each_phi_fun(block, phi,
577 live_kill.set_bit(phi->operand()->vreg_number())
578 );
579 }
580
581 LIR_OpList* instructions = block->lir()->instructions_list();
582 int num_inst = instructions->length();
583
584 // iterate all instructions of the block. skip the first because it is always a label
585 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label");
586 for (int j = 1; j < num_inst; j++) {
587 LIR_Op* op = instructions->at(j);
588
589 // visit operation to collect all operands
590 visitor.visit(op);
591
592 if (visitor.has_call()) {
593 _has_call.set_bit(op->id() >> 1);
594 local_num_calls++;
595 }
596 if (visitor.info_count() > 0) {
597 _has_info.set_bit(op->id() >> 1);
598 }
599
600 // iterate input operands of instruction
601 int k, n, reg;
602 n = visitor.opr_count(LIR_OpVisitState::inputMode);
603 for (k = 0; k < n; k++) {
604 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k);
605 assert(opr->is_register(), "visitor should only return register operands");
606
607 if (opr->is_virtual_register()) {
608 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
609 reg = opr->vreg_number();
610 if (!live_kill.at(reg)) {
611 live_gen.set_bit(reg);
612 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d", reg, op->id()));
613 }
614 if (block->loop_index() >= 0) {
615 local_interval_in_loop.set_bit(reg, block->loop_index());
616 }
617 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
618 }
619
620 #ifdef ASSERT
621 // fixed intervals are never live at block boundaries, so
622 // they need not be processed in live sets.
623 // this is checked by these assertions to be sure about it.
624 // the entry block may have incoming values in registers, which is ok.
625 if (!opr->is_virtual_register() && block != ir()->start()) {
626 reg = reg_num(opr);
627 if (is_processed_reg_num(reg)) {
628 assert(live_kill.at(reg), "using fixed register that is not defined in this block");
629 }
630 reg = reg_numHi(opr);
631 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
632 assert(live_kill.at(reg), "using fixed register that is not defined in this block");
633 }
634 }
635 #endif
636 }
637
638 // Add uses of live locals from interpreter's point of view for proper debug information generation
639 n = visitor.info_count();
640 for (k = 0; k < n; k++) {
641 CodeEmitInfo* info = visitor.info_at(k);
642 ValueStack* stack = info->stack();
643 for_each_state_value(stack, value,
644 set_live_gen_kill(value, op, live_gen, live_kill)
645 );
646 }
647
648 // iterate temp operands of instruction
649 n = visitor.opr_count(LIR_OpVisitState::tempMode);
650 for (k = 0; k < n; k++) {
651 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k);
652 assert(opr->is_register(), "visitor should only return register operands");
653
654 if (opr->is_virtual_register()) {
655 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
656 reg = opr->vreg_number();
657 live_kill.set_bit(reg);
658 if (block->loop_index() >= 0) {
659 local_interval_in_loop.set_bit(reg, block->loop_index());
660 }
661 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
662 }
663
664 #ifdef ASSERT
665 // fixed intervals are never live at block boundaries, so
666 // they need not be processed in live sets
667 // process them only in debug mode so that this can be checked
668 if (!opr->is_virtual_register()) {
669 reg = reg_num(opr);
670 if (is_processed_reg_num(reg)) {
671 live_kill.set_bit(reg_num(opr));
672 }
673 reg = reg_numHi(opr);
674 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
675 live_kill.set_bit(reg);
676 }
677 }
678 #endif
679 }
680
681 // iterate output operands of instruction
682 n = visitor.opr_count(LIR_OpVisitState::outputMode);
683 for (k = 0; k < n; k++) {
684 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k);
685 assert(opr->is_register(), "visitor should only return register operands");
686
687 if (opr->is_virtual_register()) {
688 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
689 reg = opr->vreg_number();
690 live_kill.set_bit(reg);
691 if (block->loop_index() >= 0) {
692 local_interval_in_loop.set_bit(reg, block->loop_index());
693 }
694 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
695 }
696
697 #ifdef ASSERT
698 // fixed intervals are never live at block boundaries, so
699 // they need not be processed in live sets
700 // process them only in debug mode so that this can be checked
701 if (!opr->is_virtual_register()) {
702 reg = reg_num(opr);
703 if (is_processed_reg_num(reg)) {
704 live_kill.set_bit(reg_num(opr));
705 }
706 reg = reg_numHi(opr);
707 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
708 live_kill.set_bit(reg);
709 }
710 }
711 #endif
712 }
713 } // end of instruction iteration
714
715 block->set_live_gen (live_gen);
716 block->set_live_kill(live_kill);
717 block->set_live_in (ResourceBitMap(live_size));
718 block->set_live_out (ResourceBitMap(live_size));
719
720 TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d ", block->block_id()); print_bitmap(block->live_gen()));
721 TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill()));
722 } // end of block iteration
723
724 // propagate local calculated information into LinearScan object
725 _has_fpu_registers = local_has_fpu_registers;
726 compilation()->set_has_fpu_code(local_has_fpu_registers);
727
728 _num_calls = local_num_calls;
729 _interval_in_loop = local_interval_in_loop;
730 }
731
732
733 // ********** Phase 3: perform a backward dataflow analysis to compute global live sets
734 // (sets live_in and live_out for each block)
735
736 void LinearScan::compute_global_live_sets() {
737 TIME_LINEAR_SCAN(timer_compute_global_live_sets);
738
739 int num_blocks = block_count();
740 bool change_occurred;
741 bool change_occurred_in_block;
742 int iteration_count = 0;
743 ResourceBitMap live_out(live_set_size()); // scratch set for calculations
744
745 // Perform a backward dataflow analysis to compute live_out and live_in for each block.
746 // The loop is executed until a fixpoint is reached (no changes in an iteration)
747 // Exception handlers must be processed because not all live values are
748 // present in the state array, e.g. because of global value numbering
749 do {
750 change_occurred = false;
751
752 // iterate all blocks in reverse order
753 for (int i = num_blocks - 1; i >= 0; i--) {
754 BlockBegin* block = block_at(i);
755
756 change_occurred_in_block = false;
757
758 // live_out(block) is the union of live_in(sux), for successors sux of block
759 int n = block->number_of_sux();
760 int e = block->number_of_exception_handlers();
761 if (n + e > 0) {
762 // block has successors
763 if (n > 0) {
764 live_out.set_from(block->sux_at(0)->live_in());
765 for (int j = 1; j < n; j++) {
766 live_out.set_union(block->sux_at(j)->live_in());
767 }
768 } else {
769 live_out.clear();
770 }
771 for (int j = 0; j < e; j++) {
772 live_out.set_union(block->exception_handler_at(j)->live_in());
773 }
774
775 if (!block->live_out().is_same(live_out)) {
776 // A change occurred. Swap the old and new live out sets to avoid copying.
777 ResourceBitMap temp = block->live_out();
778 block->set_live_out(live_out);
779 live_out = temp;
780
781 change_occurred = true;
782 change_occurred_in_block = true;
783 }
784 }
785
786 if (iteration_count == 0 || change_occurred_in_block) {
787 // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block))
788 // note: live_in has to be computed only in first iteration or if live_out has changed!
789 ResourceBitMap live_in = block->live_in();
790 live_in.set_from(block->live_out());
791 live_in.set_difference(block->live_kill());
792 live_in.set_union(block->live_gen());
793 }
794
795 #ifndef PRODUCT
796 if (TraceLinearScanLevel >= 4) {
797 char c = ' ';
798 if (iteration_count == 0 || change_occurred_in_block) {
799 c = '*';
800 }
801 tty->print("(%d) live_in%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in());
802 tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out());
803 }
804 #endif
805 }
806 iteration_count++;
807
808 if (change_occurred && iteration_count > 50) {
809 BAILOUT("too many iterations in compute_global_live_sets");
810 }
811 } while (change_occurred);
812
813
814 #ifdef ASSERT
815 // check that fixed intervals are not live at block boundaries
816 // (live set must be empty at fixed intervals)
817 for (int i = 0; i < num_blocks; i++) {
818 BlockBegin* block = block_at(i);
819 for (int j = 0; j < LIR_OprDesc::vreg_base; j++) {
820 assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty");
821 assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty");
822 assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty");
823 }
824 }
825 #endif
826
827 // check that the live_in set of the first block is empty
828 ResourceBitMap live_in_args(ir()->start()->live_in().size());
829 if (!ir()->start()->live_in().is_same(live_in_args)) {
830 #ifdef ASSERT
831 tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)");
832 tty->print_cr("affected registers:");
833 print_bitmap(ir()->start()->live_in());
834
835 // print some additional information to simplify debugging
836 for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) {
837 if (ir()->start()->live_in().at(i)) {
838 Instruction* instr = gen()->instruction_for_vreg(i);
839 tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id());
840
841 for (int j = 0; j < num_blocks; j++) {
842 BlockBegin* block = block_at(j);
843 if (block->live_gen().at(i)) {
844 tty->print_cr(" used in block B%d", block->block_id());
845 }
846 if (block->live_kill().at(i)) {
847 tty->print_cr(" defined in block B%d", block->block_id());
848 }
849 }
850 }
851 }
852
853 #endif
854 // when this fails, virtual registers are used before they are defined.
855 assert(false, "live_in set of first block must be empty");
856 // bailout of if this occurs in product mode.
857 bailout("live_in set of first block not empty");
858 }
859 }
860
861
862 // ********** Phase 4: build intervals
863 // (fills the list _intervals)
864
865 void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) {
866 assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type");
867 LIR_Opr opr = value->operand();
868 Constant* con = value->as_Constant();
869
870 if ((con == NULL || con->is_pinned()) && opr->is_register()) {
871 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
872 add_use(opr, from, to, use_kind);
873 }
874 }
875
876
877 void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) {
878 TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind));
879 assert(opr->is_register(), "should not be called otherwise");
880
881 if (opr->is_virtual_register()) {
882 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
883 add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register());
884
885 } else {
886 int reg = reg_num(opr);
887 if (is_processed_reg_num(reg)) {
888 add_def(reg, def_pos, use_kind, opr->type_register());
889 }
890 reg = reg_numHi(opr);
891 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
892 add_def(reg, def_pos, use_kind, opr->type_register());
893 }
894 }
895 }
896
897 void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) {
898 TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind));
899 assert(opr->is_register(), "should not be called otherwise");
900
901 if (opr->is_virtual_register()) {
902 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
903 add_use(opr->vreg_number(), from, to, use_kind, opr->type_register());
904
905 } else {
906 int reg = reg_num(opr);
907 if (is_processed_reg_num(reg)) {
908 add_use(reg, from, to, use_kind, opr->type_register());
909 }
910 reg = reg_numHi(opr);
911 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
912 add_use(reg, from, to, use_kind, opr->type_register());
913 }
914 }
915 }
916
917 void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) {
918 TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind));
919 assert(opr->is_register(), "should not be called otherwise");
920
921 if (opr->is_virtual_register()) {
922 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
923 add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register());
924
925 } else {
926 int reg = reg_num(opr);
927 if (is_processed_reg_num(reg)) {
928 add_temp(reg, temp_pos, use_kind, opr->type_register());
929 }
930 reg = reg_numHi(opr);
931 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
932 add_temp(reg, temp_pos, use_kind, opr->type_register());
933 }
934 }
935 }
936
937
938 void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) {
939 Interval* interval = interval_at(reg_num);
940 if (interval != NULL) {
941 assert(interval->reg_num() == reg_num, "wrong interval");
942
943 if (type != T_ILLEGAL) {
944 interval->set_type(type);
945 }
946
947 Range* r = interval->first();
948 if (r->from() <= def_pos) {
949 // Update the starting point (when a range is first created for a use, its
950 // start is the beginning of the current block until a def is encountered.)
951 r->set_from(def_pos);
952 interval->add_use_pos(def_pos, use_kind);
953
954 } else {
955 // Dead value - make vacuous interval
956 // also add use_kind for dead intervals
957 interval->add_range(def_pos, def_pos + 1);
958 interval->add_use_pos(def_pos, use_kind);
959 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos));
960 }
961
962 } else {
963 // Dead value - make vacuous interval
964 // also add use_kind for dead intervals
965 interval = create_interval(reg_num);
966 if (type != T_ILLEGAL) {
967 interval->set_type(type);
968 }
969
970 interval->add_range(def_pos, def_pos + 1);
971 interval->add_use_pos(def_pos, use_kind);
972 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos));
973 }
974
975 change_spill_definition_pos(interval, def_pos);
976 if (use_kind == noUse && interval->spill_state() <= startInMemory) {
977 // detection of method-parameters and roundfp-results
978 // TODO: move this directly to position where use-kind is computed
979 interval->set_spill_state(startInMemory);
980 }
981 }
982
983 void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) {
984 Interval* interval = interval_at(reg_num);
985 if (interval == NULL) {
986 interval = create_interval(reg_num);
987 }
988 assert(interval->reg_num() == reg_num, "wrong interval");
989
990 if (type != T_ILLEGAL) {
991 interval->set_type(type);
992 }
993
994 interval->add_range(from, to);
995 interval->add_use_pos(to, use_kind);
996 }
997
998 void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) {
999 Interval* interval = interval_at(reg_num);
1000 if (interval == NULL) {
1001 interval = create_interval(reg_num);
1002 }
1003 assert(interval->reg_num() == reg_num, "wrong interval");
1004
1005 if (type != T_ILLEGAL) {
1006 interval->set_type(type);
1007 }
1008
1009 interval->add_range(temp_pos, temp_pos + 1);
1010 interval->add_use_pos(temp_pos, use_kind);
1011 }
1012
1013
1014 // the results of this functions are used for optimizing spilling and reloading
1015 // if the functions return shouldHaveRegister and the interval is spilled,
1016 // it is not reloaded to a register.
1017 IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) {
1018 if (op->code() == lir_move) {
1019 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1");
1020 LIR_Op1* move = (LIR_Op1*)op;
1021 LIR_Opr res = move->result_opr();
1022 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory);
1023
1024 if (result_in_memory) {
1025 // Begin of an interval with must_start_in_memory set.
1026 // This interval will always get a stack slot first, so return noUse.
1027 return noUse;
1028
1029 } else if (move->in_opr()->is_stack()) {
1030 // method argument (condition must be equal to handle_method_arguments)
1031 return noUse;
1032
1033 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) {
1034 // Move from register to register
1035 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) {
1036 // special handling of phi-function moves inside osr-entry blocks
1037 // input operand must have a register instead of output operand (leads to better register allocation)
1038 return shouldHaveRegister;
1039 }
1040 }
1041 }
1042
1043 if (opr->is_virtual() &&
1044 gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) {
1045 // result is a stack-slot, so prevent immediate reloading
1046 return noUse;
1047 }
1048
1049 // all other operands require a register
1050 return mustHaveRegister;
1051 }
1052
1053 IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) {
1054 if (op->code() == lir_move) {
1055 assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1");
1056 LIR_Op1* move = (LIR_Op1*)op;
1057 LIR_Opr res = move->result_opr();
1058 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory);
1059
1060 if (result_in_memory) {
1061 // Move to an interval with must_start_in_memory set.
1062 // To avoid moves from stack to stack (not allowed) force the input operand to a register
1063 return mustHaveRegister;
1064
1065 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) {
1066 // Move from register to register
1067 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) {
1068 // special handling of phi-function moves inside osr-entry blocks
1069 // input operand must have a register instead of output operand (leads to better register allocation)
1070 return mustHaveRegister;
1071 }
1072
1073 // The input operand is not forced to a register (moves from stack to register are allowed),
1074 // but it is faster if the input operand is in a register
1075 return shouldHaveRegister;
1076 }
1077 }
1078
1079
1080 #if defined(X86) || defined(S390)
1081 if (op->code() == lir_cmove) {
1082 // conditional moves can handle stack operands
1083 assert(op->result_opr()->is_register(), "result must always be in a register");
1084 return shouldHaveRegister;
1085 }
1086
1087 // optimizations for second input operand of arithmehtic operations on Intel
1088 // this operand is allowed to be on the stack in some cases
1089 BasicType opr_type = opr->type_register();
1090 if (opr_type == T_FLOAT || opr_type == T_DOUBLE) {
1091 if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2 S390_ONLY(|| true)) {
1092 // SSE float instruction (T_DOUBLE only supported with SSE2)
1093 switch (op->code()) {
1094 case lir_cmp:
1095 case lir_add:
1096 case lir_sub:
1097 case lir_mul:
1098 case lir_div:
1099 {
1100 assert(op->as_Op2() != NULL, "must be LIR_Op2");
1101 LIR_Op2* op2 = (LIR_Op2*)op;
1102 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
1103 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");
1104 return shouldHaveRegister;
1105 }
1106 }
1107 default:
1108 break;
1109 }
1110 } else {
1111 // FPU stack float instruction
1112 switch (op->code()) {
1113 case lir_add:
1114 case lir_sub:
1115 case lir_mul:
1116 case lir_div:
1117 {
1118 assert(op->as_Op2() != NULL, "must be LIR_Op2");
1119 LIR_Op2* op2 = (LIR_Op2*)op;
1120 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
1121 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");
1122 return shouldHaveRegister;
1123 }
1124 }
1125 default:
1126 break;
1127 }
1128 }
1129 // We want to sometimes use logical operations on pointers, in particular in GC barriers.
1130 // Since 64bit logical operations do not current support operands on stack, we have to make sure
1131 // T_OBJECT doesn't get spilled along with T_LONG.
1132 } else if (opr_type != T_LONG LP64_ONLY(&& opr_type != T_OBJECT)) {
1133 // integer instruction (note: long operands must always be in register)
1134 switch (op->code()) {
1135 case lir_cmp:
1136 case lir_add:
1137 case lir_sub:
1138 case lir_logic_and:
1139 case lir_logic_or:
1140 case lir_logic_xor:
1141 {
1142 assert(op->as_Op2() != NULL, "must be LIR_Op2");
1143 LIR_Op2* op2 = (LIR_Op2*)op;
1144 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
1145 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");
1146 return shouldHaveRegister;
1147 }
1148 }
1149 default:
1150 break;
1151 }
1152 }
1153 #endif // X86 S390
1154
1155 // all other operands require a register
1156 return mustHaveRegister;
1157 }
1158
1159
1160 void LinearScan::handle_method_arguments(LIR_Op* op) {
1161 // special handling for method arguments (moves from stack to virtual register):
1162 // the interval gets no register assigned, but the stack slot.
1163 // it is split before the first use by the register allocator.
1164
1165 if (op->code() == lir_move) {
1166 assert(op->as_Op1() != NULL, "must be LIR_Op1");
1167 LIR_Op1* move = (LIR_Op1*)op;
1168
1169 if (move->in_opr()->is_stack()) {
1170 #ifdef ASSERT
1171 int arg_size = compilation()->method()->arg_size();
1172 LIR_Opr o = move->in_opr();
1173 if (o->is_single_stack()) {
1174 assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range");
1175 } else if (o->is_double_stack()) {
1176 assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range");
1177 } else {
1178 ShouldNotReachHere();
1179 }
1180
1181 assert(move->id() > 0, "invalid id");
1182 assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block");
1183 assert(move->result_opr()->is_virtual(), "result of move must be a virtual register");
1184
1185 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())));
1186 #endif
1187
1188 Interval* interval = interval_at(reg_num(move->result_opr()));
1189
1190 int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix());
1191 interval->set_canonical_spill_slot(stack_slot);
1192 interval->assign_reg(stack_slot);
1193 }
1194 }
1195 }
1196
1197 void LinearScan::handle_doubleword_moves(LIR_Op* op) {
1198 // special handling for doubleword move from memory to register:
1199 // in this case the registers of the input address and the result
1200 // registers must not overlap -> add a temp range for the input registers
1201 if (op->code() == lir_move) {
1202 assert(op->as_Op1() != NULL, "must be LIR_Op1");
1203 LIR_Op1* move = (LIR_Op1*)op;
1204
1205 if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) {
1206 LIR_Address* address = move->in_opr()->as_address_ptr();
1207 if (address != NULL) {
1208 if (address->base()->is_valid()) {
1209 add_temp(address->base(), op->id(), noUse);
1210 }
1211 if (address->index()->is_valid()) {
1212 add_temp(address->index(), op->id(), noUse);
1213 }
1214 }
1215 }
1216 }
1217 }
1218
1219 void LinearScan::add_register_hints(LIR_Op* op) {
1220 switch (op->code()) {
1221 case lir_move: // fall through
1222 case lir_convert: {
1223 assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1");
1224 LIR_Op1* move = (LIR_Op1*)op;
1225
1226 LIR_Opr move_from = move->in_opr();
1227 LIR_Opr move_to = move->result_opr();
1228
1229 if (move_to->is_register() && move_from->is_register()) {
1230 Interval* from = interval_at(reg_num(move_from));
1231 Interval* to = interval_at(reg_num(move_to));
1232 if (from != NULL && to != NULL) {
1233 to->set_register_hint(from);
1234 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()));
1235 }
1236 }
1237 break;
1238 }
1239 case lir_cmove: {
1240 assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2");
1241 LIR_Op2* cmove = (LIR_Op2*)op;
1242
1243 LIR_Opr move_from = cmove->in_opr1();
1244 LIR_Opr move_to = cmove->result_opr();
1245
1246 if (move_to->is_register() && move_from->is_register()) {
1247 Interval* from = interval_at(reg_num(move_from));
1248 Interval* to = interval_at(reg_num(move_to));
1249 if (from != NULL && to != NULL) {
1250 to->set_register_hint(from);
1251 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()));
1252 }
1253 }
1254 break;
1255 }
1256 default:
1257 break;
1258 }
1259 }
1260
1261
1262 void LinearScan::build_intervals() {
1263 TIME_LINEAR_SCAN(timer_build_intervals);
1264
1265 // initialize interval list with expected number of intervals
1266 // (32 is added to have some space for split children without having to resize the list)
1267 _intervals = IntervalList(num_virtual_regs() + 32);
1268 // initialize all slots that are used by build_intervals
1269 _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL);
1270
1271 // create a list with all caller-save registers (cpu, fpu, xmm)
1272 // when an instruction is a call, a temp range is created for all these registers
1273 int num_caller_save_registers = 0;
1274 int caller_save_registers[LinearScan::nof_regs];
1275
1276 int i;
1277 for (i = 0; i < FrameMap::nof_caller_save_cpu_regs(); i++) {
1278 LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i);
1279 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
1280 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
1281 caller_save_registers[num_caller_save_registers++] = reg_num(opr);
1282 }
1283
1284 // temp ranges for fpu registers are only created when the method has
1285 // virtual fpu operands. Otherwise no allocation for fpu registers is
1286 // perfomed and so the temp ranges would be useless
1287 if (has_fpu_registers()) {
1288 #ifdef X86
1289 if (UseSSE < 2) {
1290 #endif
1291 for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) {
1292 LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i);
1293 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
1294 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
1295 caller_save_registers[num_caller_save_registers++] = reg_num(opr);
1296 }
1297 #ifdef X86
1298 }
1299 if (UseSSE > 0) {
1300 int num_caller_save_xmm_regs = FrameMap::get_num_caller_save_xmms();
1301 for (i = 0; i < num_caller_save_xmm_regs; i ++) {
1302 LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i);
1303 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
1304 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
1305 caller_save_registers[num_caller_save_registers++] = reg_num(opr);
1306 }
1307 }
1308 #endif
1309 }
1310 assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds");
1311
1312
1313 LIR_OpVisitState visitor;
1314
1315 // iterate all blocks in reverse order
1316 for (i = block_count() - 1; i >= 0; i--) {
1317 BlockBegin* block = block_at(i);
1318 LIR_OpList* instructions = block->lir()->instructions_list();
1319 int block_from = block->first_lir_instruction_id();
1320 int block_to = block->last_lir_instruction_id();
1321
1322 assert(block_from == instructions->at(0)->id(), "must be");
1323 assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be");
1324
1325 // Update intervals for registers live at the end of this block;
1326 ResourceBitMap live = block->live_out();
1327 int size = (int)live.size();
1328 for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) {
1329 assert(live.at(number), "should not stop here otherwise");
1330 assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds");
1331 TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2));
1332
1333 add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL);
1334
1335 // add special use positions for loop-end blocks when the
1336 // interval is used anywhere inside this loop. It's possible
1337 // that the block was part of a non-natural loop, so it might
1338 // have an invalid loop index.
1339 if (block->is_set(BlockBegin::linear_scan_loop_end_flag) &&
1340 block->loop_index() != -1 &&
1341 is_interval_in_loop(number, block->loop_index())) {
1342 interval_at(number)->add_use_pos(block_to + 1, loopEndMarker);
1343 }
1344 }
1345
1346 // iterate all instructions of the block in reverse order.
1347 // skip the first instruction because it is always a label
1348 // definitions of intervals are processed before uses
1349 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label");
1350 for (int j = instructions->length() - 1; j >= 1; j--) {
1351 LIR_Op* op = instructions->at(j);
1352 int op_id = op->id();
1353
1354 // visit operation to collect all operands
1355 visitor.visit(op);
1356
1357 // add a temp range for each register if operation destroys caller-save registers
1358 if (visitor.has_call()) {
1359 for (int k = 0; k < num_caller_save_registers; k++) {
1360 add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL);
1361 }
1362 TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers"));
1363 }
1364
1365 // Add any platform dependent temps
1366 pd_add_temps(op);
1367
1368 // visit definitions (output and temp operands)
1369 int k, n;
1370 n = visitor.opr_count(LIR_OpVisitState::outputMode);
1371 for (k = 0; k < n; k++) {
1372 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k);
1373 assert(opr->is_register(), "visitor should only return register operands");
1374 add_def(opr, op_id, use_kind_of_output_operand(op, opr));
1375 }
1376
1377 n = visitor.opr_count(LIR_OpVisitState::tempMode);
1378 for (k = 0; k < n; k++) {
1379 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k);
1380 assert(opr->is_register(), "visitor should only return register operands");
1381 add_temp(opr, op_id, mustHaveRegister);
1382 }
1383
1384 // visit uses (input operands)
1385 n = visitor.opr_count(LIR_OpVisitState::inputMode);
1386 for (k = 0; k < n; k++) {
1387 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k);
1388 assert(opr->is_register(), "visitor should only return register operands");
1389 add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr));
1390 }
1391
1392 // Add uses of live locals from interpreter's point of view for proper
1393 // debug information generation
1394 // Treat these operands as temp values (if the life range is extended
1395 // to a call site, the value would be in a register at the call otherwise)
1396 n = visitor.info_count();
1397 for (k = 0; k < n; k++) {
1398 CodeEmitInfo* info = visitor.info_at(k);
1399 ValueStack* stack = info->stack();
1400 for_each_state_value(stack, value,
1401 add_use(value, block_from, op_id + 1, noUse);
1402 );
1403 }
1404
1405 // special steps for some instructions (especially moves)
1406 handle_method_arguments(op);
1407 handle_doubleword_moves(op);
1408 add_register_hints(op);
1409
1410 } // end of instruction iteration
1411 } // end of block iteration
1412
1413
1414 // add the range [0, 1[ to all fixed intervals
1415 // -> the register allocator need not handle unhandled fixed intervals
1416 for (int n = 0; n < LinearScan::nof_regs; n++) {
1417 Interval* interval = interval_at(n);
1418 if (interval != NULL) {
1419 interval->add_range(0, 1);
1420 }
1421 }
1422 }
1423
1424
1425 // ********** Phase 5: actual register allocation
1426
1427 int LinearScan::interval_cmp(Interval** a, Interval** b) {
1428 if (*a != NULL) {
1429 if (*b != NULL) {
1430 return (*a)->from() - (*b)->from();
1431 } else {
1432 return -1;
1433 }
1434 } else {
1435 if (*b != NULL) {
1436 return 1;
1437 } else {
1438 return 0;
1439 }
1440 }
1441 }
1442
1443 #ifndef PRODUCT
1444 int interval_cmp(Interval* const& l, Interval* const& r) {
1445 return l->from() - r->from();
1446 }
1447
1448 bool find_interval(Interval* interval, IntervalArray* intervals) {
1449 bool found;
1450 int idx = intervals->find_sorted<Interval*, interval_cmp>(interval, found);
1451
1452 if (!found) {
1453 return false;
1454 }
1455
1456 int from = interval->from();
1457
1458 // The index we've found using binary search is pointing to an interval
1459 // that is defined in the same place as the interval we were looking for.
1460 // So now we have to look around that index and find exact interval.
1461 for (int i = idx; i >= 0; i--) {
1462 if (intervals->at(i) == interval) {
1463 return true;
1464 }
1465 if (intervals->at(i)->from() != from) {
1466 break;
1467 }
1468 }
1469
1470 for (int i = idx + 1; i < intervals->length(); i++) {
1471 if (intervals->at(i) == interval) {
1472 return true;
1473 }
1474 if (intervals->at(i)->from() != from) {
1475 break;
1476 }
1477 }
1478
1479 return false;
1480 }
1481
1482 bool LinearScan::is_sorted(IntervalArray* intervals) {
1483 int from = -1;
1484 int null_count = 0;
1485
1486 for (int i = 0; i < intervals->length(); i++) {
1487 Interval* it = intervals->at(i);
1488 if (it != NULL) {
1489 assert(from <= it->from(), "Intervals are unordered");
1490 from = it->from();
1491 } else {
1492 null_count++;
1493 }
1494 }
1495
1496 assert(null_count == 0, "Sorted intervals should not contain nulls");
1497
1498 null_count = 0;
1499
1500 for (int i = 0; i < interval_count(); i++) {
1501 Interval* interval = interval_at(i);
1502 if (interval != NULL) {
1503 assert(find_interval(interval, intervals), "Lists do not contain same intervals");
1504 } else {
1505 null_count++;
1506 }
1507 }
1508
1509 assert(interval_count() - null_count == intervals->length(),
1510 "Sorted list should contain the same amount of non-NULL intervals as unsorted list");
1511
1512 return true;
1513 }
1514 #endif
1515
1516 void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) {
1517 if (*prev != NULL) {
1518 (*prev)->set_next(interval);
1519 } else {
1520 *first = interval;
1521 }
1522 *prev = interval;
1523 }
1524
1525 void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) {
1526 assert(is_sorted(_sorted_intervals), "interval list is not sorted");
1527
1528 *list1 = *list2 = Interval::end();
1529
1530 Interval* list1_prev = NULL;
1531 Interval* list2_prev = NULL;
1532 Interval* v;
1533
1534 const int n = _sorted_intervals->length();
1535 for (int i = 0; i < n; i++) {
1536 v = _sorted_intervals->at(i);
1537 if (v == NULL) continue;
1538
1539 if (is_list1(v)) {
1540 add_to_list(list1, &list1_prev, v);
1541 } else if (is_list2 == NULL || is_list2(v)) {
1542 add_to_list(list2, &list2_prev, v);
1543 }
1544 }
1545
1546 if (list1_prev != NULL) list1_prev->set_next(Interval::end());
1547 if (list2_prev != NULL) list2_prev->set_next(Interval::end());
1548
1549 assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel");
1550 assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel");
1551 }
1552
1553
1554 void LinearScan::sort_intervals_before_allocation() {
1555 TIME_LINEAR_SCAN(timer_sort_intervals_before);
1556
1557 if (_needs_full_resort) {
1558 // There is no known reason why this should occur but just in case...
1559 assert(false, "should never occur");
1560 // Re-sort existing interval list because an Interval::from() has changed
1561 _sorted_intervals->sort(interval_cmp);
1562 _needs_full_resort = false;
1563 }
1564
1565 IntervalList* unsorted_list = &_intervals;
1566 int unsorted_len = unsorted_list->length();
1567 int sorted_len = 0;
1568 int unsorted_idx;
1569 int sorted_idx = 0;
1570 int sorted_from_max = -1;
1571
1572 // calc number of items for sorted list (sorted list must not contain NULL values)
1573 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) {
1574 if (unsorted_list->at(unsorted_idx) != NULL) {
1575 sorted_len++;
1576 }
1577 }
1578 IntervalArray* sorted_list = new IntervalArray(sorted_len, sorted_len, NULL);
1579
1580 // special sorting algorithm: the original interval-list is almost sorted,
1581 // only some intervals are swapped. So this is much faster than a complete QuickSort
1582 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) {
1583 Interval* cur_interval = unsorted_list->at(unsorted_idx);
1584
1585 if (cur_interval != NULL) {
1586 int cur_from = cur_interval->from();
1587
1588 if (sorted_from_max <= cur_from) {
1589 sorted_list->at_put(sorted_idx++, cur_interval);
1590 sorted_from_max = cur_interval->from();
1591 } else {
1592 // the asumption that the intervals are already sorted failed,
1593 // so this interval must be sorted in manually
1594 int j;
1595 for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) {
1596 sorted_list->at_put(j + 1, sorted_list->at(j));
1597 }
1598 sorted_list->at_put(j + 1, cur_interval);
1599 sorted_idx++;
1600 }
1601 }
1602 }
1603 _sorted_intervals = sorted_list;
1604 assert(is_sorted(_sorted_intervals), "intervals unsorted");
1605 }
1606
1607 void LinearScan::sort_intervals_after_allocation() {
1608 TIME_LINEAR_SCAN(timer_sort_intervals_after);
1609
1610 if (_needs_full_resort) {
1611 // Re-sort existing interval list because an Interval::from() has changed
1612 _sorted_intervals->sort(interval_cmp);
1613 _needs_full_resort = false;
1614 }
1615
1616 IntervalArray* old_list = _sorted_intervals;
1617 IntervalList* new_list = _new_intervals_from_allocation;
1618 int old_len = old_list->length();
1619 int new_len = new_list->length();
1620
1621 if (new_len == 0) {
1622 // no intervals have been added during allocation, so sorted list is already up to date
1623 assert(is_sorted(_sorted_intervals), "intervals unsorted");
1624 return;
1625 }
1626
1627 // conventional sort-algorithm for new intervals
1628 new_list->sort(interval_cmp);
1629
1630 // merge old and new list (both already sorted) into one combined list
1631 int combined_list_len = old_len + new_len;
1632 IntervalArray* combined_list = new IntervalArray(combined_list_len, combined_list_len, NULL);
1633 int old_idx = 0;
1634 int new_idx = 0;
1635
1636 while (old_idx + new_idx < old_len + new_len) {
1637 if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) {
1638 combined_list->at_put(old_idx + new_idx, old_list->at(old_idx));
1639 old_idx++;
1640 } else {
1641 combined_list->at_put(old_idx + new_idx, new_list->at(new_idx));
1642 new_idx++;
1643 }
1644 }
1645
1646 _sorted_intervals = combined_list;
1647 assert(is_sorted(_sorted_intervals), "intervals unsorted");
1648 }
1649
1650
1651 void LinearScan::allocate_registers() {
1652 TIME_LINEAR_SCAN(timer_allocate_registers);
1653
1654 Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals;
1655 Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals;
1656
1657 // allocate cpu registers
1658 create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals,
1659 is_precolored_cpu_interval, is_virtual_cpu_interval);
1660
1661 // allocate fpu registers
1662 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals,
1663 is_precolored_fpu_interval, is_virtual_fpu_interval);
1664
1665 // the fpu interval allocation cannot be moved down below with the fpu section as
1666 // the cpu_lsw.walk() changes interval positions.
1667
1668 LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals);
1669 cpu_lsw.walk();
1670 cpu_lsw.finish_allocation();
1671
1672 if (has_fpu_registers()) {
1673 LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals);
1674 fpu_lsw.walk();
1675 fpu_lsw.finish_allocation();
1676 }
1677 }
1678
1679
1680 // ********** Phase 6: resolve data flow
1681 // (insert moves at edges between blocks if intervals have been split)
1682
1683 // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode
1684 // instead of returning NULL
1685 Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) {
1686 Interval* result = interval->split_child_at_op_id(op_id, mode);
1687 if (result != NULL) {
1688 return result;
1689 }
1690
1691 assert(false, "must find an interval, but do a clean bailout in product mode");
1692 result = new Interval(LIR_OprDesc::vreg_base);
1693 result->assign_reg(0);
1694 result->set_type(T_INT);
1695 BAILOUT_("LinearScan: interval is NULL", result);
1696 }
1697
1698
1699 Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) {
1700 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
1701 assert(interval_at(reg_num) != NULL, "no interval found");
1702
1703 return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode);
1704 }
1705
1706 Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) {
1707 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
1708 assert(interval_at(reg_num) != NULL, "no interval found");
1709
1710 return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode);
1711 }
1712
1713 Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) {
1714 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
1715 assert(interval_at(reg_num) != NULL, "no interval found");
1716
1717 return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode);
1718 }
1719
1720
1721 void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) {
1722 DEBUG_ONLY(move_resolver.check_empty());
1723
1724 const int num_regs = num_virtual_regs();
1725 const int size = live_set_size();
1726 const ResourceBitMap live_at_edge = to_block->live_in();
1727
1728 // visit all registers where the live_at_edge bit is set
1729 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)) {
1730 assert(r < num_regs, "live information set for not exisiting interval");
1731 assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge");
1732
1733 Interval* from_interval = interval_at_block_end(from_block, r);
1734 Interval* to_interval = interval_at_block_begin(to_block, r);
1735
1736 if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) {
1737 // need to insert move instruction
1738 move_resolver.add_mapping(from_interval, to_interval);
1739 }
1740 }
1741 }
1742
1743
1744 void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) {
1745 if (from_block->number_of_sux() <= 1) {
1746 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id()));
1747
1748 LIR_OpList* instructions = from_block->lir()->instructions_list();
1749 LIR_OpBranch* branch = instructions->last()->as_OpBranch();
1750 if (branch != NULL) {
1751 // insert moves before branch
1752 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump");
1753 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2);
1754 } else {
1755 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1);
1756 }
1757
1758 } else {
1759 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id()));
1760 #ifdef ASSERT
1761 assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label");
1762
1763 // because the number of predecessor edges matches the number of
1764 // successor edges, blocks which are reached by switch statements
1765 // may have be more than one predecessor but it will be guaranteed
1766 // that all predecessors will be the same.
1767 for (int i = 0; i < to_block->number_of_preds(); i++) {
1768 assert(from_block == to_block->pred_at(i), "all critical edges must be broken");
1769 }
1770 #endif
1771
1772 move_resolver.set_insert_position(to_block->lir(), 0);
1773 }
1774 }
1775
1776
1777 // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split
1778 void LinearScan::resolve_data_flow() {
1779 TIME_LINEAR_SCAN(timer_resolve_data_flow);
1780
1781 int num_blocks = block_count();
1782 MoveResolver move_resolver(this);
1783 ResourceBitMap block_completed(num_blocks);
1784 ResourceBitMap already_resolved(num_blocks);
1785
1786 int i;
1787 for (i = 0; i < num_blocks; i++) {
1788 BlockBegin* block = block_at(i);
1789
1790 // check if block has only one predecessor and only one successor
1791 if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) {
1792 LIR_OpList* instructions = block->lir()->instructions_list();
1793 assert(instructions->at(0)->code() == lir_label, "block must start with label");
1794 assert(instructions->last()->code() == lir_branch, "block with successors must end with branch");
1795 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch");
1796
1797 // check if block is empty (only label and branch)
1798 if (instructions->length() == 2) {
1799 BlockBegin* pred = block->pred_at(0);
1800 BlockBegin* sux = block->sux_at(0);
1801
1802 // prevent optimization of two consecutive blocks
1803 if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) {
1804 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()));
1805 block_completed.set_bit(block->linear_scan_number());
1806
1807 // directly resolve between pred and sux (without looking at the empty block between)
1808 resolve_collect_mappings(pred, sux, move_resolver);
1809 if (move_resolver.has_mappings()) {
1810 move_resolver.set_insert_position(block->lir(), 0);
1811 move_resolver.resolve_and_append_moves();
1812 }
1813 }
1814 }
1815 }
1816 }
1817
1818
1819 for (i = 0; i < num_blocks; i++) {
1820 if (!block_completed.at(i)) {
1821 BlockBegin* from_block = block_at(i);
1822 already_resolved.set_from(block_completed);
1823
1824 int num_sux = from_block->number_of_sux();
1825 for (int s = 0; s < num_sux; s++) {
1826 BlockBegin* to_block = from_block->sux_at(s);
1827
1828 // check for duplicate edges between the same blocks (can happen with switch blocks)
1829 if (!already_resolved.at(to_block->linear_scan_number())) {
1830 TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id()));
1831 already_resolved.set_bit(to_block->linear_scan_number());
1832
1833 // collect all intervals that have been split between from_block and to_block
1834 resolve_collect_mappings(from_block, to_block, move_resolver);
1835 if (move_resolver.has_mappings()) {
1836 resolve_find_insert_pos(from_block, to_block, move_resolver);
1837 move_resolver.resolve_and_append_moves();
1838 }
1839 }
1840 }
1841 }
1842 }
1843 }
1844
1845
1846 void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, 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 Interval* interval = interval_at_block_begin(block, reg_num);
1853 int reg = interval->assigned_reg();
1854 int regHi = interval->assigned_regHi();
1855
1856 if ((reg < nof_regs && interval->always_in_memory()) ||
1857 (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) {
1858 // the interval is split to get a short range that is located on the stack
1859 // in the following two cases:
1860 // * the interval started in memory (e.g. method parameter), but is currently in a register
1861 // this is an optimization for exception handling that reduces the number of moves that
1862 // are necessary for resolving the states when an exception uses this exception handler
1863 // * the interval would be on the fpu stack at the begin of the exception handler
1864 // this is not allowed because of the complicated fpu stack handling on Intel
1865
1866 // range that will be spilled to memory
1867 int from_op_id = block->first_lir_instruction_id();
1868 int to_op_id = from_op_id + 1; // short live range of length 1
1869 assert(interval->from() <= from_op_id && interval->to() >= to_op_id,
1870 "no split allowed between exception entry and first instruction");
1871
1872 if (interval->from() != from_op_id) {
1873 // the part before from_op_id is unchanged
1874 interval = interval->split(from_op_id);
1875 interval->assign_reg(reg, regHi);
1876 append_interval(interval);
1877 } else {
1878 _needs_full_resort = true;
1879 }
1880 assert(interval->from() == from_op_id, "must be true now");
1881
1882 Interval* spilled_part = interval;
1883 if (interval->to() != to_op_id) {
1884 // the part after to_op_id is unchanged
1885 spilled_part = interval->split_from_start(to_op_id);
1886 append_interval(spilled_part);
1887 move_resolver.add_mapping(spilled_part, interval);
1888 }
1889 assign_spill_slot(spilled_part);
1890
1891 assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking");
1892 }
1893 }
1894
1895 void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) {
1896 assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise");
1897 DEBUG_ONLY(move_resolver.check_empty());
1898
1899 // visit all registers where the live_in bit is set
1900 int size = live_set_size();
1901 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)) {
1902 resolve_exception_entry(block, r, move_resolver);
1903 }
1904
1905 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately
1906 for_each_phi_fun(block, phi,
1907 resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver)
1908 );
1909
1910 if (move_resolver.has_mappings()) {
1911 // insert moves after first instruction
1912 move_resolver.set_insert_position(block->lir(), 0);
1913 move_resolver.resolve_and_append_moves();
1914 }
1915 }
1916
1917
1918 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) {
1919 if (interval_at(reg_num) == NULL) {
1920 // if a phi function is never used, no interval is created -> ignore this
1921 return;
1922 }
1923
1924 // the computation of to_interval is equal to resolve_collect_mappings,
1925 // but from_interval is more complicated because of phi functions
1926 BlockBegin* to_block = handler->entry_block();
1927 Interval* to_interval = interval_at_block_begin(to_block, reg_num);
1928
1929 if (phi != NULL) {
1930 // phi function of the exception entry block
1931 // no moves are created for this phi function in the LIR_Generator, so the
1932 // interval at the throwing instruction must be searched using the operands
1933 // of the phi function
1934 Value from_value = phi->operand_at(handler->phi_operand());
1935
1936 // with phi functions it can happen that the same from_value is used in
1937 // multiple mappings, so notify move-resolver that this is allowed
1938 move_resolver.set_multiple_reads_allowed();
1939
1940 Constant* con = from_value->as_Constant();
1941 if (con != NULL && !con->is_pinned()) {
1942 // unpinned constants may have no register, so add mapping from constant to interval
1943 move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval);
1944 } else {
1945 // search split child at the throwing op_id
1946 Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id);
1947 move_resolver.add_mapping(from_interval, to_interval);
1948 }
1949
1950 } else {
1951 // no phi function, so use reg_num also for from_interval
1952 // search split child at the throwing op_id
1953 Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id);
1954 if (from_interval != to_interval) {
1955 // optimization to reduce number of moves: when to_interval is on stack and
1956 // the stack slot is known to be always correct, then no move is necessary
1957 if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) {
1958 move_resolver.add_mapping(from_interval, to_interval);
1959 }
1960 }
1961 }
1962 }
1963
1964 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) {
1965 TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id));
1966
1967 DEBUG_ONLY(move_resolver.check_empty());
1968 assert(handler->lir_op_id() == -1, "already processed this xhandler");
1969 DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id));
1970 assert(handler->entry_code() == NULL, "code already present");
1971
1972 // visit all registers where the live_in bit is set
1973 BlockBegin* block = handler->entry_block();
1974 int size = live_set_size();
1975 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)) {
1976 resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver);
1977 }
1978
1979 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately
1980 for_each_phi_fun(block, phi,
1981 resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver)
1982 );
1983
1984 if (move_resolver.has_mappings()) {
1985 LIR_List* entry_code = new LIR_List(compilation());
1986 move_resolver.set_insert_position(entry_code, 0);
1987 move_resolver.resolve_and_append_moves();
1988
1989 entry_code->jump(handler->entry_block());
1990 handler->set_entry_code(entry_code);
1991 }
1992 }
1993
1994
1995 void LinearScan::resolve_exception_handlers() {
1996 MoveResolver move_resolver(this);
1997 LIR_OpVisitState visitor;
1998 int num_blocks = block_count();
1999
2000 int i;
2001 for (i = 0; i < num_blocks; i++) {
2002 BlockBegin* block = block_at(i);
2003 if (block->is_set(BlockBegin::exception_entry_flag)) {
2004 resolve_exception_entry(block, move_resolver);
2005 }
2006 }
2007
2008 for (i = 0; i < num_blocks; i++) {
2009 BlockBegin* block = block_at(i);
2010 LIR_List* ops = block->lir();
2011 int num_ops = ops->length();
2012
2013 // iterate all instructions of the block. skip the first because it is always a label
2014 assert(visitor.no_operands(ops->at(0)), "first operation must always be a label");
2015 for (int j = 1; j < num_ops; j++) {
2016 LIR_Op* op = ops->at(j);
2017 int op_id = op->id();
2018
2019 if (op_id != -1 && has_info(op_id)) {
2020 // visit operation to collect all operands
2021 visitor.visit(op);
2022 assert(visitor.info_count() > 0, "should not visit otherwise");
2023
2024 XHandlers* xhandlers = visitor.all_xhandler();
2025 int n = xhandlers->length();
2026 for (int k = 0; k < n; k++) {
2027 resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver);
2028 }
2029
2030 #ifdef ASSERT
2031 } else {
2032 visitor.visit(op);
2033 assert(visitor.all_xhandler()->length() == 0, "missed exception handler");
2034 #endif
2035 }
2036 }
2037 }
2038 }
2039
2040
2041 // ********** Phase 7: assign register numbers back to LIR
2042 // (includes computation of debug information and oop maps)
2043
2044 VMReg LinearScan::vm_reg_for_interval(Interval* interval) {
2045 VMReg reg = interval->cached_vm_reg();
2046 if (!reg->is_valid() ) {
2047 reg = vm_reg_for_operand(operand_for_interval(interval));
2048 interval->set_cached_vm_reg(reg);
2049 }
2050 assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value");
2051 return reg;
2052 }
2053
2054 VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) {
2055 assert(opr->is_oop(), "currently only implemented for oop operands");
2056 return frame_map()->regname(opr);
2057 }
2058
2059
2060 LIR_Opr LinearScan::operand_for_interval(Interval* interval) {
2061 LIR_Opr opr = interval->cached_opr();
2062 if (opr->is_illegal()) {
2063 opr = calc_operand_for_interval(interval);
2064 interval->set_cached_opr(opr);
2065 }
2066
2067 assert(opr == calc_operand_for_interval(interval), "wrong cached value");
2068 return opr;
2069 }
2070
2071 LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) {
2072 int assigned_reg = interval->assigned_reg();
2073 BasicType type = interval->type();
2074
2075 if (assigned_reg >= nof_regs) {
2076 // stack slot
2077 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2078 return LIR_OprFact::stack(assigned_reg - nof_regs, type);
2079
2080 } else {
2081 // register
2082 switch (type) {
2083 case T_OBJECT: {
2084 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2085 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2086 return LIR_OprFact::single_cpu_oop(assigned_reg);
2087 }
2088
2089 case T_ADDRESS: {
2090 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2091 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2092 return LIR_OprFact::single_cpu_address(assigned_reg);
2093 }
2094
2095 case T_METADATA: {
2096 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2097 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2098 return LIR_OprFact::single_cpu_metadata(assigned_reg);
2099 }
2100
2101 #ifdef __SOFTFP__
2102 case T_FLOAT: // fall through
2103 #endif // __SOFTFP__
2104 case T_INT: {
2105 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2106 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2107 return LIR_OprFact::single_cpu(assigned_reg);
2108 }
2109
2110 #ifdef __SOFTFP__
2111 case T_DOUBLE: // fall through
2112 #endif // __SOFTFP__
2113 case T_LONG: {
2114 int assigned_regHi = interval->assigned_regHi();
2115 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2116 assert(num_physical_regs(T_LONG) == 1 ||
2117 (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register");
2118
2119 assert(assigned_reg != assigned_regHi, "invalid allocation");
2120 assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi,
2121 "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)");
2122 assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match");
2123 if (requires_adjacent_regs(T_LONG)) {
2124 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even");
2125 }
2126
2127 #ifdef _LP64
2128 return LIR_OprFact::double_cpu(assigned_reg, assigned_reg);
2129 #else
2130 #if defined(SPARC) || defined(PPC32)
2131 return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg);
2132 #else
2133 return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi);
2134 #endif // SPARC
2135 #endif // LP64
2136 }
2137
2138 #ifndef __SOFTFP__
2139 case T_FLOAT: {
2140 #ifdef X86
2141 if (UseSSE >= 1) {
2142 int last_xmm_reg = pd_last_xmm_reg;
2143 #ifdef _LP64
2144 if (UseAVX < 3) {
2145 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1;
2146 }
2147 #endif
2148 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register");
2149 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2150 return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg);
2151 }
2152 #endif
2153
2154 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
2155 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2156 return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg);
2157 }
2158
2159 case T_DOUBLE: {
2160 #ifdef X86
2161 if (UseSSE >= 2) {
2162 int last_xmm_reg = pd_last_xmm_reg;
2163 #ifdef _LP64
2164 if (UseAVX < 3) {
2165 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1;
2166 }
2167 #endif
2168 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register");
2169 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)");
2170 return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg);
2171 }
2172 #endif
2173
2174 #ifdef SPARC
2175 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
2176 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register");
2177 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even");
2178 LIR_Opr result = LIR_OprFact::double_fpu(interval->assigned_regHi() - pd_first_fpu_reg, assigned_reg - pd_first_fpu_reg);
2179 #elif defined(ARM32)
2180 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
2181 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register");
2182 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even");
2183 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg, interval->assigned_regHi() - pd_first_fpu_reg);
2184 #else
2185 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
2186 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)");
2187 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg);
2188 #endif
2189 return result;
2190 }
2191 #endif // __SOFTFP__
2192
2193 default: {
2194 ShouldNotReachHere();
2195 return LIR_OprFact::illegalOpr;
2196 }
2197 }
2198 }
2199 }
2200
2201 LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) {
2202 assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set");
2203 return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type());
2204 }
2205
2206 LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) {
2207 assert(opr->is_virtual(), "should not call this otherwise");
2208
2209 Interval* interval = interval_at(opr->vreg_number());
2210 assert(interval != NULL, "interval must exist");
2211
2212 if (op_id != -1) {
2213 #ifdef ASSERT
2214 BlockBegin* block = block_of_op_with_id(op_id);
2215 if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) {
2216 // check if spill moves could have been appended at the end of this block, but
2217 // before the branch instruction. So the split child information for this branch would
2218 // be incorrect.
2219 LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch();
2220 if (branch != NULL) {
2221 if (block->live_out().at(opr->vreg_number())) {
2222 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump");
2223 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)");
2224 }
2225 }
2226 }
2227 #endif
2228
2229 // operands are not changed when an interval is split during allocation,
2230 // so search the right interval here
2231 interval = split_child_at_op_id(interval, op_id, mode);
2232 }
2233
2234 LIR_Opr res = operand_for_interval(interval);
2235
2236 #ifdef X86
2237 // new semantic for is_last_use: not only set on definite end of interval,
2238 // but also before hole
2239 // This may still miss some cases (e.g. for dead values), but it is not necessary that the
2240 // last use information is completely correct
2241 // information is only needed for fpu stack allocation
2242 if (res->is_fpu_register()) {
2243 if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) {
2244 assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow");
2245 res = res->make_last_use();
2246 }
2247 }
2248 #endif
2249
2250 assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation");
2251
2252 return res;
2253 }
2254
2255
2256 #ifdef ASSERT
2257 // some methods used to check correctness of debug information
2258
2259 void assert_no_register_values(GrowableArray<ScopeValue*>* values) {
2260 if (values == NULL) {
2261 return;
2262 }
2263
2264 for (int i = 0; i < values->length(); i++) {
2265 ScopeValue* value = values->at(i);
2266
2267 if (value->is_location()) {
2268 Location location = ((LocationValue*)value)->location();
2269 assert(location.where() == Location::on_stack, "value is in register");
2270 }
2271 }
2272 }
2273
2274 void assert_no_register_values(GrowableArray<MonitorValue*>* values) {
2275 if (values == NULL) {
2276 return;
2277 }
2278
2279 for (int i = 0; i < values->length(); i++) {
2280 MonitorValue* value = values->at(i);
2281
2282 if (value->owner()->is_location()) {
2283 Location location = ((LocationValue*)value->owner())->location();
2284 assert(location.where() == Location::on_stack, "owner is in register");
2285 }
2286 assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register");
2287 }
2288 }
2289
2290 void assert_equal(Location l1, Location l2) {
2291 assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), "");
2292 }
2293
2294 void assert_equal(ScopeValue* v1, ScopeValue* v2) {
2295 if (v1->is_location()) {
2296 assert(v2->is_location(), "");
2297 assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location());
2298 } else if (v1->is_constant_int()) {
2299 assert(v2->is_constant_int(), "");
2300 assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), "");
2301 } else if (v1->is_constant_double()) {
2302 assert(v2->is_constant_double(), "");
2303 assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), "");
2304 } else if (v1->is_constant_long()) {
2305 assert(v2->is_constant_long(), "");
2306 assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), "");
2307 } else if (v1->is_constant_oop()) {
2308 assert(v2->is_constant_oop(), "");
2309 assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), "");
2310 } else {
2311 ShouldNotReachHere();
2312 }
2313 }
2314
2315 void assert_equal(MonitorValue* m1, MonitorValue* m2) {
2316 assert_equal(m1->owner(), m2->owner());
2317 assert_equal(m1->basic_lock(), m2->basic_lock());
2318 }
2319
2320 void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) {
2321 assert(d1->scope() == d2->scope(), "not equal");
2322 assert(d1->bci() == d2->bci(), "not equal");
2323
2324 if (d1->locals() != NULL) {
2325 assert(d1->locals() != NULL && d2->locals() != NULL, "not equal");
2326 assert(d1->locals()->length() == d2->locals()->length(), "not equal");
2327 for (int i = 0; i < d1->locals()->length(); i++) {
2328 assert_equal(d1->locals()->at(i), d2->locals()->at(i));
2329 }
2330 } else {
2331 assert(d1->locals() == NULL && d2->locals() == NULL, "not equal");
2332 }
2333
2334 if (d1->expressions() != NULL) {
2335 assert(d1->expressions() != NULL && d2->expressions() != NULL, "not equal");
2336 assert(d1->expressions()->length() == d2->expressions()->length(), "not equal");
2337 for (int i = 0; i < d1->expressions()->length(); i++) {
2338 assert_equal(d1->expressions()->at(i), d2->expressions()->at(i));
2339 }
2340 } else {
2341 assert(d1->expressions() == NULL && d2->expressions() == NULL, "not equal");
2342 }
2343
2344 if (d1->monitors() != NULL) {
2345 assert(d1->monitors() != NULL && d2->monitors() != NULL, "not equal");
2346 assert(d1->monitors()->length() == d2->monitors()->length(), "not equal");
2347 for (int i = 0; i < d1->monitors()->length(); i++) {
2348 assert_equal(d1->monitors()->at(i), d2->monitors()->at(i));
2349 }
2350 } else {
2351 assert(d1->monitors() == NULL && d2->monitors() == NULL, "not equal");
2352 }
2353
2354 if (d1->caller() != NULL) {
2355 assert(d1->caller() != NULL && d2->caller() != NULL, "not equal");
2356 assert_equal(d1->caller(), d2->caller());
2357 } else {
2358 assert(d1->caller() == NULL && d2->caller() == NULL, "not equal");
2359 }
2360 }
2361
2362 void check_stack_depth(CodeEmitInfo* info, int stack_end) {
2363 if (info->stack()->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) {
2364 Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->stack()->bci());
2365 switch (code) {
2366 case Bytecodes::_ifnull : // fall through
2367 case Bytecodes::_ifnonnull : // fall through
2368 case Bytecodes::_ifeq : // fall through
2369 case Bytecodes::_ifne : // fall through
2370 case Bytecodes::_iflt : // fall through
2371 case Bytecodes::_ifge : // fall through
2372 case Bytecodes::_ifgt : // fall through
2373 case Bytecodes::_ifle : // fall through
2374 case Bytecodes::_if_icmpeq : // fall through
2375 case Bytecodes::_if_icmpne : // fall through
2376 case Bytecodes::_if_icmplt : // fall through
2377 case Bytecodes::_if_icmpge : // fall through
2378 case Bytecodes::_if_icmpgt : // fall through
2379 case Bytecodes::_if_icmple : // fall through
2380 case Bytecodes::_if_acmpeq : // fall through
2381 case Bytecodes::_if_acmpne :
2382 assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode");
2383 break;
2384 default:
2385 break;
2386 }
2387 }
2388 }
2389
2390 #endif // ASSERT
2391
2392
2393 IntervalWalker* LinearScan::init_compute_oop_maps() {
2394 // setup lists of potential oops for walking
2395 Interval* oop_intervals;
2396 Interval* non_oop_intervals;
2397
2398 create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, NULL);
2399
2400 // intervals that have no oops inside need not to be processed
2401 // to ensure a walking until the last instruction id, add a dummy interval
2402 // with a high operation id
2403 non_oop_intervals = new Interval(any_reg);
2404 non_oop_intervals->add_range(max_jint - 2, max_jint - 1);
2405
2406 return new IntervalWalker(this, oop_intervals, non_oop_intervals);
2407 }
2408
2409
2410 OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) {
2411 TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d", op->id()));
2412
2413 // walk before the current operation -> intervals that start at
2414 // the operation (= output operands of the operation) are not
2415 // included in the oop map
2416 iw->walk_before(op->id());
2417
2418 int frame_size = frame_map()->framesize();
2419 int arg_count = frame_map()->oop_map_arg_count();
2420 OopMap* map = new OopMap(frame_size, arg_count);
2421
2422 // Iterate through active intervals
2423 for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) {
2424 int assigned_reg = interval->assigned_reg();
2425
2426 assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise");
2427 assert(interval->assigned_regHi() == any_reg, "oop must be single word");
2428 assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found");
2429
2430 // Check if this range covers the instruction. Intervals that
2431 // start or end at the current operation are not included in the
2432 // oop map, except in the case of patching moves. For patching
2433 // moves, any intervals which end at this instruction are included
2434 // in the oop map since we may safepoint while doing the patch
2435 // before we've consumed the inputs.
2436 if (op->is_patching() || op->id() < interval->current_to()) {
2437
2438 // caller-save registers must not be included into oop-maps at calls
2439 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");
2440
2441 VMReg name = vm_reg_for_interval(interval);
2442 set_oop(map, name);
2443
2444 // Spill optimization: when the stack value is guaranteed to be always correct,
2445 // then it must be added to the oop map even if the interval is currently in a register
2446 if (interval->always_in_memory() &&
2447 op->id() > interval->spill_definition_pos() &&
2448 interval->assigned_reg() != interval->canonical_spill_slot()) {
2449 assert(interval->spill_definition_pos() > 0, "position not set correctly");
2450 assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned");
2451 assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice");
2452
2453 set_oop(map, frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs));
2454 }
2455 }
2456 }
2457
2458 // add oops from lock stack
2459 assert(info->stack() != NULL, "CodeEmitInfo must always have a stack");
2460 int locks_count = info->stack()->total_locks_size();
2461 for (int i = 0; i < locks_count; i++) {
2462 set_oop(map, frame_map()->monitor_object_regname(i));
2463 }
2464
2465 return map;
2466 }
2467
2468
2469 void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) {
2470 assert(visitor.info_count() > 0, "no oop map needed");
2471
2472 // compute oop_map only for first CodeEmitInfo
2473 // because it is (in most cases) equal for all other infos of the same operation
2474 CodeEmitInfo* first_info = visitor.info_at(0);
2475 OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call());
2476
2477 for (int i = 0; i < visitor.info_count(); i++) {
2478 CodeEmitInfo* info = visitor.info_at(i);
2479 OopMap* oop_map = first_oop_map;
2480
2481 // compute worst case interpreter size in case of a deoptimization
2482 _compilation->update_interpreter_frame_size(info->interpreter_frame_size());
2483
2484 if (info->stack()->locks_size() != first_info->stack()->locks_size()) {
2485 // this info has a different number of locks then the precomputed oop map
2486 // (possible for lock and unlock instructions) -> compute oop map with
2487 // correct lock information
2488 oop_map = compute_oop_map(iw, op, info, visitor.has_call());
2489 }
2490
2491 if (info->_oop_map == NULL) {
2492 info->_oop_map = oop_map;
2493 } else {
2494 // a CodeEmitInfo can not be shared between different LIR-instructions
2495 // because interval splitting can occur anywhere between two instructions
2496 // and so the oop maps must be different
2497 // -> check if the already set oop_map is exactly the one calculated for this operation
2498 assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions");
2499 }
2500 }
2501 }
2502
2503
2504 // frequently used constants
2505 // Allocate them with new so they are never destroyed (otherwise, a
2506 // forced exit could destroy these objects while they are still in
2507 // use).
2508 ConstantOopWriteValue* LinearScan::_oop_null_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantOopWriteValue(NULL);
2509 ConstantIntValue* LinearScan::_int_m1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(-1);
2510 ConstantIntValue* LinearScan::_int_0_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue((jint)0);
2511 ConstantIntValue* LinearScan::_int_1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(1);
2512 ConstantIntValue* LinearScan::_int_2_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(2);
2513 LocationValue* _illegal_value = new (ResourceObj::C_HEAP, mtCompiler) LocationValue(Location());
2514
2515 void LinearScan::init_compute_debug_info() {
2516 // cache for frequently used scope values
2517 // (cpu registers and stack slots)
2518 int cache_size = (LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2;
2519 _scope_value_cache = ScopeValueArray(cache_size, cache_size, NULL);
2520 }
2521
2522 MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) {
2523 Location loc;
2524 if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) {
2525 bailout("too large frame");
2526 }
2527 ScopeValue* object_scope_value = new LocationValue(loc);
2528
2529 if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) {
2530 bailout("too large frame");
2531 }
2532 return new MonitorValue(object_scope_value, loc);
2533 }
2534
2535 LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) {
2536 Location loc;
2537 if (!frame_map()->locations_for_slot(name, loc_type, &loc)) {
2538 bailout("too large frame");
2539 }
2540 return new LocationValue(loc);
2541 }
2542
2543
2544 int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) {
2545 assert(opr->is_constant(), "should not be called otherwise");
2546
2547 LIR_Const* c = opr->as_constant_ptr();
2548 BasicType t = c->type();
2549 switch (t) {
2550 case T_OBJECT: {
2551 jobject value = c->as_jobject();
2552 if (value == NULL) {
2553 scope_values->append(_oop_null_scope_value);
2554 } else {
2555 scope_values->append(new ConstantOopWriteValue(c->as_jobject()));
2556 }
2557 return 1;
2558 }
2559
2560 case T_INT: // fall through
2561 case T_FLOAT: {
2562 int value = c->as_jint_bits();
2563 switch (value) {
2564 case -1: scope_values->append(_int_m1_scope_value); break;
2565 case 0: scope_values->append(_int_0_scope_value); break;
2566 case 1: scope_values->append(_int_1_scope_value); break;
2567 case 2: scope_values->append(_int_2_scope_value); break;
2568 default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break;
2569 }
2570 return 1;
2571 }
2572
2573 case T_LONG: // fall through
2574 case T_DOUBLE: {
2575 #ifdef _LP64
2576 scope_values->append(_int_0_scope_value);
2577 scope_values->append(new ConstantLongValue(c->as_jlong_bits()));
2578 #else
2579 if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) {
2580 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits()));
2581 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits()));
2582 } else {
2583 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits()));
2584 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits()));
2585 }
2586 #endif
2587 return 2;
2588 }
2589
2590 case T_ADDRESS: {
2591 #ifdef _LP64
2592 scope_values->append(new ConstantLongValue(c->as_jint()));
2593 #else
2594 scope_values->append(new ConstantIntValue(c->as_jint()));
2595 #endif
2596 return 1;
2597 }
2598
2599 default:
2600 ShouldNotReachHere();
2601 return -1;
2602 }
2603 }
2604
2605 int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) {
2606 if (opr->is_single_stack()) {
2607 int stack_idx = opr->single_stack_ix();
2608 bool is_oop = opr->is_oop_register();
2609 int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0);
2610
2611 ScopeValue* sv = _scope_value_cache.at(cache_idx);
2612 if (sv == NULL) {
2613 Location::Type loc_type = is_oop ? Location::oop : Location::normal;
2614 sv = location_for_name(stack_idx, loc_type);
2615 _scope_value_cache.at_put(cache_idx, sv);
2616 }
2617
2618 // check if cached value is correct
2619 DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal)));
2620
2621 scope_values->append(sv);
2622 return 1;
2623
2624 } else if (opr->is_single_cpu()) {
2625 bool is_oop = opr->is_oop_register();
2626 int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0);
2627 Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long);
2628
2629 ScopeValue* sv = _scope_value_cache.at(cache_idx);
2630 if (sv == NULL) {
2631 Location::Type loc_type = is_oop ? Location::oop : int_loc_type;
2632 VMReg rname = frame_map()->regname(opr);
2633 sv = new LocationValue(Location::new_reg_loc(loc_type, rname));
2634 _scope_value_cache.at_put(cache_idx, sv);
2635 }
2636
2637 // check if cached value is correct
2638 DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr)))));
2639
2640 scope_values->append(sv);
2641 return 1;
2642
2643 #ifdef X86
2644 } else if (opr->is_single_xmm()) {
2645 VMReg rname = opr->as_xmm_float_reg()->as_VMReg();
2646 LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname));
2647
2648 scope_values->append(sv);
2649 return 1;
2650 #endif
2651
2652 } else if (opr->is_single_fpu()) {
2653 #ifdef X86
2654 // the exact location of fpu stack values is only known
2655 // during fpu stack allocation, so the stack allocator object
2656 // must be present
2657 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)");
2658 assert(_fpu_stack_allocator != NULL, "must be present");
2659 opr = _fpu_stack_allocator->to_fpu_stack(opr);
2660 #endif
2661
2662 Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal;
2663 VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr());
2664 #ifndef __SOFTFP__
2665 #ifndef VM_LITTLE_ENDIAN
2666 // On S390 a (single precision) float value occupies only the high
2667 // word of the full double register. So when the double register is
2668 // stored to memory (e.g. by the RegisterSaver), then the float value
2669 // is found at offset 0. I.e. the code below is not needed on S390.
2670 #ifndef S390
2671 if (! float_saved_as_double) {
2672 // On big endian system, we may have an issue if float registers use only
2673 // the low half of the (same) double registers.
2674 // Both the float and the double could have the same regnr but would correspond
2675 // to two different addresses once saved.
2676
2677 // get next safely (no assertion checks)
2678 VMReg next = VMRegImpl::as_VMReg(1+rname->value());
2679 if (next->is_reg() &&
2680 (next->as_FloatRegister() == rname->as_FloatRegister())) {
2681 // the back-end does use the same numbering for the double and the float
2682 rname = next; // VMReg for the low bits, e.g. the real VMReg for the float
2683 }
2684 }
2685 #endif // !S390
2686 #endif
2687 #endif
2688 LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname));
2689
2690 scope_values->append(sv);
2691 return 1;
2692
2693 } else {
2694 // double-size operands
2695
2696 ScopeValue* first;
2697 ScopeValue* second;
2698
2699 if (opr->is_double_stack()) {
2700 #ifdef _LP64
2701 Location loc1;
2702 Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl;
2703 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) {
2704 bailout("too large frame");
2705 }
2706 // Does this reverse on x86 vs. sparc?
2707 first = new LocationValue(loc1);
2708 second = _int_0_scope_value;
2709 #else
2710 Location loc1, loc2;
2711 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) {
2712 bailout("too large frame");
2713 }
2714 first = new LocationValue(loc1);
2715 second = new LocationValue(loc2);
2716 #endif // _LP64
2717
2718 } else if (opr->is_double_cpu()) {
2719 #ifdef _LP64
2720 VMReg rname_first = opr->as_register_lo()->as_VMReg();
2721 first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first));
2722 second = _int_0_scope_value;
2723 #else
2724 VMReg rname_first = opr->as_register_lo()->as_VMReg();
2725 VMReg rname_second = opr->as_register_hi()->as_VMReg();
2726
2727 if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) {
2728 // lo/hi and swapped relative to first and second, so swap them
2729 VMReg tmp = rname_first;
2730 rname_first = rname_second;
2731 rname_second = tmp;
2732 }
2733
2734 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
2735 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
2736 #endif //_LP64
2737
2738
2739 #ifdef X86
2740 } else if (opr->is_double_xmm()) {
2741 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation");
2742 VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg();
2743 # ifdef _LP64
2744 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first));
2745 second = _int_0_scope_value;
2746 # else
2747 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
2748 // %%% This is probably a waste but we'll keep things as they were for now
2749 if (true) {
2750 VMReg rname_second = rname_first->next();
2751 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
2752 }
2753 # endif
2754 #endif
2755
2756 } else if (opr->is_double_fpu()) {
2757 // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of
2758 // the double as float registers in the native ordering. On X86,
2759 // fpu_regnrLo is a FPU stack slot whose VMReg represents
2760 // the low-order word of the double and fpu_regnrLo + 1 is the
2761 // name for the other half. *first and *second must represent the
2762 // least and most significant words, respectively.
2763
2764 #ifdef X86
2765 // the exact location of fpu stack values is only known
2766 // during fpu stack allocation, so the stack allocator object
2767 // must be present
2768 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)");
2769 assert(_fpu_stack_allocator != NULL, "must be present");
2770 opr = _fpu_stack_allocator->to_fpu_stack(opr);
2771
2772 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrLo is used)");
2773 #endif
2774 #ifdef SPARC
2775 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)");
2776 #endif
2777 #ifdef ARM32
2778 assert(opr->fpu_regnrHi() == opr->fpu_regnrLo() + 1, "assumed in calculation (only fpu_regnrLo is used)");
2779 #endif
2780 #ifdef PPC32
2781 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)");
2782 #endif
2783
2784 #ifdef VM_LITTLE_ENDIAN
2785 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrLo());
2786 #else
2787 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi());
2788 #endif
2789
2790 #ifdef _LP64
2791 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first));
2792 second = _int_0_scope_value;
2793 #else
2794 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
2795 // %%% This is probably a waste but we'll keep things as they were for now
2796 if (true) {
2797 VMReg rname_second = rname_first->next();
2798 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
2799 }
2800 #endif
2801
2802 } else {
2803 ShouldNotReachHere();
2804 first = NULL;
2805 second = NULL;
2806 }
2807
2808 assert(first != NULL && second != NULL, "must be set");
2809 // The convention the interpreter uses is that the second local
2810 // holds the first raw word of the native double representation.
2811 // This is actually reasonable, since locals and stack arrays
2812 // grow downwards in all implementations.
2813 // (If, on some machine, the interpreter's Java locals or stack
2814 // were to grow upwards, the embedded doubles would be word-swapped.)
2815 scope_values->append(second);
2816 scope_values->append(first);
2817 return 2;
2818 }
2819 }
2820
2821
2822 int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) {
2823 if (value != NULL) {
2824 LIR_Opr opr = value->operand();
2825 Constant* con = value->as_Constant();
2826
2827 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)");
2828 assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands");
2829
2830 if (con != NULL && !con->is_pinned() && !opr->is_constant()) {
2831 // Unpinned constants may have a virtual operand for a part of the lifetime
2832 // or may be illegal when it was optimized away,
2833 // so always use a constant operand
2834 opr = LIR_OprFact::value_type(con->type());
2835 }
2836 assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here");
2837
2838 if (opr->is_virtual()) {
2839 LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode;
2840
2841 BlockBegin* block = block_of_op_with_id(op_id);
2842 if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) {
2843 // generating debug information for the last instruction of a block.
2844 // if this instruction is a branch, spill moves are inserted before this branch
2845 // and so the wrong operand would be returned (spill moves at block boundaries are not
2846 // considered in the live ranges of intervals)
2847 // Solution: use the first op_id of the branch target block instead.
2848 if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) {
2849 if (block->live_out().at(opr->vreg_number())) {
2850 op_id = block->sux_at(0)->first_lir_instruction_id();
2851 mode = LIR_OpVisitState::outputMode;
2852 }
2853 }
2854 }
2855
2856 // Get current location of operand
2857 // The operand must be live because debug information is considered when building the intervals
2858 // if the interval is not live, color_lir_opr will cause an assertion failure
2859 opr = color_lir_opr(opr, op_id, mode);
2860 assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls");
2861
2862 // Append to ScopeValue array
2863 return append_scope_value_for_operand(opr, scope_values);
2864
2865 } else {
2866 assert(value->as_Constant() != NULL, "all other instructions have only virtual operands");
2867 assert(opr->is_constant(), "operand must be constant");
2868
2869 return append_scope_value_for_constant(opr, scope_values);
2870 }
2871 } else {
2872 // append a dummy value because real value not needed
2873 scope_values->append(_illegal_value);
2874 return 1;
2875 }
2876 }
2877
2878
2879 IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state) {
2880 IRScopeDebugInfo* caller_debug_info = NULL;
2881
2882 ValueStack* caller_state = cur_state->caller_state();
2883 if (caller_state != NULL) {
2884 // process recursively to compute outermost scope first
2885 caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state);
2886 }
2887
2888 // initialize these to null.
2889 // If we don't need deopt info or there are no locals, expressions or monitors,
2890 // then these get recorded as no information and avoids the allocation of 0 length arrays.
2891 GrowableArray<ScopeValue*>* locals = NULL;
2892 GrowableArray<ScopeValue*>* expressions = NULL;
2893 GrowableArray<MonitorValue*>* monitors = NULL;
2894
2895 // describe local variable values
2896 int nof_locals = cur_state->locals_size();
2897 if (nof_locals > 0) {
2898 locals = new GrowableArray<ScopeValue*>(nof_locals);
2899
2900 int pos = 0;
2901 while (pos < nof_locals) {
2902 assert(pos < cur_state->locals_size(), "why not?");
2903
2904 Value local = cur_state->local_at(pos);
2905 pos += append_scope_value(op_id, local, locals);
2906
2907 assert(locals->length() == pos, "must match");
2908 }
2909 assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals");
2910 assert(locals->length() == cur_state->locals_size(), "wrong number of locals");
2911 } else if (cur_scope->method()->max_locals() > 0) {
2912 assert(cur_state->kind() == ValueStack::EmptyExceptionState, "should be");
2913 nof_locals = cur_scope->method()->max_locals();
2914 locals = new GrowableArray<ScopeValue*>(nof_locals);
2915 for(int i = 0; i < nof_locals; i++) {
2916 locals->append(_illegal_value);
2917 }
2918 }
2919
2920 // describe expression stack
2921 int nof_stack = cur_state->stack_size();
2922 if (nof_stack > 0) {
2923 expressions = new GrowableArray<ScopeValue*>(nof_stack);
2924
2925 int pos = 0;
2926 while (pos < nof_stack) {
2927 Value expression = cur_state->stack_at_inc(pos);
2928 append_scope_value(op_id, expression, expressions);
2929
2930 assert(expressions->length() == pos, "must match");
2931 }
2932 assert(expressions->length() == cur_state->stack_size(), "wrong number of stack entries");
2933 }
2934
2935 // describe monitors
2936 int nof_locks = cur_state->locks_size();
2937 if (nof_locks > 0) {
2938 int lock_offset = cur_state->caller_state() != NULL ? cur_state->caller_state()->total_locks_size() : 0;
2939 monitors = new GrowableArray<MonitorValue*>(nof_locks);
2940 for (int i = 0; i < nof_locks; i++) {
2941 monitors->append(location_for_monitor_index(lock_offset + i));
2942 }
2943 }
2944
2945 return new IRScopeDebugInfo(cur_scope, cur_state->bci(), locals, expressions, monitors, caller_debug_info);
2946 }
2947
2948
2949 void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) {
2950 TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id));
2951
2952 IRScope* innermost_scope = info->scope();
2953 ValueStack* innermost_state = info->stack();
2954
2955 assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?");
2956
2957 DEBUG_ONLY(check_stack_depth(info, innermost_state->stack_size()));
2958
2959 if (info->_scope_debug_info == NULL) {
2960 // compute debug information
2961 info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state);
2962 } else {
2963 // debug information already set. Check that it is correct from the current point of view
2964 DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state)));
2965 }
2966 }
2967
2968
2969 void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) {
2970 LIR_OpVisitState visitor;
2971 int num_inst = instructions->length();
2972 bool has_dead = false;
2973
2974 for (int j = 0; j < num_inst; j++) {
2975 LIR_Op* op = instructions->at(j);
2976 if (op == NULL) { // this can happen when spill-moves are removed in eliminate_spill_moves
2977 has_dead = true;
2978 continue;
2979 }
2980 int op_id = op->id();
2981
2982 // visit instruction to get list of operands
2983 visitor.visit(op);
2984
2985 // iterate all modes of the visitor and process all virtual operands
2986 for_each_visitor_mode(mode) {
2987 int n = visitor.opr_count(mode);
2988 for (int k = 0; k < n; k++) {
2989 LIR_Opr opr = visitor.opr_at(mode, k);
2990 if (opr->is_virtual_register()) {
2991 visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode));
2992 }
2993 }
2994 }
2995
2996 if (visitor.info_count() > 0) {
2997 // exception handling
2998 if (compilation()->has_exception_handlers()) {
2999 XHandlers* xhandlers = visitor.all_xhandler();
3000 int n = xhandlers->length();
3001 for (int k = 0; k < n; k++) {
3002 XHandler* handler = xhandlers->handler_at(k);
3003 if (handler->entry_code() != NULL) {
3004 assign_reg_num(handler->entry_code()->instructions_list(), NULL);
3005 }
3006 }
3007 } else {
3008 assert(visitor.all_xhandler()->length() == 0, "missed exception handler");
3009 }
3010
3011 // compute oop map
3012 assert(iw != NULL, "needed for compute_oop_map");
3013 compute_oop_map(iw, visitor, op);
3014
3015 // compute debug information
3016 if (!use_fpu_stack_allocation()) {
3017 // compute debug information if fpu stack allocation is not needed.
3018 // when fpu stack allocation is needed, the debug information can not
3019 // be computed here because the exact location of fpu operands is not known
3020 // -> debug information is created inside the fpu stack allocator
3021 int n = visitor.info_count();
3022 for (int k = 0; k < n; k++) {
3023 compute_debug_info(visitor.info_at(k), op_id);
3024 }
3025 }
3026 }
3027
3028 #ifdef ASSERT
3029 // make sure we haven't made the op invalid.
3030 op->verify();
3031 #endif
3032
3033 // remove useless moves
3034 if (op->code() == lir_move) {
3035 assert(op->as_Op1() != NULL, "move must be LIR_Op1");
3036 LIR_Op1* move = (LIR_Op1*)op;
3037 LIR_Opr src = move->in_opr();
3038 LIR_Opr dst = move->result_opr();
3039 if (dst == src ||
3040 (!dst->is_pointer() && !src->is_pointer() &&
3041 src->is_same_register(dst))) {
3042 instructions->at_put(j, NULL);
3043 has_dead = true;
3044 }
3045 }
3046 }
3047
3048 if (has_dead) {
3049 // iterate all instructions of the block and remove all null-values.
3050 int insert_point = 0;
3051 for (int j = 0; j < num_inst; j++) {
3052 LIR_Op* op = instructions->at(j);
3053 if (op != NULL) {
3054 if (insert_point != j) {
3055 instructions->at_put(insert_point, op);
3056 }
3057 insert_point++;
3058 }
3059 }
3060 instructions->trunc_to(insert_point);
3061 }
3062 }
3063
3064 void LinearScan::assign_reg_num() {
3065 TIME_LINEAR_SCAN(timer_assign_reg_num);
3066
3067 init_compute_debug_info();
3068 IntervalWalker* iw = init_compute_oop_maps();
3069
3070 int num_blocks = block_count();
3071 for (int i = 0; i < num_blocks; i++) {
3072 BlockBegin* block = block_at(i);
3073 assign_reg_num(block->lir()->instructions_list(), iw);
3074 }
3075 }
3076
3077
3078 void LinearScan::do_linear_scan() {
3079 NOT_PRODUCT(_total_timer.begin_method());
3080
3081 number_instructions();
3082
3083 NOT_PRODUCT(print_lir(1, "Before Register Allocation"));
3084
3085 compute_local_live_sets();
3086 compute_global_live_sets();
3087 CHECK_BAILOUT();
3088
3089 build_intervals();
3090 CHECK_BAILOUT();
3091 sort_intervals_before_allocation();
3092
3093 NOT_PRODUCT(print_intervals("Before Register Allocation"));
3094 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc));
3095
3096 allocate_registers();
3097 CHECK_BAILOUT();
3098
3099 resolve_data_flow();
3100 if (compilation()->has_exception_handlers()) {
3101 resolve_exception_handlers();
3102 }
3103 // fill in number of spill slots into frame_map
3104 propagate_spill_slots();
3105 CHECK_BAILOUT();
3106
3107 NOT_PRODUCT(print_intervals("After Register Allocation"));
3108 NOT_PRODUCT(print_lir(2, "LIR after register allocation:"));
3109
3110 sort_intervals_after_allocation();
3111
3112 DEBUG_ONLY(verify());
3113
3114 eliminate_spill_moves();
3115 assign_reg_num();
3116 CHECK_BAILOUT();
3117
3118 NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:"));
3119 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign));
3120
3121 { TIME_LINEAR_SCAN(timer_allocate_fpu_stack);
3122
3123 if (use_fpu_stack_allocation()) {
3124 allocate_fpu_stack(); // Only has effect on Intel
3125 NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:"));
3126 }
3127 }
3128
3129 { TIME_LINEAR_SCAN(timer_optimize_lir);
3130
3131 EdgeMoveOptimizer::optimize(ir()->code());
3132 ControlFlowOptimizer::optimize(ir()->code());
3133 // check that cfg is still correct after optimizations
3134 ir()->verify();
3135 }
3136
3137 NOT_PRODUCT(print_lir(1, "Before Code Generation", false));
3138 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final));
3139 NOT_PRODUCT(_total_timer.end_method(this));
3140 }
3141
3142
3143 // ********** Printing functions
3144
3145 #ifndef PRODUCT
3146
3147 void LinearScan::print_timers(double total) {
3148 _total_timer.print(total);
3149 }
3150
3151 void LinearScan::print_statistics() {
3152 _stat_before_alloc.print("before allocation");
3153 _stat_after_asign.print("after assignment of register");
3154 _stat_final.print("after optimization");
3155 }
3156
3157 void LinearScan::print_bitmap(BitMap& b) {
3158 for (unsigned int i = 0; i < b.size(); i++) {
3159 if (b.at(i)) tty->print("%d ", i);
3160 }
3161 tty->cr();
3162 }
3163
3164 void LinearScan::print_intervals(const char* label) {
3165 if (TraceLinearScanLevel >= 1) {
3166 int i;
3167 tty->cr();
3168 tty->print_cr("%s", label);
3169
3170 for (i = 0; i < interval_count(); i++) {
3171 Interval* interval = interval_at(i);
3172 if (interval != NULL) {
3173 interval->print();
3174 }
3175 }
3176
3177 tty->cr();
3178 tty->print_cr("--- Basic Blocks ---");
3179 for (i = 0; i < block_count(); i++) {
3180 BlockBegin* block = block_at(i);
3181 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());
3182 }
3183 tty->cr();
3184 tty->cr();
3185 }
3186
3187 if (PrintCFGToFile) {
3188 CFGPrinter::print_intervals(&_intervals, label);
3189 }
3190 }
3191
3192 void LinearScan::print_lir(int level, const char* label, bool hir_valid) {
3193 if (TraceLinearScanLevel >= level) {
3194 tty->cr();
3195 tty->print_cr("%s", label);
3196 print_LIR(ir()->linear_scan_order());
3197 tty->cr();
3198 }
3199
3200 if (level == 1 && PrintCFGToFile) {
3201 CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true);
3202 }
3203 }
3204
3205 #endif //PRODUCT
3206
3207
3208 // ********** verification functions for allocation
3209 // (check that all intervals have a correct register and that no registers are overwritten)
3210 #ifdef ASSERT
3211
3212 void LinearScan::verify() {
3213 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************"));
3214 verify_intervals();
3215
3216 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************"));
3217 verify_no_oops_in_fixed_intervals();
3218
3219 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries"));
3220 verify_constants();
3221
3222 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************"));
3223 verify_registers();
3224
3225 TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************"));
3226 }
3227
3228 void LinearScan::verify_intervals() {
3229 int len = interval_count();
3230 bool has_error = false;
3231
3232 for (int i = 0; i < len; i++) {
3233 Interval* i1 = interval_at(i);
3234 if (i1 == NULL) continue;
3235
3236 i1->check_split_children();
3237
3238 if (i1->reg_num() != i) {
3239 tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr();
3240 has_error = true;
3241 }
3242
3243 if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) {
3244 tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr();
3245 has_error = true;
3246 }
3247
3248 if (i1->assigned_reg() == any_reg) {
3249 tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr();
3250 has_error = true;
3251 }
3252
3253 if (i1->assigned_reg() == i1->assigned_regHi()) {
3254 tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr();
3255 has_error = true;
3256 }
3257
3258 if (!is_processed_reg_num(i1->assigned_reg())) {
3259 tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr();
3260 has_error = true;
3261 }
3262
3263 // special intervals that are created in MoveResolver
3264 // -> ignore them because the range information has no meaning there
3265 if (i1->from() == 1 && i1->to() == 2) continue;
3266
3267 if (i1->first() == Range::end()) {
3268 tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr();
3269 has_error = true;
3270 }
3271
3272 for (Range* r = i1->first(); r != Range::end(); r = r->next()) {
3273 if (r->from() >= r->to()) {
3274 tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr();
3275 has_error = true;
3276 }
3277 }
3278
3279 for (int j = i + 1; j < len; j++) {
3280 Interval* i2 = interval_at(j);
3281 if (i2 == NULL || (i2->from() == 1 && i2->to() == 2)) continue;
3282
3283 int r1 = i1->assigned_reg();
3284 int r1Hi = i1->assigned_regHi();
3285 int r2 = i2->assigned_reg();
3286 int r2Hi = i2->assigned_regHi();
3287 if ((r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi))) && i1->intersects(i2)) {
3288 tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num());
3289 i1->print(); tty->cr();
3290 i2->print(); tty->cr();
3291 has_error = true;
3292 }
3293 }
3294 }
3295
3296 assert(has_error == false, "register allocation invalid");
3297 }
3298
3299
3300 void LinearScan::verify_no_oops_in_fixed_intervals() {
3301 Interval* fixed_intervals;
3302 Interval* other_intervals;
3303 create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL);
3304
3305 // to ensure a walking until the last instruction id, add a dummy interval
3306 // with a high operation id
3307 other_intervals = new Interval(any_reg);
3308 other_intervals->add_range(max_jint - 2, max_jint - 1);
3309 IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals);
3310
3311 LIR_OpVisitState visitor;
3312 for (int i = 0; i < block_count(); i++) {
3313 BlockBegin* block = block_at(i);
3314
3315 LIR_OpList* instructions = block->lir()->instructions_list();
3316
3317 for (int j = 0; j < instructions->length(); j++) {
3318 LIR_Op* op = instructions->at(j);
3319 int op_id = op->id();
3320
3321 visitor.visit(op);
3322
3323 if (visitor.info_count() > 0) {
3324 iw->walk_before(op->id());
3325 bool check_live = true;
3326 if (op->code() == lir_move) {
3327 LIR_Op1* move = (LIR_Op1*)op;
3328 check_live = (move->patch_code() == lir_patch_none);
3329 }
3330 LIR_OpBranch* branch = op->as_OpBranch();
3331 if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) {
3332 // Don't bother checking the stub in this case since the
3333 // exception stub will never return to normal control flow.
3334 check_live = false;
3335 }
3336
3337 // Make sure none of the fixed registers is live across an
3338 // oopmap since we can't handle that correctly.
3339 if (check_live) {
3340 for (Interval* interval = iw->active_first(fixedKind);
3341 interval != Interval::end();
3342 interval = interval->next()) {
3343 if (interval->current_to() > op->id() + 1) {
3344 // This interval is live out of this op so make sure
3345 // that this interval represents some value that's
3346 // referenced by this op either as an input or output.
3347 bool ok = false;
3348 for_each_visitor_mode(mode) {
3349 int n = visitor.opr_count(mode);
3350 for (int k = 0; k < n; k++) {
3351 LIR_Opr opr = visitor.opr_at(mode, k);
3352 if (opr->is_fixed_cpu()) {
3353 if (interval_at(reg_num(opr)) == interval) {
3354 ok = true;
3355 break;
3356 }
3357 int hi = reg_numHi(opr);
3358 if (hi != -1 && interval_at(hi) == interval) {
3359 ok = true;
3360 break;
3361 }
3362 }
3363 }
3364 }
3365 assert(ok, "fixed intervals should never be live across an oopmap point");
3366 }
3367 }
3368 }
3369 }
3370
3371 // oop-maps at calls do not contain registers, so check is not needed
3372 if (!visitor.has_call()) {
3373
3374 for_each_visitor_mode(mode) {
3375 int n = visitor.opr_count(mode);
3376 for (int k = 0; k < n; k++) {
3377 LIR_Opr opr = visitor.opr_at(mode, k);
3378
3379 if (opr->is_fixed_cpu() && opr->is_oop()) {
3380 // operand is a non-virtual cpu register and contains an oop
3381 TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr());
3382
3383 Interval* interval = interval_at(reg_num(opr));
3384 assert(interval != NULL, "no interval");
3385
3386 if (mode == LIR_OpVisitState::inputMode) {
3387 if (interval->to() >= op_id + 1) {
3388 assert(interval->to() < op_id + 2 ||
3389 interval->has_hole_between(op_id, op_id + 2),
3390 "oop input operand live after instruction");
3391 }
3392 } else if (mode == LIR_OpVisitState::outputMode) {
3393 if (interval->from() <= op_id - 1) {
3394 assert(interval->has_hole_between(op_id - 1, op_id),
3395 "oop input operand live after instruction");
3396 }
3397 }
3398 }
3399 }
3400 }
3401 }
3402 }
3403 }
3404 }
3405
3406
3407 void LinearScan::verify_constants() {
3408 int num_regs = num_virtual_regs();
3409 int size = live_set_size();
3410 int num_blocks = block_count();
3411
3412 for (int i = 0; i < num_blocks; i++) {
3413 BlockBegin* block = block_at(i);
3414 ResourceBitMap live_at_edge = block->live_in();
3415
3416 // visit all registers where the live_at_edge bit is set
3417 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)) {
3418 TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id()));
3419
3420 Value value = gen()->instruction_for_vreg(r);
3421
3422 assert(value != NULL, "all intervals live across block boundaries must have Value");
3423 assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand");
3424 assert(value->operand()->vreg_number() == r, "register number must match");
3425 // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries");
3426 }
3427 }
3428 }
3429
3430
3431 class RegisterVerifier: public StackObj {
3432 private:
3433 LinearScan* _allocator;
3434 BlockList _work_list; // all blocks that must be processed
3435 IntervalsList _saved_states; // saved information of previous check
3436
3437 // simplified access to methods of LinearScan
3438 Compilation* compilation() const { return _allocator->compilation(); }
3439 Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); }
3440 int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); }
3441
3442 // currently, only registers are processed
3443 int state_size() { return LinearScan::nof_regs; }
3444
3445 // accessors
3446 IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); }
3447 void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); }
3448 void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); }
3449
3450 // helper functions
3451 IntervalList* copy(IntervalList* input_state);
3452 void state_put(IntervalList* input_state, int reg, Interval* interval);
3453 bool check_state(IntervalList* input_state, int reg, Interval* interval);
3454
3455 void process_block(BlockBegin* block);
3456 void process_xhandler(XHandler* xhandler, IntervalList* input_state);
3457 void process_successor(BlockBegin* block, IntervalList* input_state);
3458 void process_operations(LIR_List* ops, IntervalList* input_state);
3459
3460 public:
3461 RegisterVerifier(LinearScan* allocator)
3462 : _allocator(allocator)
3463 , _work_list(16)
3464 , _saved_states(BlockBegin::number_of_blocks(), BlockBegin::number_of_blocks(), NULL)
3465 { }
3466
3467 void verify(BlockBegin* start);
3468 };
3469
3470
3471 // entry function from LinearScan that starts the verification
3472 void LinearScan::verify_registers() {
3473 RegisterVerifier verifier(this);
3474 verifier.verify(block_at(0));
3475 }
3476
3477
3478 void RegisterVerifier::verify(BlockBegin* start) {
3479 // setup input registers (method arguments) for first block
3480 int input_state_len = state_size();
3481 IntervalList* input_state = new IntervalList(input_state_len, input_state_len, NULL);
3482 CallingConvention* args = compilation()->frame_map()->incoming_arguments();
3483 for (int n = 0; n < args->length(); n++) {
3484 LIR_Opr opr = args->at(n);
3485 if (opr->is_register()) {
3486 Interval* interval = interval_at(reg_num(opr));
3487
3488 if (interval->assigned_reg() < state_size()) {
3489 input_state->at_put(interval->assigned_reg(), interval);
3490 }
3491 if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) {
3492 input_state->at_put(interval->assigned_regHi(), interval);
3493 }
3494 }
3495 }
3496
3497 set_state_for_block(start, input_state);
3498 add_to_work_list(start);
3499
3500 // main loop for verification
3501 do {
3502 BlockBegin* block = _work_list.at(0);
3503 _work_list.remove_at(0);
3504
3505 process_block(block);
3506 } while (!_work_list.is_empty());
3507 }
3508
3509 void RegisterVerifier::process_block(BlockBegin* block) {
3510 TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id()));
3511
3512 // must copy state because it is modified
3513 IntervalList* input_state = copy(state_for_block(block));
3514
3515 if (TraceLinearScanLevel >= 4) {
3516 tty->print_cr("Input-State of intervals:");
3517 tty->print(" ");
3518 for (int i = 0; i < state_size(); i++) {
3519 if (input_state->at(i) != NULL) {
3520 tty->print(" %4d", input_state->at(i)->reg_num());
3521 } else {
3522 tty->print(" __");
3523 }
3524 }
3525 tty->cr();
3526 tty->cr();
3527 }
3528
3529 // process all operations of the block
3530 process_operations(block->lir(), input_state);
3531
3532 // iterate all successors
3533 for (int i = 0; i < block->number_of_sux(); i++) {
3534 process_successor(block->sux_at(i), input_state);
3535 }
3536 }
3537
3538 void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) {
3539 TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id()));
3540
3541 // must copy state because it is modified
3542 input_state = copy(input_state);
3543
3544 if (xhandler->entry_code() != NULL) {
3545 process_operations(xhandler->entry_code(), input_state);
3546 }
3547 process_successor(xhandler->entry_block(), input_state);
3548 }
3549
3550 void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) {
3551 IntervalList* saved_state = state_for_block(block);
3552
3553 if (saved_state != NULL) {
3554 // this block was already processed before.
3555 // check if new input_state is consistent with saved_state
3556
3557 bool saved_state_correct = true;
3558 for (int i = 0; i < state_size(); i++) {
3559 if (input_state->at(i) != saved_state->at(i)) {
3560 // current input_state and previous saved_state assume a different
3561 // interval in this register -> assume that this register is invalid
3562 if (saved_state->at(i) != NULL) {
3563 // invalidate old calculation only if it assumed that
3564 // register was valid. when the register was already invalid,
3565 // then the old calculation was correct.
3566 saved_state_correct = false;
3567 saved_state->at_put(i, NULL);
3568
3569 TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i));
3570 }
3571 }
3572 }
3573
3574 if (saved_state_correct) {
3575 // already processed block with correct input_state
3576 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id()));
3577 } else {
3578 // must re-visit this block
3579 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id()));
3580 add_to_work_list(block);
3581 }
3582
3583 } else {
3584 // block was not processed before, so set initial input_state
3585 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id()));
3586
3587 set_state_for_block(block, copy(input_state));
3588 add_to_work_list(block);
3589 }
3590 }
3591
3592
3593 IntervalList* RegisterVerifier::copy(IntervalList* input_state) {
3594 IntervalList* copy_state = new IntervalList(input_state->length());
3595 copy_state->appendAll(input_state);
3596 return copy_state;
3597 }
3598
3599 void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) {
3600 if (reg != LinearScan::any_reg && reg < state_size()) {
3601 if (interval != NULL) {
3602 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d", reg, interval->reg_num()));
3603 } else if (input_state->at(reg) != NULL) {
3604 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = NULL", reg));
3605 }
3606
3607 input_state->at_put(reg, interval);
3608 }
3609 }
3610
3611 bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) {
3612 if (reg != LinearScan::any_reg && reg < state_size()) {
3613 if (input_state->at(reg) != interval) {
3614 tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num());
3615 return true;
3616 }
3617 }
3618 return false;
3619 }
3620
3621 void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) {
3622 // visit all instructions of the block
3623 LIR_OpVisitState visitor;
3624 bool has_error = false;
3625
3626 for (int i = 0; i < ops->length(); i++) {
3627 LIR_Op* op = ops->at(i);
3628 visitor.visit(op);
3629
3630 TRACE_LINEAR_SCAN(4, op->print_on(tty));
3631
3632 // check if input operands are correct
3633 int j;
3634 int n = visitor.opr_count(LIR_OpVisitState::inputMode);
3635 for (j = 0; j < n; j++) {
3636 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j);
3637 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
3638 Interval* interval = interval_at(reg_num(opr));
3639 if (op->id() != -1) {
3640 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode);
3641 }
3642
3643 has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent());
3644 has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent());
3645
3646 // When an operand is marked with is_last_use, then the fpu stack allocator
3647 // removes the register from the fpu stack -> the register contains no value
3648 if (opr->is_last_use()) {
3649 state_put(input_state, interval->assigned_reg(), NULL);
3650 state_put(input_state, interval->assigned_regHi(), NULL);
3651 }
3652 }
3653 }
3654
3655 // invalidate all caller save registers at calls
3656 if (visitor.has_call()) {
3657 for (j = 0; j < FrameMap::nof_caller_save_cpu_regs(); j++) {
3658 state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL);
3659 }
3660 for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) {
3661 state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL);
3662 }
3663
3664 #ifdef X86
3665 int num_caller_save_xmm_regs = FrameMap::get_num_caller_save_xmms();
3666 for (j = 0; j < num_caller_save_xmm_regs; j++) {
3667 state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL);
3668 }
3669 #endif
3670 }
3671
3672 // process xhandler before output and temp operands
3673 XHandlers* xhandlers = visitor.all_xhandler();
3674 n = xhandlers->length();
3675 for (int k = 0; k < n; k++) {
3676 process_xhandler(xhandlers->handler_at(k), input_state);
3677 }
3678
3679 // set temp operands (some operations use temp operands also as output operands, so can't set them NULL)
3680 n = visitor.opr_count(LIR_OpVisitState::tempMode);
3681 for (j = 0; j < n; j++) {
3682 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j);
3683 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
3684 Interval* interval = interval_at(reg_num(opr));
3685 if (op->id() != -1) {
3686 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode);
3687 }
3688
3689 state_put(input_state, interval->assigned_reg(), interval->split_parent());
3690 state_put(input_state, interval->assigned_regHi(), interval->split_parent());
3691 }
3692 }
3693
3694 // set output operands
3695 n = visitor.opr_count(LIR_OpVisitState::outputMode);
3696 for (j = 0; j < n; j++) {
3697 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j);
3698 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
3699 Interval* interval = interval_at(reg_num(opr));
3700 if (op->id() != -1) {
3701 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode);
3702 }
3703
3704 state_put(input_state, interval->assigned_reg(), interval->split_parent());
3705 state_put(input_state, interval->assigned_regHi(), interval->split_parent());
3706 }
3707 }
3708 }
3709 assert(has_error == false, "Error in register allocation");
3710 }
3711
3712 #endif // ASSERT
3713
3714
3715
3716 // **** Implementation of MoveResolver ******************************
3717
3718 MoveResolver::MoveResolver(LinearScan* allocator) :
3719 _allocator(allocator),
3720 _multiple_reads_allowed(false),
3721 _mapping_from(8),
3722 _mapping_from_opr(8),
3723 _mapping_to(8),
3724 _insert_list(NULL),
3725 _insert_idx(-1),
3726 _insertion_buffer()
3727 {
3728 for (int i = 0; i < LinearScan::nof_regs; i++) {
3729 _register_blocked[i] = 0;
3730 }
3731 DEBUG_ONLY(check_empty());
3732 }
3733
3734
3735 #ifdef ASSERT
3736
3737 void MoveResolver::check_empty() {
3738 assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing");
3739 for (int i = 0; i < LinearScan::nof_regs; i++) {
3740 assert(register_blocked(i) == 0, "register map must be empty before and after processing");
3741 }
3742 assert(_multiple_reads_allowed == false, "must have default value");
3743 }
3744
3745 void MoveResolver::verify_before_resolve() {
3746 assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal");
3747 assert(_mapping_from.length() == _mapping_to.length(), "length must be equal");
3748 assert(_insert_list != NULL && _insert_idx != -1, "insert position not set");
3749
3750 int i, j;
3751 if (!_multiple_reads_allowed) {
3752 for (i = 0; i < _mapping_from.length(); i++) {
3753 for (j = i + 1; j < _mapping_from.length(); j++) {
3754 assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice");
3755 }
3756 }
3757 }
3758
3759 for (i = 0; i < _mapping_to.length(); i++) {
3760 for (j = i + 1; j < _mapping_to.length(); j++) {
3761 assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice");
3762 }
3763 }
3764
3765
3766 ResourceBitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills());
3767 if (!_multiple_reads_allowed) {
3768 for (i = 0; i < _mapping_from.length(); i++) {
3769 Interval* it = _mapping_from.at(i);
3770 if (it != NULL) {
3771 assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice");
3772 used_regs.set_bit(it->assigned_reg());
3773
3774 if (it->assigned_regHi() != LinearScan::any_reg) {
3775 assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice");
3776 used_regs.set_bit(it->assigned_regHi());
3777 }
3778 }
3779 }
3780 }
3781
3782 used_regs.clear();
3783 for (i = 0; i < _mapping_to.length(); i++) {
3784 Interval* it = _mapping_to.at(i);
3785 assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice");
3786 used_regs.set_bit(it->assigned_reg());
3787
3788 if (it->assigned_regHi() != LinearScan::any_reg) {
3789 assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice");
3790 used_regs.set_bit(it->assigned_regHi());
3791 }
3792 }
3793
3794 used_regs.clear();
3795 for (i = 0; i < _mapping_from.length(); i++) {
3796 Interval* it = _mapping_from.at(i);
3797 if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) {
3798 used_regs.set_bit(it->assigned_reg());
3799 }
3800 }
3801 for (i = 0; i < _mapping_to.length(); i++) {
3802 Interval* it = _mapping_to.at(i);
3803 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");
3804 }
3805 }
3806
3807 #endif // ASSERT
3808
3809
3810 // mark assigned_reg and assigned_regHi of the interval as blocked
3811 void MoveResolver::block_registers(Interval* it) {
3812 int reg = it->assigned_reg();
3813 if (reg < LinearScan::nof_regs) {
3814 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used");
3815 set_register_blocked(reg, 1);
3816 }
3817 reg = it->assigned_regHi();
3818 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
3819 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used");
3820 set_register_blocked(reg, 1);
3821 }
3822 }
3823
3824 // mark assigned_reg and assigned_regHi of the interval as unblocked
3825 void MoveResolver::unblock_registers(Interval* it) {
3826 int reg = it->assigned_reg();
3827 if (reg < LinearScan::nof_regs) {
3828 assert(register_blocked(reg) > 0, "register already marked as unused");
3829 set_register_blocked(reg, -1);
3830 }
3831 reg = it->assigned_regHi();
3832 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
3833 assert(register_blocked(reg) > 0, "register already marked as unused");
3834 set_register_blocked(reg, -1);
3835 }
3836 }
3837
3838 // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from)
3839 bool MoveResolver::save_to_process_move(Interval* from, Interval* to) {
3840 int from_reg = -1;
3841 int from_regHi = -1;
3842 if (from != NULL) {
3843 from_reg = from->assigned_reg();
3844 from_regHi = from->assigned_regHi();
3845 }
3846
3847 int reg = to->assigned_reg();
3848 if (reg < LinearScan::nof_regs) {
3849 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) {
3850 return false;
3851 }
3852 }
3853 reg = to->assigned_regHi();
3854 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
3855 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) {
3856 return false;
3857 }
3858 }
3859
3860 return true;
3861 }
3862
3863
3864 void MoveResolver::create_insertion_buffer(LIR_List* list) {
3865 assert(!_insertion_buffer.initialized(), "overwriting existing buffer");
3866 _insertion_buffer.init(list);
3867 }
3868
3869 void MoveResolver::append_insertion_buffer() {
3870 if (_insertion_buffer.initialized()) {
3871 _insertion_buffer.lir_list()->append(&_insertion_buffer);
3872 }
3873 assert(!_insertion_buffer.initialized(), "must be uninitialized now");
3874
3875 _insert_list = NULL;
3876 _insert_idx = -1;
3877 }
3878
3879 void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) {
3880 assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal");
3881 assert(from_interval->type() == to_interval->type(), "move between different types");
3882 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first");
3883 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer");
3884
3885 LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type());
3886 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type());
3887
3888 if (!_multiple_reads_allowed) {
3889 // the last_use flag is an optimization for FPU stack allocation. When the same
3890 // input interval is used in more than one move, then it is too difficult to determine
3891 // if this move is really the last use.
3892 from_opr = from_opr->make_last_use();
3893 }
3894 _insertion_buffer.move(_insert_idx, from_opr, to_opr);
3895
3896 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()));
3897 }
3898
3899 void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) {
3900 assert(from_opr->type() == to_interval->type(), "move between different types");
3901 assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first");
3902 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer");
3903
3904 LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type());
3905 _insertion_buffer.move(_insert_idx, from_opr, to_opr);
3906
3907 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()));
3908 }
3909
3910
3911 void MoveResolver::resolve_mappings() {
3912 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));
3913 DEBUG_ONLY(verify_before_resolve());
3914
3915 // Block all registers that are used as input operands of a move.
3916 // When a register is blocked, no move to this register is emitted.
3917 // This is necessary for detecting cycles in moves.
3918 int i;
3919 for (i = _mapping_from.length() - 1; i >= 0; i--) {
3920 Interval* from_interval = _mapping_from.at(i);
3921 if (from_interval != NULL) {
3922 block_registers(from_interval);
3923 }
3924 }
3925
3926 int spill_candidate = -1;
3927 while (_mapping_from.length() > 0) {
3928 bool processed_interval = false;
3929
3930 for (i = _mapping_from.length() - 1; i >= 0; i--) {
3931 Interval* from_interval = _mapping_from.at(i);
3932 Interval* to_interval = _mapping_to.at(i);
3933
3934 if (save_to_process_move(from_interval, to_interval)) {
3935 // this inverval can be processed because target is free
3936 if (from_interval != NULL) {
3937 insert_move(from_interval, to_interval);
3938 unblock_registers(from_interval);
3939 } else {
3940 insert_move(_mapping_from_opr.at(i), to_interval);
3941 }
3942 _mapping_from.remove_at(i);
3943 _mapping_from_opr.remove_at(i);
3944 _mapping_to.remove_at(i);
3945
3946 processed_interval = true;
3947 } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) {
3948 // this interval cannot be processed now because target is not free
3949 // it starts in a register, so it is a possible candidate for spilling
3950 spill_candidate = i;
3951 }
3952 }
3953
3954 if (!processed_interval) {
3955 // no move could be processed because there is a cycle in the move list
3956 // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory
3957 assert(spill_candidate != -1, "no interval in register for spilling found");
3958
3959 // create a new spill interval and assign a stack slot to it
3960 Interval* from_interval = _mapping_from.at(spill_candidate);
3961 Interval* spill_interval = new Interval(-1);
3962 spill_interval->set_type(from_interval->type());
3963
3964 // add a dummy range because real position is difficult to calculate
3965 // Note: this range is a special case when the integrity of the allocation is checked
3966 spill_interval->add_range(1, 2);
3967
3968 // do not allocate a new spill slot for temporary interval, but
3969 // use spill slot assigned to from_interval. Otherwise moves from
3970 // one stack slot to another can happen (not allowed by LIR_Assembler
3971 int spill_slot = from_interval->canonical_spill_slot();
3972 if (spill_slot < 0) {
3973 spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2);
3974 from_interval->set_canonical_spill_slot(spill_slot);
3975 }
3976 spill_interval->assign_reg(spill_slot);
3977 allocator()->append_interval(spill_interval);
3978
3979 TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num()));
3980
3981 // insert a move from register to stack and update the mapping
3982 insert_move(from_interval, spill_interval);
3983 _mapping_from.at_put(spill_candidate, spill_interval);
3984 unblock_registers(from_interval);
3985 }
3986 }
3987
3988 // reset to default value
3989 _multiple_reads_allowed = false;
3990
3991 // check that all intervals have been processed
3992 DEBUG_ONLY(check_empty());
3993 }
3994
3995
3996 void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) {
3997 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));
3998 assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set");
3999
4000 create_insertion_buffer(insert_list);
4001 _insert_list = insert_list;
4002 _insert_idx = insert_idx;
4003 }
4004
4005 void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) {
4006 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));
4007
4008 if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) {
4009 // insert position changed -> resolve current mappings
4010 resolve_mappings();
4011 }
4012
4013 if (insert_list != _insert_list) {
4014 // block changed -> append insertion_buffer because it is
4015 // bound to a specific block and create a new insertion_buffer
4016 append_insertion_buffer();
4017 create_insertion_buffer(insert_list);
4018 }
4019
4020 _insert_list = insert_list;
4021 _insert_idx = insert_idx;
4022 }
4023
4024 void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) {
4025 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()));
4026
4027 _mapping_from.append(from_interval);
4028 _mapping_from_opr.append(LIR_OprFact::illegalOpr);
4029 _mapping_to.append(to_interval);
4030 }
4031
4032
4033 void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) {
4034 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()));
4035 assert(from_opr->is_constant(), "only for constants");
4036
4037 _mapping_from.append(NULL);
4038 _mapping_from_opr.append(from_opr);
4039 _mapping_to.append(to_interval);
4040 }
4041
4042 void MoveResolver::resolve_and_append_moves() {
4043 if (has_mappings()) {
4044 resolve_mappings();
4045 }
4046 append_insertion_buffer();
4047 }
4048
4049
4050
4051 // **** Implementation of Range *************************************
4052
4053 Range::Range(int from, int to, Range* next) :
4054 _from(from),
4055 _to(to),
4056 _next(next)
4057 {
4058 }
4059
4060 // initialize sentinel
4061 Range* Range::_end = NULL;
4062 void Range::initialize(Arena* arena) {
4063 _end = new (arena) Range(max_jint, max_jint, NULL);
4064 }
4065
4066 int Range::intersects_at(Range* r2) const {
4067 const Range* r1 = this;
4068
4069 assert(r1 != NULL && r2 != NULL, "null ranges not allowed");
4070 assert(r1 != _end && r2 != _end, "empty ranges not allowed");
4071
4072 do {
4073 if (r1->from() < r2->from()) {
4074 if (r1->to() <= r2->from()) {
4075 r1 = r1->next(); if (r1 == _end) return -1;
4076 } else {
4077 return r2->from();
4078 }
4079 } else if (r2->from() < r1->from()) {
4080 if (r2->to() <= r1->from()) {
4081 r2 = r2->next(); if (r2 == _end) return -1;
4082 } else {
4083 return r1->from();
4084 }
4085 } else { // r1->from() == r2->from()
4086 if (r1->from() == r1->to()) {
4087 r1 = r1->next(); if (r1 == _end) return -1;
4088 } else if (r2->from() == r2->to()) {
4089 r2 = r2->next(); if (r2 == _end) return -1;
4090 } else {
4091 return r1->from();
4092 }
4093 }
4094 } while (true);
4095 }
4096
4097 #ifndef PRODUCT
4098 void Range::print(outputStream* out) const {
4099 out->print("[%d, %d[ ", _from, _to);
4100 }
4101 #endif
4102
4103
4104
4105 // **** Implementation of Interval **********************************
4106
4107 // initialize sentinel
4108 Interval* Interval::_end = NULL;
4109 void Interval::initialize(Arena* arena) {
4110 Range::initialize(arena);
4111 _end = new (arena) Interval(-1);
4112 }
4113
4114 Interval::Interval(int reg_num) :
4115 _reg_num(reg_num),
4116 _type(T_ILLEGAL),
4117 _first(Range::end()),
4118 _use_pos_and_kinds(12),
4119 _current(Range::end()),
4120 _next(_end),
4121 _state(invalidState),
4122 _assigned_reg(LinearScan::any_reg),
4123 _assigned_regHi(LinearScan::any_reg),
4124 _cached_to(-1),
4125 _cached_opr(LIR_OprFact::illegalOpr),
4126 _cached_vm_reg(VMRegImpl::Bad()),
4127 _split_children(0),
4128 _canonical_spill_slot(-1),
4129 _insert_move_when_activated(false),
4130 _register_hint(NULL),
4131 _spill_state(noDefinitionFound),
4132 _spill_definition_pos(-1)
4133 {
4134 _split_parent = this;
4135 _current_split_child = this;
4136 }
4137
4138 int Interval::calc_to() {
4139 assert(_first != Range::end(), "interval has no range");
4140
4141 Range* r = _first;
4142 while (r->next() != Range::end()) {
4143 r = r->next();
4144 }
4145 return r->to();
4146 }
4147
4148
4149 #ifdef ASSERT
4150 // consistency check of split-children
4151 void Interval::check_split_children() {
4152 if (_split_children.length() > 0) {
4153 assert(is_split_parent(), "only split parents can have children");
4154
4155 for (int i = 0; i < _split_children.length(); i++) {
4156 Interval* i1 = _split_children.at(i);
4157
4158 assert(i1->split_parent() == this, "not a split child of this interval");
4159 assert(i1->type() == type(), "must be equal for all split children");
4160 assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children");
4161
4162 for (int j = i + 1; j < _split_children.length(); j++) {
4163 Interval* i2 = _split_children.at(j);
4164
4165 assert(i1->reg_num() != i2->reg_num(), "same register number");
4166
4167 if (i1->from() < i2->from()) {
4168 assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping");
4169 } else {
4170 assert(i2->from() < i1->from(), "intervals start at same op_id");
4171 assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping");
4172 }
4173 }
4174 }
4175 }
4176 }
4177 #endif // ASSERT
4178
4179 Interval* Interval::register_hint(bool search_split_child) const {
4180 if (!search_split_child) {
4181 return _register_hint;
4182 }
4183
4184 if (_register_hint != NULL) {
4185 assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers");
4186
4187 if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) {
4188 return _register_hint;
4189
4190 } else if (_register_hint->_split_children.length() > 0) {
4191 // search the first split child that has a register assigned
4192 int len = _register_hint->_split_children.length();
4193 for (int i = 0; i < len; i++) {
4194 Interval* cur = _register_hint->_split_children.at(i);
4195
4196 if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) {
4197 return cur;
4198 }
4199 }
4200 }
4201 }
4202
4203 // no hint interval found that has a register assigned
4204 return NULL;
4205 }
4206
4207
4208 Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) {
4209 assert(is_split_parent(), "can only be called for split parents");
4210 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)");
4211
4212 Interval* result;
4213 if (_split_children.length() == 0) {
4214 result = this;
4215 } else {
4216 result = NULL;
4217 int len = _split_children.length();
4218
4219 // in outputMode, the end of the interval (op_id == cur->to()) is not valid
4220 int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1);
4221
4222 int i;
4223 for (i = 0; i < len; i++) {
4224 Interval* cur = _split_children.at(i);
4225 if (cur->from() <= op_id && op_id < cur->to() + to_offset) {
4226 if (i > 0) {
4227 // exchange current split child to start of list (faster access for next call)
4228 _split_children.at_put(i, _split_children.at(0));
4229 _split_children.at_put(0, cur);
4230 }
4231
4232 // interval found
4233 result = cur;
4234 break;
4235 }
4236 }
4237
4238 #ifdef ASSERT
4239 for (i = 0; i < len; i++) {
4240 Interval* tmp = _split_children.at(i);
4241 if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) {
4242 tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num());
4243 result->print();
4244 tmp->print();
4245 assert(false, "two valid result intervals found");
4246 }
4247 }
4248 #endif
4249 }
4250
4251 assert(result != NULL, "no matching interval found");
4252 assert(result->covers(op_id, mode), "op_id not covered by interval");
4253
4254 return result;
4255 }
4256
4257
4258 // returns the last split child that ends before the given op_id
4259 Interval* Interval::split_child_before_op_id(int op_id) {
4260 assert(op_id >= 0, "invalid op_id");
4261
4262 Interval* parent = split_parent();
4263 Interval* result = NULL;
4264
4265 int len = parent->_split_children.length();
4266 assert(len > 0, "no split children available");
4267
4268 for (int i = len - 1; i >= 0; i--) {
4269 Interval* cur = parent->_split_children.at(i);
4270 if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) {
4271 result = cur;
4272 }
4273 }
4274
4275 assert(result != NULL, "no split child found");
4276 return result;
4277 }
4278
4279
4280 // checks if op_id is covered by any split child
4281 bool Interval::split_child_covers(int op_id, LIR_OpVisitState::OprMode mode) {
4282 assert(is_split_parent(), "can only be called for split parents");
4283 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)");
4284
4285 if (_split_children.length() == 0) {
4286 // simple case if interval was not split
4287 return covers(op_id, mode);
4288
4289 } else {
4290 // extended case: check all split children
4291 int len = _split_children.length();
4292 for (int i = 0; i < len; i++) {
4293 Interval* cur = _split_children.at(i);
4294 if (cur->covers(op_id, mode)) {
4295 return true;
4296 }
4297 }
4298 return false;
4299 }
4300 }
4301
4302
4303 // Note: use positions are sorted descending -> first use has highest index
4304 int Interval::first_usage(IntervalUseKind min_use_kind) const {
4305 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4306
4307 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4308 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) {
4309 return _use_pos_and_kinds.at(i);
4310 }
4311 }
4312 return max_jint;
4313 }
4314
4315 int Interval::next_usage(IntervalUseKind min_use_kind, int from) const {
4316 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4317
4318 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4319 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) {
4320 return _use_pos_and_kinds.at(i);
4321 }
4322 }
4323 return max_jint;
4324 }
4325
4326 int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const {
4327 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4328
4329 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4330 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) {
4331 return _use_pos_and_kinds.at(i);
4332 }
4333 }
4334 return max_jint;
4335 }
4336
4337 int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const {
4338 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4339
4340 int prev = 0;
4341 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4342 if (_use_pos_and_kinds.at(i) > from) {
4343 return prev;
4344 }
4345 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) {
4346 prev = _use_pos_and_kinds.at(i);
4347 }
4348 }
4349 return prev;
4350 }
4351
4352 void Interval::add_use_pos(int pos, IntervalUseKind use_kind) {
4353 assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range");
4354
4355 // do not add use positions for precolored intervals because
4356 // they are never used
4357 if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) {
4358 #ifdef ASSERT
4359 assert(_use_pos_and_kinds.length() % 2 == 0, "must be");
4360 for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) {
4361 assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position");
4362 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4363 if (i > 0) {
4364 assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending");
4365 }
4366 }
4367 #endif
4368
4369 // Note: add_use is called in descending order, so list gets sorted
4370 // automatically by just appending new use positions
4371 int len = _use_pos_and_kinds.length();
4372 if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) {
4373 _use_pos_and_kinds.append(pos);
4374 _use_pos_and_kinds.append(use_kind);
4375 } else if (_use_pos_and_kinds.at(len - 1) < use_kind) {
4376 assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly");
4377 _use_pos_and_kinds.at_put(len - 1, use_kind);
4378 }
4379 }
4380 }
4381
4382 void Interval::add_range(int from, int to) {
4383 assert(from < to, "invalid range");
4384 assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval");
4385 assert(from <= first()->to(), "not inserting at begin of interval");
4386
4387 if (first()->from() <= to) {
4388 // join intersecting ranges
4389 first()->set_from(MIN2(from, first()->from()));
4390 first()->set_to (MAX2(to, first()->to()));
4391 } else {
4392 // insert new range
4393 _first = new Range(from, to, first());
4394 }
4395 }
4396
4397 Interval* Interval::new_split_child() {
4398 // allocate new interval
4399 Interval* result = new Interval(-1);
4400 result->set_type(type());
4401
4402 Interval* parent = split_parent();
4403 result->_split_parent = parent;
4404 result->set_register_hint(parent);
4405
4406 // insert new interval in children-list of parent
4407 if (parent->_split_children.length() == 0) {
4408 assert(is_split_parent(), "list must be initialized at first split");
4409
4410 parent->_split_children = IntervalList(4);
4411 parent->_split_children.append(this);
4412 }
4413 parent->_split_children.append(result);
4414
4415 return result;
4416 }
4417
4418 // split this interval at the specified position and return
4419 // the remainder as a new interval.
4420 //
4421 // when an interval is split, a bi-directional link is established between the original interval
4422 // (the split parent) and the intervals that are split off this interval (the split children)
4423 // When a split child is split again, the new created interval is also a direct child
4424 // of the original parent (there is no tree of split children stored, but a flat list)
4425 // All split children are spilled to the same stack slot (stored in _canonical_spill_slot)
4426 //
4427 // Note: The new interval has no valid reg_num
4428 Interval* Interval::split(int split_pos) {
4429 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals");
4430
4431 // allocate new interval
4432 Interval* result = new_split_child();
4433
4434 // split the ranges
4435 Range* prev = NULL;
4436 Range* cur = _first;
4437 while (cur != Range::end() && cur->to() <= split_pos) {
4438 prev = cur;
4439 cur = cur->next();
4440 }
4441 assert(cur != Range::end(), "split interval after end of last range");
4442
4443 if (cur->from() < split_pos) {
4444 result->_first = new Range(split_pos, cur->to(), cur->next());
4445 cur->set_to(split_pos);
4446 cur->set_next(Range::end());
4447
4448 } else {
4449 assert(prev != NULL, "split before start of first range");
4450 result->_first = cur;
4451 prev->set_next(Range::end());
4452 }
4453 result->_current = result->_first;
4454 _cached_to = -1; // clear cached value
4455
4456 // split list of use positions
4457 int total_len = _use_pos_and_kinds.length();
4458 int start_idx = total_len - 2;
4459 while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) {
4460 start_idx -= 2;
4461 }
4462
4463 intStack new_use_pos_and_kinds(total_len - start_idx);
4464 int i;
4465 for (i = start_idx + 2; i < total_len; i++) {
4466 new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i));
4467 }
4468
4469 _use_pos_and_kinds.trunc_to(start_idx + 2);
4470 result->_use_pos_and_kinds = _use_pos_and_kinds;
4471 _use_pos_and_kinds = new_use_pos_and_kinds;
4472
4473 #ifdef ASSERT
4474 assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos");
4475 assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos");
4476 assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries");
4477
4478 for (i = 0; i < _use_pos_and_kinds.length(); i += 2) {
4479 assert(_use_pos_and_kinds.at(i) < split_pos, "must be");
4480 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4481 }
4482 for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) {
4483 assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be");
4484 assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4485 }
4486 #endif
4487
4488 return result;
4489 }
4490
4491 // split this interval at the specified position and return
4492 // the head as a new interval (the original interval is the tail)
4493 //
4494 // Currently, only the first range can be split, and the new interval
4495 // must not have split positions
4496 Interval* Interval::split_from_start(int split_pos) {
4497 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals");
4498 assert(split_pos > from() && split_pos < to(), "can only split inside interval");
4499 assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range");
4500 assert(first_usage(noUse) > split_pos, "can not split when use positions are present");
4501
4502 // allocate new interval
4503 Interval* result = new_split_child();
4504
4505 // the new created interval has only one range (checked by assertion above),
4506 // so the splitting of the ranges is very simple
4507 result->add_range(_first->from(), split_pos);
4508
4509 if (split_pos == _first->to()) {
4510 assert(_first->next() != Range::end(), "must not be at end");
4511 _first = _first->next();
4512 } else {
4513 _first->set_from(split_pos);
4514 }
4515
4516 return result;
4517 }
4518
4519
4520 // returns true if the op_id is inside the interval
4521 bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const {
4522 Range* cur = _first;
4523
4524 while (cur != Range::end() && cur->to() < op_id) {
4525 cur = cur->next();
4526 }
4527 if (cur != Range::end()) {
4528 assert(cur->to() != cur->next()->from(), "ranges not separated");
4529
4530 if (mode == LIR_OpVisitState::outputMode) {
4531 return cur->from() <= op_id && op_id < cur->to();
4532 } else {
4533 return cur->from() <= op_id && op_id <= cur->to();
4534 }
4535 }
4536 return false;
4537 }
4538
4539 // returns true if the interval has any hole between hole_from and hole_to
4540 // (even if the hole has only the length 1)
4541 bool Interval::has_hole_between(int hole_from, int hole_to) {
4542 assert(hole_from < hole_to, "check");
4543 assert(from() <= hole_from && hole_to <= to(), "index out of interval");
4544
4545 Range* cur = _first;
4546 while (cur != Range::end()) {
4547 assert(cur->to() < cur->next()->from(), "no space between ranges");
4548
4549 // hole-range starts before this range -> hole
4550 if (hole_from < cur->from()) {
4551 return true;
4552
4553 // hole-range completely inside this range -> no hole
4554 } else if (hole_to <= cur->to()) {
4555 return false;
4556
4557 // overlapping of hole-range with this range -> hole
4558 } else if (hole_from <= cur->to()) {
4559 return true;
4560 }
4561
4562 cur = cur->next();
4563 }
4564
4565 return false;
4566 }
4567
4568
4569 #ifndef PRODUCT
4570 void Interval::print(outputStream* out) const {
4571 const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" };
4572 const char* UseKind2Name[] = { "N", "L", "S", "M" };
4573
4574 const char* type_name;
4575 LIR_Opr opr = LIR_OprFact::illegal();
4576 if (reg_num() < LIR_OprDesc::vreg_base) {
4577 type_name = "fixed";
4578 // need a temporary operand for fixed intervals because type() cannot be called
4579 #ifdef X86
4580 int last_xmm_reg = pd_last_xmm_reg;
4581 #ifdef _LP64
4582 if (UseAVX < 3) {
4583 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1;
4584 }
4585 #endif
4586 #endif
4587 if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) {
4588 opr = LIR_OprFact::single_cpu(assigned_reg());
4589 } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) {
4590 opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg);
4591 #ifdef X86
4592 } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= last_xmm_reg) {
4593 opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg);
4594 #endif
4595 } else {
4596 ShouldNotReachHere();
4597 }
4598 } else {
4599 type_name = type2name(type());
4600 if (assigned_reg() != -1 &&
4601 (LinearScan::num_physical_regs(type()) == 1 || assigned_regHi() != -1)) {
4602 opr = LinearScan::calc_operand_for_interval(this);
4603 }
4604 }
4605
4606 out->print("%d %s ", reg_num(), type_name);
4607 if (opr->is_valid()) {
4608 out->print("\"");
4609 opr->print(out);
4610 out->print("\" ");
4611 }
4612 out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1));
4613
4614 // print ranges
4615 Range* cur = _first;
4616 while (cur != Range::end()) {
4617 cur->print(out);
4618 cur = cur->next();
4619 assert(cur != NULL, "range list not closed with range sentinel");
4620 }
4621
4622 // print use positions
4623 int prev = 0;
4624 assert(_use_pos_and_kinds.length() % 2 == 0, "must be");
4625 for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4626 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4627 assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted");
4628
4629 out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]);
4630 prev = _use_pos_and_kinds.at(i);
4631 }
4632
4633 out->print(" \"%s\"", SpillState2Name[spill_state()]);
4634 out->cr();
4635 }
4636 #endif
4637
4638
4639
4640 // **** Implementation of IntervalWalker ****************************
4641
4642 IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first)
4643 : _compilation(allocator->compilation())
4644 , _allocator(allocator)
4645 {
4646 _unhandled_first[fixedKind] = unhandled_fixed_first;
4647 _unhandled_first[anyKind] = unhandled_any_first;
4648 _active_first[fixedKind] = Interval::end();
4649 _inactive_first[fixedKind] = Interval::end();
4650 _active_first[anyKind] = Interval::end();
4651 _inactive_first[anyKind] = Interval::end();
4652 _current_position = -1;
4653 _current = NULL;
4654 next_interval();
4655 }
4656
4657
4658 // append interval at top of list
4659 void IntervalWalker::append_unsorted(Interval** list, Interval* interval) {
4660 interval->set_next(*list); *list = interval;
4661 }
4662
4663
4664 // append interval in order of current range from()
4665 void IntervalWalker::append_sorted(Interval** list, Interval* interval) {
4666 Interval* prev = NULL;
4667 Interval* cur = *list;
4668 while (cur->current_from() < interval->current_from()) {
4669 prev = cur; cur = cur->next();
4670 }
4671 if (prev == NULL) {
4672 *list = interval;
4673 } else {
4674 prev->set_next(interval);
4675 }
4676 interval->set_next(cur);
4677 }
4678
4679 void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) {
4680 assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position");
4681
4682 Interval* prev = NULL;
4683 Interval* cur = *list;
4684 while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) {
4685 prev = cur; cur = cur->next();
4686 }
4687 if (prev == NULL) {
4688 *list = interval;
4689 } else {
4690 prev->set_next(interval);
4691 }
4692 interval->set_next(cur);
4693 }
4694
4695
4696 inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) {
4697 while (*list != Interval::end() && *list != i) {
4698 list = (*list)->next_addr();
4699 }
4700 if (*list != Interval::end()) {
4701 assert(*list == i, "check");
4702 *list = (*list)->next();
4703 return true;
4704 } else {
4705 return false;
4706 }
4707 }
4708
4709 void IntervalWalker::remove_from_list(Interval* i) {
4710 bool deleted;
4711
4712 if (i->state() == activeState) {
4713 deleted = remove_from_list(active_first_addr(anyKind), i);
4714 } else {
4715 assert(i->state() == inactiveState, "invalid state");
4716 deleted = remove_from_list(inactive_first_addr(anyKind), i);
4717 }
4718
4719 assert(deleted, "interval has not been found in list");
4720 }
4721
4722
4723 void IntervalWalker::walk_to(IntervalState state, int from) {
4724 assert (state == activeState || state == inactiveState, "wrong state");
4725 for_each_interval_kind(kind) {
4726 Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind);
4727 Interval* next = *prev;
4728 while (next->current_from() <= from) {
4729 Interval* cur = next;
4730 next = cur->next();
4731
4732 bool range_has_changed = false;
4733 while (cur->current_to() <= from) {
4734 cur->next_range();
4735 range_has_changed = true;
4736 }
4737
4738 // also handle move from inactive list to active list
4739 range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from);
4740
4741 if (range_has_changed) {
4742 // remove cur from list
4743 *prev = next;
4744 if (cur->current_at_end()) {
4745 // move to handled state (not maintained as a list)
4746 cur->set_state(handledState);
4747 interval_moved(cur, kind, state, handledState);
4748 } else if (cur->current_from() <= from){
4749 // sort into active list
4750 append_sorted(active_first_addr(kind), cur);
4751 cur->set_state(activeState);
4752 if (*prev == cur) {
4753 assert(state == activeState, "check");
4754 prev = cur->next_addr();
4755 }
4756 interval_moved(cur, kind, state, activeState);
4757 } else {
4758 // sort into inactive list
4759 append_sorted(inactive_first_addr(kind), cur);
4760 cur->set_state(inactiveState);
4761 if (*prev == cur) {
4762 assert(state == inactiveState, "check");
4763 prev = cur->next_addr();
4764 }
4765 interval_moved(cur, kind, state, inactiveState);
4766 }
4767 } else {
4768 prev = cur->next_addr();
4769 continue;
4770 }
4771 }
4772 }
4773 }
4774
4775
4776 void IntervalWalker::next_interval() {
4777 IntervalKind kind;
4778 Interval* any = _unhandled_first[anyKind];
4779 Interval* fixed = _unhandled_first[fixedKind];
4780
4781 if (any != Interval::end()) {
4782 // intervals may start at same position -> prefer fixed interval
4783 kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind;
4784
4785 assert (kind == fixedKind && fixed->from() <= any->from() ||
4786 kind == anyKind && any->from() <= fixed->from(), "wrong interval!!!");
4787 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");
4788
4789 } else if (fixed != Interval::end()) {
4790 kind = fixedKind;
4791 } else {
4792 _current = NULL; return;
4793 }
4794 _current_kind = kind;
4795 _current = _unhandled_first[kind];
4796 _unhandled_first[kind] = _current->next();
4797 _current->set_next(Interval::end());
4798 _current->rewind_range();
4799 }
4800
4801
4802 void IntervalWalker::walk_to(int lir_op_id) {
4803 assert(_current_position <= lir_op_id, "can not walk backwards");
4804 while (current() != NULL) {
4805 bool is_active = current()->from() <= lir_op_id;
4806 int id = is_active ? current()->from() : lir_op_id;
4807
4808 TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); })
4809
4810 // set _current_position prior to call of walk_to
4811 _current_position = id;
4812
4813 // call walk_to even if _current_position == id
4814 walk_to(activeState, id);
4815 walk_to(inactiveState, id);
4816
4817 if (is_active) {
4818 current()->set_state(activeState);
4819 if (activate_current()) {
4820 append_sorted(active_first_addr(current_kind()), current());
4821 interval_moved(current(), current_kind(), unhandledState, activeState);
4822 }
4823
4824 next_interval();
4825 } else {
4826 return;
4827 }
4828 }
4829 }
4830
4831 void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) {
4832 #ifndef PRODUCT
4833 if (TraceLinearScanLevel >= 4) {
4834 #define print_state(state) \
4835 switch(state) {\
4836 case unhandledState: tty->print("unhandled"); break;\
4837 case activeState: tty->print("active"); break;\
4838 case inactiveState: tty->print("inactive"); break;\
4839 case handledState: tty->print("handled"); break;\
4840 default: ShouldNotReachHere(); \
4841 }
4842
4843 print_state(from); tty->print(" to "); print_state(to);
4844 tty->fill_to(23);
4845 interval->print();
4846
4847 #undef print_state
4848 }
4849 #endif
4850 }
4851
4852
4853
4854 // **** Implementation of LinearScanWalker **************************
4855
4856 LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first)
4857 : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first)
4858 , _move_resolver(allocator)
4859 {
4860 for (int i = 0; i < LinearScan::nof_regs; i++) {
4861 _spill_intervals[i] = new IntervalList(2);
4862 }
4863 }
4864
4865
4866 inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) {
4867 for (int i = _first_reg; i <= _last_reg; i++) {
4868 _use_pos[i] = max_jint;
4869
4870 if (!only_process_use_pos) {
4871 _block_pos[i] = max_jint;
4872 _spill_intervals[i]->clear();
4873 }
4874 }
4875 }
4876
4877 inline void LinearScanWalker::exclude_from_use(int reg) {
4878 assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)");
4879 if (reg >= _first_reg && reg <= _last_reg) {
4880 _use_pos[reg] = 0;
4881 }
4882 }
4883 inline void LinearScanWalker::exclude_from_use(Interval* i) {
4884 assert(i->assigned_reg() != any_reg, "interval has no register assigned");
4885
4886 exclude_from_use(i->assigned_reg());
4887 exclude_from_use(i->assigned_regHi());
4888 }
4889
4890 inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) {
4891 assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0");
4892
4893 if (reg >= _first_reg && reg <= _last_reg) {
4894 if (_use_pos[reg] > use_pos) {
4895 _use_pos[reg] = use_pos;
4896 }
4897 if (!only_process_use_pos) {
4898 _spill_intervals[reg]->append(i);
4899 }
4900 }
4901 }
4902 inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) {
4903 assert(i->assigned_reg() != any_reg, "interval has no register assigned");
4904 if (use_pos != -1) {
4905 set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos);
4906 set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos);
4907 }
4908 }
4909
4910 inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) {
4911 if (reg >= _first_reg && reg <= _last_reg) {
4912 if (_block_pos[reg] > block_pos) {
4913 _block_pos[reg] = block_pos;
4914 }
4915 if (_use_pos[reg] > block_pos) {
4916 _use_pos[reg] = block_pos;
4917 }
4918 }
4919 }
4920 inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) {
4921 assert(i->assigned_reg() != any_reg, "interval has no register assigned");
4922 if (block_pos != -1) {
4923 set_block_pos(i->assigned_reg(), i, block_pos);
4924 set_block_pos(i->assigned_regHi(), i, block_pos);
4925 }
4926 }
4927
4928
4929 void LinearScanWalker::free_exclude_active_fixed() {
4930 Interval* list = active_first(fixedKind);
4931 while (list != Interval::end()) {
4932 assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned");
4933 exclude_from_use(list);
4934 list = list->next();
4935 }
4936 }
4937
4938 void LinearScanWalker::free_exclude_active_any() {
4939 Interval* list = active_first(anyKind);
4940 while (list != Interval::end()) {
4941 exclude_from_use(list);
4942 list = list->next();
4943 }
4944 }
4945
4946 void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) {
4947 Interval* list = inactive_first(fixedKind);
4948 while (list != Interval::end()) {
4949 if (cur->to() <= list->current_from()) {
4950 assert(list->current_intersects_at(cur) == -1, "must not intersect");
4951 set_use_pos(list, list->current_from(), true);
4952 } else {
4953 set_use_pos(list, list->current_intersects_at(cur), true);
4954 }
4955 list = list->next();
4956 }
4957 }
4958
4959 void LinearScanWalker::free_collect_inactive_any(Interval* cur) {
4960 Interval* list = inactive_first(anyKind);
4961 while (list != Interval::end()) {
4962 set_use_pos(list, list->current_intersects_at(cur), true);
4963 list = list->next();
4964 }
4965 }
4966
4967 void LinearScanWalker::free_collect_unhandled(IntervalKind kind, Interval* cur) {
4968 Interval* list = unhandled_first(kind);
4969 while (list != Interval::end()) {
4970 set_use_pos(list, list->intersects_at(cur), true);
4971 if (kind == fixedKind && cur->to() <= list->from()) {
4972 set_use_pos(list, list->from(), true);
4973 }
4974 list = list->next();
4975 }
4976 }
4977
4978 void LinearScanWalker::spill_exclude_active_fixed() {
4979 Interval* list = active_first(fixedKind);
4980 while (list != Interval::end()) {
4981 exclude_from_use(list);
4982 list = list->next();
4983 }
4984 }
4985
4986 void LinearScanWalker::spill_block_unhandled_fixed(Interval* cur) {
4987 Interval* list = unhandled_first(fixedKind);
4988 while (list != Interval::end()) {
4989 set_block_pos(list, list->intersects_at(cur));
4990 list = list->next();
4991 }
4992 }
4993
4994 void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) {
4995 Interval* list = inactive_first(fixedKind);
4996 while (list != Interval::end()) {
4997 if (cur->to() > list->current_from()) {
4998 set_block_pos(list, list->current_intersects_at(cur));
4999 } else {
5000 assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect");
5001 }
5002
5003 list = list->next();
5004 }
5005 }
5006
5007 void LinearScanWalker::spill_collect_active_any() {
5008 Interval* list = active_first(anyKind);
5009 while (list != Interval::end()) {
5010 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false);
5011 list = list->next();
5012 }
5013 }
5014
5015 void LinearScanWalker::spill_collect_inactive_any(Interval* cur) {
5016 Interval* list = inactive_first(anyKind);
5017 while (list != Interval::end()) {
5018 if (list->current_intersects(cur)) {
5019 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false);
5020 }
5021 list = list->next();
5022 }
5023 }
5024
5025
5026 void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) {
5027 // output all moves here. When source and target are equal, the move is
5028 // optimized away later in assign_reg_nums
5029
5030 op_id = (op_id + 1) & ~1;
5031 BlockBegin* op_block = allocator()->block_of_op_with_id(op_id);
5032 assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary");
5033
5034 // calculate index of instruction inside instruction list of current block
5035 // the minimal index (for a block with no spill moves) can be calculated because the
5036 // numbering of instructions is known.
5037 // When the block already contains spill moves, the index must be increased until the
5038 // correct index is reached.
5039 LIR_OpList* list = op_block->lir()->instructions_list();
5040 int index = (op_id - list->at(0)->id()) / 2;
5041 assert(list->at(index)->id() <= op_id, "error in calculation");
5042
5043 while (list->at(index)->id() != op_id) {
5044 index++;
5045 assert(0 <= index && index < list->length(), "index out of bounds");
5046 }
5047 assert(1 <= index && index < list->length(), "index out of bounds");
5048 assert(list->at(index)->id() == op_id, "error in calculation");
5049
5050 // insert new instruction before instruction at position index
5051 _move_resolver.move_insert_position(op_block->lir(), index - 1);
5052 _move_resolver.add_mapping(src_it, dst_it);
5053 }
5054
5055
5056 int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) {
5057 int from_block_nr = min_block->linear_scan_number();
5058 int to_block_nr = max_block->linear_scan_number();
5059
5060 assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range");
5061 assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range");
5062 assert(from_block_nr < to_block_nr, "must cross block boundary");
5063
5064 // Try to split at end of max_block. If this would be after
5065 // max_split_pos, then use the begin of max_block
5066 int optimal_split_pos = max_block->last_lir_instruction_id() + 2;
5067 if (optimal_split_pos > max_split_pos) {
5068 optimal_split_pos = max_block->first_lir_instruction_id();
5069 }
5070
5071 int min_loop_depth = max_block->loop_depth();
5072 for (int i = to_block_nr - 1; i >= from_block_nr; i--) {
5073 BlockBegin* cur = block_at(i);
5074
5075 if (cur->loop_depth() < min_loop_depth) {
5076 // block with lower loop-depth found -> split at the end of this block
5077 min_loop_depth = cur->loop_depth();
5078 optimal_split_pos = cur->last_lir_instruction_id() + 2;
5079 }
5080 }
5081 assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary");
5082
5083 return optimal_split_pos;
5084 }
5085
5086
5087 int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) {
5088 int optimal_split_pos = -1;
5089 if (min_split_pos == max_split_pos) {
5090 // trivial case, no optimization of split position possible
5091 TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible"));
5092 optimal_split_pos = min_split_pos;
5093
5094 } else {
5095 assert(min_split_pos < max_split_pos, "must be true then");
5096 assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise");
5097
5098 // reason for using min_split_pos - 1: when the minimal split pos is exactly at the
5099 // beginning of a block, then min_split_pos is also a possible split position.
5100 // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos
5101 BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1);
5102
5103 // reason for using max_split_pos - 1: otherwise there would be an assertion failure
5104 // when an interval ends at the end of the last block of the method
5105 // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no
5106 // block at this op_id)
5107 BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1);
5108
5109 assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order");
5110 if (min_block == max_block) {
5111 // split position cannot be moved to block boundary, so split as late as possible
5112 TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block"));
5113 optimal_split_pos = max_split_pos;
5114
5115 } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) {
5116 // Do not move split position if the interval has a hole before max_split_pos.
5117 // Intervals resulting from Phi-Functions have more than one definition (marked
5118 // as mustHaveRegister) with a hole before each definition. When the register is needed
5119 // for the second definition, an earlier reloading is unnecessary.
5120 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos"));
5121 optimal_split_pos = max_split_pos;
5122
5123 } else {
5124 // seach optimal block boundary between min_split_pos and max_split_pos
5125 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()));
5126
5127 if (do_loop_optimization) {
5128 // Loop optimization: if a loop-end marker is found between min- and max-position,
5129 // then split before this loop
5130 int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2);
5131 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d", loop_end_pos));
5132
5133 assert(loop_end_pos > min_split_pos, "invalid order");
5134 if (loop_end_pos < max_split_pos) {
5135 // loop-end marker found between min- and max-position
5136 // if it is not the end marker for the same loop as the min-position, then move
5137 // the max-position to this loop block.
5138 // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading
5139 // of the interval (normally, only mustHaveRegister causes a reloading)
5140 BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos);
5141
5142 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()));
5143 assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between");
5144
5145 optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2);
5146 if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) {
5147 optimal_split_pos = -1;
5148 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary"));
5149 } else {
5150 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful"));
5151 }
5152 }
5153 }
5154
5155 if (optimal_split_pos == -1) {
5156 // not calculated by loop optimization
5157 optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos);
5158 }
5159 }
5160 }
5161 TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d", optimal_split_pos));
5162
5163 return optimal_split_pos;
5164 }
5165
5166
5167 /*
5168 split an interval at the optimal position between min_split_pos and
5169 max_split_pos in two parts:
5170 1) the left part has already a location assigned
5171 2) the right part is sorted into to the unhandled-list
5172 */
5173 void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) {
5174 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: "); it->print());
5175 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos));
5176
5177 assert(it->from() < min_split_pos, "cannot split at start of interval");
5178 assert(current_position() < min_split_pos, "cannot split before current position");
5179 assert(min_split_pos <= max_split_pos, "invalid order");
5180 assert(max_split_pos <= it->to(), "cannot split after end of interval");
5181
5182 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true);
5183
5184 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range");
5185 assert(optimal_split_pos <= it->to(), "cannot split after end of interval");
5186 assert(optimal_split_pos > it->from(), "cannot split at start of interval");
5187
5188 if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) {
5189 // the split position would be just before the end of the interval
5190 // -> no split at all necessary
5191 TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval"));
5192 return;
5193 }
5194
5195 // must calculate this before the actual split is performed and before split position is moved to odd op_id
5196 bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos);
5197
5198 if (!allocator()->is_block_begin(optimal_split_pos)) {
5199 // move position before actual instruction (odd op_id)
5200 optimal_split_pos = (optimal_split_pos - 1) | 1;
5201 }
5202
5203 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos));
5204 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary");
5205 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary");
5206
5207 Interval* split_part = it->split(optimal_split_pos);
5208
5209 allocator()->append_interval(split_part);
5210 allocator()->copy_register_flags(it, split_part);
5211 split_part->set_insert_move_when_activated(move_necessary);
5212 append_to_unhandled(unhandled_first_addr(anyKind), split_part);
5213
5214 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)", move_necessary));
5215 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print());
5216 TRACE_LINEAR_SCAN(2, tty->print (" "); split_part->print());
5217 }
5218
5219 /*
5220 split an interval at the optimal position between min_split_pos and
5221 max_split_pos in two parts:
5222 1) the left part has already a location assigned
5223 2) the right part is always on the stack and therefore ignored in further processing
5224 */
5225 void LinearScanWalker::split_for_spilling(Interval* it) {
5226 // calculate allowed range of splitting position
5227 int max_split_pos = current_position();
5228 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from());
5229
5230 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: "); it->print());
5231 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos));
5232
5233 assert(it->state() == activeState, "why spill interval that is not active?");
5234 assert(it->from() <= min_split_pos, "cannot split before start of interval");
5235 assert(min_split_pos <= max_split_pos, "invalid order");
5236 assert(max_split_pos < it->to(), "cannot split at end end of interval");
5237 assert(current_position() < it->to(), "interval must not end before current position");
5238
5239 if (min_split_pos == it->from()) {
5240 // the whole interval is never used, so spill it entirely to memory
5241 TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval"));
5242 assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position");
5243
5244 allocator()->assign_spill_slot(it);
5245 allocator()->change_spill_state(it, min_split_pos);
5246
5247 // Also kick parent intervals out of register to memory when they have no use
5248 // position. This avoids short interval in register surrounded by intervals in
5249 // memory -> avoid useless moves from memory to register and back
5250 Interval* parent = it;
5251 while (parent != NULL && parent->is_split_child()) {
5252 parent = parent->split_child_before_op_id(parent->from());
5253
5254 if (parent->assigned_reg() < LinearScan::nof_regs) {
5255 if (parent->first_usage(shouldHaveRegister) == max_jint) {
5256 // parent is never used, so kick it out of its assigned register
5257 TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used", parent->reg_num()));
5258 allocator()->assign_spill_slot(parent);
5259 } else {
5260 // do not go further back because the register is actually used by the interval
5261 parent = NULL;
5262 }
5263 }
5264 }
5265
5266 } else {
5267 // search optimal split pos, split interval and spill only the right hand part
5268 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false);
5269
5270 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range");
5271 assert(optimal_split_pos < it->to(), "cannot split at end of interval");
5272 assert(optimal_split_pos >= it->from(), "cannot split before start of interval");
5273
5274 if (!allocator()->is_block_begin(optimal_split_pos)) {
5275 // move position before actual instruction (odd op_id)
5276 optimal_split_pos = (optimal_split_pos - 1) | 1;
5277 }
5278
5279 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos));
5280 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary");
5281 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary");
5282
5283 Interval* spilled_part = it->split(optimal_split_pos);
5284 allocator()->append_interval(spilled_part);
5285 allocator()->assign_spill_slot(spilled_part);
5286 allocator()->change_spill_state(spilled_part, optimal_split_pos);
5287
5288 if (!allocator()->is_block_begin(optimal_split_pos)) {
5289 TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num()));
5290 insert_move(optimal_split_pos, it, spilled_part);
5291 }
5292
5293 // the current_split_child is needed later when moves are inserted for reloading
5294 assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child");
5295 spilled_part->make_current_split_child();
5296
5297 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts"));
5298 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print());
5299 TRACE_LINEAR_SCAN(2, tty->print (" "); spilled_part->print());
5300 }
5301 }
5302
5303
5304 void LinearScanWalker::split_stack_interval(Interval* it) {
5305 int min_split_pos = current_position() + 1;
5306 int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to());
5307
5308 split_before_usage(it, min_split_pos, max_split_pos);
5309 }
5310
5311 void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) {
5312 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1);
5313 int max_split_pos = register_available_until;
5314
5315 split_before_usage(it, min_split_pos, max_split_pos);
5316 }
5317
5318 void LinearScanWalker::split_and_spill_interval(Interval* it) {
5319 assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed");
5320
5321 int current_pos = current_position();
5322 if (it->state() == inactiveState) {
5323 // the interval is currently inactive, so no spill slot is needed for now.
5324 // when the split part is activated, the interval has a new chance to get a register,
5325 // so in the best case no stack slot is necessary
5326 assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise");
5327 split_before_usage(it, current_pos + 1, current_pos + 1);
5328
5329 } else {
5330 // search the position where the interval must have a register and split
5331 // at the optimal position before.
5332 // The new created part is added to the unhandled list and will get a register
5333 // when it is activated
5334 int min_split_pos = current_pos + 1;
5335 int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to());
5336
5337 split_before_usage(it, min_split_pos, max_split_pos);
5338
5339 assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register");
5340 split_for_spilling(it);
5341 }
5342 }
5343
5344
5345 int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) {
5346 int min_full_reg = any_reg;
5347 int max_partial_reg = any_reg;
5348
5349 for (int i = _first_reg; i <= _last_reg; i++) {
5350 if (i == ignore_reg) {
5351 // this register must be ignored
5352
5353 } else if (_use_pos[i] >= interval_to) {
5354 // this register is free for the full interval
5355 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) {
5356 min_full_reg = i;
5357 }
5358 } else if (_use_pos[i] > reg_needed_until) {
5359 // this register is at least free until reg_needed_until
5360 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) {
5361 max_partial_reg = i;
5362 }
5363 }
5364 }
5365
5366 if (min_full_reg != any_reg) {
5367 return min_full_reg;
5368 } else if (max_partial_reg != any_reg) {
5369 *need_split = true;
5370 return max_partial_reg;
5371 } else {
5372 return any_reg;
5373 }
5374 }
5375
5376 int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) {
5377 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm");
5378
5379 int min_full_reg = any_reg;
5380 int max_partial_reg = any_reg;
5381
5382 for (int i = _first_reg; i < _last_reg; i+=2) {
5383 if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) {
5384 // this register is free for the full interval
5385 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) {
5386 min_full_reg = i;
5387 }
5388 } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) {
5389 // this register is at least free until reg_needed_until
5390 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) {
5391 max_partial_reg = i;
5392 }
5393 }
5394 }
5395
5396 if (min_full_reg != any_reg) {
5397 return min_full_reg;
5398 } else if (max_partial_reg != any_reg) {
5399 *need_split = true;
5400 return max_partial_reg;
5401 } else {
5402 return any_reg;
5403 }
5404 }
5405
5406
5407 bool LinearScanWalker::alloc_free_reg(Interval* cur) {
5408 TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print());
5409
5410 init_use_lists(true);
5411 free_exclude_active_fixed();
5412 free_exclude_active_any();
5413 free_collect_inactive_fixed(cur);
5414 free_collect_inactive_any(cur);
5415 // free_collect_unhandled(fixedKind, cur);
5416 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0");
5417
5418 // _use_pos contains the start of the next interval that has this register assigned
5419 // (either as a fixed register or a normal allocated register in the past)
5420 // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely
5421 TRACE_LINEAR_SCAN(4, tty->print_cr(" state of registers:"));
5422 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]));
5423
5424 int hint_reg, hint_regHi;
5425 Interval* register_hint = cur->register_hint();
5426 if (register_hint != NULL) {
5427 hint_reg = register_hint->assigned_reg();
5428 hint_regHi = register_hint->assigned_regHi();
5429
5430 if (allocator()->is_precolored_cpu_interval(register_hint)) {
5431 assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals");
5432 hint_regHi = hint_reg + 1; // connect e.g. eax-edx
5433 }
5434 TRACE_LINEAR_SCAN(4, tty->print(" hint registers %d, %d from interval ", hint_reg, hint_regHi); register_hint->print());
5435
5436 } else {
5437 hint_reg = any_reg;
5438 hint_regHi = any_reg;
5439 }
5440 assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal");
5441 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval");
5442
5443 // the register must be free at least until this position
5444 int reg_needed_until = cur->from() + 1;
5445 int interval_to = cur->to();
5446
5447 bool need_split = false;
5448 int split_pos = -1;
5449 int reg = any_reg;
5450 int regHi = any_reg;
5451
5452 if (_adjacent_regs) {
5453 reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split);
5454 regHi = reg + 1;
5455 if (reg == any_reg) {
5456 return false;
5457 }
5458 split_pos = MIN2(_use_pos[reg], _use_pos[regHi]);
5459
5460 } else {
5461 reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split);
5462 if (reg == any_reg) {
5463 return false;
5464 }
5465 split_pos = _use_pos[reg];
5466
5467 if (_num_phys_regs == 2) {
5468 regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split);
5469
5470 if (_use_pos[reg] < interval_to && regHi == any_reg) {
5471 // do not split interval if only one register can be assigned until the split pos
5472 // (when one register is found for the whole interval, split&spill is only
5473 // performed for the hi register)
5474 return false;
5475
5476 } else if (regHi != any_reg) {
5477 split_pos = MIN2(split_pos, _use_pos[regHi]);
5478
5479 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax
5480 if (reg > regHi) {
5481 int temp = reg;
5482 reg = regHi;
5483 regHi = temp;
5484 }
5485 }
5486 }
5487 }
5488
5489 cur->assign_reg(reg, regHi);
5490 TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d", reg, regHi));
5491
5492 assert(split_pos > 0, "invalid split_pos");
5493 if (need_split) {
5494 // register not available for full interval, so split it
5495 split_when_partial_register_available(cur, split_pos);
5496 }
5497
5498 // only return true if interval is completely assigned
5499 return _num_phys_regs == 1 || regHi != any_reg;
5500 }
5501
5502
5503 int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) {
5504 int max_reg = any_reg;
5505
5506 for (int i = _first_reg; i <= _last_reg; i++) {
5507 if (i == ignore_reg) {
5508 // this register must be ignored
5509
5510 } else if (_use_pos[i] > reg_needed_until) {
5511 if (max_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_reg] && max_reg != hint_reg)) {
5512 max_reg = i;
5513 }
5514 }
5515 }
5516
5517 if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) {
5518 *need_split = true;
5519 }
5520
5521 return max_reg;
5522 }
5523
5524 int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) {
5525 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm");
5526
5527 int max_reg = any_reg;
5528
5529 for (int i = _first_reg; i < _last_reg; i+=2) {
5530 if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) {
5531 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) {
5532 max_reg = i;
5533 }
5534 }
5535 }
5536
5537 if (max_reg != any_reg &&
5538 (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to)) {
5539 *need_split = true;
5540 }
5541
5542 return max_reg;
5543 }
5544
5545 void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) {
5546 assert(reg != any_reg, "no register assigned");
5547
5548 for (int i = 0; i < _spill_intervals[reg]->length(); i++) {
5549 Interval* it = _spill_intervals[reg]->at(i);
5550 remove_from_list(it);
5551 split_and_spill_interval(it);
5552 }
5553
5554 if (regHi != any_reg) {
5555 IntervalList* processed = _spill_intervals[reg];
5556 for (int i = 0; i < _spill_intervals[regHi]->length(); i++) {
5557 Interval* it = _spill_intervals[regHi]->at(i);
5558 if (processed->find(it) == -1) {
5559 remove_from_list(it);
5560 split_and_spill_interval(it);
5561 }
5562 }
5563 }
5564 }
5565
5566
5567 // Split an Interval and spill it to memory so that cur can be placed in a register
5568 void LinearScanWalker::alloc_locked_reg(Interval* cur) {
5569 TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print());
5570
5571 // collect current usage of registers
5572 init_use_lists(false);
5573 spill_exclude_active_fixed();
5574 // spill_block_unhandled_fixed(cur);
5575 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0");
5576 spill_block_inactive_fixed(cur);
5577 spill_collect_active_any();
5578 spill_collect_inactive_any(cur);
5579
5580 #ifndef PRODUCT
5581 if (TraceLinearScanLevel >= 4) {
5582 tty->print_cr(" state of registers:");
5583 for (int i = _first_reg; i <= _last_reg; i++) {
5584 tty->print(" reg %d: use_pos: %d, block_pos: %d, intervals: ", i, _use_pos[i], _block_pos[i]);
5585 for (int j = 0; j < _spill_intervals[i]->length(); j++) {
5586 tty->print("%d ", _spill_intervals[i]->at(j)->reg_num());
5587 }
5588 tty->cr();
5589 }
5590 }
5591 #endif
5592
5593 // the register must be free at least until this position
5594 int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1);
5595 int interval_to = cur->to();
5596 assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use");
5597
5598 int split_pos = 0;
5599 int use_pos = 0;
5600 bool need_split = false;
5601 int reg, regHi;
5602
5603 if (_adjacent_regs) {
5604 reg = find_locked_double_reg(reg_needed_until, interval_to, any_reg, &need_split);
5605 regHi = reg + 1;
5606
5607 if (reg != any_reg) {
5608 use_pos = MIN2(_use_pos[reg], _use_pos[regHi]);
5609 split_pos = MIN2(_block_pos[reg], _block_pos[regHi]);
5610 }
5611 } else {
5612 reg = find_locked_reg(reg_needed_until, interval_to, any_reg, cur->assigned_reg(), &need_split);
5613 regHi = any_reg;
5614
5615 if (reg != any_reg) {
5616 use_pos = _use_pos[reg];
5617 split_pos = _block_pos[reg];
5618
5619 if (_num_phys_regs == 2) {
5620 if (cur->assigned_reg() != any_reg) {
5621 regHi = reg;
5622 reg = cur->assigned_reg();
5623 } else {
5624 regHi = find_locked_reg(reg_needed_until, interval_to, any_reg, reg, &need_split);
5625 if (regHi != any_reg) {
5626 use_pos = MIN2(use_pos, _use_pos[regHi]);
5627 split_pos = MIN2(split_pos, _block_pos[regHi]);
5628 }
5629 }
5630
5631 if (regHi != any_reg && reg > regHi) {
5632 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax
5633 int temp = reg;
5634 reg = regHi;
5635 regHi = temp;
5636 }
5637 }
5638 }
5639 }
5640
5641 if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) {
5642 // the first use of cur is later than the spilling position -> spill cur
5643 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));
5644
5645 if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) {
5646 assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)");
5647 // assign a reasonable register and do a bailout in product mode to avoid errors
5648 allocator()->assign_spill_slot(cur);
5649 BAILOUT("LinearScan: no register found");
5650 }
5651
5652 split_and_spill_interval(cur);
5653 } else {
5654 TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d", reg, regHi));
5655 assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found");
5656 assert(split_pos > 0, "invalid split_pos");
5657 assert(need_split == false || split_pos > cur->from(), "splitting interval at from");
5658
5659 cur->assign_reg(reg, regHi);
5660 if (need_split) {
5661 // register not available for full interval, so split it
5662 split_when_partial_register_available(cur, split_pos);
5663 }
5664
5665 // perform splitting and spilling for all affected intervalls
5666 split_and_spill_intersecting_intervals(reg, regHi);
5667 }
5668 }
5669
5670 bool LinearScanWalker::no_allocation_possible(Interval* cur) {
5671 #ifdef X86
5672 // fast calculation of intervals that can never get a register because the
5673 // the next instruction is a call that blocks all registers
5674 // Note: this does not work if callee-saved registers are available (e.g. on Sparc)
5675
5676 // check if this interval is the result of a split operation
5677 // (an interval got a register until this position)
5678 int pos = cur->from();
5679 if ((pos & 1) == 1) {
5680 // the current instruction is a call that blocks all registers
5681 if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) {
5682 TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call"));
5683
5684 // safety check that there is really no register available
5685 assert(alloc_free_reg(cur) == false, "found a register for this interval");
5686 return true;
5687 }
5688
5689 }
5690 #endif
5691 return false;
5692 }
5693
5694 void LinearScanWalker::init_vars_for_alloc(Interval* cur) {
5695 BasicType type = cur->type();
5696 _num_phys_regs = LinearScan::num_physical_regs(type);
5697 _adjacent_regs = LinearScan::requires_adjacent_regs(type);
5698
5699 if (pd_init_regs_for_alloc(cur)) {
5700 // the appropriate register range was selected.
5701 } else if (type == T_FLOAT || type == T_DOUBLE) {
5702 _first_reg = pd_first_fpu_reg;
5703 _last_reg = pd_last_fpu_reg;
5704 } else {
5705 _first_reg = pd_first_cpu_reg;
5706 _last_reg = FrameMap::last_cpu_reg();
5707 }
5708
5709 assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range");
5710 assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range");
5711 }
5712
5713
5714 bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) {
5715 if (op->code() != lir_move) {
5716 return false;
5717 }
5718 assert(op->as_Op1() != NULL, "move must be LIR_Op1");
5719
5720 LIR_Opr in = ((LIR_Op1*)op)->in_opr();
5721 LIR_Opr res = ((LIR_Op1*)op)->result_opr();
5722 return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num();
5723 }
5724
5725 // optimization (especially for phi functions of nested loops):
5726 // assign same spill slot to non-intersecting intervals
5727 void LinearScanWalker::combine_spilled_intervals(Interval* cur) {
5728 if (cur->is_split_child()) {
5729 // optimization is only suitable for split parents
5730 return;
5731 }
5732
5733 Interval* register_hint = cur->register_hint(false);
5734 if (register_hint == NULL) {
5735 // cur is not the target of a move, otherwise register_hint would be set
5736 return;
5737 }
5738 assert(register_hint->is_split_parent(), "register hint must be split parent");
5739
5740 if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) {
5741 // combining the stack slots for intervals where spill move optimization is applied
5742 // is not benefitial and would cause problems
5743 return;
5744 }
5745
5746 int begin_pos = cur->from();
5747 int end_pos = cur->to();
5748 if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) {
5749 // safety check that lir_op_with_id is allowed
5750 return;
5751 }
5752
5753 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)) {
5754 // cur and register_hint are not connected with two moves
5755 return;
5756 }
5757
5758 Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode);
5759 Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode);
5760 if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) {
5761 // register_hint must be split, otherwise the re-writing of use positions does not work
5762 return;
5763 }
5764
5765 assert(begin_hint->assigned_reg() != any_reg, "must have register assigned");
5766 assert(end_hint->assigned_reg() == any_reg, "must not have register assigned");
5767 assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move");
5768 assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move");
5769
5770 if (begin_hint->assigned_reg() < LinearScan::nof_regs) {
5771 // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur
5772 return;
5773 }
5774 assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled");
5775
5776 // modify intervals such that cur gets the same stack slot as register_hint
5777 // delete use positions to prevent the intervals to get a register at beginning
5778 cur->set_canonical_spill_slot(register_hint->canonical_spill_slot());
5779 cur->remove_first_use_pos();
5780 end_hint->remove_first_use_pos();
5781 }
5782
5783
5784 // allocate a physical register or memory location to an interval
5785 bool LinearScanWalker::activate_current() {
5786 Interval* cur = current();
5787 bool result = true;
5788
5789 TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval "); cur->print());
5790 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()));
5791
5792 if (cur->assigned_reg() >= LinearScan::nof_regs) {
5793 // activating an interval that has a stack slot assigned -> split it at first use position
5794 // used for method parameters
5795 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use"));
5796
5797 split_stack_interval(cur);
5798 result = false;
5799
5800 } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) {
5801 // activating an interval that must start in a stack slot, but may get a register later
5802 // used for lir_roundfp: rounding is done by store to stack and reload later
5803 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use"));
5804 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned");
5805
5806 allocator()->assign_spill_slot(cur);
5807 split_stack_interval(cur);
5808 result = false;
5809
5810 } else if (cur->assigned_reg() == any_reg) {
5811 // interval has not assigned register -> normal allocation
5812 // (this is the normal case for most intervals)
5813 TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register"));
5814
5815 // assign same spill slot to non-intersecting intervals
5816 combine_spilled_intervals(cur);
5817
5818 init_vars_for_alloc(cur);
5819 if (no_allocation_possible(cur) || !alloc_free_reg(cur)) {
5820 // no empty register available.
5821 // split and spill another interval so that this interval gets a register
5822 alloc_locked_reg(cur);
5823 }
5824
5825 // spilled intervals need not be move to active-list
5826 if (cur->assigned_reg() >= LinearScan::nof_regs) {
5827 result = false;
5828 }
5829 }
5830
5831 // load spilled values that become active from stack slot to register
5832 if (cur->insert_move_when_activated()) {
5833 assert(cur->is_split_child(), "must be");
5834 assert(cur->current_split_child() != NULL, "must be");
5835 assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval");
5836 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()));
5837
5838 insert_move(cur->from(), cur->current_split_child(), cur);
5839 }
5840 cur->make_current_split_child();
5841
5842 return result; // true = interval is moved to active list
5843 }
5844
5845
5846 // Implementation of EdgeMoveOptimizer
5847
5848 EdgeMoveOptimizer::EdgeMoveOptimizer() :
5849 _edge_instructions(4),
5850 _edge_instructions_idx(4)
5851 {
5852 }
5853
5854 void EdgeMoveOptimizer::optimize(BlockList* code) {
5855 EdgeMoveOptimizer optimizer = EdgeMoveOptimizer();
5856
5857 // ignore the first block in the list (index 0 is not processed)
5858 for (int i = code->length() - 1; i >= 1; i--) {
5859 BlockBegin* block = code->at(i);
5860
5861 if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) {
5862 optimizer.optimize_moves_at_block_end(block);
5863 }
5864 if (block->number_of_sux() == 2) {
5865 optimizer.optimize_moves_at_block_begin(block);
5866 }
5867 }
5868 }
5869
5870
5871 // clear all internal data structures
5872 void EdgeMoveOptimizer::init_instructions() {
5873 _edge_instructions.clear();
5874 _edge_instructions_idx.clear();
5875 }
5876
5877 // append a lir-instruction-list and the index of the current operation in to the list
5878 void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) {
5879 _edge_instructions.append(instructions);
5880 _edge_instructions_idx.append(instructions_idx);
5881 }
5882
5883 // return the current operation of the given edge (predecessor or successor)
5884 LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) {
5885 LIR_OpList* instructions = _edge_instructions.at(edge);
5886 int idx = _edge_instructions_idx.at(edge);
5887
5888 if (idx < instructions->length()) {
5889 return instructions->at(idx);
5890 } else {
5891 return NULL;
5892 }
5893 }
5894
5895 // removes the current operation of the given edge (predecessor or successor)
5896 void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) {
5897 LIR_OpList* instructions = _edge_instructions.at(edge);
5898 int idx = _edge_instructions_idx.at(edge);
5899 instructions->remove_at(idx);
5900
5901 if (decrement_index) {
5902 _edge_instructions_idx.at_put(edge, idx - 1);
5903 }
5904 }
5905
5906
5907 bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) {
5908 if (op1 == NULL || op2 == NULL) {
5909 // at least one block is already empty -> no optimization possible
5910 return true;
5911 }
5912
5913 if (op1->code() == lir_move && op2->code() == lir_move) {
5914 assert(op1->as_Op1() != NULL, "move must be LIR_Op1");
5915 assert(op2->as_Op1() != NULL, "move must be LIR_Op1");
5916 LIR_Op1* move1 = (LIR_Op1*)op1;
5917 LIR_Op1* move2 = (LIR_Op1*)op2;
5918 if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) {
5919 // these moves are exactly equal and can be optimized
5920 return false;
5921 }
5922
5923 } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) {
5924 assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1");
5925 assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1");
5926 LIR_Op1* fxch1 = (LIR_Op1*)op1;
5927 LIR_Op1* fxch2 = (LIR_Op1*)op2;
5928 if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) {
5929 // equal FPU stack operations can be optimized
5930 return false;
5931 }
5932
5933 } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) {
5934 // equal FPU stack operations can be optimized
5935 return false;
5936 }
5937
5938 // no optimization possible
5939 return true;
5940 }
5941
5942 void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) {
5943 TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id()));
5944
5945 if (block->is_predecessor(block)) {
5946 // currently we can't handle this correctly.
5947 return;
5948 }
5949
5950 init_instructions();
5951 int num_preds = block->number_of_preds();
5952 assert(num_preds > 1, "do not call otherwise");
5953 assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed");
5954
5955 // setup a list with the lir-instructions of all predecessors
5956 int i;
5957 for (i = 0; i < num_preds; i++) {
5958 BlockBegin* pred = block->pred_at(i);
5959 LIR_OpList* pred_instructions = pred->lir()->instructions_list();
5960
5961 if (pred->number_of_sux() != 1) {
5962 // this can happen with switch-statements where multiple edges are between
5963 // the same blocks.
5964 return;
5965 }
5966
5967 assert(pred->number_of_sux() == 1, "can handle only one successor");
5968 assert(pred->sux_at(0) == block, "invalid control flow");
5969 assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch");
5970 assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch");
5971 assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch");
5972
5973 if (pred_instructions->last()->info() != NULL) {
5974 // can not optimize instructions when debug info is needed
5975 return;
5976 }
5977
5978 // ignore the unconditional branch at the end of the block
5979 append_instructions(pred_instructions, pred_instructions->length() - 2);
5980 }
5981
5982
5983 // process lir-instructions while all predecessors end with the same instruction
5984 while (true) {
5985 LIR_Op* op = instruction_at(0);
5986 for (i = 1; i < num_preds; i++) {
5987 if (operations_different(op, instruction_at(i))) {
5988 // these instructions are different and cannot be optimized ->
5989 // no further optimization possible
5990 return;
5991 }
5992 }
5993
5994 TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print());
5995
5996 // insert the instruction at the beginning of the current block
5997 block->lir()->insert_before(1, op);
5998
5999 // delete the instruction at the end of all predecessors
6000 for (i = 0; i < num_preds; i++) {
6001 remove_cur_instruction(i, true);
6002 }
6003 }
6004 }
6005
6006
6007 void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) {
6008 TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id()));
6009
6010 init_instructions();
6011 int num_sux = block->number_of_sux();
6012
6013 LIR_OpList* cur_instructions = block->lir()->instructions_list();
6014
6015 assert(num_sux == 2, "method should not be called otherwise");
6016 assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch");
6017 assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch");
6018 assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch");
6019
6020 if (cur_instructions->last()->info() != NULL) {
6021 // can no optimize instructions when debug info is needed
6022 return;
6023 }
6024
6025 LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2);
6026 if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) {
6027 // not a valid case for optimization
6028 // currently, only blocks that end with two branches (conditional branch followed
6029 // by unconditional branch) are optimized
6030 return;
6031 }
6032
6033 // now it is guaranteed that the block ends with two branch instructions.
6034 // the instructions are inserted at the end of the block before these two branches
6035 int insert_idx = cur_instructions->length() - 2;
6036
6037 int i;
6038 #ifdef ASSERT
6039 for (i = insert_idx - 1; i >= 0; i--) {
6040 LIR_Op* op = cur_instructions->at(i);
6041 if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) {
6042 assert(false, "block with two successors can have only two branch instructions");
6043 }
6044 }
6045 #endif
6046
6047 // setup a list with the lir-instructions of all successors
6048 for (i = 0; i < num_sux; i++) {
6049 BlockBegin* sux = block->sux_at(i);
6050 LIR_OpList* sux_instructions = sux->lir()->instructions_list();
6051
6052 assert(sux_instructions->at(0)->code() == lir_label, "block must start with label");
6053
6054 if (sux->number_of_preds() != 1) {
6055 // this can happen with switch-statements where multiple edges are between
6056 // the same blocks.
6057 return;
6058 }
6059 assert(sux->pred_at(0) == block, "invalid control flow");
6060 assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed");
6061
6062 // ignore the label at the beginning of the block
6063 append_instructions(sux_instructions, 1);
6064 }
6065
6066 // process lir-instructions while all successors begin with the same instruction
6067 while (true) {
6068 LIR_Op* op = instruction_at(0);
6069 for (i = 1; i < num_sux; i++) {
6070 if (operations_different(op, instruction_at(i))) {
6071 // these instructions are different and cannot be optimized ->
6072 // no further optimization possible
6073 return;
6074 }
6075 }
6076
6077 TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print());
6078
6079 // insert instruction at end of current block
6080 block->lir()->insert_before(insert_idx, op);
6081 insert_idx++;
6082
6083 // delete the instructions at the beginning of all successors
6084 for (i = 0; i < num_sux; i++) {
6085 remove_cur_instruction(i, false);
6086 }
6087 }
6088 }
6089
6090
6091 // Implementation of ControlFlowOptimizer
6092
6093 ControlFlowOptimizer::ControlFlowOptimizer() :
6094 _original_preds(4)
6095 {
6096 }
6097
6098 void ControlFlowOptimizer::optimize(BlockList* code) {
6099 ControlFlowOptimizer optimizer = ControlFlowOptimizer();
6100
6101 // push the OSR entry block to the end so that we're not jumping over it.
6102 BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry();
6103 if (osr_entry) {
6104 int index = osr_entry->linear_scan_number();
6105 assert(code->at(index) == osr_entry, "wrong index");
6106 code->remove_at(index);
6107 code->append(osr_entry);
6108 }
6109
6110 optimizer.reorder_short_loops(code);
6111 optimizer.delete_empty_blocks(code);
6112 optimizer.delete_unnecessary_jumps(code);
6113 optimizer.delete_jumps_to_return(code);
6114 }
6115
6116 void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) {
6117 int i = header_idx + 1;
6118 int max_end = MIN2(header_idx + ShortLoopSize, code->length());
6119 while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) {
6120 i++;
6121 }
6122
6123 if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) {
6124 int end_idx = i - 1;
6125 BlockBegin* end_block = code->at(end_idx);
6126
6127 if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) {
6128 // short loop from header_idx to end_idx found -> reorder blocks such that
6129 // the header_block is the last block instead of the first block of the loop
6130 TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d",
6131 end_idx - header_idx + 1,
6132 header_block->block_id(), end_block->block_id()));
6133
6134 for (int j = header_idx; j < end_idx; j++) {
6135 code->at_put(j, code->at(j + 1));
6136 }
6137 code->at_put(end_idx, header_block);
6138
6139 // correct the flags so that any loop alignment occurs in the right place.
6140 assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target");
6141 code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag);
6142 code->at(header_idx)->set(BlockBegin::backward_branch_target_flag);
6143 }
6144 }
6145 }
6146
6147 void ControlFlowOptimizer::reorder_short_loops(BlockList* code) {
6148 for (int i = code->length() - 1; i >= 0; i--) {
6149 BlockBegin* block = code->at(i);
6150
6151 if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) {
6152 reorder_short_loop(code, block, i);
6153 }
6154 }
6155
6156 DEBUG_ONLY(verify(code));
6157 }
6158
6159 // only blocks with exactly one successor can be deleted. Such blocks
6160 // must always end with an unconditional branch to this successor
6161 bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) {
6162 if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) {
6163 return false;
6164 }
6165
6166 LIR_OpList* instructions = block->lir()->instructions_list();
6167
6168 assert(instructions->length() >= 2, "block must have label and branch");
6169 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label");
6170 assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch");
6171 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional");
6172 assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor");
6173
6174 // block must have exactly one successor
6175
6176 if (instructions->length() == 2 && instructions->last()->info() == NULL) {
6177 return true;
6178 }
6179 return false;
6180 }
6181
6182 // substitute branch targets in all branch-instructions of this blocks
6183 void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) {
6184 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()));
6185
6186 LIR_OpList* instructions = block->lir()->instructions_list();
6187
6188 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label");
6189 for (int i = instructions->length() - 1; i >= 1; i--) {
6190 LIR_Op* op = instructions->at(i);
6191
6192 if (op->code() == lir_branch || op->code() == lir_cond_float_branch) {
6193 assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch");
6194 LIR_OpBranch* branch = (LIR_OpBranch*)op;
6195
6196 if (branch->block() == target_from) {
6197 branch->change_block(target_to);
6198 }
6199 if (branch->ublock() == target_from) {
6200 branch->change_ublock(target_to);
6201 }
6202 }
6203 }
6204 }
6205
6206 void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) {
6207 int old_pos = 0;
6208 int new_pos = 0;
6209 int num_blocks = code->length();
6210
6211 while (old_pos < num_blocks) {
6212 BlockBegin* block = code->at(old_pos);
6213
6214 if (can_delete_block(block)) {
6215 BlockBegin* new_target = block->sux_at(0);
6216
6217 // propagate backward branch target flag for correct code alignment
6218 if (block->is_set(BlockBegin::backward_branch_target_flag)) {
6219 new_target->set(BlockBegin::backward_branch_target_flag);
6220 }
6221
6222 // collect a list with all predecessors that contains each predecessor only once
6223 // the predecessors of cur are changed during the substitution, so a copy of the
6224 // predecessor list is necessary
6225 int j;
6226 _original_preds.clear();
6227 for (j = block->number_of_preds() - 1; j >= 0; j--) {
6228 BlockBegin* pred = block->pred_at(j);
6229 if (_original_preds.find(pred) == -1) {
6230 _original_preds.append(pred);
6231 }
6232 }
6233
6234 for (j = _original_preds.length() - 1; j >= 0; j--) {
6235 BlockBegin* pred = _original_preds.at(j);
6236 substitute_branch_target(pred, block, new_target);
6237 pred->substitute_sux(block, new_target);
6238 }
6239 } else {
6240 // adjust position of this block in the block list if blocks before
6241 // have been deleted
6242 if (new_pos != old_pos) {
6243 code->at_put(new_pos, code->at(old_pos));
6244 }
6245 new_pos++;
6246 }
6247 old_pos++;
6248 }
6249 code->trunc_to(new_pos);
6250
6251 DEBUG_ONLY(verify(code));
6252 }
6253
6254 void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) {
6255 // skip the last block because there a branch is always necessary
6256 for (int i = code->length() - 2; i >= 0; i--) {
6257 BlockBegin* block = code->at(i);
6258 LIR_OpList* instructions = block->lir()->instructions_list();
6259
6260 LIR_Op* last_op = instructions->last();
6261 if (last_op->code() == lir_branch) {
6262 assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch");
6263 LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op;
6264
6265 assert(last_branch->block() != NULL, "last branch must always have a block as target");
6266 assert(last_branch->label() == last_branch->block()->label(), "must be equal");
6267
6268 if (last_branch->info() == NULL) {
6269 if (last_branch->block() == code->at(i + 1)) {
6270
6271 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id()));
6272
6273 // delete last branch instruction
6274 instructions->trunc_to(instructions->length() - 1);
6275
6276 } else {
6277 LIR_Op* prev_op = instructions->at(instructions->length() - 2);
6278 if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) {
6279 assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch");
6280 LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op;
6281
6282 if (prev_branch->stub() == NULL) {
6283
6284 LIR_Op2* prev_cmp = NULL;
6285 // There might be a cmove inserted for profiling which depends on the same
6286 // compare. If we change the condition of the respective compare, we have
6287 // to take care of this cmove as well.
6288 LIR_Op2* prev_cmove = NULL;
6289
6290 for(int j = instructions->length() - 3; j >= 0 && prev_cmp == NULL; j--) {
6291 prev_op = instructions->at(j);
6292 // check for the cmove
6293 if (prev_op->code() == lir_cmove) {
6294 assert(prev_op->as_Op2() != NULL, "cmove must be of type LIR_Op2");
6295 prev_cmove = (LIR_Op2*)prev_op;
6296 assert(prev_branch->cond() == prev_cmove->condition(), "should be the same");
6297 }
6298 if (prev_op->code() == lir_cmp) {
6299 assert(prev_op->as_Op2() != NULL, "branch must be of type LIR_Op2");
6300 prev_cmp = (LIR_Op2*)prev_op;
6301 assert(prev_branch->cond() == prev_cmp->condition(), "should be the same");
6302 }
6303 }
6304 assert(prev_cmp != NULL, "should have found comp instruction for branch");
6305 if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) {
6306
6307 TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d", block->block_id()));
6308
6309 // eliminate a conditional branch to the immediate successor
6310 prev_branch->change_block(last_branch->block());
6311 prev_branch->negate_cond();
6312 prev_cmp->set_condition(prev_branch->cond());
6313 instructions->trunc_to(instructions->length() - 1);
6314 // if we do change the condition, we have to change the cmove as well
6315 if (prev_cmove != NULL) {
6316 prev_cmove->set_condition(prev_branch->cond());
6317 LIR_Opr t = prev_cmove->in_opr1();
6318 prev_cmove->set_in_opr1(prev_cmove->in_opr2());
6319 prev_cmove->set_in_opr2(t);
6320 }
6321 }
6322 }
6323 }
6324 }
6325 }
6326 }
6327 }
6328
6329 DEBUG_ONLY(verify(code));
6330 }
6331
6332 void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) {
6333 #ifdef ASSERT
6334 ResourceBitMap return_converted(BlockBegin::number_of_blocks());
6335 #endif
6336
6337 for (int i = code->length() - 1; i >= 0; i--) {
6338 BlockBegin* block = code->at(i);
6339 LIR_OpList* cur_instructions = block->lir()->instructions_list();
6340 LIR_Op* cur_last_op = cur_instructions->last();
6341
6342 assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label");
6343 if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) {
6344 // the block contains only a label and a return
6345 // if a predecessor ends with an unconditional jump to this block, then the jump
6346 // can be replaced with a return instruction
6347 //
6348 // Note: the original block with only a return statement cannot be deleted completely
6349 // because the predecessors might have other (conditional) jumps to this block
6350 // -> this may lead to unnecesary return instructions in the final code
6351
6352 assert(cur_last_op->info() == NULL, "return instructions do not have debug information");
6353 assert(block->number_of_sux() == 0 ||
6354 (return_converted.at(block->block_id()) && block->number_of_sux() == 1),
6355 "blocks that end with return must not have successors");
6356
6357 assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1");
6358 LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr();
6359
6360 for (int j = block->number_of_preds() - 1; j >= 0; j--) {
6361 BlockBegin* pred = block->pred_at(j);
6362 LIR_OpList* pred_instructions = pred->lir()->instructions_list();
6363 LIR_Op* pred_last_op = pred_instructions->last();
6364
6365 if (pred_last_op->code() == lir_branch) {
6366 assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch");
6367 LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op;
6368
6369 if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) {
6370 // replace the jump to a return with a direct return
6371 // Note: currently the edge between the blocks is not deleted
6372 pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr));
6373 #ifdef ASSERT
6374 return_converted.set_bit(pred->block_id());
6375 #endif
6376 }
6377 }
6378 }
6379 }
6380 }
6381 }
6382
6383
6384 #ifdef ASSERT
6385 void ControlFlowOptimizer::verify(BlockList* code) {
6386 for (int i = 0; i < code->length(); i++) {
6387 BlockBegin* block = code->at(i);
6388 LIR_OpList* instructions = block->lir()->instructions_list();
6389
6390 int j;
6391 for (j = 0; j < instructions->length(); j++) {
6392 LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch();
6393
6394 if (op_branch != NULL) {
6395 assert(op_branch->block() == NULL || code->find(op_branch->block()) != -1, "branch target not valid");
6396 assert(op_branch->ublock() == NULL || code->find(op_branch->ublock()) != -1, "branch target not valid");
6397 }
6398 }
6399
6400 for (j = 0; j < block->number_of_sux() - 1; j++) {
6401 BlockBegin* sux = block->sux_at(j);
6402 assert(code->find(sux) != -1, "successor not valid");
6403 }
6404
6405 for (j = 0; j < block->number_of_preds() - 1; j++) {
6406 BlockBegin* pred = block->pred_at(j);
6407 assert(code->find(pred) != -1, "successor not valid");
6408 }
6409 }
6410 }
6411 #endif
6412
6413
6414 #ifndef PRODUCT
6415
6416 // Implementation of LinearStatistic
6417
6418 const char* LinearScanStatistic::counter_name(int counter_idx) {
6419 switch (counter_idx) {
6420 case counter_method: return "compiled methods";
6421 case counter_fpu_method: return "methods using fpu";
6422 case counter_loop_method: return "methods with loops";
6423 case counter_exception_method:return "methods with xhandler";
6424
6425 case counter_loop: return "loops";
6426 case counter_block: return "blocks";
6427 case counter_loop_block: return "blocks inside loop";
6428 case counter_exception_block: return "exception handler entries";
6429 case counter_interval: return "intervals";
6430 case counter_fixed_interval: return "fixed intervals";
6431 case counter_range: return "ranges";
6432 case counter_fixed_range: return "fixed ranges";
6433 case counter_use_pos: return "use positions";
6434 case counter_fixed_use_pos: return "fixed use positions";
6435 case counter_spill_slots: return "spill slots";
6436
6437 // counter for classes of lir instructions
6438 case counter_instruction: return "total instructions";
6439 case counter_label: return "labels";
6440 case counter_entry: return "method entries";
6441 case counter_return: return "method returns";
6442 case counter_call: return "method calls";
6443 case counter_move: return "moves";
6444 case counter_cmp: return "compare";
6445 case counter_cond_branch: return "conditional branches";
6446 case counter_uncond_branch: return "unconditional branches";
6447 case counter_stub_branch: return "branches to stub";
6448 case counter_alu: return "artithmetic + logic";
6449 case counter_alloc: return "allocations";
6450 case counter_sync: return "synchronisation";
6451 case counter_throw: return "throw";
6452 case counter_unwind: return "unwind";
6453 case counter_typecheck: return "type+null-checks";
6454 case counter_fpu_stack: return "fpu-stack";
6455 case counter_misc_inst: return "other instructions";
6456 case counter_other_inst: return "misc. instructions";
6457
6458 // counter for different types of moves
6459 case counter_move_total: return "total moves";
6460 case counter_move_reg_reg: return "register->register";
6461 case counter_move_reg_stack: return "register->stack";
6462 case counter_move_stack_reg: return "stack->register";
6463 case counter_move_stack_stack:return "stack->stack";
6464 case counter_move_reg_mem: return "register->memory";
6465 case counter_move_mem_reg: return "memory->register";
6466 case counter_move_const_any: return "constant->any";
6467
6468 case blank_line_1: return "";
6469 case blank_line_2: return "";
6470
6471 default: ShouldNotReachHere(); return "";
6472 }
6473 }
6474
6475 LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) {
6476 if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) {
6477 return counter_method;
6478 } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) {
6479 return counter_block;
6480 } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) {
6481 return counter_instruction;
6482 } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) {
6483 return counter_move_total;
6484 }
6485 return invalid_counter;
6486 }
6487
6488 LinearScanStatistic::LinearScanStatistic() {
6489 for (int i = 0; i < number_of_counters; i++) {
6490 _counters_sum[i] = 0;
6491 _counters_max[i] = -1;
6492 }
6493
6494 }
6495
6496 // add the method-local numbers to the total sum
6497 void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) {
6498 for (int i = 0; i < number_of_counters; i++) {
6499 _counters_sum[i] += method_statistic._counters_sum[i];
6500 _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]);
6501 }
6502 }
6503
6504 void LinearScanStatistic::print(const char* title) {
6505 if (CountLinearScan || TraceLinearScanLevel > 0) {
6506 tty->cr();
6507 tty->print_cr("***** LinearScan statistic - %s *****", title);
6508
6509 for (int i = 0; i < number_of_counters; i++) {
6510 if (_counters_sum[i] > 0 || _counters_max[i] >= 0) {
6511 tty->print("%25s: %8d", counter_name(i), _counters_sum[i]);
6512
6513 LinearScanStatistic::Counter cntr = base_counter(i);
6514 if (cntr != invalid_counter) {
6515 tty->print(" (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[cntr]);
6516 } else {
6517 tty->print(" ");
6518 }
6519
6520 if (_counters_max[i] >= 0) {
6521 tty->print("%8d", _counters_max[i]);
6522 }
6523 }
6524 tty->cr();
6525 }
6526 }
6527 }
6528
6529 void LinearScanStatistic::collect(LinearScan* allocator) {
6530 inc_counter(counter_method);
6531 if (allocator->has_fpu_registers()) {
6532 inc_counter(counter_fpu_method);
6533 }
6534 if (allocator->num_loops() > 0) {
6535 inc_counter(counter_loop_method);
6536 }
6537 inc_counter(counter_loop, allocator->num_loops());
6538 inc_counter(counter_spill_slots, allocator->max_spills());
6539
6540 int i;
6541 for (i = 0; i < allocator->interval_count(); i++) {
6542 Interval* cur = allocator->interval_at(i);
6543
6544 if (cur != NULL) {
6545 inc_counter(counter_interval);
6546 inc_counter(counter_use_pos, cur->num_use_positions());
6547 if (LinearScan::is_precolored_interval(cur)) {
6548 inc_counter(counter_fixed_interval);
6549 inc_counter(counter_fixed_use_pos, cur->num_use_positions());
6550 }
6551
6552 Range* range = cur->first();
6553 while (range != Range::end()) {
6554 inc_counter(counter_range);
6555 if (LinearScan::is_precolored_interval(cur)) {
6556 inc_counter(counter_fixed_range);
6557 }
6558 range = range->next();
6559 }
6560 }
6561 }
6562
6563 bool has_xhandlers = false;
6564 // Note: only count blocks that are in code-emit order
6565 for (i = 0; i < allocator->ir()->code()->length(); i++) {
6566 BlockBegin* cur = allocator->ir()->code()->at(i);
6567
6568 inc_counter(counter_block);
6569 if (cur->loop_depth() > 0) {
6570 inc_counter(counter_loop_block);
6571 }
6572 if (cur->is_set(BlockBegin::exception_entry_flag)) {
6573 inc_counter(counter_exception_block);
6574 has_xhandlers = true;
6575 }
6576
6577 LIR_OpList* instructions = cur->lir()->instructions_list();
6578 for (int j = 0; j < instructions->length(); j++) {
6579 LIR_Op* op = instructions->at(j);
6580
6581 inc_counter(counter_instruction);
6582
6583 switch (op->code()) {
6584 case lir_label: inc_counter(counter_label); break;
6585 case lir_std_entry:
6586 case lir_osr_entry: inc_counter(counter_entry); break;
6587 case lir_return: inc_counter(counter_return); break;
6588
6589 case lir_rtcall:
6590 case lir_static_call:
6591 case lir_optvirtual_call:
6592 case lir_virtual_call: inc_counter(counter_call); break;
6593
6594 case lir_move: {
6595 inc_counter(counter_move);
6596 inc_counter(counter_move_total);
6597
6598 LIR_Opr in = op->as_Op1()->in_opr();
6599 LIR_Opr res = op->as_Op1()->result_opr();
6600 if (in->is_register()) {
6601 if (res->is_register()) {
6602 inc_counter(counter_move_reg_reg);
6603 } else if (res->is_stack()) {
6604 inc_counter(counter_move_reg_stack);
6605 } else if (res->is_address()) {
6606 inc_counter(counter_move_reg_mem);
6607 } else {
6608 ShouldNotReachHere();
6609 }
6610 } else if (in->is_stack()) {
6611 if (res->is_register()) {
6612 inc_counter(counter_move_stack_reg);
6613 } else {
6614 inc_counter(counter_move_stack_stack);
6615 }
6616 } else if (in->is_address()) {
6617 assert(res->is_register(), "must be");
6618 inc_counter(counter_move_mem_reg);
6619 } else if (in->is_constant()) {
6620 inc_counter(counter_move_const_any);
6621 } else {
6622 ShouldNotReachHere();
6623 }
6624 break;
6625 }
6626
6627 case lir_cmp: inc_counter(counter_cmp); break;
6628
6629 case lir_branch:
6630 case lir_cond_float_branch: {
6631 LIR_OpBranch* branch = op->as_OpBranch();
6632 if (branch->block() == NULL) {
6633 inc_counter(counter_stub_branch);
6634 } else if (branch->cond() == lir_cond_always) {
6635 inc_counter(counter_uncond_branch);
6636 } else {
6637 inc_counter(counter_cond_branch);
6638 }
6639 break;
6640 }
6641
6642 case lir_neg:
6643 case lir_add:
6644 case lir_sub:
6645 case lir_mul:
6646 case lir_mul_strictfp:
6647 case lir_div:
6648 case lir_div_strictfp:
6649 case lir_rem:
6650 case lir_sqrt:
6651 case lir_abs:
6652 case lir_log10:
6653 case lir_logic_and:
6654 case lir_logic_or:
6655 case lir_logic_xor:
6656 case lir_shl:
6657 case lir_shr:
6658 case lir_ushr: inc_counter(counter_alu); break;
6659
6660 case lir_alloc_object:
6661 case lir_alloc_array: inc_counter(counter_alloc); break;
6662
6663 case lir_monaddr:
6664 case lir_lock:
6665 case lir_unlock: inc_counter(counter_sync); break;
6666
6667 case lir_throw: inc_counter(counter_throw); break;
6668
6669 case lir_unwind: inc_counter(counter_unwind); break;
6670
6671 case lir_null_check:
6672 case lir_leal:
6673 case lir_instanceof:
6674 case lir_checkcast:
6675 case lir_store_check: inc_counter(counter_typecheck); break;
6676
6677 case lir_fpop_raw:
6678 case lir_fxch:
6679 case lir_fld: inc_counter(counter_fpu_stack); break;
6680
6681 case lir_nop:
6682 case lir_push:
6683 case lir_pop:
6684 case lir_convert:
6685 case lir_roundfp:
6686 case lir_cmove: inc_counter(counter_misc_inst); break;
6687
6688 default: inc_counter(counter_other_inst); break;
6689 }
6690 }
6691 }
6692
6693 if (has_xhandlers) {
6694 inc_counter(counter_exception_method);
6695 }
6696 }
6697
6698 void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) {
6699 if (CountLinearScan || TraceLinearScanLevel > 0) {
6700
6701 LinearScanStatistic local_statistic = LinearScanStatistic();
6702
6703 local_statistic.collect(allocator);
6704 global_statistic.sum_up(local_statistic);
6705
6706 if (TraceLinearScanLevel > 2) {
6707 local_statistic.print("current local statistic");
6708 }
6709 }
6710 }
6711
6712
6713 // Implementation of LinearTimers
6714
6715 LinearScanTimers::LinearScanTimers() {
6716 for (int i = 0; i < number_of_timers; i++) {
6717 timer(i)->reset();
6718 }
6719 }
6720
6721 const char* LinearScanTimers::timer_name(int idx) {
6722 switch (idx) {
6723 case timer_do_nothing: return "Nothing (Time Check)";
6724 case timer_number_instructions: return "Number Instructions";
6725 case timer_compute_local_live_sets: return "Local Live Sets";
6726 case timer_compute_global_live_sets: return "Global Live Sets";
6727 case timer_build_intervals: return "Build Intervals";
6728 case timer_sort_intervals_before: return "Sort Intervals Before";
6729 case timer_allocate_registers: return "Allocate Registers";
6730 case timer_resolve_data_flow: return "Resolve Data Flow";
6731 case timer_sort_intervals_after: return "Sort Intervals After";
6732 case timer_eliminate_spill_moves: return "Spill optimization";
6733 case timer_assign_reg_num: return "Assign Reg Num";
6734 case timer_allocate_fpu_stack: return "Allocate FPU Stack";
6735 case timer_optimize_lir: return "Optimize LIR";
6736 default: ShouldNotReachHere(); return "";
6737 }
6738 }
6739
6740 void LinearScanTimers::begin_method() {
6741 if (TimeEachLinearScan) {
6742 // reset all timers to measure only current method
6743 for (int i = 0; i < number_of_timers; i++) {
6744 timer(i)->reset();
6745 }
6746 }
6747 }
6748
6749 void LinearScanTimers::end_method(LinearScan* allocator) {
6750 if (TimeEachLinearScan) {
6751
6752 double c = timer(timer_do_nothing)->seconds();
6753 double total = 0;
6754 for (int i = 1; i < number_of_timers; i++) {
6755 total += timer(i)->seconds() - c;
6756 }
6757
6758 if (total >= 0.0005) {
6759 // print all information in one line for automatic processing
6760 tty->print("@"); allocator->compilation()->method()->print_name();
6761
6762 tty->print("@ %d ", allocator->compilation()->method()->code_size());
6763 tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2);
6764 tty->print("@ %d ", allocator->block_count());
6765 tty->print("@ %d ", allocator->num_virtual_regs());
6766 tty->print("@ %d ", allocator->interval_count());
6767 tty->print("@ %d ", allocator->_num_calls);
6768 tty->print("@ %d ", allocator->num_loops());
6769
6770 tty->print("@ %6.6f ", total);
6771 for (int i = 1; i < number_of_timers; i++) {
6772 tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100);
6773 }
6774 tty->cr();
6775 }
6776 }
6777 }
6778
6779 void LinearScanTimers::print(double total_time) {
6780 if (TimeLinearScan) {
6781 // correction value: sum of dummy-timer that only measures the time that
6782 // is necesary to start and stop itself
6783 double c = timer(timer_do_nothing)->seconds();
6784
6785 for (int i = 0; i < number_of_timers; i++) {
6786 double t = timer(i)->seconds();
6787 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);
6788 }
6789 }
6790 }
6791
6792 #endif // #ifndef PRODUCT