1 /*
2 * Copyright (c) 2005, 2024, 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 #else
50 #define TIME_LINEAR_SCAN(timer_name)
51 #endif
52
53 #ifdef ASSERT
54
55 // helper macro for short definition of trace-output inside code
56 #define TRACE_LINEAR_SCAN(level, code) \
57 if (TraceLinearScanLevel >= level) { \
58 code; \
59 }
60 #else
61 #define TRACE_LINEAR_SCAN(level, code)
62 #endif
63
64 // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words
65 #ifdef _LP64
66 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};
67 #else
68 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};
69 #endif
70
71
72 // Implementation of LinearScan
73
74 LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map)
75 : _compilation(ir->compilation())
76 , _ir(ir)
77 , _gen(gen)
78 , _frame_map(frame_map)
79 , _cached_blocks(*ir->linear_scan_order())
80 , _num_virtual_regs(gen->max_virtual_register_number())
81 , _has_fpu_registers(false)
82 , _num_calls(-1)
83 , _max_spills(0)
84 , _unused_spill_slot(-1)
85 , _intervals(0) // initialized later with correct length
86 , _new_intervals_from_allocation(nullptr)
87 , _sorted_intervals(nullptr)
88 , _needs_full_resort(false)
89 , _lir_ops(0) // initialized later with correct length
90 , _block_of_op(0) // initialized later with correct length
91 , _has_info(0)
92 , _has_call(0)
93 , _interval_in_loop(0) // initialized later with correct length
94 , _scope_value_cache(0) // initialized later with correct length
95 #ifdef IA32
96 , _fpu_stack_allocator(nullptr)
97 #endif
98 {
99 assert(this->ir() != nullptr, "check if valid");
100 assert(this->compilation() != nullptr, "check if valid");
101 assert(this->gen() != nullptr, "check if valid");
102 assert(this->frame_map() != nullptr, "check if valid");
103 }
104
105
106 // ********** functions for converting LIR-Operands to register numbers
107 //
108 // Emulate a flat register file comprising physical integer registers,
109 // physical floating-point registers and virtual registers, in that order.
110 // Virtual registers already have appropriate numbers, since V0 is
111 // the number of physical registers.
112 // Returns -1 for hi word if opr is a single word operand.
113 //
114 // Note: the inverse operation (calculating an operand for register numbers)
115 // is done in calc_operand_for_interval()
116
117 int LinearScan::reg_num(LIR_Opr opr) {
118 assert(opr->is_register(), "should not call this otherwise");
119
120 if (opr->is_virtual_register()) {
121 assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number");
122 return opr->vreg_number();
123 } else if (opr->is_single_cpu()) {
124 return opr->cpu_regnr();
125 } else if (opr->is_double_cpu()) {
126 return opr->cpu_regnrLo();
127 #ifdef X86
128 } else if (opr->is_single_xmm()) {
129 return opr->fpu_regnr() + pd_first_xmm_reg;
130 } else if (opr->is_double_xmm()) {
131 return opr->fpu_regnrLo() + pd_first_xmm_reg;
132 #endif
133 } else if (opr->is_single_fpu()) {
134 return opr->fpu_regnr() + pd_first_fpu_reg;
135 } else if (opr->is_double_fpu()) {
136 return opr->fpu_regnrLo() + pd_first_fpu_reg;
137 } else {
138 ShouldNotReachHere();
139 return -1;
140 }
141 }
142
143 int LinearScan::reg_numHi(LIR_Opr opr) {
144 assert(opr->is_register(), "should not call this otherwise");
145
146 if (opr->is_virtual_register()) {
147 return -1;
148 } else if (opr->is_single_cpu()) {
149 return -1;
150 } else if (opr->is_double_cpu()) {
151 return opr->cpu_regnrHi();
152 #ifdef X86
153 } else if (opr->is_single_xmm()) {
154 return -1;
155 } else if (opr->is_double_xmm()) {
156 return -1;
157 #endif
158 } else if (opr->is_single_fpu()) {
159 return -1;
160 } else if (opr->is_double_fpu()) {
161 return opr->fpu_regnrHi() + pd_first_fpu_reg;
162 } else {
163 ShouldNotReachHere();
164 return -1;
165 }
166 }
167
168
169 // ********** functions for classification of intervals
170
171 bool LinearScan::is_precolored_interval(const Interval* i) {
172 return i->reg_num() < LinearScan::nof_regs;
173 }
174
175 bool LinearScan::is_virtual_interval(const Interval* i) {
176 return i->reg_num() >= LIR_Opr::vreg_base;
177 }
178
179 bool LinearScan::is_precolored_cpu_interval(const Interval* i) {
180 return i->reg_num() < LinearScan::nof_cpu_regs;
181 }
182
183 bool LinearScan::is_virtual_cpu_interval(const Interval* i) {
184 #if defined(__SOFTFP__) || defined(E500V2)
185 return i->reg_num() >= LIR_Opr::vreg_base;
186 #else
187 return i->reg_num() >= LIR_Opr::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE);
188 #endif // __SOFTFP__ or E500V2
189 }
190
191 bool LinearScan::is_precolored_fpu_interval(const Interval* i) {
192 return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs;
193 }
194
195 bool LinearScan::is_virtual_fpu_interval(const Interval* i) {
196 #if defined(__SOFTFP__) || defined(E500V2)
197 return false;
198 #else
199 return i->reg_num() >= LIR_Opr::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE);
200 #endif // __SOFTFP__ or E500V2
201 }
202
203 bool LinearScan::is_in_fpu_register(const Interval* i) {
204 // fixed intervals not needed for FPU stack allocation
205 return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg;
206 }
207
208 bool LinearScan::is_oop_interval(const Interval* i) {
209 // fixed intervals never contain oops
210 return i->reg_num() >= nof_regs && i->type() == T_OBJECT;
211 }
212
213
214 // ********** General helper functions
215
216 // compute next unused stack index that can be used for spilling
217 int LinearScan::allocate_spill_slot(bool double_word) {
218 int spill_slot;
219 if (double_word) {
220 if ((_max_spills & 1) == 1) {
221 // alignment of double-word values
222 // the hole because of the alignment is filled with the next single-word value
223 assert(_unused_spill_slot == -1, "wasting a spill slot");
224 _unused_spill_slot = _max_spills;
225 _max_spills++;
226 }
227 spill_slot = _max_spills;
228 _max_spills += 2;
229
230 } else if (_unused_spill_slot != -1) {
231 // re-use hole that was the result of a previous double-word alignment
232 spill_slot = _unused_spill_slot;
233 _unused_spill_slot = -1;
234
235 } else {
236 spill_slot = _max_spills;
237 _max_spills++;
238 }
239
240 int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount();
241
242 // if too many slots used, bailout compilation.
243 if (result > 2000) {
244 bailout("too many stack slots used");
245 }
246
247 return result;
248 }
249
250 void LinearScan::assign_spill_slot(Interval* it) {
251 // assign the canonical spill slot of the parent (if a part of the interval
252 // is already spilled) or allocate a new spill slot
253 if (it->canonical_spill_slot() >= 0) {
254 it->assign_reg(it->canonical_spill_slot());
255 } else {
256 int spill = allocate_spill_slot(type2spill_size[it->type()] == 2);
257 it->set_canonical_spill_slot(spill);
258 it->assign_reg(spill);
259 }
260 }
261
262 void LinearScan::propagate_spill_slots() {
263 if (!frame_map()->finalize_frame(max_spills())) {
264 bailout("frame too large");
265 }
266 }
267
268 // create a new interval with a predefined reg_num
269 // (only used for parent intervals that are created during the building phase)
270 Interval* LinearScan::create_interval(int reg_num) {
271 assert(_intervals.at(reg_num) == nullptr, "overwriting existing interval");
272
273 Interval* interval = new Interval(reg_num);
274 _intervals.at_put(reg_num, interval);
275
276 // assign register number for precolored intervals
277 if (reg_num < LIR_Opr::vreg_base) {
278 interval->assign_reg(reg_num);
279 }
280 return interval;
281 }
282
283 // assign a new reg_num to the interval and append it to the list of intervals
284 // (only used for child intervals that are created during register allocation)
285 void LinearScan::append_interval(Interval* it) {
286 it->set_reg_num(_intervals.length());
287 _intervals.append(it);
288 IntervalList* new_intervals = _new_intervals_from_allocation;
289 if (new_intervals == nullptr) {
290 new_intervals = _new_intervals_from_allocation = new IntervalList();
291 }
292 new_intervals->append(it);
293 }
294
295 // copy the vreg-flags if an interval is split
296 void LinearScan::copy_register_flags(Interval* from, Interval* to) {
297 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) {
298 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg);
299 }
300 if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) {
301 gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved);
302 }
303
304 // Note: do not copy the must_start_in_memory flag because it is not necessary for child
305 // intervals (only the very beginning of the interval must be in memory)
306 }
307
308
309 // ********** spill move optimization
310 // eliminate moves from register to stack if stack slot is known to be correct
311
312 // called during building of intervals
313 void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) {
314 assert(interval->is_split_parent(), "can only be called for split parents");
315
316 switch (interval->spill_state()) {
317 case noDefinitionFound:
318 assert(interval->spill_definition_pos() == -1, "must no be set before");
319 interval->set_spill_definition_pos(def_pos);
320 interval->set_spill_state(oneDefinitionFound);
321 break;
322
323 case oneDefinitionFound:
324 assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created");
325 if (def_pos < interval->spill_definition_pos() - 2) {
326 // second definition found, so no spill optimization possible for this interval
327 interval->set_spill_state(noOptimization);
328 } else {
329 // two consecutive definitions (because of two-operand LIR form)
330 assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal");
331 }
332 break;
333
334 case noOptimization:
335 // nothing to do
336 break;
337
338 default:
339 assert(false, "other states not allowed at this time");
340 }
341 }
342
343 // called during register allocation
344 void LinearScan::change_spill_state(Interval* interval, int spill_pos) {
345 switch (interval->spill_state()) {
346 case oneDefinitionFound: {
347 int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth();
348 int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth();
349
350 if (def_loop_depth < spill_loop_depth) {
351 // the loop depth of the spilling position is higher then the loop depth
352 // at the definition of the interval -> move write to memory out of loop
353 // by storing at definitin of the interval
354 interval->set_spill_state(storeAtDefinition);
355 } else {
356 // the interval is currently spilled only once, so for now there is no
357 // reason to store the interval at the definition
358 interval->set_spill_state(oneMoveInserted);
359 }
360 break;
361 }
362
363 case oneMoveInserted: {
364 // the interval is spilled more then once, so it is better to store it to
365 // memory at the definition
366 interval->set_spill_state(storeAtDefinition);
367 break;
368 }
369
370 case storeAtDefinition:
371 case startInMemory:
372 case noOptimization:
373 case noDefinitionFound:
374 // nothing to do
375 break;
376
377 default:
378 assert(false, "other states not allowed at this time");
379 }
380 }
381
382
383 bool LinearScan::must_store_at_definition(const Interval* i) {
384 return i->is_split_parent() && i->spill_state() == storeAtDefinition;
385 }
386
387 // called once before assignment of register numbers
388 void LinearScan::eliminate_spill_moves() {
389 TIME_LINEAR_SCAN(timer_eliminate_spill_moves);
390 TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves"));
391
392 // collect all intervals that must be stored after their definion.
393 // the list is sorted by Interval::spill_definition_pos
394 Interval* interval;
395 Interval* temp_list;
396 create_unhandled_lists(&interval, &temp_list, must_store_at_definition, nullptr);
397
398 #ifdef ASSERT
399 Interval* prev = nullptr;
400 Interval* temp = interval;
401 while (temp != Interval::end()) {
402 assert(temp->spill_definition_pos() > 0, "invalid spill definition pos");
403 if (prev != nullptr) {
404 assert(temp->from() >= prev->from(), "intervals not sorted");
405 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");
406 }
407
408 assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned");
409 assert(temp->spill_definition_pos() >= temp->from(), "invalid order");
410 assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized");
411
412 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()));
413
414 temp = temp->next();
415 }
416 #endif
417
418 LIR_InsertionBuffer insertion_buffer;
419 int num_blocks = block_count();
420 for (int i = 0; i < num_blocks; i++) {
421 BlockBegin* block = block_at(i);
422 LIR_OpList* instructions = block->lir()->instructions_list();
423 int num_inst = instructions->length();
424 bool has_new = false;
425
426 // iterate all instructions of the block. skip the first because it is always a label
427 for (int j = 1; j < num_inst; j++) {
428 LIR_Op* op = instructions->at(j);
429 int op_id = op->id();
430
431 if (op_id == -1) {
432 // remove move from register to stack if the stack slot is guaranteed to be correct.
433 // only moves that have been inserted by LinearScan can be removed.
434 assert(op->code() == lir_move, "only moves can have a op_id of -1");
435 assert(op->as_Op1() != nullptr, "move must be LIR_Op1");
436 assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers");
437
438 LIR_Op1* op1 = (LIR_Op1*)op;
439 Interval* interval = interval_at(op1->result_opr()->vreg_number());
440
441 if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) {
442 // move target is a stack slot that is always correct, so eliminate instruction
443 TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d", op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number()));
444 instructions->at_put(j, nullptr); // null-instructions are deleted by assign_reg_num
445 }
446
447 } else {
448 // insert move from register to stack just after the beginning of the interval
449 assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order");
450 assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval");
451
452 while (interval != Interval::end() && interval->spill_definition_pos() == op_id) {
453 if (!has_new) {
454 // prepare insertion buffer (appended when all instructions of the block are processed)
455 insertion_buffer.init(block->lir());
456 has_new = true;
457 }
458
459 LIR_Opr from_opr = operand_for_interval(interval);
460 LIR_Opr to_opr = canonical_spill_opr(interval);
461 assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register");
462 assert(to_opr->is_stack(), "to operand must be a stack slot");
463
464 insertion_buffer.move(j, from_opr, to_opr);
465 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));
466
467 interval = interval->next();
468 }
469 }
470 } // end of instruction iteration
471
472 if (has_new) {
473 block->lir()->append(&insertion_buffer);
474 }
475 } // end of block iteration
476
477 assert(interval == Interval::end(), "missed an interval");
478 }
479
480
481 // ********** Phase 1: number all instructions in all blocks
482 // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan.
483
484 void LinearScan::number_instructions() {
485 {
486 // dummy-timer to measure the cost of the timer itself
487 // (this time is then subtracted from all other timers to get the real value)
488 TIME_LINEAR_SCAN(timer_do_nothing);
489 }
490 TIME_LINEAR_SCAN(timer_number_instructions);
491
492 // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node.
493 int num_blocks = block_count();
494 int num_instructions = 0;
495 int i;
496 for (i = 0; i < num_blocks; i++) {
497 num_instructions += block_at(i)->lir()->instructions_list()->length();
498 }
499
500 // initialize with correct length
501 _lir_ops = LIR_OpArray(num_instructions, num_instructions, nullptr);
502 _block_of_op = BlockBeginArray(num_instructions, num_instructions, nullptr);
503
504 int op_id = 0;
505 int idx = 0;
506
507 for (i = 0; i < num_blocks; i++) {
508 BlockBegin* block = block_at(i);
509 block->set_first_lir_instruction_id(op_id);
510 LIR_OpList* instructions = block->lir()->instructions_list();
511
512 int num_inst = instructions->length();
513 for (int j = 0; j < num_inst; j++) {
514 LIR_Op* op = instructions->at(j);
515 op->set_id(op_id);
516
517 _lir_ops.at_put(idx, op);
518 _block_of_op.at_put(idx, block);
519 assert(lir_op_with_id(op_id) == op, "must match");
520
521 idx++;
522 op_id += 2; // numbering of lir_ops by two
523 }
524 block->set_last_lir_instruction_id(op_id - 2);
525 }
526 assert(idx == num_instructions, "must match");
527 assert(idx * 2 == op_id, "must match");
528
529 _has_call.initialize(num_instructions);
530 _has_info.initialize(num_instructions);
531 }
532
533
534 // ********** Phase 2: compute local live sets separately for each block
535 // (sets live_gen and live_kill for each block)
536
537 void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) {
538 LIR_Opr opr = value->operand();
539 Constant* con = value->as_Constant();
540
541 // check some asumptions about debug information
542 assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type");
543 assert(con == nullptr || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "assumption: Constant instructions have only constant operands");
544 assert(con != nullptr || opr->is_virtual(), "assumption: non-Constant instructions have only virtual operands");
545
546 if ((con == nullptr || con->is_pinned()) && opr->is_register()) {
547 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
548 int reg = opr->vreg_number();
549 if (!live_kill.at(reg)) {
550 live_gen.set_bit(reg);
551 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));
552 }
553 }
554 }
555
556
557 void LinearScan::compute_local_live_sets() {
558 TIME_LINEAR_SCAN(timer_compute_local_live_sets);
559
560 int num_blocks = block_count();
561 int live_size = live_set_size();
562 bool local_has_fpu_registers = false;
563 int local_num_calls = 0;
564 LIR_OpVisitState visitor;
565
566 BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops());
567
568 // iterate all blocks
569 for (int i = 0; i < num_blocks; i++) {
570 BlockBegin* block = block_at(i);
571
572 ResourceBitMap live_gen(live_size);
573 ResourceBitMap live_kill(live_size);
574
575 if (block->is_set(BlockBegin::exception_entry_flag)) {
576 // Phi functions at the begin of an exception handler are
577 // implicitly defined (= killed) at the beginning of the block.
578 for_each_phi_fun(block, phi,
579 if (!phi->is_illegal()) { live_kill.set_bit(phi->operand()->vreg_number()); }
580 );
581 }
582
583 LIR_OpList* instructions = block->lir()->instructions_list();
584 int num_inst = instructions->length();
585
586 // iterate all instructions of the block. skip the first because it is always a label
587 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label");
588 for (int j = 1; j < num_inst; j++) {
589 LIR_Op* op = instructions->at(j);
590
591 // visit operation to collect all operands
592 visitor.visit(op);
593
594 if (visitor.has_call()) {
595 _has_call.set_bit(op->id() >> 1);
596 local_num_calls++;
597 }
598 if (visitor.info_count() > 0) {
599 _has_info.set_bit(op->id() >> 1);
600 }
601
602 // iterate input operands of instruction
603 int k, n, reg;
604 n = visitor.opr_count(LIR_OpVisitState::inputMode);
605 for (k = 0; k < n; k++) {
606 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k);
607 assert(opr->is_register(), "visitor should only return register operands");
608
609 if (opr->is_virtual_register()) {
610 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
611 reg = opr->vreg_number();
612 if (!live_kill.at(reg)) {
613 live_gen.set_bit(reg);
614 TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d", reg, op->id()));
615 }
616 if (block->loop_index() >= 0) {
617 local_interval_in_loop.set_bit(reg, block->loop_index());
618 }
619 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
620 }
621
622 #ifdef ASSERT
623 // fixed intervals are never live at block boundaries, so
624 // they need not be processed in live sets.
625 // this is checked by these assertions to be sure about it.
626 // the entry block may have incoming values in registers, which is ok.
627 if (!opr->is_virtual_register() && block != ir()->start()) {
628 reg = reg_num(opr);
629 if (is_processed_reg_num(reg)) {
630 assert(live_kill.at(reg), "using fixed register that is not defined in this block");
631 }
632 reg = reg_numHi(opr);
633 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
634 assert(live_kill.at(reg), "using fixed register that is not defined in this block");
635 }
636 }
637 #endif
638 }
639
640 // Add uses of live locals from interpreter's point of view for proper debug information generation
641 n = visitor.info_count();
642 for (k = 0; k < n; k++) {
643 CodeEmitInfo* info = visitor.info_at(k);
644 ValueStack* stack = info->stack();
645 for_each_state_value(stack, value,
646 set_live_gen_kill(value, op, live_gen, live_kill);
647 local_has_fpu_registers = local_has_fpu_registers || value->type()->is_float_kind();
648 );
649 }
650
651 // iterate temp operands of instruction
652 n = visitor.opr_count(LIR_OpVisitState::tempMode);
653 for (k = 0; k < n; k++) {
654 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k);
655 assert(opr->is_register(), "visitor should only return register operands");
656
657 if (opr->is_virtual_register()) {
658 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
659 reg = opr->vreg_number();
660 live_kill.set_bit(reg);
661 if (block->loop_index() >= 0) {
662 local_interval_in_loop.set_bit(reg, block->loop_index());
663 }
664 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
665 }
666
667 #ifdef ASSERT
668 // fixed intervals are never live at block boundaries, so
669 // they need not be processed in live sets
670 // process them only in debug mode so that this can be checked
671 if (!opr->is_virtual_register()) {
672 reg = reg_num(opr);
673 if (is_processed_reg_num(reg)) {
674 live_kill.set_bit(reg_num(opr));
675 }
676 reg = reg_numHi(opr);
677 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
678 live_kill.set_bit(reg);
679 }
680 }
681 #endif
682 }
683
684 // iterate output operands of instruction
685 n = visitor.opr_count(LIR_OpVisitState::outputMode);
686 for (k = 0; k < n; k++) {
687 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k);
688 assert(opr->is_register(), "visitor should only return register operands");
689
690 if (opr->is_virtual_register()) {
691 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
692 reg = opr->vreg_number();
693 live_kill.set_bit(reg);
694 if (block->loop_index() >= 0) {
695 local_interval_in_loop.set_bit(reg, block->loop_index());
696 }
697 local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu();
698 }
699
700 #ifdef ASSERT
701 // fixed intervals are never live at block boundaries, so
702 // they need not be processed in live sets
703 // process them only in debug mode so that this can be checked
704 if (!opr->is_virtual_register()) {
705 reg = reg_num(opr);
706 if (is_processed_reg_num(reg)) {
707 live_kill.set_bit(reg_num(opr));
708 }
709 reg = reg_numHi(opr);
710 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
711 live_kill.set_bit(reg);
712 }
713 }
714 #endif
715 }
716 } // end of instruction iteration
717
718 block->set_live_gen (live_gen);
719 block->set_live_kill(live_kill);
720 block->set_live_in (ResourceBitMap(live_size));
721 block->set_live_out (ResourceBitMap(live_size));
722
723 TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d ", block->block_id()); print_bitmap(block->live_gen()));
724 TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d ", block->block_id()); print_bitmap(block->live_kill()));
725 } // end of block iteration
726
727 // propagate local calculated information into LinearScan object
728 _has_fpu_registers = local_has_fpu_registers;
729 compilation()->set_has_fpu_code(local_has_fpu_registers);
730
731 _num_calls = local_num_calls;
732 _interval_in_loop = local_interval_in_loop;
733 }
734
735
736 // ********** Phase 3: perform a backward dataflow analysis to compute global live sets
737 // (sets live_in and live_out for each block)
738
739 void LinearScan::compute_global_live_sets() {
740 TIME_LINEAR_SCAN(timer_compute_global_live_sets);
741
742 int num_blocks = block_count();
743 bool change_occurred;
744 bool change_occurred_in_block;
745 int iteration_count = 0;
746 ResourceBitMap live_out(live_set_size()); // scratch set for calculations
747
748 // Perform a backward dataflow analysis to compute live_out and live_in for each block.
749 // The loop is executed until a fixpoint is reached (no changes in an iteration)
750 // Exception handlers must be processed because not all live values are
751 // present in the state array, e.g. because of global value numbering
752 do {
753 change_occurred = false;
754
755 // iterate all blocks in reverse order
756 for (int i = num_blocks - 1; i >= 0; i--) {
757 BlockBegin* block = block_at(i);
758
759 change_occurred_in_block = false;
760
761 // live_out(block) is the union of live_in(sux), for successors sux of block
762 int n = block->number_of_sux();
763 int e = block->number_of_exception_handlers();
764 if (n + e > 0) {
765 // block has successors
766 if (n > 0) {
767 live_out.set_from(block->sux_at(0)->live_in());
768 for (int j = 1; j < n; j++) {
769 live_out.set_union(block->sux_at(j)->live_in());
770 }
771 } else {
772 live_out.clear();
773 }
774 for (int j = 0; j < e; j++) {
775 live_out.set_union(block->exception_handler_at(j)->live_in());
776 }
777
778 if (!block->live_out().is_same(live_out)) {
779 // A change occurred. Swap the old and new live out sets to avoid copying.
780 ResourceBitMap temp = block->live_out();
781 block->set_live_out(live_out);
782 live_out = temp;
783
784 change_occurred = true;
785 change_occurred_in_block = true;
786 }
787 }
788
789 if (iteration_count == 0 || change_occurred_in_block) {
790 // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block))
791 // note: live_in has to be computed only in first iteration or if live_out has changed!
792 ResourceBitMap live_in = block->live_in();
793 live_in.set_from(block->live_out());
794 live_in.set_difference(block->live_kill());
795 live_in.set_union(block->live_gen());
796 }
797
798 #ifdef ASSERT
799 if (TraceLinearScanLevel >= 4) {
800 char c = ' ';
801 if (iteration_count == 0 || change_occurred_in_block) {
802 c = '*';
803 }
804 tty->print("(%d) live_in%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_in());
805 tty->print("(%d) live_out%c B%d ", iteration_count, c, block->block_id()); print_bitmap(block->live_out());
806 }
807 #endif
808 }
809 iteration_count++;
810
811 if (change_occurred && iteration_count > 50) {
812 BAILOUT("too many iterations in compute_global_live_sets");
813 }
814 } while (change_occurred);
815
816
817 #ifdef ASSERT
818 // check that fixed intervals are not live at block boundaries
819 // (live set must be empty at fixed intervals)
820 for (int i = 0; i < num_blocks; i++) {
821 BlockBegin* block = block_at(i);
822 for (int j = 0; j < LIR_Opr::vreg_base; j++) {
823 assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty");
824 assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty");
825 assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty");
826 }
827 }
828 #endif
829
830 // check that the live_in set of the first block is empty
831 ResourceBitMap live_in_args(ir()->start()->live_in().size());
832 if (!ir()->start()->live_in().is_same(live_in_args)) {
833 #ifdef ASSERT
834 tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)");
835 tty->print_cr("affected registers:");
836 print_bitmap(ir()->start()->live_in());
837
838 // print some additional information to simplify debugging
839 for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) {
840 if (ir()->start()->live_in().at(i)) {
841 Instruction* instr = gen()->instruction_for_vreg(i);
842 tty->print_cr("* vreg %d (HIR instruction %c%d)", i, instr == nullptr ? ' ' : instr->type()->tchar(), instr == nullptr ? 0 : instr->id());
843
844 for (int j = 0; j < num_blocks; j++) {
845 BlockBegin* block = block_at(j);
846 if (block->live_gen().at(i)) {
847 tty->print_cr(" used in block B%d", block->block_id());
848 }
849 if (block->live_kill().at(i)) {
850 tty->print_cr(" defined in block B%d", block->block_id());
851 }
852 }
853 }
854 }
855
856 #endif
857 // when this fails, virtual registers are used before they are defined.
858 assert(false, "live_in set of first block must be empty");
859 // bailout of if this occurs in product mode.
860 bailout("live_in set of first block not empty");
861 }
862 }
863
864
865 // ********** Phase 4: build intervals
866 // (fills the list _intervals)
867
868 void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) {
869 assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type");
870 LIR_Opr opr = value->operand();
871 Constant* con = value->as_Constant();
872
873 if ((con == nullptr || con->is_pinned()) && opr->is_register()) {
874 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
875 add_use(opr, from, to, use_kind);
876 }
877 }
878
879
880 void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) {
881 TRACE_LINEAR_SCAN(2, tty->print(" def "); opr->print(tty); tty->print_cr(" def_pos %d (%d)", def_pos, use_kind));
882 assert(opr->is_register(), "should not be called otherwise");
883
884 if (opr->is_virtual_register()) {
885 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
886 add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register());
887
888 } else {
889 int reg = reg_num(opr);
890 if (is_processed_reg_num(reg)) {
891 add_def(reg, def_pos, use_kind, opr->type_register());
892 }
893 reg = reg_numHi(opr);
894 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
895 add_def(reg, def_pos, use_kind, opr->type_register());
896 }
897 }
898 }
899
900 void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) {
901 TRACE_LINEAR_SCAN(2, tty->print(" use "); opr->print(tty); tty->print_cr(" from %d to %d (%d)", from, to, use_kind));
902 assert(opr->is_register(), "should not be called otherwise");
903
904 if (opr->is_virtual_register()) {
905 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
906 add_use(opr->vreg_number(), from, to, use_kind, opr->type_register());
907
908 } else {
909 int reg = reg_num(opr);
910 if (is_processed_reg_num(reg)) {
911 add_use(reg, from, to, use_kind, opr->type_register());
912 }
913 reg = reg_numHi(opr);
914 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
915 add_use(reg, from, to, use_kind, opr->type_register());
916 }
917 }
918 }
919
920 void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) {
921 TRACE_LINEAR_SCAN(2, tty->print(" temp "); opr->print(tty); tty->print_cr(" temp_pos %d (%d)", temp_pos, use_kind));
922 assert(opr->is_register(), "should not be called otherwise");
923
924 if (opr->is_virtual_register()) {
925 assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below");
926 add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register());
927
928 } else {
929 int reg = reg_num(opr);
930 if (is_processed_reg_num(reg)) {
931 add_temp(reg, temp_pos, use_kind, opr->type_register());
932 }
933 reg = reg_numHi(opr);
934 if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) {
935 add_temp(reg, temp_pos, use_kind, opr->type_register());
936 }
937 }
938 }
939
940
941 void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) {
942 Interval* interval = interval_at(reg_num);
943 if (interval != nullptr) {
944 assert(interval->reg_num() == reg_num, "wrong interval");
945
946 if (type != T_ILLEGAL) {
947 interval->set_type(type);
948 }
949
950 Range* r = interval->first();
951 if (r->from() <= def_pos) {
952 // Update the starting point (when a range is first created for a use, its
953 // start is the beginning of the current block until a def is encountered.)
954 r->set_from(def_pos);
955 interval->add_use_pos(def_pos, use_kind);
956
957 } else {
958 // Dead value - make vacuous interval
959 // also add use_kind for dead intervals
960 interval->add_range(def_pos, def_pos + 1);
961 interval->add_use_pos(def_pos, use_kind);
962 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use", reg_num, def_pos));
963 }
964
965 } else {
966 // Dead value - make vacuous interval
967 // also add use_kind for dead intervals
968 interval = create_interval(reg_num);
969 if (type != T_ILLEGAL) {
970 interval->set_type(type);
971 }
972
973 interval->add_range(def_pos, def_pos + 1);
974 interval->add_use_pos(def_pos, use_kind);
975 TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals", reg_num, def_pos));
976 }
977
978 change_spill_definition_pos(interval, def_pos);
979 if (use_kind == noUse && interval->spill_state() <= startInMemory) {
980 // detection of method-parameters and roundfp-results
981 // TODO: move this directly to position where use-kind is computed
982 interval->set_spill_state(startInMemory);
983 }
984 }
985
986 void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) {
987 Interval* interval = interval_at(reg_num);
988 if (interval == nullptr) {
989 interval = create_interval(reg_num);
990 }
991 assert(interval->reg_num() == reg_num, "wrong interval");
992
993 if (type != T_ILLEGAL) {
994 interval->set_type(type);
995 }
996
997 interval->add_range(from, to);
998 interval->add_use_pos(to, use_kind);
999 }
1000
1001 void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) {
1002 Interval* interval = interval_at(reg_num);
1003 if (interval == nullptr) {
1004 interval = create_interval(reg_num);
1005 }
1006 assert(interval->reg_num() == reg_num, "wrong interval");
1007
1008 if (type != T_ILLEGAL) {
1009 interval->set_type(type);
1010 }
1011
1012 interval->add_range(temp_pos, temp_pos + 1);
1013 interval->add_use_pos(temp_pos, use_kind);
1014 }
1015
1016
1017 // the results of this functions are used for optimizing spilling and reloading
1018 // if the functions return shouldHaveRegister and the interval is spilled,
1019 // it is not reloaded to a register.
1020 IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) {
1021 if (op->code() == lir_move) {
1022 assert(op->as_Op1() != nullptr, "lir_move must be LIR_Op1");
1023 LIR_Op1* move = (LIR_Op1*)op;
1024 LIR_Opr res = move->result_opr();
1025 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory);
1026
1027 if (result_in_memory) {
1028 // Begin of an interval with must_start_in_memory set.
1029 // This interval will always get a stack slot first, so return noUse.
1030 return noUse;
1031
1032 } else if (move->in_opr()->is_stack()) {
1033 // method argument (condition must be equal to handle_method_arguments)
1034 return noUse;
1035
1036 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) {
1037 // Move from register to register
1038 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) {
1039 // special handling of phi-function moves inside osr-entry blocks
1040 // input operand must have a register instead of output operand (leads to better register allocation)
1041 return shouldHaveRegister;
1042 }
1043 }
1044 }
1045
1046 if (opr->is_virtual() &&
1047 gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) {
1048 // result is a stack-slot, so prevent immediate reloading
1049 return noUse;
1050 }
1051
1052 // all other operands require a register
1053 return mustHaveRegister;
1054 }
1055
1056 IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) {
1057 if (op->code() == lir_move) {
1058 assert(op->as_Op1() != nullptr, "lir_move must be LIR_Op1");
1059 LIR_Op1* move = (LIR_Op1*)op;
1060 LIR_Opr res = move->result_opr();
1061 bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory);
1062
1063 if (result_in_memory) {
1064 // Move to an interval with must_start_in_memory set.
1065 // To avoid moves from stack to stack (not allowed) force the input operand to a register
1066 return mustHaveRegister;
1067
1068 } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) {
1069 // Move from register to register
1070 if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) {
1071 // special handling of phi-function moves inside osr-entry blocks
1072 // input operand must have a register instead of output operand (leads to better register allocation)
1073 return mustHaveRegister;
1074 }
1075
1076 // The input operand is not forced to a register (moves from stack to register are allowed),
1077 // but it is faster if the input operand is in a register
1078 return shouldHaveRegister;
1079 }
1080 }
1081
1082
1083 #if defined(X86) || defined(S390)
1084 if (op->code() == lir_cmove) {
1085 // conditional moves can handle stack operands
1086 assert(op->result_opr()->is_register(), "result must always be in a register");
1087 return shouldHaveRegister;
1088 }
1089
1090 // optimizations for second input operand of arithmehtic operations on Intel
1091 // this operand is allowed to be on the stack in some cases
1092 BasicType opr_type = opr->type_register();
1093 if (opr_type == T_FLOAT || opr_type == T_DOUBLE) {
1094 if (IA32_ONLY( (UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2 ) NOT_IA32( true )) {
1095 // SSE float instruction (T_DOUBLE only supported with SSE2)
1096 switch (op->code()) {
1097 case lir_cmp:
1098 case lir_add:
1099 case lir_sub:
1100 case lir_mul:
1101 case lir_div:
1102 {
1103 assert(op->as_Op2() != nullptr, "must be LIR_Op2");
1104 LIR_Op2* op2 = (LIR_Op2*)op;
1105 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
1106 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");
1107 return shouldHaveRegister;
1108 }
1109 }
1110 default:
1111 break;
1112 }
1113 } else {
1114 // FPU stack float instruction
1115 switch (op->code()) {
1116 case lir_add:
1117 case lir_sub:
1118 case lir_mul:
1119 case lir_div:
1120 {
1121 assert(op->as_Op2() != nullptr, "must be LIR_Op2");
1122 LIR_Op2* op2 = (LIR_Op2*)op;
1123 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
1124 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");
1125 return shouldHaveRegister;
1126 }
1127 }
1128 default:
1129 break;
1130 }
1131 }
1132 // We want to sometimes use logical operations on pointers, in particular in GC barriers.
1133 // Since 64bit logical operations do not current support operands on stack, we have to make sure
1134 // T_OBJECT doesn't get spilled along with T_LONG.
1135 } else if (opr_type != T_LONG LP64_ONLY(&& opr_type != T_OBJECT)) {
1136 // integer instruction (note: long operands must always be in register)
1137 switch (op->code()) {
1138 case lir_cmp:
1139 case lir_add:
1140 case lir_sub:
1141 case lir_logic_and:
1142 case lir_logic_or:
1143 case lir_logic_xor:
1144 {
1145 assert(op->as_Op2() != nullptr, "must be LIR_Op2");
1146 LIR_Op2* op2 = (LIR_Op2*)op;
1147 if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) {
1148 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");
1149 return shouldHaveRegister;
1150 }
1151 }
1152 default:
1153 break;
1154 }
1155 }
1156 #endif // X86 || S390
1157
1158 // all other operands require a register
1159 return mustHaveRegister;
1160 }
1161
1162
1163 void LinearScan::handle_method_arguments(LIR_Op* op) {
1164 // special handling for method arguments (moves from stack to virtual register):
1165 // the interval gets no register assigned, but the stack slot.
1166 // it is split before the first use by the register allocator.
1167
1168 if (op->code() == lir_move) {
1169 assert(op->as_Op1() != nullptr, "must be LIR_Op1");
1170 LIR_Op1* move = (LIR_Op1*)op;
1171
1172 if (move->in_opr()->is_stack()) {
1173 #ifdef ASSERT
1174 int arg_size = compilation()->method()->arg_size();
1175 LIR_Opr o = move->in_opr();
1176 if (o->is_single_stack()) {
1177 assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range");
1178 } else if (o->is_double_stack()) {
1179 assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range");
1180 } else {
1181 ShouldNotReachHere();
1182 }
1183
1184 assert(move->id() > 0, "invalid id");
1185 assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block");
1186 assert(move->result_opr()->is_virtual(), "result of move must be a virtual register");
1187
1188 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())));
1189 #endif
1190
1191 Interval* interval = interval_at(reg_num(move->result_opr()));
1192
1193 int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix());
1194 interval->set_canonical_spill_slot(stack_slot);
1195 interval->assign_reg(stack_slot);
1196 }
1197 }
1198 }
1199
1200 void LinearScan::handle_doubleword_moves(LIR_Op* op) {
1201 // special handling for doubleword move from memory to register:
1202 // in this case the registers of the input address and the result
1203 // registers must not overlap -> add a temp range for the input registers
1204 if (op->code() == lir_move) {
1205 assert(op->as_Op1() != nullptr, "must be LIR_Op1");
1206 LIR_Op1* move = (LIR_Op1*)op;
1207
1208 if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) {
1209 LIR_Address* address = move->in_opr()->as_address_ptr();
1210 if (address != nullptr) {
1211 if (address->base()->is_valid()) {
1212 add_temp(address->base(), op->id(), noUse);
1213 }
1214 if (address->index()->is_valid()) {
1215 add_temp(address->index(), op->id(), noUse);
1216 }
1217 }
1218 }
1219 }
1220 }
1221
1222 void LinearScan::add_register_hints(LIR_Op* op) {
1223 switch (op->code()) {
1224 case lir_move: // fall through
1225 case lir_convert: {
1226 assert(op->as_Op1() != nullptr, "lir_move, lir_convert must be LIR_Op1");
1227 LIR_Op1* move = (LIR_Op1*)op;
1228
1229 LIR_Opr move_from = move->in_opr();
1230 LIR_Opr move_to = move->result_opr();
1231
1232 if (move_to->is_register() && move_from->is_register()) {
1233 Interval* from = interval_at(reg_num(move_from));
1234 Interval* to = interval_at(reg_num(move_to));
1235 if (from != nullptr && to != nullptr) {
1236 to->set_register_hint(from);
1237 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()));
1238 }
1239 }
1240 break;
1241 }
1242 case lir_cmove: {
1243 assert(op->as_Op4() != nullptr, "lir_cmove must be LIR_Op4");
1244 LIR_Op4* cmove = (LIR_Op4*)op;
1245
1246 LIR_Opr move_from = cmove->in_opr1();
1247 LIR_Opr move_to = cmove->result_opr();
1248
1249 if (move_to->is_register() && move_from->is_register()) {
1250 Interval* from = interval_at(reg_num(move_from));
1251 Interval* to = interval_at(reg_num(move_to));
1252 if (from != nullptr && to != nullptr) {
1253 to->set_register_hint(from);
1254 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()));
1255 }
1256 }
1257 break;
1258 }
1259 default:
1260 break;
1261 }
1262 }
1263
1264
1265 void LinearScan::build_intervals() {
1266 TIME_LINEAR_SCAN(timer_build_intervals);
1267
1268 // initialize interval list with expected number of intervals
1269 // (32 is added to have some space for split children without having to resize the list)
1270 _intervals = IntervalList(num_virtual_regs() + 32);
1271 // initialize all slots that are used by build_intervals
1272 _intervals.at_put_grow(num_virtual_regs() - 1, nullptr, nullptr);
1273
1274 // create a list with all caller-save registers (cpu, fpu, xmm)
1275 // when an instruction is a call, a temp range is created for all these registers
1276 int num_caller_save_registers = 0;
1277 int caller_save_registers[LinearScan::nof_regs];
1278
1279 int i;
1280 for (i = 0; i < FrameMap::nof_caller_save_cpu_regs(); i++) {
1281 LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i);
1282 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
1283 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
1284 caller_save_registers[num_caller_save_registers++] = reg_num(opr);
1285 }
1286
1287 // temp ranges for fpu registers are only created when the method has
1288 // virtual fpu operands. Otherwise no allocation for fpu registers is
1289 // performed and so the temp ranges would be useless
1290 if (has_fpu_registers()) {
1291 #ifdef X86
1292 if (UseSSE < 2) {
1293 #endif // X86
1294 for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) {
1295 LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i);
1296 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
1297 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
1298 caller_save_registers[num_caller_save_registers++] = reg_num(opr);
1299 }
1300 #ifdef X86
1301 }
1302 #endif // X86
1303
1304 #ifdef X86
1305 if (UseSSE > 0) {
1306 int num_caller_save_xmm_regs = FrameMap::get_num_caller_save_xmms();
1307 for (i = 0; i < num_caller_save_xmm_regs; i ++) {
1308 LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i);
1309 assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands");
1310 assert(reg_numHi(opr) == -1, "missing addition of range for hi-register");
1311 caller_save_registers[num_caller_save_registers++] = reg_num(opr);
1312 }
1313 }
1314 #endif // X86
1315 }
1316 assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds");
1317
1318
1319 LIR_OpVisitState visitor;
1320
1321 // iterate all blocks in reverse order
1322 for (i = block_count() - 1; i >= 0; i--) {
1323 BlockBegin* block = block_at(i);
1324 LIR_OpList* instructions = block->lir()->instructions_list();
1325 int block_from = block->first_lir_instruction_id();
1326 int block_to = block->last_lir_instruction_id();
1327
1328 assert(block_from == instructions->at(0)->id(), "must be");
1329 assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be");
1330
1331 // Update intervals for registers live at the end of this block;
1332 ResourceBitMap& live = block->live_out();
1333 auto updater = [&](BitMap::idx_t index) {
1334 int number = static_cast<int>(index);
1335 assert(number >= LIR_Opr::vreg_base, "fixed intervals must not be live on block bounds");
1336 TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d", number, block_to + 2));
1337
1338 add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL);
1339
1340 // add special use positions for loop-end blocks when the
1341 // interval is used anywhere inside this loop. It's possible
1342 // that the block was part of a non-natural loop, so it might
1343 // have an invalid loop index.
1344 if (block->is_set(BlockBegin::linear_scan_loop_end_flag) &&
1345 block->loop_index() != -1 &&
1346 is_interval_in_loop(number, block->loop_index())) {
1347 interval_at(number)->add_use_pos(block_to + 1, loopEndMarker);
1348 }
1349 };
1350 live.iterate(updater);
1351
1352 // iterate all instructions of the block in reverse order.
1353 // skip the first instruction because it is always a label
1354 // definitions of intervals are processed before uses
1355 assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label");
1356 for (int j = instructions->length() - 1; j >= 1; j--) {
1357 LIR_Op* op = instructions->at(j);
1358 int op_id = op->id();
1359
1360 // visit operation to collect all operands
1361 visitor.visit(op);
1362
1363 // add a temp range for each register if operation destroys caller-save registers
1364 if (visitor.has_call()) {
1365 for (int k = 0; k < num_caller_save_registers; k++) {
1366 add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL);
1367 }
1368 TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers"));
1369 }
1370
1371 // Add any platform dependent temps
1372 pd_add_temps(op);
1373
1374 // visit definitions (output and temp operands)
1375 int k, n;
1376 n = visitor.opr_count(LIR_OpVisitState::outputMode);
1377 for (k = 0; k < n; k++) {
1378 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k);
1379 assert(opr->is_register(), "visitor should only return register operands");
1380 add_def(opr, op_id, use_kind_of_output_operand(op, opr));
1381 }
1382
1383 n = visitor.opr_count(LIR_OpVisitState::tempMode);
1384 for (k = 0; k < n; k++) {
1385 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k);
1386 assert(opr->is_register(), "visitor should only return register operands");
1387 add_temp(opr, op_id, mustHaveRegister);
1388 }
1389
1390 // visit uses (input operands)
1391 n = visitor.opr_count(LIR_OpVisitState::inputMode);
1392 for (k = 0; k < n; k++) {
1393 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k);
1394 assert(opr->is_register(), "visitor should only return register operands");
1395 add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr));
1396 }
1397
1398 // Add uses of live locals from interpreter's point of view for proper
1399 // debug information generation
1400 // Treat these operands as temp values (if the life range is extended
1401 // to a call site, the value would be in a register at the call otherwise)
1402 n = visitor.info_count();
1403 for (k = 0; k < n; k++) {
1404 CodeEmitInfo* info = visitor.info_at(k);
1405 ValueStack* stack = info->stack();
1406 for_each_state_value(stack, value,
1407 add_use(value, block_from, op_id + 1, noUse);
1408 );
1409 }
1410
1411 // special steps for some instructions (especially moves)
1412 handle_method_arguments(op);
1413 handle_doubleword_moves(op);
1414 add_register_hints(op);
1415
1416 } // end of instruction iteration
1417 } // end of block iteration
1418
1419
1420 // add the range [0, 1[ to all fixed intervals
1421 // -> the register allocator need not handle unhandled fixed intervals
1422 for (int n = 0; n < LinearScan::nof_regs; n++) {
1423 Interval* interval = interval_at(n);
1424 if (interval != nullptr) {
1425 interval->add_range(0, 1);
1426 }
1427 }
1428 }
1429
1430
1431 // ********** Phase 5: actual register allocation
1432
1433 int LinearScan::interval_cmp(Interval** a, Interval** b) {
1434 if (*a != nullptr) {
1435 if (*b != nullptr) {
1436 return (*a)->from() - (*b)->from();
1437 } else {
1438 return -1;
1439 }
1440 } else {
1441 if (*b != nullptr) {
1442 return 1;
1443 } else {
1444 return 0;
1445 }
1446 }
1447 }
1448
1449 #ifdef ASSERT
1450 static int interval_cmp(Interval* const& l, Interval* const& r) {
1451 return l->from() - r->from();
1452 }
1453
1454 static bool find_interval(Interval* interval, IntervalArray* intervals) {
1455 bool found;
1456 int idx = intervals->find_sorted<Interval*, interval_cmp>(interval, found);
1457
1458 if (!found) {
1459 return false;
1460 }
1461
1462 int from = interval->from();
1463
1464 // The index we've found using binary search is pointing to an interval
1465 // that is defined in the same place as the interval we were looking for.
1466 // So now we have to look around that index and find exact interval.
1467 for (int i = idx; i >= 0; i--) {
1468 if (intervals->at(i) == interval) {
1469 return true;
1470 }
1471 if (intervals->at(i)->from() != from) {
1472 break;
1473 }
1474 }
1475
1476 for (int i = idx + 1; i < intervals->length(); i++) {
1477 if (intervals->at(i) == interval) {
1478 return true;
1479 }
1480 if (intervals->at(i)->from() != from) {
1481 break;
1482 }
1483 }
1484
1485 return false;
1486 }
1487
1488 bool LinearScan::is_sorted(IntervalArray* intervals) {
1489 int from = -1;
1490 int null_count = 0;
1491
1492 for (int i = 0; i < intervals->length(); i++) {
1493 Interval* it = intervals->at(i);
1494 if (it != nullptr) {
1495 assert(from <= it->from(), "Intervals are unordered");
1496 from = it->from();
1497 } else {
1498 null_count++;
1499 }
1500 }
1501
1502 assert(null_count == 0, "Sorted intervals should not contain nulls");
1503
1504 null_count = 0;
1505
1506 for (int i = 0; i < interval_count(); i++) {
1507 Interval* interval = interval_at(i);
1508 if (interval != nullptr) {
1509 assert(find_interval(interval, intervals), "Lists do not contain same intervals");
1510 } else {
1511 null_count++;
1512 }
1513 }
1514
1515 assert(interval_count() - null_count == intervals->length(),
1516 "Sorted list should contain the same amount of non-null intervals as unsorted list");
1517
1518 return true;
1519 }
1520 #endif
1521
1522 void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) {
1523 if (*prev != nullptr) {
1524 (*prev)->set_next(interval);
1525 } else {
1526 *first = interval;
1527 }
1528 *prev = interval;
1529 }
1530
1531 void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) {
1532 assert(is_sorted(_sorted_intervals), "interval list is not sorted");
1533
1534 *list1 = *list2 = Interval::end();
1535
1536 Interval* list1_prev = nullptr;
1537 Interval* list2_prev = nullptr;
1538 Interval* v;
1539
1540 const int n = _sorted_intervals->length();
1541 for (int i = 0; i < n; i++) {
1542 v = _sorted_intervals->at(i);
1543 if (v == nullptr) continue;
1544
1545 if (is_list1(v)) {
1546 add_to_list(list1, &list1_prev, v);
1547 } else if (is_list2 == nullptr || is_list2(v)) {
1548 add_to_list(list2, &list2_prev, v);
1549 }
1550 }
1551
1552 if (list1_prev != nullptr) list1_prev->set_next(Interval::end());
1553 if (list2_prev != nullptr) list2_prev->set_next(Interval::end());
1554
1555 assert(list1_prev == nullptr || list1_prev->next() == Interval::end(), "linear list ends not with sentinel");
1556 assert(list2_prev == nullptr || list2_prev->next() == Interval::end(), "linear list ends not with sentinel");
1557 }
1558
1559
1560 void LinearScan::sort_intervals_before_allocation() {
1561 TIME_LINEAR_SCAN(timer_sort_intervals_before);
1562
1563 if (_needs_full_resort) {
1564 // There is no known reason why this should occur but just in case...
1565 assert(false, "should never occur");
1566 // Re-sort existing interval list because an Interval::from() has changed
1567 _sorted_intervals->sort(interval_cmp);
1568 _needs_full_resort = false;
1569 }
1570
1571 IntervalList* unsorted_list = &_intervals;
1572 int unsorted_len = unsorted_list->length();
1573 int sorted_len = 0;
1574 int unsorted_idx;
1575 int sorted_idx = 0;
1576 int sorted_from_max = -1;
1577
1578 // calc number of items for sorted list (sorted list must not contain null values)
1579 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) {
1580 if (unsorted_list->at(unsorted_idx) != nullptr) {
1581 sorted_len++;
1582 }
1583 }
1584 IntervalArray* sorted_list = new IntervalArray(sorted_len, sorted_len, nullptr);
1585
1586 // special sorting algorithm: the original interval-list is almost sorted,
1587 // only some intervals are swapped. So this is much faster than a complete QuickSort
1588 for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) {
1589 Interval* cur_interval = unsorted_list->at(unsorted_idx);
1590
1591 if (cur_interval != nullptr) {
1592 int cur_from = cur_interval->from();
1593
1594 if (sorted_from_max <= cur_from) {
1595 sorted_list->at_put(sorted_idx++, cur_interval);
1596 sorted_from_max = cur_interval->from();
1597 } else {
1598 // the assumption that the intervals are already sorted failed,
1599 // so this interval must be sorted in manually
1600 int j;
1601 for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) {
1602 sorted_list->at_put(j + 1, sorted_list->at(j));
1603 }
1604 sorted_list->at_put(j + 1, cur_interval);
1605 sorted_idx++;
1606 }
1607 }
1608 }
1609 _sorted_intervals = sorted_list;
1610 assert(is_sorted(_sorted_intervals), "intervals unsorted");
1611 }
1612
1613 void LinearScan::sort_intervals_after_allocation() {
1614 TIME_LINEAR_SCAN(timer_sort_intervals_after);
1615
1616 if (_needs_full_resort) {
1617 // Re-sort existing interval list because an Interval::from() has changed
1618 _sorted_intervals->sort(interval_cmp);
1619 _needs_full_resort = false;
1620 }
1621
1622 IntervalArray* old_list = _sorted_intervals;
1623 IntervalList* new_list = _new_intervals_from_allocation;
1624 int old_len = old_list->length();
1625 int new_len = new_list == nullptr ? 0 : new_list->length();
1626
1627 if (new_len == 0) {
1628 // no intervals have been added during allocation, so sorted list is already up to date
1629 assert(is_sorted(_sorted_intervals), "intervals unsorted");
1630 return;
1631 }
1632
1633 // conventional sort-algorithm for new intervals
1634 new_list->sort(interval_cmp);
1635
1636 // merge old and new list (both already sorted) into one combined list
1637 int combined_list_len = old_len + new_len;
1638 IntervalArray* combined_list = new IntervalArray(combined_list_len, combined_list_len, nullptr);
1639 int old_idx = 0;
1640 int new_idx = 0;
1641
1642 while (old_idx + new_idx < old_len + new_len) {
1643 if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) {
1644 combined_list->at_put(old_idx + new_idx, old_list->at(old_idx));
1645 old_idx++;
1646 } else {
1647 combined_list->at_put(old_idx + new_idx, new_list->at(new_idx));
1648 new_idx++;
1649 }
1650 }
1651
1652 _sorted_intervals = combined_list;
1653 assert(is_sorted(_sorted_intervals), "intervals unsorted");
1654 }
1655
1656
1657 void LinearScan::allocate_registers() {
1658 TIME_LINEAR_SCAN(timer_allocate_registers);
1659
1660 Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals;
1661 Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals;
1662
1663 // collect cpu intervals
1664 create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals,
1665 is_precolored_cpu_interval, is_virtual_cpu_interval);
1666
1667 // collect fpu intervals
1668 create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals,
1669 is_precolored_fpu_interval, is_virtual_fpu_interval);
1670 // this fpu interval collection cannot be moved down below with the allocation section as
1671 // the cpu_lsw.walk() changes interval positions.
1672
1673 if (!has_fpu_registers()) {
1674 #ifdef ASSERT
1675 assert(not_precolored_fpu_intervals == Interval::end(), "missed an uncolored fpu interval");
1676 #else
1677 if (not_precolored_fpu_intervals != Interval::end()) {
1678 BAILOUT("missed an uncolored fpu interval");
1679 }
1680 #endif
1681 }
1682
1683 // allocate cpu registers
1684 LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals);
1685 cpu_lsw.walk();
1686 cpu_lsw.finish_allocation();
1687
1688 if (has_fpu_registers()) {
1689 // allocate fpu registers
1690 LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals);
1691 fpu_lsw.walk();
1692 fpu_lsw.finish_allocation();
1693 }
1694 }
1695
1696
1697 // ********** Phase 6: resolve data flow
1698 // (insert moves at edges between blocks if intervals have been split)
1699
1700 // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode
1701 // instead of returning null
1702 Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) {
1703 Interval* result = interval->split_child_at_op_id(op_id, mode);
1704 if (result != nullptr) {
1705 return result;
1706 }
1707
1708 assert(false, "must find an interval, but do a clean bailout in product mode");
1709 result = new Interval(LIR_Opr::vreg_base);
1710 result->assign_reg(0);
1711 result->set_type(T_INT);
1712 BAILOUT_("LinearScan: interval is null", result);
1713 }
1714
1715
1716 Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) {
1717 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
1718 assert(interval_at(reg_num) != nullptr, "no interval found");
1719
1720 return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode);
1721 }
1722
1723 Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) {
1724 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
1725 assert(interval_at(reg_num) != nullptr, "no interval found");
1726
1727 return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode);
1728 }
1729
1730 Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) {
1731 assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds");
1732 assert(interval_at(reg_num) != nullptr, "no interval found");
1733
1734 return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode);
1735 }
1736
1737
1738 void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) {
1739 DEBUG_ONLY(move_resolver.check_empty());
1740
1741 // visit all registers where the live_at_edge bit is set
1742 const ResourceBitMap& live_at_edge = to_block->live_in();
1743 auto visitor = [&](BitMap::idx_t index) {
1744 int r = static_cast<int>(index);
1745 assert(r < num_virtual_regs(), "live information set for not existing interval");
1746 assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge");
1747
1748 Interval* from_interval = interval_at_block_end(from_block, r);
1749 Interval* to_interval = interval_at_block_begin(to_block, r);
1750
1751 if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) {
1752 // need to insert move instruction
1753 move_resolver.add_mapping(from_interval, to_interval);
1754 }
1755 };
1756 live_at_edge.iterate(visitor, 0, live_set_size());
1757 }
1758
1759
1760 void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) {
1761 if (from_block->number_of_sux() <= 1) {
1762 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d", from_block->block_id()));
1763
1764 LIR_OpList* instructions = from_block->lir()->instructions_list();
1765 LIR_OpBranch* branch = instructions->last()->as_OpBranch();
1766 if (branch != nullptr) {
1767 // insert moves before branch
1768 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump");
1769 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2);
1770 } else {
1771 move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1);
1772 }
1773
1774 } else {
1775 TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d", to_block->block_id()));
1776 #ifdef ASSERT
1777 assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != nullptr, "block does not start with a label");
1778
1779 // because the number of predecessor edges matches the number of
1780 // successor edges, blocks which are reached by switch statements
1781 // may have be more than one predecessor but it will be guaranteed
1782 // that all predecessors will be the same.
1783 for (int i = 0; i < to_block->number_of_preds(); i++) {
1784 assert(from_block == to_block->pred_at(i), "all critical edges must be broken");
1785 }
1786 #endif
1787
1788 move_resolver.set_insert_position(to_block->lir(), 0);
1789 }
1790 }
1791
1792
1793 // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split
1794 void LinearScan::resolve_data_flow() {
1795 TIME_LINEAR_SCAN(timer_resolve_data_flow);
1796
1797 int num_blocks = block_count();
1798 MoveResolver move_resolver(this);
1799 ResourceBitMap block_completed(num_blocks);
1800 ResourceBitMap already_resolved(num_blocks);
1801
1802 int i;
1803 for (i = 0; i < num_blocks; i++) {
1804 BlockBegin* block = block_at(i);
1805
1806 // check if block has only one predecessor and only one successor
1807 if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) {
1808 LIR_OpList* instructions = block->lir()->instructions_list();
1809 assert(instructions->at(0)->code() == lir_label, "block must start with label");
1810 assert(instructions->last()->code() == lir_branch, "block with successors must end with branch");
1811 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch");
1812
1813 // check if block is empty (only label and branch)
1814 if (instructions->length() == 2) {
1815 BlockBegin* pred = block->pred_at(0);
1816 BlockBegin* sux = block->sux_at(0);
1817
1818 // prevent optimization of two consecutive blocks
1819 if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) {
1820 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()));
1821 block_completed.set_bit(block->linear_scan_number());
1822
1823 // directly resolve between pred and sux (without looking at the empty block between)
1824 resolve_collect_mappings(pred, sux, move_resolver);
1825 if (move_resolver.has_mappings()) {
1826 move_resolver.set_insert_position(block->lir(), 0);
1827 move_resolver.resolve_and_append_moves();
1828 }
1829 }
1830 }
1831 }
1832 }
1833
1834
1835 for (i = 0; i < num_blocks; i++) {
1836 if (!block_completed.at(i)) {
1837 BlockBegin* from_block = block_at(i);
1838 already_resolved.set_from(block_completed);
1839
1840 int num_sux = from_block->number_of_sux();
1841 for (int s = 0; s < num_sux; s++) {
1842 BlockBegin* to_block = from_block->sux_at(s);
1843
1844 // check for duplicate edges between the same blocks (can happen with switch blocks)
1845 if (!already_resolved.at(to_block->linear_scan_number())) {
1846 TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d", from_block->block_id(), to_block->block_id()));
1847 already_resolved.set_bit(to_block->linear_scan_number());
1848
1849 // collect all intervals that have been split between from_block and to_block
1850 resolve_collect_mappings(from_block, to_block, move_resolver);
1851 if (move_resolver.has_mappings()) {
1852 resolve_find_insert_pos(from_block, to_block, move_resolver);
1853 move_resolver.resolve_and_append_moves();
1854 }
1855 }
1856 }
1857 }
1858 }
1859 }
1860
1861
1862 void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) {
1863 if (interval_at(reg_num) == nullptr) {
1864 // if a phi function is never used, no interval is created -> ignore this
1865 return;
1866 }
1867
1868 Interval* interval = interval_at_block_begin(block, reg_num);
1869 int reg = interval->assigned_reg();
1870 int regHi = interval->assigned_regHi();
1871
1872 if ((reg < nof_regs && interval->always_in_memory()) ||
1873 (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) {
1874 // the interval is split to get a short range that is located on the stack
1875 // in the following two cases:
1876 // * the interval started in memory (e.g. method parameter), but is currently in a register
1877 // this is an optimization for exception handling that reduces the number of moves that
1878 // are necessary for resolving the states when an exception uses this exception handler
1879 // * the interval would be on the fpu stack at the begin of the exception handler
1880 // this is not allowed because of the complicated fpu stack handling on Intel
1881
1882 // range that will be spilled to memory
1883 int from_op_id = block->first_lir_instruction_id();
1884 int to_op_id = from_op_id + 1; // short live range of length 1
1885 assert(interval->from() <= from_op_id && interval->to() >= to_op_id,
1886 "no split allowed between exception entry and first instruction");
1887
1888 if (interval->from() != from_op_id) {
1889 // the part before from_op_id is unchanged
1890 interval = interval->split(from_op_id);
1891 interval->assign_reg(reg, regHi);
1892 append_interval(interval);
1893 } else {
1894 _needs_full_resort = true;
1895 }
1896 assert(interval->from() == from_op_id, "must be true now");
1897
1898 Interval* spilled_part = interval;
1899 if (interval->to() != to_op_id) {
1900 // the part after to_op_id is unchanged
1901 spilled_part = interval->split_from_start(to_op_id);
1902 append_interval(spilled_part);
1903 move_resolver.add_mapping(spilled_part, interval);
1904 }
1905 assign_spill_slot(spilled_part);
1906
1907 assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking");
1908 }
1909 }
1910
1911 void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) {
1912 assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise");
1913 DEBUG_ONLY(move_resolver.check_empty());
1914
1915 // visit all registers where the live_in bit is set
1916 auto resolver = [&](BitMap::idx_t index) {
1917 int r = static_cast<int>(index);
1918 resolve_exception_entry(block, r, move_resolver);
1919 };
1920 block->live_in().iterate(resolver, 0, live_set_size());
1921
1922 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately
1923 for_each_phi_fun(block, phi,
1924 if (!phi->is_illegal()) { resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver); }
1925 );
1926
1927 if (move_resolver.has_mappings()) {
1928 // insert moves after first instruction
1929 move_resolver.set_insert_position(block->lir(), 0);
1930 move_resolver.resolve_and_append_moves();
1931 }
1932 }
1933
1934
1935 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) {
1936 if (interval_at(reg_num) == nullptr) {
1937 // if a phi function is never used, no interval is created -> ignore this
1938 return;
1939 }
1940
1941 // the computation of to_interval is equal to resolve_collect_mappings,
1942 // but from_interval is more complicated because of phi functions
1943 BlockBegin* to_block = handler->entry_block();
1944 Interval* to_interval = interval_at_block_begin(to_block, reg_num);
1945
1946 if (phi != nullptr) {
1947 // phi function of the exception entry block
1948 // no moves are created for this phi function in the LIR_Generator, so the
1949 // interval at the throwing instruction must be searched using the operands
1950 // of the phi function
1951 Value from_value = phi->operand_at(handler->phi_operand());
1952 if (from_value == nullptr) {
1953 // We have reached here in a kotlin application running with JVMTI
1954 // capability "can_access_local_variables".
1955 // The illegal state is not yet propagated to this phi. Do it here.
1956 phi->make_illegal();
1957 // We can skip the illegal phi edge.
1958 return;
1959 }
1960
1961 // with phi functions it can happen that the same from_value is used in
1962 // multiple mappings, so notify move-resolver that this is allowed
1963 move_resolver.set_multiple_reads_allowed();
1964
1965 Constant* con = from_value->as_Constant();
1966 if (con != nullptr && (!con->is_pinned() || con->operand()->is_constant())) {
1967 // Need a mapping from constant to interval if unpinned (may have no register) or if the operand is a constant (no register).
1968 move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval);
1969 } else {
1970 // search split child at the throwing op_id
1971 Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id);
1972 move_resolver.add_mapping(from_interval, to_interval);
1973 }
1974 } else {
1975 // no phi function, so use reg_num also for from_interval
1976 // search split child at the throwing op_id
1977 Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id);
1978 if (from_interval != to_interval) {
1979 // optimization to reduce number of moves: when to_interval is on stack and
1980 // the stack slot is known to be always correct, then no move is necessary
1981 if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) {
1982 move_resolver.add_mapping(from_interval, to_interval);
1983 }
1984 }
1985 }
1986 }
1987
1988 void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) {
1989 TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d", handler->entry_block()->block_id(), throwing_op_id));
1990
1991 DEBUG_ONLY(move_resolver.check_empty());
1992 assert(handler->lir_op_id() == -1, "already processed this xhandler");
1993 DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id));
1994 assert(handler->entry_code() == nullptr, "code already present");
1995
1996 // visit all registers where the live_in bit is set
1997 BlockBegin* block = handler->entry_block();
1998 auto resolver = [&](BitMap::idx_t index) {
1999 int r = static_cast<int>(index);
2000 resolve_exception_edge(handler, throwing_op_id, r, nullptr, move_resolver);
2001 };
2002 block->live_in().iterate(resolver, 0, live_set_size());
2003
2004 // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately
2005 for_each_phi_fun(block, phi,
2006 if (!phi->is_illegal()) { resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver); }
2007 );
2008
2009 if (move_resolver.has_mappings()) {
2010 LIR_List* entry_code = new LIR_List(compilation());
2011 move_resolver.set_insert_position(entry_code, 0);
2012 move_resolver.resolve_and_append_moves();
2013
2014 entry_code->jump(handler->entry_block());
2015 handler->set_entry_code(entry_code);
2016 }
2017 }
2018
2019
2020 void LinearScan::resolve_exception_handlers() {
2021 MoveResolver move_resolver(this);
2022 LIR_OpVisitState visitor;
2023 int num_blocks = block_count();
2024
2025 int i;
2026 for (i = 0; i < num_blocks; i++) {
2027 BlockBegin* block = block_at(i);
2028 if (block->is_set(BlockBegin::exception_entry_flag)) {
2029 resolve_exception_entry(block, move_resolver);
2030 }
2031 }
2032
2033 for (i = 0; i < num_blocks; i++) {
2034 BlockBegin* block = block_at(i);
2035 LIR_List* ops = block->lir();
2036 int num_ops = ops->length();
2037
2038 // iterate all instructions of the block. skip the first because it is always a label
2039 assert(visitor.no_operands(ops->at(0)), "first operation must always be a label");
2040 for (int j = 1; j < num_ops; j++) {
2041 LIR_Op* op = ops->at(j);
2042 int op_id = op->id();
2043
2044 if (op_id != -1 && has_info(op_id)) {
2045 // visit operation to collect all operands
2046 visitor.visit(op);
2047 assert(visitor.info_count() > 0, "should not visit otherwise");
2048
2049 XHandlers* xhandlers = visitor.all_xhandler();
2050 int n = xhandlers->length();
2051 for (int k = 0; k < n; k++) {
2052 resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver);
2053 }
2054
2055 #ifdef ASSERT
2056 } else {
2057 visitor.visit(op);
2058 assert(visitor.all_xhandler()->length() == 0, "missed exception handler");
2059 #endif
2060 }
2061 }
2062 }
2063 }
2064
2065
2066 // ********** Phase 7: assign register numbers back to LIR
2067 // (includes computation of debug information and oop maps)
2068
2069 VMReg LinearScan::vm_reg_for_interval(Interval* interval) {
2070 VMReg reg = interval->cached_vm_reg();
2071 if (!reg->is_valid() ) {
2072 reg = vm_reg_for_operand(operand_for_interval(interval));
2073 interval->set_cached_vm_reg(reg);
2074 }
2075 assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value");
2076 return reg;
2077 }
2078
2079 VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) {
2080 assert(opr->is_oop(), "currently only implemented for oop operands");
2081 return frame_map()->regname(opr);
2082 }
2083
2084
2085 LIR_Opr LinearScan::operand_for_interval(Interval* interval) {
2086 LIR_Opr opr = interval->cached_opr();
2087 if (opr->is_illegal()) {
2088 opr = calc_operand_for_interval(interval);
2089 interval->set_cached_opr(opr);
2090 }
2091
2092 assert(opr == calc_operand_for_interval(interval), "wrong cached value");
2093 return opr;
2094 }
2095
2096 LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) {
2097 int assigned_reg = interval->assigned_reg();
2098 BasicType type = interval->type();
2099
2100 if (assigned_reg >= nof_regs) {
2101 // stack slot
2102 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2103 return LIR_OprFact::stack(assigned_reg - nof_regs, type);
2104
2105 } else {
2106 // register
2107 switch (type) {
2108 case T_OBJECT: {
2109 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2110 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2111 return LIR_OprFact::single_cpu_oop(assigned_reg);
2112 }
2113
2114 case T_ADDRESS: {
2115 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2116 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2117 return LIR_OprFact::single_cpu_address(assigned_reg);
2118 }
2119
2120 case T_METADATA: {
2121 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2122 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2123 return LIR_OprFact::single_cpu_metadata(assigned_reg);
2124 }
2125
2126 #ifdef __SOFTFP__
2127 case T_FLOAT: // fall through
2128 #endif // __SOFTFP__
2129 case T_INT: {
2130 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2131 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2132 return LIR_OprFact::single_cpu(assigned_reg);
2133 }
2134
2135 #ifdef __SOFTFP__
2136 case T_DOUBLE: // fall through
2137 #endif // __SOFTFP__
2138 case T_LONG: {
2139 int assigned_regHi = interval->assigned_regHi();
2140 assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register");
2141 assert(num_physical_regs(T_LONG) == 1 ||
2142 (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register");
2143
2144 assert(assigned_reg != assigned_regHi, "invalid allocation");
2145 assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi,
2146 "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)");
2147 assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match");
2148 if (requires_adjacent_regs(T_LONG)) {
2149 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even");
2150 }
2151
2152 #ifdef _LP64
2153 return LIR_OprFact::double_cpu(assigned_reg, assigned_reg);
2154 #else
2155 return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi);
2156 #endif // LP64
2157 }
2158
2159 #ifndef __SOFTFP__
2160 case T_FLOAT: {
2161 #ifdef X86
2162 if (UseSSE >= 1) {
2163 int last_xmm_reg = pd_last_xmm_reg;
2164 #ifdef _LP64
2165 if (UseAVX < 3) {
2166 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1;
2167 }
2168 #endif // LP64
2169 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register");
2170 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2171 return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg);
2172 }
2173 #endif // X86
2174
2175 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
2176 assert(interval->assigned_regHi() == any_reg, "must not have hi register");
2177 return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg);
2178 }
2179
2180 case T_DOUBLE: {
2181 #ifdef X86
2182 if (UseSSE >= 2) {
2183 int last_xmm_reg = pd_last_xmm_reg;
2184 #ifdef _LP64
2185 if (UseAVX < 3) {
2186 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1;
2187 }
2188 #endif // LP64
2189 assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register");
2190 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)");
2191 return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg);
2192 }
2193 #endif // X86
2194
2195 #if defined(ARM32)
2196 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
2197 assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register");
2198 assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even");
2199 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg, interval->assigned_regHi() - pd_first_fpu_reg);
2200 #else
2201 assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register");
2202 assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)");
2203 LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg);
2204 #endif
2205 return result;
2206 }
2207 #endif // __SOFTFP__
2208
2209 default: {
2210 ShouldNotReachHere();
2211 return LIR_OprFact::illegalOpr;
2212 }
2213 }
2214 }
2215 }
2216
2217 LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) {
2218 assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set");
2219 return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type());
2220 }
2221
2222 LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) {
2223 assert(opr->is_virtual(), "should not call this otherwise");
2224
2225 Interval* interval = interval_at(opr->vreg_number());
2226 assert(interval != nullptr, "interval must exist");
2227
2228 if (op_id != -1) {
2229 #ifdef ASSERT
2230 BlockBegin* block = block_of_op_with_id(op_id);
2231 if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) {
2232 // check if spill moves could have been appended at the end of this block, but
2233 // before the branch instruction. So the split child information for this branch would
2234 // be incorrect.
2235 LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch();
2236 if (branch != nullptr) {
2237 if (block->live_out().at(opr->vreg_number())) {
2238 assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump");
2239 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)");
2240 }
2241 }
2242 }
2243 #endif
2244
2245 // operands are not changed when an interval is split during allocation,
2246 // so search the right interval here
2247 interval = split_child_at_op_id(interval, op_id, mode);
2248 }
2249
2250 LIR_Opr res = operand_for_interval(interval);
2251
2252 #ifdef X86
2253 // new semantic for is_last_use: not only set on definite end of interval,
2254 // but also before hole
2255 // This may still miss some cases (e.g. for dead values), but it is not necessary that the
2256 // last use information is completely correct
2257 // information is only needed for fpu stack allocation
2258 if (res->is_fpu_register()) {
2259 if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) {
2260 assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow");
2261 res = res->make_last_use();
2262 }
2263 }
2264 #endif
2265
2266 assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation");
2267
2268 return res;
2269 }
2270
2271
2272 #ifdef ASSERT
2273 // some methods used to check correctness of debug information
2274
2275 void assert_no_register_values(GrowableArray<ScopeValue*>* values) {
2276 if (values == nullptr) {
2277 return;
2278 }
2279
2280 for (int i = 0; i < values->length(); i++) {
2281 ScopeValue* value = values->at(i);
2282
2283 if (value->is_location()) {
2284 Location location = ((LocationValue*)value)->location();
2285 assert(location.where() == Location::on_stack, "value is in register");
2286 }
2287 }
2288 }
2289
2290 void assert_no_register_values(GrowableArray<MonitorValue*>* values) {
2291 if (values == nullptr) {
2292 return;
2293 }
2294
2295 for (int i = 0; i < values->length(); i++) {
2296 MonitorValue* value = values->at(i);
2297
2298 if (value->owner()->is_location()) {
2299 Location location = ((LocationValue*)value->owner())->location();
2300 assert(location.where() == Location::on_stack, "owner is in register");
2301 }
2302 assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register");
2303 }
2304 }
2305
2306 static void assert_equal(Location l1, Location l2) {
2307 assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), "");
2308 }
2309
2310 static void assert_equal(ScopeValue* v1, ScopeValue* v2) {
2311 if (v1->is_location()) {
2312 assert(v2->is_location(), "");
2313 assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location());
2314 } else if (v1->is_constant_int()) {
2315 assert(v2->is_constant_int(), "");
2316 assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), "");
2317 } else if (v1->is_constant_double()) {
2318 assert(v2->is_constant_double(), "");
2319 assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), "");
2320 } else if (v1->is_constant_long()) {
2321 assert(v2->is_constant_long(), "");
2322 assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), "");
2323 } else if (v1->is_constant_oop()) {
2324 assert(v2->is_constant_oop(), "");
2325 assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), "");
2326 } else {
2327 ShouldNotReachHere();
2328 }
2329 }
2330
2331 static void assert_equal(MonitorValue* m1, MonitorValue* m2) {
2332 assert_equal(m1->owner(), m2->owner());
2333 assert_equal(m1->basic_lock(), m2->basic_lock());
2334 }
2335
2336 static void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) {
2337 assert(d1->scope() == d2->scope(), "not equal");
2338 assert(d1->bci() == d2->bci(), "not equal");
2339
2340 if (d1->locals() != nullptr) {
2341 assert(d1->locals() != nullptr && d2->locals() != nullptr, "not equal");
2342 assert(d1->locals()->length() == d2->locals()->length(), "not equal");
2343 for (int i = 0; i < d1->locals()->length(); i++) {
2344 assert_equal(d1->locals()->at(i), d2->locals()->at(i));
2345 }
2346 } else {
2347 assert(d1->locals() == nullptr && d2->locals() == nullptr, "not equal");
2348 }
2349
2350 if (d1->expressions() != nullptr) {
2351 assert(d1->expressions() != nullptr && d2->expressions() != nullptr, "not equal");
2352 assert(d1->expressions()->length() == d2->expressions()->length(), "not equal");
2353 for (int i = 0; i < d1->expressions()->length(); i++) {
2354 assert_equal(d1->expressions()->at(i), d2->expressions()->at(i));
2355 }
2356 } else {
2357 assert(d1->expressions() == nullptr && d2->expressions() == nullptr, "not equal");
2358 }
2359
2360 if (d1->monitors() != nullptr) {
2361 assert(d1->monitors() != nullptr && d2->monitors() != nullptr, "not equal");
2362 assert(d1->monitors()->length() == d2->monitors()->length(), "not equal");
2363 for (int i = 0; i < d1->monitors()->length(); i++) {
2364 assert_equal(d1->monitors()->at(i), d2->monitors()->at(i));
2365 }
2366 } else {
2367 assert(d1->monitors() == nullptr && d2->monitors() == nullptr, "not equal");
2368 }
2369
2370 if (d1->caller() != nullptr) {
2371 assert(d1->caller() != nullptr && d2->caller() != nullptr, "not equal");
2372 assert_equal(d1->caller(), d2->caller());
2373 } else {
2374 assert(d1->caller() == nullptr && d2->caller() == nullptr, "not equal");
2375 }
2376 }
2377
2378 static void check_stack_depth(CodeEmitInfo* info, int stack_end) {
2379 if (info->stack()->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) {
2380 Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->stack()->bci());
2381 switch (code) {
2382 case Bytecodes::_ifnull : // fall through
2383 case Bytecodes::_ifnonnull : // fall through
2384 case Bytecodes::_ifeq : // fall through
2385 case Bytecodes::_ifne : // fall through
2386 case Bytecodes::_iflt : // fall through
2387 case Bytecodes::_ifge : // fall through
2388 case Bytecodes::_ifgt : // fall through
2389 case Bytecodes::_ifle : // fall through
2390 case Bytecodes::_if_icmpeq : // fall through
2391 case Bytecodes::_if_icmpne : // fall through
2392 case Bytecodes::_if_icmplt : // fall through
2393 case Bytecodes::_if_icmpge : // fall through
2394 case Bytecodes::_if_icmpgt : // fall through
2395 case Bytecodes::_if_icmple : // fall through
2396 case Bytecodes::_if_acmpeq : // fall through
2397 case Bytecodes::_if_acmpne :
2398 assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode");
2399 break;
2400 default:
2401 break;
2402 }
2403 }
2404 }
2405
2406 #endif // ASSERT
2407
2408
2409 IntervalWalker* LinearScan::init_compute_oop_maps() {
2410 // setup lists of potential oops for walking
2411 Interval* oop_intervals;
2412 Interval* non_oop_intervals;
2413
2414 create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, nullptr);
2415
2416 // intervals that have no oops inside need not to be processed
2417 // to ensure a walking until the last instruction id, add a dummy interval
2418 // with a high operation id
2419 non_oop_intervals = new Interval(any_reg);
2420 non_oop_intervals->add_range(max_jint - 2, max_jint - 1);
2421
2422 return new IntervalWalker(this, oop_intervals, non_oop_intervals);
2423 }
2424
2425
2426 OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) {
2427 TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d", op->id()));
2428
2429 // walk before the current operation -> intervals that start at
2430 // the operation (= output operands of the operation) are not
2431 // included in the oop map
2432 iw->walk_before(op->id());
2433
2434 int frame_size = frame_map()->framesize();
2435 int arg_count = frame_map()->oop_map_arg_count();
2436 OopMap* map = new OopMap(frame_size, arg_count);
2437
2438 // Iterate through active intervals
2439 for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) {
2440 int assigned_reg = interval->assigned_reg();
2441
2442 assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise");
2443 assert(interval->assigned_regHi() == any_reg, "oop must be single word");
2444 assert(interval->reg_num() >= LIR_Opr::vreg_base, "fixed interval found");
2445
2446 // Check if this range covers the instruction. Intervals that
2447 // start or end at the current operation are not included in the
2448 // oop map, except in the case of patching moves. For patching
2449 // moves, any intervals which end at this instruction are included
2450 // in the oop map since we may safepoint while doing the patch
2451 // before we've consumed the inputs.
2452 if (op->is_patching() || op->id() < interval->current_to()) {
2453
2454 // caller-save registers must not be included into oop-maps at calls
2455 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");
2456
2457 VMReg name = vm_reg_for_interval(interval);
2458 set_oop(map, name);
2459
2460 // Spill optimization: when the stack value is guaranteed to be always correct,
2461 // then it must be added to the oop map even if the interval is currently in a register
2462 if (interval->always_in_memory() &&
2463 op->id() > interval->spill_definition_pos() &&
2464 interval->assigned_reg() != interval->canonical_spill_slot()) {
2465 assert(interval->spill_definition_pos() > 0, "position not set correctly");
2466 assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned");
2467 assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice");
2468
2469 set_oop(map, frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs));
2470 }
2471 }
2472 }
2473
2474 // add oops from lock stack
2475 assert(info->stack() != nullptr, "CodeEmitInfo must always have a stack");
2476 int locks_count = info->stack()->total_locks_size();
2477 for (int i = 0; i < locks_count; i++) {
2478 set_oop(map, frame_map()->monitor_object_regname(i));
2479 }
2480
2481 return map;
2482 }
2483
2484
2485 void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) {
2486 assert(visitor.info_count() > 0, "no oop map needed");
2487
2488 // compute oop_map only for first CodeEmitInfo
2489 // because it is (in most cases) equal for all other infos of the same operation
2490 CodeEmitInfo* first_info = visitor.info_at(0);
2491 OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call());
2492
2493 for (int i = 0; i < visitor.info_count(); i++) {
2494 CodeEmitInfo* info = visitor.info_at(i);
2495 OopMap* oop_map = first_oop_map;
2496
2497 // compute worst case interpreter size in case of a deoptimization
2498 _compilation->update_interpreter_frame_size(info->interpreter_frame_size());
2499
2500 if (info->stack()->locks_size() != first_info->stack()->locks_size()) {
2501 // this info has a different number of locks then the precomputed oop map
2502 // (possible for lock and unlock instructions) -> compute oop map with
2503 // correct lock information
2504 oop_map = compute_oop_map(iw, op, info, visitor.has_call());
2505 }
2506
2507 if (info->_oop_map == nullptr) {
2508 info->_oop_map = oop_map;
2509 } else {
2510 // a CodeEmitInfo can not be shared between different LIR-instructions
2511 // because interval splitting can occur anywhere between two instructions
2512 // and so the oop maps must be different
2513 // -> check if the already set oop_map is exactly the one calculated for this operation
2514 assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions");
2515 }
2516 }
2517 }
2518
2519
2520 // frequently used constants
2521 // Allocate them with new so they are never destroyed (otherwise, a
2522 // forced exit could destroy these objects while they are still in
2523 // use).
2524 ConstantOopWriteValue* LinearScan::_oop_null_scope_value = new (mtCompiler) ConstantOopWriteValue(nullptr);
2525 ConstantIntValue* LinearScan::_int_m1_scope_value = new (mtCompiler) ConstantIntValue(-1);
2526 ConstantIntValue* LinearScan::_int_0_scope_value = new (mtCompiler) ConstantIntValue((jint)0);
2527 ConstantIntValue* LinearScan::_int_1_scope_value = new (mtCompiler) ConstantIntValue(1);
2528 ConstantIntValue* LinearScan::_int_2_scope_value = new (mtCompiler) ConstantIntValue(2);
2529 LocationValue* _illegal_value = new (mtCompiler) LocationValue(Location());
2530
2531 void LinearScan::init_compute_debug_info() {
2532 // cache for frequently used scope values
2533 // (cpu registers and stack slots)
2534 int cache_size = (LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2;
2535 _scope_value_cache = ScopeValueArray(cache_size, cache_size, nullptr);
2536 }
2537
2538 MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) {
2539 Location loc;
2540 if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) {
2541 bailout("too large frame");
2542 }
2543 ScopeValue* object_scope_value = new LocationValue(loc);
2544
2545 if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) {
2546 bailout("too large frame");
2547 }
2548 return new MonitorValue(object_scope_value, loc);
2549 }
2550
2551 LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) {
2552 Location loc;
2553 if (!frame_map()->locations_for_slot(name, loc_type, &loc)) {
2554 bailout("too large frame");
2555 }
2556 return new LocationValue(loc);
2557 }
2558
2559
2560 int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) {
2561 assert(opr->is_constant(), "should not be called otherwise");
2562
2563 LIR_Const* c = opr->as_constant_ptr();
2564 BasicType t = c->type();
2565 switch (t) {
2566 case T_OBJECT: {
2567 jobject value = c->as_jobject();
2568 if (value == nullptr) {
2569 scope_values->append(_oop_null_scope_value);
2570 } else {
2571 scope_values->append(new ConstantOopWriteValue(c->as_jobject()));
2572 }
2573 return 1;
2574 }
2575
2576 case T_INT: // fall through
2577 case T_FLOAT: {
2578 int value = c->as_jint_bits();
2579 switch (value) {
2580 case -1: scope_values->append(_int_m1_scope_value); break;
2581 case 0: scope_values->append(_int_0_scope_value); break;
2582 case 1: scope_values->append(_int_1_scope_value); break;
2583 case 2: scope_values->append(_int_2_scope_value); break;
2584 default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break;
2585 }
2586 return 1;
2587 }
2588
2589 case T_LONG: // fall through
2590 case T_DOUBLE: {
2591 #ifdef _LP64
2592 scope_values->append(_int_0_scope_value);
2593 scope_values->append(new ConstantLongValue(c->as_jlong_bits()));
2594 #else
2595 if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) {
2596 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits()));
2597 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits()));
2598 } else {
2599 scope_values->append(new ConstantIntValue(c->as_jint_lo_bits()));
2600 scope_values->append(new ConstantIntValue(c->as_jint_hi_bits()));
2601 }
2602 #endif
2603 return 2;
2604 }
2605
2606 case T_ADDRESS: {
2607 #ifdef _LP64
2608 scope_values->append(new ConstantLongValue(c->as_jint()));
2609 #else
2610 scope_values->append(new ConstantIntValue(c->as_jint()));
2611 #endif
2612 return 1;
2613 }
2614
2615 default:
2616 ShouldNotReachHere();
2617 return -1;
2618 }
2619 }
2620
2621 int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) {
2622 if (opr->is_single_stack()) {
2623 int stack_idx = opr->single_stack_ix();
2624 bool is_oop = opr->is_oop_register();
2625 int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0);
2626
2627 ScopeValue* sv = _scope_value_cache.at(cache_idx);
2628 if (sv == nullptr) {
2629 Location::Type loc_type = is_oop ? Location::oop : Location::normal;
2630 sv = location_for_name(stack_idx, loc_type);
2631 _scope_value_cache.at_put(cache_idx, sv);
2632 }
2633
2634 // check if cached value is correct
2635 DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal)));
2636
2637 scope_values->append(sv);
2638 return 1;
2639
2640 } else if (opr->is_single_cpu()) {
2641 bool is_oop = opr->is_oop_register();
2642 int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0);
2643 Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long);
2644
2645 ScopeValue* sv = _scope_value_cache.at(cache_idx);
2646 if (sv == nullptr) {
2647 Location::Type loc_type = is_oop ? Location::oop : int_loc_type;
2648 VMReg rname = frame_map()->regname(opr);
2649 sv = new LocationValue(Location::new_reg_loc(loc_type, rname));
2650 _scope_value_cache.at_put(cache_idx, sv);
2651 }
2652
2653 // check if cached value is correct
2654 DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr)))));
2655
2656 scope_values->append(sv);
2657 return 1;
2658
2659 #ifdef X86
2660 } else if (opr->is_single_xmm()) {
2661 VMReg rname = opr->as_xmm_float_reg()->as_VMReg();
2662 LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname));
2663
2664 scope_values->append(sv);
2665 return 1;
2666 #endif
2667
2668 } else if (opr->is_single_fpu()) {
2669 #ifdef IA32
2670 // the exact location of fpu stack values is only known
2671 // during fpu stack allocation, so the stack allocator object
2672 // must be present
2673 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)");
2674 assert(_fpu_stack_allocator != nullptr, "must be present");
2675 opr = _fpu_stack_allocator->to_fpu_stack(opr);
2676 #elif defined(AMD64)
2677 assert(false, "FPU not used on x86-64");
2678 #endif
2679
2680 Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal;
2681 VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr());
2682 #ifndef __SOFTFP__
2683 #ifndef VM_LITTLE_ENDIAN
2684 // On S390 a (single precision) float value occupies only the high
2685 // word of the full double register. So when the double register is
2686 // stored to memory (e.g. by the RegisterSaver), then the float value
2687 // is found at offset 0. I.e. the code below is not needed on S390.
2688 #ifndef S390
2689 if (! float_saved_as_double) {
2690 // On big endian system, we may have an issue if float registers use only
2691 // the low half of the (same) double registers.
2692 // Both the float and the double could have the same regnr but would correspond
2693 // to two different addresses once saved.
2694
2695 // get next safely (no assertion checks)
2696 VMReg next = VMRegImpl::as_VMReg(1+rname->value());
2697 if (next->is_reg() &&
2698 (next->as_FloatRegister() == rname->as_FloatRegister())) {
2699 // the back-end does use the same numbering for the double and the float
2700 rname = next; // VMReg for the low bits, e.g. the real VMReg for the float
2701 }
2702 }
2703 #endif // !S390
2704 #endif
2705 #endif
2706 LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname));
2707
2708 scope_values->append(sv);
2709 return 1;
2710
2711 } else {
2712 // double-size operands
2713
2714 ScopeValue* first;
2715 ScopeValue* second;
2716
2717 if (opr->is_double_stack()) {
2718 #ifdef _LP64
2719 Location loc1;
2720 Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl;
2721 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, nullptr)) {
2722 bailout("too large frame");
2723 }
2724
2725 first = new LocationValue(loc1);
2726 second = _int_0_scope_value;
2727 #else
2728 Location loc1, loc2;
2729 if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) {
2730 bailout("too large frame");
2731 }
2732 first = new LocationValue(loc1);
2733 second = new LocationValue(loc2);
2734 #endif // _LP64
2735
2736 } else if (opr->is_double_cpu()) {
2737 #ifdef _LP64
2738 VMReg rname_first = opr->as_register_lo()->as_VMReg();
2739 first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first));
2740 second = _int_0_scope_value;
2741 #else
2742 VMReg rname_first = opr->as_register_lo()->as_VMReg();
2743 VMReg rname_second = opr->as_register_hi()->as_VMReg();
2744
2745 if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) {
2746 // lo/hi and swapped relative to first and second, so swap them
2747 VMReg tmp = rname_first;
2748 rname_first = rname_second;
2749 rname_second = tmp;
2750 }
2751
2752 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
2753 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
2754 #endif //_LP64
2755
2756
2757 #ifdef X86
2758 } else if (opr->is_double_xmm()) {
2759 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation");
2760 VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg();
2761 # ifdef _LP64
2762 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first));
2763 second = _int_0_scope_value;
2764 # else
2765 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
2766 // %%% This is probably a waste but we'll keep things as they were for now
2767 if (true) {
2768 VMReg rname_second = rname_first->next();
2769 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
2770 }
2771 # endif
2772 #endif
2773
2774 } else if (opr->is_double_fpu()) {
2775 // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of
2776 // the double as float registers in the native ordering. On X86,
2777 // fpu_regnrLo is a FPU stack slot whose VMReg represents
2778 // the low-order word of the double and fpu_regnrLo + 1 is the
2779 // name for the other half. *first and *second must represent the
2780 // least and most significant words, respectively.
2781
2782 #ifdef IA32
2783 // the exact location of fpu stack values is only known
2784 // during fpu stack allocation, so the stack allocator object
2785 // must be present
2786 assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)");
2787 assert(_fpu_stack_allocator != nullptr, "must be present");
2788 opr = _fpu_stack_allocator->to_fpu_stack(opr);
2789
2790 assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrLo is used)");
2791 #endif
2792 #ifdef AMD64
2793 assert(false, "FPU not used on x86-64");
2794 #endif
2795 #ifdef ARM32
2796 assert(opr->fpu_regnrHi() == opr->fpu_regnrLo() + 1, "assumed in calculation (only fpu_regnrLo is used)");
2797 #endif
2798
2799 #ifdef VM_LITTLE_ENDIAN
2800 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrLo());
2801 #else
2802 VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi());
2803 #endif
2804
2805 #ifdef _LP64
2806 first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first));
2807 second = _int_0_scope_value;
2808 #else
2809 first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first));
2810 // %%% This is probably a waste but we'll keep things as they were for now
2811 if (true) {
2812 VMReg rname_second = rname_first->next();
2813 second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second));
2814 }
2815 #endif
2816
2817 } else {
2818 ShouldNotReachHere();
2819 first = nullptr;
2820 second = nullptr;
2821 }
2822
2823 assert(first != nullptr && second != nullptr, "must be set");
2824 // The convention the interpreter uses is that the second local
2825 // holds the first raw word of the native double representation.
2826 // This is actually reasonable, since locals and stack arrays
2827 // grow downwards in all implementations.
2828 // (If, on some machine, the interpreter's Java locals or stack
2829 // were to grow upwards, the embedded doubles would be word-swapped.)
2830 scope_values->append(second);
2831 scope_values->append(first);
2832 return 2;
2833 }
2834 }
2835
2836
2837 int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) {
2838 if (value != nullptr) {
2839 LIR_Opr opr = value->operand();
2840 Constant* con = value->as_Constant();
2841
2842 assert(con == nullptr || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "assumption: Constant instructions have only constant operands (or illegal if constant is optimized away)");
2843 assert(con != nullptr || opr->is_virtual(), "assumption: non-Constant instructions have only virtual operands");
2844
2845 if (con != nullptr && !con->is_pinned() && !opr->is_constant()) {
2846 // Unpinned constants may have a virtual operand for a part of the lifetime
2847 // or may be illegal when it was optimized away,
2848 // so always use a constant operand
2849 opr = LIR_OprFact::value_type(con->type());
2850 }
2851 assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here");
2852
2853 if (opr->is_virtual()) {
2854 LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode;
2855
2856 BlockBegin* block = block_of_op_with_id(op_id);
2857 if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) {
2858 // generating debug information for the last instruction of a block.
2859 // if this instruction is a branch, spill moves are inserted before this branch
2860 // and so the wrong operand would be returned (spill moves at block boundaries are not
2861 // considered in the live ranges of intervals)
2862 // Solution: use the first op_id of the branch target block instead.
2863 if (block->lir()->instructions_list()->last()->as_OpBranch() != nullptr) {
2864 if (block->live_out().at(opr->vreg_number())) {
2865 op_id = block->sux_at(0)->first_lir_instruction_id();
2866 mode = LIR_OpVisitState::outputMode;
2867 }
2868 }
2869 }
2870
2871 // Get current location of operand
2872 // The operand must be live because debug information is considered when building the intervals
2873 // if the interval is not live, color_lir_opr will cause an assertion failure
2874 opr = color_lir_opr(opr, op_id, mode);
2875 assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls");
2876
2877 // Append to ScopeValue array
2878 return append_scope_value_for_operand(opr, scope_values);
2879
2880 } else {
2881 assert(value->as_Constant() != nullptr, "all other instructions have only virtual operands");
2882 assert(opr->is_constant(), "operand must be constant");
2883
2884 return append_scope_value_for_constant(opr, scope_values);
2885 }
2886 } else {
2887 // append a dummy value because real value not needed
2888 scope_values->append(_illegal_value);
2889 return 1;
2890 }
2891 }
2892
2893
2894 IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state) {
2895 IRScopeDebugInfo* caller_debug_info = nullptr;
2896
2897 ValueStack* caller_state = cur_state->caller_state();
2898 if (caller_state != nullptr) {
2899 // process recursively to compute outermost scope first
2900 caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state);
2901 }
2902
2903 // initialize these to null.
2904 // If we don't need deopt info or there are no locals, expressions or monitors,
2905 // then these get recorded as no information and avoids the allocation of 0 length arrays.
2906 GrowableArray<ScopeValue*>* locals = nullptr;
2907 GrowableArray<ScopeValue*>* expressions = nullptr;
2908 GrowableArray<MonitorValue*>* monitors = nullptr;
2909
2910 // describe local variable values
2911 int nof_locals = cur_state->locals_size();
2912 if (nof_locals > 0) {
2913 locals = new GrowableArray<ScopeValue*>(nof_locals);
2914
2915 int pos = 0;
2916 while (pos < nof_locals) {
2917 assert(pos < cur_state->locals_size(), "why not?");
2918
2919 Value local = cur_state->local_at(pos);
2920 pos += append_scope_value(op_id, local, locals);
2921
2922 assert(locals->length() == pos, "must match");
2923 }
2924 assert(locals->length() == nof_locals, "wrong number of locals");
2925 }
2926 assert(nof_locals == cur_scope->method()->max_locals(), "wrong number of locals");
2927 assert(nof_locals == cur_state->locals_size(), "wrong number of locals");
2928
2929 // describe expression stack
2930 int nof_stack = cur_state->stack_size();
2931 if (nof_stack > 0) {
2932 expressions = new GrowableArray<ScopeValue*>(nof_stack);
2933
2934 int pos = 0;
2935 while (pos < nof_stack) {
2936 Value expression = cur_state->stack_at(pos);
2937 pos += append_scope_value(op_id, expression, expressions);
2938
2939 assert(expressions->length() == pos, "must match");
2940 }
2941 assert(expressions->length() == cur_state->stack_size(), "wrong number of stack entries");
2942 }
2943
2944 // describe monitors
2945 int nof_locks = cur_state->locks_size();
2946 if (nof_locks > 0) {
2947 int lock_offset = cur_state->caller_state() != nullptr ? cur_state->caller_state()->total_locks_size() : 0;
2948 monitors = new GrowableArray<MonitorValue*>(nof_locks);
2949 for (int i = 0; i < nof_locks; i++) {
2950 monitors->append(location_for_monitor_index(lock_offset + i));
2951 }
2952 }
2953
2954 return new IRScopeDebugInfo(cur_scope, cur_state->bci(), locals, expressions, monitors, caller_debug_info);
2955 }
2956
2957
2958 void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) {
2959 TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d", op_id));
2960
2961 IRScope* innermost_scope = info->scope();
2962 ValueStack* innermost_state = info->stack();
2963
2964 assert(innermost_scope != nullptr && innermost_state != nullptr, "why is it missing?");
2965
2966 DEBUG_ONLY(check_stack_depth(info, innermost_state->stack_size()));
2967
2968 if (info->_scope_debug_info == nullptr) {
2969 // compute debug information
2970 info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state);
2971 } else {
2972 // debug information already set. Check that it is correct from the current point of view
2973 DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state)));
2974 }
2975 }
2976
2977
2978 void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) {
2979 LIR_OpVisitState visitor;
2980 int num_inst = instructions->length();
2981 bool has_dead = false;
2982
2983 for (int j = 0; j < num_inst; j++) {
2984 LIR_Op* op = instructions->at(j);
2985 if (op == nullptr) { // this can happen when spill-moves are removed in eliminate_spill_moves
2986 has_dead = true;
2987 continue;
2988 }
2989 int op_id = op->id();
2990
2991 // visit instruction to get list of operands
2992 visitor.visit(op);
2993
2994 // iterate all modes of the visitor and process all virtual operands
2995 for_each_visitor_mode(mode) {
2996 int n = visitor.opr_count(mode);
2997 for (int k = 0; k < n; k++) {
2998 LIR_Opr opr = visitor.opr_at(mode, k);
2999 if (opr->is_virtual_register()) {
3000 visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode));
3001 }
3002 }
3003 }
3004
3005 if (visitor.info_count() > 0) {
3006 // exception handling
3007 if (compilation()->has_exception_handlers()) {
3008 XHandlers* xhandlers = visitor.all_xhandler();
3009 int n = xhandlers->length();
3010 for (int k = 0; k < n; k++) {
3011 XHandler* handler = xhandlers->handler_at(k);
3012 if (handler->entry_code() != nullptr) {
3013 assign_reg_num(handler->entry_code()->instructions_list(), nullptr);
3014 }
3015 }
3016 } else {
3017 assert(visitor.all_xhandler()->length() == 0, "missed exception handler");
3018 }
3019
3020 // compute oop map
3021 assert(iw != nullptr, "needed for compute_oop_map");
3022 compute_oop_map(iw, visitor, op);
3023
3024 // compute debug information
3025 if (!use_fpu_stack_allocation()) {
3026 // compute debug information if fpu stack allocation is not needed.
3027 // when fpu stack allocation is needed, the debug information can not
3028 // be computed here because the exact location of fpu operands is not known
3029 // -> debug information is created inside the fpu stack allocator
3030 int n = visitor.info_count();
3031 for (int k = 0; k < n; k++) {
3032 compute_debug_info(visitor.info_at(k), op_id);
3033 }
3034 }
3035 }
3036
3037 #ifdef ASSERT
3038 // make sure we haven't made the op invalid.
3039 op->verify();
3040 #endif
3041
3042 // remove useless moves
3043 if (op->code() == lir_move) {
3044 assert(op->as_Op1() != nullptr, "move must be LIR_Op1");
3045 LIR_Op1* move = (LIR_Op1*)op;
3046 LIR_Opr src = move->in_opr();
3047 LIR_Opr dst = move->result_opr();
3048 if (dst == src ||
3049 (!dst->is_pointer() && !src->is_pointer() &&
3050 src->is_same_register(dst))) {
3051 instructions->at_put(j, nullptr);
3052 has_dead = true;
3053 }
3054 }
3055 }
3056
3057 if (has_dead) {
3058 // iterate all instructions of the block and remove all null-values.
3059 int insert_point = 0;
3060 for (int j = 0; j < num_inst; j++) {
3061 LIR_Op* op = instructions->at(j);
3062 if (op != nullptr) {
3063 if (insert_point != j) {
3064 instructions->at_put(insert_point, op);
3065 }
3066 insert_point++;
3067 }
3068 }
3069 instructions->trunc_to(insert_point);
3070 }
3071 }
3072
3073 void LinearScan::assign_reg_num() {
3074 TIME_LINEAR_SCAN(timer_assign_reg_num);
3075
3076 init_compute_debug_info();
3077 IntervalWalker* iw = init_compute_oop_maps();
3078
3079 int num_blocks = block_count();
3080 for (int i = 0; i < num_blocks; i++) {
3081 BlockBegin* block = block_at(i);
3082 assign_reg_num(block->lir()->instructions_list(), iw);
3083 }
3084 }
3085
3086
3087 void LinearScan::do_linear_scan() {
3088 NOT_PRODUCT(_total_timer.begin_method());
3089
3090 number_instructions();
3091
3092 NOT_PRODUCT(print_lir(1, "Before Register Allocation"));
3093
3094 compute_local_live_sets();
3095 compute_global_live_sets();
3096 CHECK_BAILOUT();
3097
3098 build_intervals();
3099 CHECK_BAILOUT();
3100 sort_intervals_before_allocation();
3101
3102 NOT_PRODUCT(print_intervals("Before Register Allocation"));
3103 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc));
3104
3105 allocate_registers();
3106 CHECK_BAILOUT();
3107
3108 resolve_data_flow();
3109 if (compilation()->has_exception_handlers()) {
3110 resolve_exception_handlers();
3111 }
3112 // fill in number of spill slots into frame_map
3113 propagate_spill_slots();
3114 CHECK_BAILOUT();
3115
3116 NOT_PRODUCT(print_intervals("After Register Allocation"));
3117 NOT_PRODUCT(print_lir(2, "LIR after register allocation:"));
3118
3119 sort_intervals_after_allocation();
3120
3121 DEBUG_ONLY(verify());
3122
3123 eliminate_spill_moves();
3124 assign_reg_num();
3125 CHECK_BAILOUT();
3126
3127 NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:"));
3128 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign));
3129
3130 { TIME_LINEAR_SCAN(timer_allocate_fpu_stack);
3131
3132 if (use_fpu_stack_allocation()) {
3133 allocate_fpu_stack(); // Only has effect on Intel
3134 NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:"));
3135 }
3136 }
3137
3138 #ifndef RISCV
3139 // Disable these optimizations on riscv temporarily, because it does not
3140 // work when the comparison operands are bound to branches or cmoves.
3141 { TIME_LINEAR_SCAN(timer_optimize_lir);
3142
3143 EdgeMoveOptimizer::optimize(ir()->code());
3144 ControlFlowOptimizer::optimize(ir()->code());
3145 // check that cfg is still correct after optimizations
3146 ir()->verify();
3147 }
3148 #endif
3149
3150 NOT_PRODUCT(print_lir(1, "Before Code Generation", false));
3151 NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final));
3152 NOT_PRODUCT(_total_timer.end_method(this));
3153 }
3154
3155
3156 // ********** Printing functions
3157
3158 #ifndef PRODUCT
3159
3160 void LinearScan::print_timers(double total) {
3161 _total_timer.print(total);
3162 }
3163
3164 void LinearScan::print_statistics() {
3165 _stat_before_alloc.print("before allocation");
3166 _stat_after_asign.print("after assignment of register");
3167 _stat_final.print("after optimization");
3168 }
3169
3170 void LinearScan::print_bitmap(BitMap& b) {
3171 for (unsigned int i = 0; i < b.size(); i++) {
3172 if (b.at(i)) tty->print("%d ", i);
3173 }
3174 tty->cr();
3175 }
3176
3177 void LinearScan::print_intervals(const char* label) {
3178 if (TraceLinearScanLevel >= 1) {
3179 int i;
3180 tty->cr();
3181 tty->print_cr("%s", label);
3182
3183 for (i = 0; i < interval_count(); i++) {
3184 Interval* interval = interval_at(i);
3185 if (interval != nullptr) {
3186 interval->print();
3187 }
3188 }
3189
3190 tty->cr();
3191 tty->print_cr("--- Basic Blocks ---");
3192 for (i = 0; i < block_count(); i++) {
3193 BlockBegin* block = block_at(i);
3194 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());
3195 }
3196 tty->cr();
3197 tty->cr();
3198 }
3199
3200 if (PrintCFGToFile) {
3201 CFGPrinter::print_intervals(&_intervals, label);
3202 }
3203 }
3204
3205 void LinearScan::print_lir(int level, const char* label, bool hir_valid) {
3206 if (TraceLinearScanLevel >= level) {
3207 tty->cr();
3208 tty->print_cr("%s", label);
3209 print_LIR(ir()->linear_scan_order());
3210 tty->cr();
3211 }
3212
3213 if (level == 1 && PrintCFGToFile) {
3214 CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true);
3215 }
3216 }
3217
3218 void LinearScan::print_reg_num(outputStream* out, int reg_num) {
3219 if (reg_num == -1) {
3220 out->print("[ANY]");
3221 return;
3222 } else if (reg_num >= LIR_Opr::vreg_base) {
3223 out->print("[VREG %d]", reg_num);
3224 return;
3225 }
3226
3227 LIR_Opr opr = get_operand(reg_num);
3228 assert(opr->is_valid(), "unknown register");
3229 opr->print(out);
3230 }
3231
3232 LIR_Opr LinearScan::get_operand(int reg_num) {
3233 LIR_Opr opr = LIR_OprFact::illegal();
3234
3235 #ifdef X86
3236 int last_xmm_reg = pd_last_xmm_reg;
3237 #ifdef _LP64
3238 if (UseAVX < 3) {
3239 last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1;
3240 }
3241 #endif
3242 #endif
3243 if (reg_num >= pd_first_cpu_reg && reg_num <= pd_last_cpu_reg) {
3244 opr = LIR_OprFact::single_cpu(reg_num);
3245 } else if (reg_num >= pd_first_fpu_reg && reg_num <= pd_last_fpu_reg) {
3246 opr = LIR_OprFact::single_fpu(reg_num - pd_first_fpu_reg);
3247 #ifdef X86
3248 } else if (reg_num >= pd_first_xmm_reg && reg_num <= last_xmm_reg) {
3249 opr = LIR_OprFact::single_xmm(reg_num - pd_first_xmm_reg);
3250 #endif
3251 } else {
3252 // reg_num == -1 or a virtual register, return the illegal operand
3253 }
3254 return opr;
3255 }
3256
3257 Interval* LinearScan::find_interval_at(int reg_num) const {
3258 if (reg_num < 0 || reg_num >= _intervals.length()) {
3259 return nullptr;
3260 }
3261 return interval_at(reg_num);
3262 }
3263
3264 #endif // PRODUCT
3265
3266
3267 // ********** verification functions for allocation
3268 // (check that all intervals have a correct register and that no registers are overwritten)
3269 #ifdef ASSERT
3270
3271 void LinearScan::verify() {
3272 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************"));
3273 verify_intervals();
3274
3275 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************"));
3276 verify_no_oops_in_fixed_intervals();
3277
3278 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries"));
3279 verify_constants();
3280
3281 TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************"));
3282 verify_registers();
3283
3284 TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************"));
3285 }
3286
3287 void LinearScan::verify_intervals() {
3288 int len = interval_count();
3289 bool has_error = false;
3290
3291 for (int i = 0; i < len; i++) {
3292 Interval* i1 = interval_at(i);
3293 if (i1 == nullptr) continue;
3294
3295 i1->check_split_children();
3296
3297 if (i1->reg_num() != i) {
3298 tty->print_cr("Interval %d is on position %d in list", i1->reg_num(), i); i1->print(); tty->cr();
3299 has_error = true;
3300 }
3301
3302 if (i1->reg_num() >= LIR_Opr::vreg_base && i1->type() == T_ILLEGAL) {
3303 tty->print_cr("Interval %d has no type assigned", i1->reg_num()); i1->print(); tty->cr();
3304 has_error = true;
3305 }
3306
3307 if (i1->assigned_reg() == any_reg) {
3308 tty->print_cr("Interval %d has no register assigned", i1->reg_num()); i1->print(); tty->cr();
3309 has_error = true;
3310 }
3311
3312 if (i1->assigned_reg() == i1->assigned_regHi()) {
3313 tty->print_cr("Interval %d: low and high register equal", i1->reg_num()); i1->print(); tty->cr();
3314 has_error = true;
3315 }
3316
3317 if (!is_processed_reg_num(i1->assigned_reg())) {
3318 tty->print_cr("Can not have an Interval for an ignored register"); i1->print(); tty->cr();
3319 has_error = true;
3320 }
3321
3322 // special intervals that are created in MoveResolver
3323 // -> ignore them because the range information has no meaning there
3324 if (i1->from() == 1 && i1->to() == 2) continue;
3325
3326 if (i1->first() == Range::end()) {
3327 tty->print_cr("Interval %d has no Range", i1->reg_num()); i1->print(); tty->cr();
3328 has_error = true;
3329 }
3330
3331 for (Range* r = i1->first(); r != Range::end(); r = r->next()) {
3332 if (r->from() >= r->to()) {
3333 tty->print_cr("Interval %d has zero length range", i1->reg_num()); i1->print(); tty->cr();
3334 has_error = true;
3335 }
3336 }
3337
3338 for (int j = i + 1; j < len; j++) {
3339 Interval* i2 = interval_at(j);
3340 if (i2 == nullptr || (i2->from() == 1 && i2->to() == 2)) continue;
3341
3342 int r1 = i1->assigned_reg();
3343 int r1Hi = i1->assigned_regHi();
3344 int r2 = i2->assigned_reg();
3345 int r2Hi = i2->assigned_regHi();
3346 if ((r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi))) && i1->intersects(i2)) {
3347 tty->print_cr("Intervals %d and %d overlap and have the same register assigned", i1->reg_num(), i2->reg_num());
3348 i1->print(); tty->cr();
3349 i2->print(); tty->cr();
3350 has_error = true;
3351 }
3352 }
3353 }
3354
3355 assert(has_error == false, "register allocation invalid");
3356 }
3357
3358
3359 void LinearScan::verify_no_oops_in_fixed_intervals() {
3360 Interval* fixed_intervals;
3361 Interval* other_intervals;
3362 create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, nullptr);
3363
3364 // to ensure a walking until the last instruction id, add a dummy interval
3365 // with a high operation id
3366 other_intervals = new Interval(any_reg);
3367 other_intervals->add_range(max_jint - 2, max_jint - 1);
3368 IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals);
3369
3370 LIR_OpVisitState visitor;
3371 for (int i = 0; i < block_count(); i++) {
3372 BlockBegin* block = block_at(i);
3373
3374 LIR_OpList* instructions = block->lir()->instructions_list();
3375
3376 for (int j = 0; j < instructions->length(); j++) {
3377 LIR_Op* op = instructions->at(j);
3378 int op_id = op->id();
3379
3380 visitor.visit(op);
3381
3382 if (visitor.info_count() > 0) {
3383 iw->walk_before(op->id());
3384 bool check_live = true;
3385 if (op->code() == lir_move) {
3386 LIR_Op1* move = (LIR_Op1*)op;
3387 check_live = (move->patch_code() == lir_patch_none);
3388 }
3389 LIR_OpBranch* branch = op->as_OpBranch();
3390 if (branch != nullptr && branch->stub() != nullptr && branch->stub()->is_exception_throw_stub()) {
3391 // Don't bother checking the stub in this case since the
3392 // exception stub will never return to normal control flow.
3393 check_live = false;
3394 }
3395
3396 // Make sure none of the fixed registers is live across an
3397 // oopmap since we can't handle that correctly.
3398 if (check_live) {
3399 for (Interval* interval = iw->active_first(fixedKind);
3400 interval != Interval::end();
3401 interval = interval->next()) {
3402 if (interval->current_to() > op->id() + 1) {
3403 // This interval is live out of this op so make sure
3404 // that this interval represents some value that's
3405 // referenced by this op either as an input or output.
3406 bool ok = false;
3407 for_each_visitor_mode(mode) {
3408 int n = visitor.opr_count(mode);
3409 for (int k = 0; k < n; k++) {
3410 LIR_Opr opr = visitor.opr_at(mode, k);
3411 if (opr->is_fixed_cpu()) {
3412 if (interval_at(reg_num(opr)) == interval) {
3413 ok = true;
3414 break;
3415 }
3416 int hi = reg_numHi(opr);
3417 if (hi != -1 && interval_at(hi) == interval) {
3418 ok = true;
3419 break;
3420 }
3421 }
3422 }
3423 }
3424 assert(ok, "fixed intervals should never be live across an oopmap point");
3425 }
3426 }
3427 }
3428 }
3429
3430 // oop-maps at calls do not contain registers, so check is not needed
3431 if (!visitor.has_call()) {
3432
3433 for_each_visitor_mode(mode) {
3434 int n = visitor.opr_count(mode);
3435 for (int k = 0; k < n; k++) {
3436 LIR_Opr opr = visitor.opr_at(mode, k);
3437
3438 if (opr->is_fixed_cpu() && opr->is_oop()) {
3439 // operand is a non-virtual cpu register and contains an oop
3440 TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand "); opr->print(); tty->cr());
3441
3442 Interval* interval = interval_at(reg_num(opr));
3443 assert(interval != nullptr, "no interval");
3444
3445 if (mode == LIR_OpVisitState::inputMode) {
3446 if (interval->to() >= op_id + 1) {
3447 assert(interval->to() < op_id + 2 ||
3448 interval->has_hole_between(op_id, op_id + 2),
3449 "oop input operand live after instruction");
3450 }
3451 } else if (mode == LIR_OpVisitState::outputMode) {
3452 if (interval->from() <= op_id - 1) {
3453 assert(interval->has_hole_between(op_id - 1, op_id),
3454 "oop input operand live after instruction");
3455 }
3456 }
3457 }
3458 }
3459 }
3460 }
3461 }
3462 }
3463 }
3464
3465
3466 void LinearScan::verify_constants() {
3467 int num_regs = num_virtual_regs();
3468 int size = live_set_size();
3469 int num_blocks = block_count();
3470
3471 for (int i = 0; i < num_blocks; i++) {
3472 BlockBegin* block = block_at(i);
3473 ResourceBitMap& live_at_edge = block->live_in();
3474
3475 // visit all registers where the live_at_edge bit is set
3476 auto visitor = [&](BitMap::idx_t index) {
3477 int r = static_cast<int>(index);
3478 TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d", r, block->block_id()));
3479
3480 Value value = gen()->instruction_for_vreg(r);
3481
3482 assert(value != nullptr, "all intervals live across block boundaries must have Value");
3483 assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand");
3484 assert(value->operand()->vreg_number() == r, "register number must match");
3485 // TKR assert(value->as_Constant() == nullptr || value->is_pinned(), "only pinned constants can be alive across block boundaries");
3486 };
3487 live_at_edge.iterate(visitor, 0, size);
3488 }
3489 }
3490
3491
3492 class RegisterVerifier: public StackObj {
3493 private:
3494 LinearScan* _allocator;
3495 BlockList _work_list; // all blocks that must be processed
3496 IntervalsList _saved_states; // saved information of previous check
3497
3498 // simplified access to methods of LinearScan
3499 Compilation* compilation() const { return _allocator->compilation(); }
3500 Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); }
3501 int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); }
3502
3503 // currently, only registers are processed
3504 int state_size() { return LinearScan::nof_regs; }
3505
3506 // accessors
3507 IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); }
3508 void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); }
3509 void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); }
3510
3511 // helper functions
3512 IntervalList* copy(IntervalList* input_state);
3513 void state_put(IntervalList* input_state, int reg, Interval* interval);
3514 bool check_state(IntervalList* input_state, int reg, Interval* interval);
3515
3516 void process_block(BlockBegin* block);
3517 void process_xhandler(XHandler* xhandler, IntervalList* input_state);
3518 void process_successor(BlockBegin* block, IntervalList* input_state);
3519 void process_operations(LIR_List* ops, IntervalList* input_state);
3520
3521 public:
3522 RegisterVerifier(LinearScan* allocator)
3523 : _allocator(allocator)
3524 , _work_list(16)
3525 , _saved_states(BlockBegin::number_of_blocks(), BlockBegin::number_of_blocks(), nullptr)
3526 { }
3527
3528 void verify(BlockBegin* start);
3529 };
3530
3531
3532 // entry function from LinearScan that starts the verification
3533 void LinearScan::verify_registers() {
3534 RegisterVerifier verifier(this);
3535 verifier.verify(block_at(0));
3536 }
3537
3538
3539 void RegisterVerifier::verify(BlockBegin* start) {
3540 // setup input registers (method arguments) for first block
3541 int input_state_len = state_size();
3542 IntervalList* input_state = new IntervalList(input_state_len, input_state_len, nullptr);
3543 CallingConvention* args = compilation()->frame_map()->incoming_arguments();
3544 for (int n = 0; n < args->length(); n++) {
3545 LIR_Opr opr = args->at(n);
3546 if (opr->is_register()) {
3547 Interval* interval = interval_at(reg_num(opr));
3548
3549 if (interval->assigned_reg() < state_size()) {
3550 input_state->at_put(interval->assigned_reg(), interval);
3551 }
3552 if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) {
3553 input_state->at_put(interval->assigned_regHi(), interval);
3554 }
3555 }
3556 }
3557
3558 set_state_for_block(start, input_state);
3559 add_to_work_list(start);
3560
3561 // main loop for verification
3562 do {
3563 BlockBegin* block = _work_list.at(0);
3564 _work_list.remove_at(0);
3565
3566 process_block(block);
3567 } while (!_work_list.is_empty());
3568 }
3569
3570 void RegisterVerifier::process_block(BlockBegin* block) {
3571 TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d", block->block_id()));
3572
3573 // must copy state because it is modified
3574 IntervalList* input_state = copy(state_for_block(block));
3575
3576 if (TraceLinearScanLevel >= 4) {
3577 tty->print_cr("Input-State of intervals:");
3578 tty->print(" ");
3579 for (int i = 0; i < state_size(); i++) {
3580 if (input_state->at(i) != nullptr) {
3581 tty->print(" %4d", input_state->at(i)->reg_num());
3582 } else {
3583 tty->print(" __");
3584 }
3585 }
3586 tty->cr();
3587 tty->cr();
3588 }
3589
3590 // process all operations of the block
3591 process_operations(block->lir(), input_state);
3592
3593 // iterate all successors
3594 for (int i = 0; i < block->number_of_sux(); i++) {
3595 process_successor(block->sux_at(i), input_state);
3596 }
3597 }
3598
3599 void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) {
3600 TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d", xhandler->entry_block()->block_id()));
3601
3602 // must copy state because it is modified
3603 input_state = copy(input_state);
3604
3605 if (xhandler->entry_code() != nullptr) {
3606 process_operations(xhandler->entry_code(), input_state);
3607 }
3608 process_successor(xhandler->entry_block(), input_state);
3609 }
3610
3611 void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) {
3612 IntervalList* saved_state = state_for_block(block);
3613
3614 if (saved_state != nullptr) {
3615 // this block was already processed before.
3616 // check if new input_state is consistent with saved_state
3617
3618 bool saved_state_correct = true;
3619 for (int i = 0; i < state_size(); i++) {
3620 if (input_state->at(i) != saved_state->at(i)) {
3621 // current input_state and previous saved_state assume a different
3622 // interval in this register -> assume that this register is invalid
3623 if (saved_state->at(i) != nullptr) {
3624 // invalidate old calculation only if it assumed that
3625 // register was valid. when the register was already invalid,
3626 // then the old calculation was correct.
3627 saved_state_correct = false;
3628 saved_state->at_put(i, nullptr);
3629
3630 TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d", block->block_id(), i));
3631 }
3632 }
3633 }
3634
3635 if (saved_state_correct) {
3636 // already processed block with correct input_state
3637 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct", block->block_id()));
3638 } else {
3639 // must re-visit this block
3640 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed", block->block_id()));
3641 add_to_work_list(block);
3642 }
3643
3644 } else {
3645 // block was not processed before, so set initial input_state
3646 TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit", block->block_id()));
3647
3648 set_state_for_block(block, copy(input_state));
3649 add_to_work_list(block);
3650 }
3651 }
3652
3653
3654 IntervalList* RegisterVerifier::copy(IntervalList* input_state) {
3655 IntervalList* copy_state = new IntervalList(input_state->length());
3656 copy_state->appendAll(input_state);
3657 return copy_state;
3658 }
3659
3660 void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) {
3661 if (reg != LinearScan::any_reg && reg < state_size()) {
3662 if (interval != nullptr) {
3663 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d", reg, interval->reg_num()));
3664 } else if (input_state->at(reg) != nullptr) {
3665 TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = null", reg));
3666 }
3667
3668 input_state->at_put(reg, interval);
3669 }
3670 }
3671
3672 bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) {
3673 if (reg != LinearScan::any_reg && reg < state_size()) {
3674 if (input_state->at(reg) != interval) {
3675 tty->print_cr("!! Error in register allocation: register %d does not contain interval %d", reg, interval->reg_num());
3676 return true;
3677 }
3678 }
3679 return false;
3680 }
3681
3682 void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) {
3683 // visit all instructions of the block
3684 LIR_OpVisitState visitor;
3685 bool has_error = false;
3686
3687 for (int i = 0; i < ops->length(); i++) {
3688 LIR_Op* op = ops->at(i);
3689 visitor.visit(op);
3690
3691 TRACE_LINEAR_SCAN(4, op->print_on(tty));
3692
3693 // check if input operands are correct
3694 int j;
3695 int n = visitor.opr_count(LIR_OpVisitState::inputMode);
3696 for (j = 0; j < n; j++) {
3697 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, 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::inputMode);
3702 }
3703
3704 has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent());
3705 has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent());
3706
3707 // When an operand is marked with is_last_use, then the fpu stack allocator
3708 // removes the register from the fpu stack -> the register contains no value
3709 if (opr->is_last_use()) {
3710 state_put(input_state, interval->assigned_reg(), nullptr);
3711 state_put(input_state, interval->assigned_regHi(), nullptr);
3712 }
3713 }
3714 }
3715
3716 // invalidate all caller save registers at calls
3717 if (visitor.has_call()) {
3718 for (j = 0; j < FrameMap::nof_caller_save_cpu_regs(); j++) {
3719 state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), nullptr);
3720 }
3721 for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) {
3722 state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), nullptr);
3723 }
3724
3725 #ifdef X86
3726 int num_caller_save_xmm_regs = FrameMap::get_num_caller_save_xmms();
3727 for (j = 0; j < num_caller_save_xmm_regs; j++) {
3728 state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), nullptr);
3729 }
3730 #endif
3731 }
3732
3733 // process xhandler before output and temp operands
3734 XHandlers* xhandlers = visitor.all_xhandler();
3735 n = xhandlers->length();
3736 for (int k = 0; k < n; k++) {
3737 process_xhandler(xhandlers->handler_at(k), input_state);
3738 }
3739
3740 // set temp operands (some operations use temp operands also as output operands, so can't set them null)
3741 n = visitor.opr_count(LIR_OpVisitState::tempMode);
3742 for (j = 0; j < n; j++) {
3743 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j);
3744 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
3745 Interval* interval = interval_at(reg_num(opr));
3746 if (op->id() != -1) {
3747 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode);
3748 }
3749
3750 state_put(input_state, interval->assigned_reg(), interval->split_parent());
3751 state_put(input_state, interval->assigned_regHi(), interval->split_parent());
3752 }
3753 }
3754
3755 // set output operands
3756 n = visitor.opr_count(LIR_OpVisitState::outputMode);
3757 for (j = 0; j < n; j++) {
3758 LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j);
3759 if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) {
3760 Interval* interval = interval_at(reg_num(opr));
3761 if (op->id() != -1) {
3762 interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode);
3763 }
3764
3765 state_put(input_state, interval->assigned_reg(), interval->split_parent());
3766 state_put(input_state, interval->assigned_regHi(), interval->split_parent());
3767 }
3768 }
3769 }
3770 assert(has_error == false, "Error in register allocation");
3771 }
3772
3773 #endif // ASSERT
3774
3775
3776
3777 // **** Implementation of MoveResolver ******************************
3778
3779 MoveResolver::MoveResolver(LinearScan* allocator) :
3780 _allocator(allocator),
3781 _insert_list(nullptr),
3782 _insert_idx(-1),
3783 _insertion_buffer(),
3784 _mapping_from(8),
3785 _mapping_from_opr(8),
3786 _mapping_to(8),
3787 _multiple_reads_allowed(false)
3788 {
3789 for (int i = 0; i < LinearScan::nof_regs; i++) {
3790 _register_blocked[i] = 0;
3791 }
3792 DEBUG_ONLY(check_empty());
3793 }
3794
3795
3796 #ifdef ASSERT
3797
3798 void MoveResolver::check_empty() {
3799 assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing");
3800 for (int i = 0; i < LinearScan::nof_regs; i++) {
3801 assert(register_blocked(i) == 0, "register map must be empty before and after processing");
3802 }
3803 assert(_multiple_reads_allowed == false, "must have default value");
3804 }
3805
3806 void MoveResolver::verify_before_resolve() {
3807 assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal");
3808 assert(_mapping_from.length() == _mapping_to.length(), "length must be equal");
3809 assert(_insert_list != nullptr && _insert_idx != -1, "insert position not set");
3810
3811 int i, j;
3812 if (!_multiple_reads_allowed) {
3813 for (i = 0; i < _mapping_from.length(); i++) {
3814 for (j = i + 1; j < _mapping_from.length(); j++) {
3815 assert(_mapping_from.at(i) == nullptr || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice");
3816 }
3817 }
3818 }
3819
3820 for (i = 0; i < _mapping_to.length(); i++) {
3821 for (j = i + 1; j < _mapping_to.length(); j++) {
3822 assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice");
3823 }
3824 }
3825
3826
3827 ResourceBitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills());
3828 if (!_multiple_reads_allowed) {
3829 for (i = 0; i < _mapping_from.length(); i++) {
3830 Interval* it = _mapping_from.at(i);
3831 if (it != nullptr) {
3832 assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice");
3833 used_regs.set_bit(it->assigned_reg());
3834
3835 if (it->assigned_regHi() != LinearScan::any_reg) {
3836 assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice");
3837 used_regs.set_bit(it->assigned_regHi());
3838 }
3839 }
3840 }
3841 }
3842
3843 used_regs.clear();
3844 for (i = 0; i < _mapping_to.length(); i++) {
3845 Interval* it = _mapping_to.at(i);
3846 assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice");
3847 used_regs.set_bit(it->assigned_reg());
3848
3849 if (it->assigned_regHi() != LinearScan::any_reg) {
3850 assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice");
3851 used_regs.set_bit(it->assigned_regHi());
3852 }
3853 }
3854
3855 used_regs.clear();
3856 for (i = 0; i < _mapping_from.length(); i++) {
3857 Interval* it = _mapping_from.at(i);
3858 if (it != nullptr && it->assigned_reg() >= LinearScan::nof_regs) {
3859 used_regs.set_bit(it->assigned_reg());
3860 }
3861 }
3862 for (i = 0; i < _mapping_to.length(); i++) {
3863 Interval* it = _mapping_to.at(i);
3864 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");
3865 }
3866 }
3867
3868 #endif // ASSERT
3869
3870
3871 // mark assigned_reg and assigned_regHi of the interval as blocked
3872 void MoveResolver::block_registers(Interval* it) {
3873 int reg = it->assigned_reg();
3874 if (reg < LinearScan::nof_regs) {
3875 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used");
3876 set_register_blocked(reg, 1);
3877 }
3878 reg = it->assigned_regHi();
3879 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
3880 assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used");
3881 set_register_blocked(reg, 1);
3882 }
3883 }
3884
3885 // mark assigned_reg and assigned_regHi of the interval as unblocked
3886 void MoveResolver::unblock_registers(Interval* it) {
3887 int reg = it->assigned_reg();
3888 if (reg < LinearScan::nof_regs) {
3889 assert(register_blocked(reg) > 0, "register already marked as unused");
3890 set_register_blocked(reg, -1);
3891 }
3892 reg = it->assigned_regHi();
3893 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
3894 assert(register_blocked(reg) > 0, "register already marked as unused");
3895 set_register_blocked(reg, -1);
3896 }
3897 }
3898
3899 // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from)
3900 bool MoveResolver::save_to_process_move(Interval* from, Interval* to) {
3901 int from_reg = -1;
3902 int from_regHi = -1;
3903 if (from != nullptr) {
3904 from_reg = from->assigned_reg();
3905 from_regHi = from->assigned_regHi();
3906 }
3907
3908 int reg = to->assigned_reg();
3909 if (reg < LinearScan::nof_regs) {
3910 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) {
3911 return false;
3912 }
3913 }
3914 reg = to->assigned_regHi();
3915 if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) {
3916 if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) {
3917 return false;
3918 }
3919 }
3920
3921 return true;
3922 }
3923
3924
3925 void MoveResolver::create_insertion_buffer(LIR_List* list) {
3926 assert(!_insertion_buffer.initialized(), "overwriting existing buffer");
3927 _insertion_buffer.init(list);
3928 }
3929
3930 void MoveResolver::append_insertion_buffer() {
3931 if (_insertion_buffer.initialized()) {
3932 _insertion_buffer.lir_list()->append(&_insertion_buffer);
3933 }
3934 assert(!_insertion_buffer.initialized(), "must be uninitialized now");
3935
3936 _insert_list = nullptr;
3937 _insert_idx = -1;
3938 }
3939
3940 void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) {
3941 assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal");
3942 assert(from_interval->type() == to_interval->type(), "move between different types");
3943 assert(_insert_list != nullptr && _insert_idx != -1, "must setup insert position first");
3944 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer");
3945
3946 LIR_Opr from_opr = get_virtual_register(from_interval);
3947 LIR_Opr to_opr = get_virtual_register(to_interval);
3948
3949 if (!_multiple_reads_allowed) {
3950 // the last_use flag is an optimization for FPU stack allocation. When the same
3951 // input interval is used in more than one move, then it is too difficult to determine
3952 // if this move is really the last use.
3953 from_opr = from_opr->make_last_use();
3954 }
3955 _insertion_buffer.move(_insert_idx, from_opr, to_opr);
3956
3957 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()));
3958 }
3959
3960 void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) {
3961 assert(from_opr->type() == to_interval->type(), "move between different types");
3962 assert(_insert_list != nullptr && _insert_idx != -1, "must setup insert position first");
3963 assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer");
3964
3965 LIR_Opr to_opr = get_virtual_register(to_interval);
3966 _insertion_buffer.move(_insert_idx, from_opr, to_opr);
3967
3968 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()));
3969 }
3970
3971 LIR_Opr MoveResolver::get_virtual_register(Interval* interval) {
3972 // Add a little fudge factor for the bailout since the bailout is only checked periodically. This allows us to hand out
3973 // a few extra registers before we really run out which helps to avoid to trip over assertions.
3974 int reg_num = interval->reg_num();
3975 if (reg_num + 20 >= LIR_Opr::vreg_max) {
3976 _allocator->bailout("out of virtual registers in linear scan");
3977 if (reg_num + 2 >= LIR_Opr::vreg_max) {
3978 // Wrap it around and continue until bailout really happens to avoid hitting assertions.
3979 reg_num = LIR_Opr::vreg_base;
3980 }
3981 }
3982 LIR_Opr vreg = LIR_OprFact::virtual_register(reg_num, interval->type());
3983 assert(vreg != LIR_OprFact::illegal(), "ran out of virtual registers");
3984 return vreg;
3985 }
3986
3987 void MoveResolver::resolve_mappings() {
3988 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: resolving mappings for Block B%d, index %d", _insert_list->block() != nullptr ? _insert_list->block()->block_id() : -1, _insert_idx));
3989 DEBUG_ONLY(verify_before_resolve());
3990
3991 // Block all registers that are used as input operands of a move.
3992 // When a register is blocked, no move to this register is emitted.
3993 // This is necessary for detecting cycles in moves.
3994 int i;
3995 for (i = _mapping_from.length() - 1; i >= 0; i--) {
3996 Interval* from_interval = _mapping_from.at(i);
3997 if (from_interval != nullptr) {
3998 block_registers(from_interval);
3999 }
4000 }
4001
4002 int spill_candidate = -1;
4003 while (_mapping_from.length() > 0) {
4004 bool processed_interval = false;
4005
4006 for (i = _mapping_from.length() - 1; i >= 0; i--) {
4007 Interval* from_interval = _mapping_from.at(i);
4008 Interval* to_interval = _mapping_to.at(i);
4009
4010 if (save_to_process_move(from_interval, to_interval)) {
4011 // this interval can be processed because target is free
4012 if (from_interval != nullptr) {
4013 insert_move(from_interval, to_interval);
4014 unblock_registers(from_interval);
4015 } else {
4016 insert_move(_mapping_from_opr.at(i), to_interval);
4017 }
4018 _mapping_from.remove_at(i);
4019 _mapping_from_opr.remove_at(i);
4020 _mapping_to.remove_at(i);
4021
4022 processed_interval = true;
4023 } else if (from_interval != nullptr && from_interval->assigned_reg() < LinearScan::nof_regs) {
4024 // this interval cannot be processed now because target is not free
4025 // it starts in a register, so it is a possible candidate for spilling
4026 spill_candidate = i;
4027 }
4028 }
4029
4030 if (!processed_interval) {
4031 // no move could be processed because there is a cycle in the move list
4032 // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory
4033 guarantee(spill_candidate != -1, "no interval in register for spilling found");
4034
4035 // create a new spill interval and assign a stack slot to it
4036 Interval* from_interval = _mapping_from.at(spill_candidate);
4037 Interval* spill_interval = new Interval(-1);
4038 spill_interval->set_type(from_interval->type());
4039
4040 // add a dummy range because real position is difficult to calculate
4041 // Note: this range is a special case when the integrity of the allocation is checked
4042 spill_interval->add_range(1, 2);
4043
4044 // do not allocate a new spill slot for temporary interval, but
4045 // use spill slot assigned to from_interval. Otherwise moves from
4046 // one stack slot to another can happen (not allowed by LIR_Assembler
4047 int spill_slot = from_interval->canonical_spill_slot();
4048 if (spill_slot < 0) {
4049 spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2);
4050 from_interval->set_canonical_spill_slot(spill_slot);
4051 }
4052 spill_interval->assign_reg(spill_slot);
4053 allocator()->append_interval(spill_interval);
4054
4055 TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling", spill_interval->reg_num()));
4056
4057 // insert a move from register to stack and update the mapping
4058 insert_move(from_interval, spill_interval);
4059 _mapping_from.at_put(spill_candidate, spill_interval);
4060 unblock_registers(from_interval);
4061 }
4062 }
4063
4064 // reset to default value
4065 _multiple_reads_allowed = false;
4066
4067 // check that all intervals have been processed
4068 DEBUG_ONLY(check_empty());
4069 }
4070
4071
4072 void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) {
4073 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: setting insert position to Block B%d, index %d", insert_list->block() != nullptr ? insert_list->block()->block_id() : -1, insert_idx));
4074 assert(_insert_list == nullptr && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set");
4075
4076 create_insertion_buffer(insert_list);
4077 _insert_list = insert_list;
4078 _insert_idx = insert_idx;
4079 }
4080
4081 void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) {
4082 TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: moving insert position to Block B%d, index %d", insert_list->block() != nullptr ? insert_list->block()->block_id() : -1, insert_idx));
4083
4084 if (_insert_list != nullptr && (insert_list != _insert_list || insert_idx != _insert_idx)) {
4085 // insert position changed -> resolve current mappings
4086 resolve_mappings();
4087 }
4088
4089 if (insert_list != _insert_list) {
4090 // block changed -> append insertion_buffer because it is
4091 // bound to a specific block and create a new insertion_buffer
4092 append_insertion_buffer();
4093 create_insertion_buffer(insert_list);
4094 }
4095
4096 _insert_list = insert_list;
4097 _insert_idx = insert_idx;
4098 }
4099
4100 void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) {
4101 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()));
4102
4103 _mapping_from.append(from_interval);
4104 _mapping_from_opr.append(LIR_OprFact::illegalOpr);
4105 _mapping_to.append(to_interval);
4106 }
4107
4108
4109 void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) {
4110 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()));
4111 assert(from_opr->is_constant(), "only for constants");
4112
4113 _mapping_from.append(nullptr);
4114 _mapping_from_opr.append(from_opr);
4115 _mapping_to.append(to_interval);
4116 }
4117
4118 void MoveResolver::resolve_and_append_moves() {
4119 if (has_mappings()) {
4120 resolve_mappings();
4121 }
4122 append_insertion_buffer();
4123 }
4124
4125
4126
4127 // **** Implementation of Range *************************************
4128
4129 Range::Range(int from, int to, Range* next) :
4130 _from(from),
4131 _to(to),
4132 _next(next)
4133 {
4134 }
4135
4136 // initialize sentinel
4137 Range* Range::_end = nullptr;
4138 void Range::initialize() {
4139 assert(_end == nullptr, "Range initialized more than once");
4140 alignas(Range) static uint8_t end_storage[sizeof(Range)];
4141 _end = ::new(static_cast<void*>(end_storage)) Range(max_jint, max_jint, nullptr);
4142 }
4143
4144 int Range::intersects_at(Range* r2) const {
4145 const Range* r1 = this;
4146
4147 assert(r1 != nullptr && r2 != nullptr, "null ranges not allowed");
4148 assert(r1 != _end && r2 != _end, "empty ranges not allowed");
4149
4150 do {
4151 if (r1->from() < r2->from()) {
4152 if (r1->to() <= r2->from()) {
4153 r1 = r1->next(); if (r1 == _end) return -1;
4154 } else {
4155 return r2->from();
4156 }
4157 } else if (r2->from() < r1->from()) {
4158 if (r2->to() <= r1->from()) {
4159 r2 = r2->next(); if (r2 == _end) return -1;
4160 } else {
4161 return r1->from();
4162 }
4163 } else { // r1->from() == r2->from()
4164 if (r1->from() == r1->to()) {
4165 r1 = r1->next(); if (r1 == _end) return -1;
4166 } else if (r2->from() == r2->to()) {
4167 r2 = r2->next(); if (r2 == _end) return -1;
4168 } else {
4169 return r1->from();
4170 }
4171 }
4172 } while (true);
4173 }
4174
4175 #ifndef PRODUCT
4176 void Range::print(outputStream* out) const {
4177 out->print("[%d, %d[ ", _from, _to);
4178 }
4179 #endif
4180
4181
4182
4183 // **** Implementation of Interval **********************************
4184
4185 // initialize sentinel
4186 Interval* Interval::_end = nullptr;
4187 void Interval::initialize() {
4188 Range::initialize();
4189 assert(_end == nullptr, "Interval initialized more than once");
4190 alignas(Interval) static uint8_t end_storage[sizeof(Interval)];
4191 _end = ::new(static_cast<void*>(end_storage)) Interval(-1);
4192 }
4193
4194 Interval::Interval(int reg_num) :
4195 _reg_num(reg_num),
4196 _type(T_ILLEGAL),
4197 _first(Range::end()),
4198 _use_pos_and_kinds(12),
4199 _current(Range::end()),
4200 _next(_end),
4201 _state(invalidState),
4202 _assigned_reg(LinearScan::any_reg),
4203 _assigned_regHi(LinearScan::any_reg),
4204 _cached_to(-1),
4205 _cached_opr(LIR_OprFact::illegalOpr),
4206 _cached_vm_reg(VMRegImpl::Bad()),
4207 _split_children(nullptr),
4208 _canonical_spill_slot(-1),
4209 _insert_move_when_activated(false),
4210 _spill_state(noDefinitionFound),
4211 _spill_definition_pos(-1),
4212 _register_hint(nullptr)
4213 {
4214 _split_parent = this;
4215 _current_split_child = this;
4216 }
4217
4218 int Interval::calc_to() {
4219 assert(_first != Range::end(), "interval has no range");
4220
4221 Range* r = _first;
4222 while (r->next() != Range::end()) {
4223 r = r->next();
4224 }
4225 return r->to();
4226 }
4227
4228
4229 #ifdef ASSERT
4230 // consistency check of split-children
4231 void Interval::check_split_children() {
4232 if (_split_children != nullptr && _split_children->length() > 0) {
4233 assert(is_split_parent(), "only split parents can have children");
4234
4235 for (int i = 0; i < _split_children->length(); i++) {
4236 Interval* i1 = _split_children->at(i);
4237
4238 assert(i1->split_parent() == this, "not a split child of this interval");
4239 assert(i1->type() == type(), "must be equal for all split children");
4240 assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children");
4241
4242 for (int j = i + 1; j < _split_children->length(); j++) {
4243 Interval* i2 = _split_children->at(j);
4244
4245 assert(i1->reg_num() != i2->reg_num(), "same register number");
4246
4247 if (i1->from() < i2->from()) {
4248 assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping");
4249 } else {
4250 assert(i2->from() < i1->from(), "intervals start at same op_id");
4251 assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping");
4252 }
4253 }
4254 }
4255 }
4256 }
4257 #endif // ASSERT
4258
4259 Interval* Interval::register_hint(bool search_split_child) const {
4260 if (!search_split_child) {
4261 return _register_hint;
4262 }
4263
4264 if (_register_hint != nullptr) {
4265 assert(_register_hint->is_split_parent(), "only split parents are valid hint registers");
4266
4267 if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) {
4268 return _register_hint;
4269
4270 } else if (_register_hint->_split_children != nullptr && _register_hint->_split_children->length() > 0) {
4271 // search the first split child that has a register assigned
4272 int len = _register_hint->_split_children->length();
4273 for (int i = 0; i < len; i++) {
4274 Interval* cur = _register_hint->_split_children->at(i);
4275
4276 if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) {
4277 return cur;
4278 }
4279 }
4280 }
4281 }
4282
4283 // no hint interval found that has a register assigned
4284 return nullptr;
4285 }
4286
4287
4288 Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) {
4289 assert(is_split_parent(), "can only be called for split parents");
4290 assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)");
4291
4292 Interval* result;
4293 if (_split_children == nullptr || _split_children->length() == 0) {
4294 result = this;
4295 } else {
4296 result = nullptr;
4297 int len = _split_children->length();
4298
4299 // in outputMode, the end of the interval (op_id == cur->to()) is not valid
4300 int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1);
4301
4302 int i;
4303 for (i = 0; i < len; i++) {
4304 Interval* cur = _split_children->at(i);
4305 if (cur->from() <= op_id && op_id < cur->to() + to_offset) {
4306 if (i > 0) {
4307 // exchange current split child to start of list (faster access for next call)
4308 _split_children->at_put(i, _split_children->at(0));
4309 _split_children->at_put(0, cur);
4310 }
4311
4312 // interval found
4313 result = cur;
4314 break;
4315 }
4316 }
4317
4318 #ifdef ASSERT
4319 for (i = 0; i < len; i++) {
4320 Interval* tmp = _split_children->at(i);
4321 if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) {
4322 tty->print_cr("two valid result intervals found for op_id %d: %d and %d", op_id, result->reg_num(), tmp->reg_num());
4323 result->print();
4324 tmp->print();
4325 assert(false, "two valid result intervals found");
4326 }
4327 }
4328 #endif
4329 }
4330
4331 assert(result != nullptr, "no matching interval found");
4332 assert(result->covers(op_id, mode), "op_id not covered by interval");
4333
4334 return result;
4335 }
4336
4337
4338 // returns the last split child that ends before the given op_id
4339 Interval* Interval::split_child_before_op_id(int op_id) {
4340 assert(op_id >= 0, "invalid op_id");
4341
4342 Interval* parent = split_parent();
4343 Interval* result = nullptr;
4344
4345 assert(parent->_split_children != nullptr, "no split children available");
4346 int len = parent->_split_children->length();
4347 assert(len > 0, "no split children available");
4348
4349 for (int i = len - 1; i >= 0; i--) {
4350 Interval* cur = parent->_split_children->at(i);
4351 if (cur->to() <= op_id && (result == nullptr || result->to() < cur->to())) {
4352 result = cur;
4353 }
4354 }
4355
4356 assert(result != nullptr, "no split child found");
4357 return result;
4358 }
4359
4360
4361 // Note: use positions are sorted descending -> first use has highest index
4362 int Interval::first_usage(IntervalUseKind min_use_kind) const {
4363 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4364
4365 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4366 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) {
4367 return _use_pos_and_kinds.at(i);
4368 }
4369 }
4370 return max_jint;
4371 }
4372
4373 int Interval::next_usage(IntervalUseKind min_use_kind, int from) const {
4374 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4375
4376 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4377 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) {
4378 return _use_pos_and_kinds.at(i);
4379 }
4380 }
4381 return max_jint;
4382 }
4383
4384 int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const {
4385 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4386
4387 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4388 if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) {
4389 return _use_pos_and_kinds.at(i);
4390 }
4391 }
4392 return max_jint;
4393 }
4394
4395 int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const {
4396 assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals");
4397
4398 int prev = 0;
4399 for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4400 if (_use_pos_and_kinds.at(i) > from) {
4401 return prev;
4402 }
4403 if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) {
4404 prev = _use_pos_and_kinds.at(i);
4405 }
4406 }
4407 return prev;
4408 }
4409
4410 void Interval::add_use_pos(int pos, IntervalUseKind use_kind) {
4411 assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range");
4412
4413 // do not add use positions for precolored intervals because
4414 // they are never used
4415 if (use_kind != noUse && reg_num() >= LIR_Opr::vreg_base) {
4416 #ifdef ASSERT
4417 assert(_use_pos_and_kinds.length() % 2 == 0, "must be");
4418 for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) {
4419 assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position");
4420 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4421 if (i > 0) {
4422 assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending");
4423 }
4424 }
4425 #endif
4426
4427 // Note: add_use is called in descending order, so list gets sorted
4428 // automatically by just appending new use positions
4429 int len = _use_pos_and_kinds.length();
4430 if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) {
4431 _use_pos_and_kinds.append(pos);
4432 _use_pos_and_kinds.append(use_kind);
4433 } else if (_use_pos_and_kinds.at(len - 1) < use_kind) {
4434 assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly");
4435 _use_pos_and_kinds.at_put(len - 1, use_kind);
4436 }
4437 }
4438 }
4439
4440 void Interval::add_range(int from, int to) {
4441 assert(from < to, "invalid range");
4442 assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval");
4443 assert(from <= first()->to(), "not inserting at begin of interval");
4444
4445 if (first()->from() <= to) {
4446 // join intersecting ranges
4447 first()->set_from(MIN2(from, first()->from()));
4448 first()->set_to (MAX2(to, first()->to()));
4449 } else {
4450 // insert new range
4451 _first = new Range(from, to, first());
4452 }
4453 }
4454
4455 Interval* Interval::new_split_child() {
4456 // allocate new interval
4457 Interval* result = new Interval(-1);
4458 result->set_type(type());
4459
4460 Interval* parent = split_parent();
4461 result->_split_parent = parent;
4462 result->set_register_hint(parent);
4463
4464 // insert new interval in children-list of parent
4465 if (parent->_split_children == nullptr) {
4466 assert(is_split_parent(), "list must be initialized at first split");
4467
4468 parent->_split_children = new IntervalList(4);
4469 parent->_split_children->append(this);
4470 }
4471 parent->_split_children->append(result);
4472
4473 return result;
4474 }
4475
4476 // split this interval at the specified position and return
4477 // the remainder as a new interval.
4478 //
4479 // when an interval is split, a bi-directional link is established between the original interval
4480 // (the split parent) and the intervals that are split off this interval (the split children)
4481 // When a split child is split again, the new created interval is also a direct child
4482 // of the original parent (there is no tree of split children stored, but a flat list)
4483 // All split children are spilled to the same stack slot (stored in _canonical_spill_slot)
4484 //
4485 // Note: The new interval has no valid reg_num
4486 Interval* Interval::split(int split_pos) {
4487 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals");
4488
4489 // allocate new interval
4490 Interval* result = new_split_child();
4491
4492 // split the ranges
4493 Range* prev = nullptr;
4494 Range* cur = _first;
4495 while (cur != Range::end() && cur->to() <= split_pos) {
4496 prev = cur;
4497 cur = cur->next();
4498 }
4499 assert(cur != Range::end(), "split interval after end of last range");
4500
4501 if (cur->from() < split_pos) {
4502 result->_first = new Range(split_pos, cur->to(), cur->next());
4503 cur->set_to(split_pos);
4504 cur->set_next(Range::end());
4505
4506 } else {
4507 assert(prev != nullptr, "split before start of first range");
4508 result->_first = cur;
4509 prev->set_next(Range::end());
4510 }
4511 result->_current = result->_first;
4512 _cached_to = -1; // clear cached value
4513
4514 // split list of use positions
4515 int total_len = _use_pos_and_kinds.length();
4516 int start_idx = total_len - 2;
4517 while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) {
4518 start_idx -= 2;
4519 }
4520
4521 intStack new_use_pos_and_kinds(total_len - start_idx);
4522 int i;
4523 for (i = start_idx + 2; i < total_len; i++) {
4524 new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i));
4525 }
4526
4527 _use_pos_and_kinds.trunc_to(start_idx + 2);
4528 result->_use_pos_and_kinds = _use_pos_and_kinds;
4529 _use_pos_and_kinds = new_use_pos_and_kinds;
4530
4531 #ifdef ASSERT
4532 assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos");
4533 assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos");
4534 assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries");
4535
4536 for (i = 0; i < _use_pos_and_kinds.length(); i += 2) {
4537 assert(_use_pos_and_kinds.at(i) < split_pos, "must be");
4538 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4539 }
4540 for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) {
4541 assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be");
4542 assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4543 }
4544 #endif
4545
4546 return result;
4547 }
4548
4549 // split this interval at the specified position and return
4550 // the head as a new interval (the original interval is the tail)
4551 //
4552 // Currently, only the first range can be split, and the new interval
4553 // must not have split positions
4554 Interval* Interval::split_from_start(int split_pos) {
4555 assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals");
4556 assert(split_pos > from() && split_pos < to(), "can only split inside interval");
4557 assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range");
4558 assert(first_usage(noUse) > split_pos, "can not split when use positions are present");
4559
4560 // allocate new interval
4561 Interval* result = new_split_child();
4562
4563 // the new created interval has only one range (checked by assertion above),
4564 // so the splitting of the ranges is very simple
4565 result->add_range(_first->from(), split_pos);
4566
4567 if (split_pos == _first->to()) {
4568 assert(_first->next() != Range::end(), "must not be at end");
4569 _first = _first->next();
4570 } else {
4571 _first->set_from(split_pos);
4572 }
4573
4574 return result;
4575 }
4576
4577
4578 // returns true if the op_id is inside the interval
4579 bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const {
4580 Range* cur = _first;
4581
4582 while (cur != Range::end() && cur->to() < op_id) {
4583 cur = cur->next();
4584 }
4585 if (cur != Range::end()) {
4586 assert(cur->to() != cur->next()->from(), "ranges not separated");
4587
4588 if (mode == LIR_OpVisitState::outputMode) {
4589 return cur->from() <= op_id && op_id < cur->to();
4590 } else {
4591 return cur->from() <= op_id && op_id <= cur->to();
4592 }
4593 }
4594 return false;
4595 }
4596
4597 // returns true if the interval has any hole between hole_from and hole_to
4598 // (even if the hole has only the length 1)
4599 bool Interval::has_hole_between(int hole_from, int hole_to) {
4600 assert(hole_from < hole_to, "check");
4601 assert(from() <= hole_from && hole_to <= to(), "index out of interval");
4602
4603 Range* cur = _first;
4604 while (cur != Range::end()) {
4605 assert(cur->to() < cur->next()->from(), "no space between ranges");
4606
4607 // hole-range starts before this range -> hole
4608 if (hole_from < cur->from()) {
4609 return true;
4610
4611 // hole-range completely inside this range -> no hole
4612 } else if (hole_to <= cur->to()) {
4613 return false;
4614
4615 // overlapping of hole-range with this range -> hole
4616 } else if (hole_from <= cur->to()) {
4617 return true;
4618 }
4619
4620 cur = cur->next();
4621 }
4622
4623 return false;
4624 }
4625
4626 // Check if there is an intersection with any of the split children of 'interval'
4627 bool Interval::intersects_any_children_of(Interval* interval) const {
4628 if (interval->_split_children != nullptr) {
4629 for (int i = 0; i < interval->_split_children->length(); i++) {
4630 if (intersects(interval->_split_children->at(i))) {
4631 return true;
4632 }
4633 }
4634 }
4635 return false;
4636 }
4637
4638
4639 #ifndef PRODUCT
4640 void Interval::print_on(outputStream* out, bool is_cfg_printer) const {
4641 const char* SpillState2Name[] = { "no definition", "no spill store", "one spill store", "store at definition", "start in memory", "no optimization" };
4642 const char* UseKind2Name[] = { "N", "L", "S", "M" };
4643
4644 const char* type_name;
4645 if (reg_num() < LIR_Opr::vreg_base) {
4646 type_name = "fixed";
4647 } else {
4648 type_name = type2name(type());
4649 }
4650 out->print("%d %s ", reg_num(), type_name);
4651
4652 if (is_cfg_printer) {
4653 // Special version for compatibility with C1 Visualizer.
4654 LIR_Opr opr = LinearScan::get_operand(reg_num());
4655 if (opr->is_valid()) {
4656 out->print("\"");
4657 opr->print(out);
4658 out->print("\" ");
4659 }
4660 } else {
4661 // Improved output for normal debugging.
4662 if (reg_num() < LIR_Opr::vreg_base) {
4663 LinearScan::print_reg_num(out, assigned_reg());
4664 } else if (assigned_reg() != -1 && (LinearScan::num_physical_regs(type()) == 1 || assigned_regHi() != -1)) {
4665 LinearScan::calc_operand_for_interval(this)->print(out);
4666 } else {
4667 // Virtual register that has no assigned register yet.
4668 out->print("[ANY]");
4669 }
4670 out->print(" ");
4671 }
4672 out->print("%d %d ", split_parent()->reg_num(), (register_hint(false) != nullptr ? register_hint(false)->reg_num() : -1));
4673
4674 // print ranges
4675 Range* cur = _first;
4676 while (cur != Range::end()) {
4677 cur->print(out);
4678 cur = cur->next();
4679 assert(cur != nullptr, "range list not closed with range sentinel");
4680 }
4681
4682 // print use positions
4683 int prev = 0;
4684 assert(_use_pos_and_kinds.length() % 2 == 0, "must be");
4685 for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) {
4686 assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind");
4687 assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted");
4688
4689 out->print("%d %s ", _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]);
4690 prev = _use_pos_and_kinds.at(i);
4691 }
4692
4693 out->print(" \"%s\"", SpillState2Name[spill_state()]);
4694 out->cr();
4695 }
4696
4697 void Interval::print_parent() const {
4698 if (_split_parent != this) {
4699 _split_parent->print_on(tty);
4700 } else {
4701 tty->print_cr("Parent: this");
4702 }
4703 }
4704
4705 void Interval::print_children() const {
4706 if (_split_children == nullptr) {
4707 tty->print_cr("Children: []");
4708 } else {
4709 tty->print_cr("Children:");
4710 for (int i = 0; i < _split_children->length(); i++) {
4711 tty->print("%d: ", i);
4712 _split_children->at(i)->print_on(tty);
4713 }
4714 }
4715 }
4716 #endif // NOT PRODUCT
4717
4718
4719
4720
4721 // **** Implementation of IntervalWalker ****************************
4722
4723 IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first)
4724 : _compilation(allocator->compilation())
4725 , _allocator(allocator)
4726 {
4727 _unhandled_first[fixedKind] = unhandled_fixed_first;
4728 _unhandled_first[anyKind] = unhandled_any_first;
4729 _active_first[fixedKind] = Interval::end();
4730 _inactive_first[fixedKind] = Interval::end();
4731 _active_first[anyKind] = Interval::end();
4732 _inactive_first[anyKind] = Interval::end();
4733 _current_position = -1;
4734 _current = nullptr;
4735 next_interval();
4736 }
4737
4738
4739 // append interval in order of current range from()
4740 void IntervalWalker::append_sorted(Interval** list, Interval* interval) {
4741 Interval* prev = nullptr;
4742 Interval* cur = *list;
4743 while (cur->current_from() < interval->current_from()) {
4744 prev = cur; cur = cur->next();
4745 }
4746 if (prev == nullptr) {
4747 *list = interval;
4748 } else {
4749 prev->set_next(interval);
4750 }
4751 interval->set_next(cur);
4752 }
4753
4754 void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) {
4755 assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position");
4756
4757 Interval* prev = nullptr;
4758 Interval* cur = *list;
4759 while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) {
4760 prev = cur; cur = cur->next();
4761 }
4762 if (prev == nullptr) {
4763 *list = interval;
4764 } else {
4765 prev->set_next(interval);
4766 }
4767 interval->set_next(cur);
4768 }
4769
4770
4771 inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) {
4772 while (*list != Interval::end() && *list != i) {
4773 list = (*list)->next_addr();
4774 }
4775 if (*list != Interval::end()) {
4776 assert(*list == i, "check");
4777 *list = (*list)->next();
4778 return true;
4779 } else {
4780 return false;
4781 }
4782 }
4783
4784 void IntervalWalker::remove_from_list(Interval* i) {
4785 bool deleted;
4786
4787 if (i->state() == activeState) {
4788 deleted = remove_from_list(active_first_addr(anyKind), i);
4789 } else {
4790 assert(i->state() == inactiveState, "invalid state");
4791 deleted = remove_from_list(inactive_first_addr(anyKind), i);
4792 }
4793
4794 assert(deleted, "interval has not been found in list");
4795 }
4796
4797
4798 void IntervalWalker::walk_to(IntervalState state, int from) {
4799 assert (state == activeState || state == inactiveState, "wrong state");
4800 for_each_interval_kind(kind) {
4801 Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind);
4802 Interval* next = *prev;
4803 while (next->current_from() <= from) {
4804 Interval* cur = next;
4805 next = cur->next();
4806
4807 bool range_has_changed = false;
4808 while (cur->current_to() <= from) {
4809 cur->next_range();
4810 range_has_changed = true;
4811 }
4812
4813 // also handle move from inactive list to active list
4814 range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from);
4815
4816 if (range_has_changed) {
4817 // remove cur from list
4818 *prev = next;
4819 if (cur->current_at_end()) {
4820 // move to handled state (not maintained as a list)
4821 cur->set_state(handledState);
4822 DEBUG_ONLY(interval_moved(cur, kind, state, handledState);)
4823 } else if (cur->current_from() <= from){
4824 // sort into active list
4825 append_sorted(active_first_addr(kind), cur);
4826 cur->set_state(activeState);
4827 if (*prev == cur) {
4828 assert(state == activeState, "check");
4829 prev = cur->next_addr();
4830 }
4831 DEBUG_ONLY(interval_moved(cur, kind, state, activeState);)
4832 } else {
4833 // sort into inactive list
4834 append_sorted(inactive_first_addr(kind), cur);
4835 cur->set_state(inactiveState);
4836 if (*prev == cur) {
4837 assert(state == inactiveState, "check");
4838 prev = cur->next_addr();
4839 }
4840 DEBUG_ONLY(interval_moved(cur, kind, state, inactiveState);)
4841 }
4842 } else {
4843 prev = cur->next_addr();
4844 continue;
4845 }
4846 }
4847 }
4848 }
4849
4850
4851 void IntervalWalker::next_interval() {
4852 IntervalKind kind;
4853 Interval* any = _unhandled_first[anyKind];
4854 Interval* fixed = _unhandled_first[fixedKind];
4855
4856 if (any != Interval::end()) {
4857 // intervals may start at same position -> prefer fixed interval
4858 kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind;
4859
4860 assert((kind == fixedKind && fixed->from() <= any->from()) ||
4861 (kind == anyKind && any->from() <= fixed->from()), "wrong interval!!!");
4862 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");
4863
4864 } else if (fixed != Interval::end()) {
4865 kind = fixedKind;
4866 } else {
4867 _current = nullptr; return;
4868 }
4869 _current_kind = kind;
4870 _current = _unhandled_first[kind];
4871 _unhandled_first[kind] = _current->next();
4872 _current->set_next(Interval::end());
4873 _current->rewind_range();
4874 }
4875
4876
4877 void IntervalWalker::walk_to(int lir_op_id) {
4878 assert(_current_position <= lir_op_id, "can not walk backwards");
4879 while (current() != nullptr) {
4880 bool is_active = current()->from() <= lir_op_id;
4881 int id = is_active ? current()->from() : lir_op_id;
4882
4883 TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************", id); })
4884
4885 // set _current_position prior to call of walk_to
4886 _current_position = id;
4887
4888 // call walk_to even if _current_position == id
4889 walk_to(activeState, id);
4890 walk_to(inactiveState, id);
4891
4892 if (is_active) {
4893 current()->set_state(activeState);
4894 if (activate_current()) {
4895 append_sorted(active_first_addr(current_kind()), current());
4896 DEBUG_ONLY(interval_moved(current(), current_kind(), unhandledState, activeState);)
4897 }
4898
4899 next_interval();
4900 } else {
4901 return;
4902 }
4903 }
4904 }
4905
4906 #ifdef ASSERT
4907 void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) {
4908 if (TraceLinearScanLevel >= 4) {
4909 #define print_state(state) \
4910 switch(state) {\
4911 case unhandledState: tty->print("unhandled"); break;\
4912 case activeState: tty->print("active"); break;\
4913 case inactiveState: tty->print("inactive"); break;\
4914 case handledState: tty->print("handled"); break;\
4915 default: ShouldNotReachHere(); \
4916 }
4917
4918 print_state(from); tty->print(" to "); print_state(to);
4919 tty->fill_to(23);
4920 interval->print();
4921
4922 #undef print_state
4923 }
4924 }
4925 #endif // ASSERT
4926
4927 // **** Implementation of LinearScanWalker **************************
4928
4929 LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first)
4930 : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first)
4931 , _move_resolver(allocator)
4932 {
4933 for (int i = 0; i < LinearScan::nof_regs; i++) {
4934 _spill_intervals[i] = new IntervalList(2);
4935 }
4936 }
4937
4938
4939 inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) {
4940 for (int i = _first_reg; i <= _last_reg; i++) {
4941 _use_pos[i] = max_jint;
4942
4943 if (!only_process_use_pos) {
4944 _block_pos[i] = max_jint;
4945 _spill_intervals[i]->clear();
4946 }
4947 }
4948 }
4949
4950 inline void LinearScanWalker::exclude_from_use(int reg) {
4951 assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)");
4952 if (reg >= _first_reg && reg <= _last_reg) {
4953 _use_pos[reg] = 0;
4954 }
4955 }
4956 inline void LinearScanWalker::exclude_from_use(Interval* i) {
4957 assert(i->assigned_reg() != any_reg, "interval has no register assigned");
4958
4959 exclude_from_use(i->assigned_reg());
4960 exclude_from_use(i->assigned_regHi());
4961 }
4962
4963 inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) {
4964 assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0");
4965
4966 if (reg >= _first_reg && reg <= _last_reg) {
4967 if (_use_pos[reg] > use_pos) {
4968 _use_pos[reg] = use_pos;
4969 }
4970 if (!only_process_use_pos) {
4971 _spill_intervals[reg]->append(i);
4972 }
4973 }
4974 }
4975 inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) {
4976 assert(i->assigned_reg() != any_reg, "interval has no register assigned");
4977 if (use_pos != -1) {
4978 set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos);
4979 set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos);
4980 }
4981 }
4982
4983 inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) {
4984 if (reg >= _first_reg && reg <= _last_reg) {
4985 if (_block_pos[reg] > block_pos) {
4986 _block_pos[reg] = block_pos;
4987 }
4988 if (_use_pos[reg] > block_pos) {
4989 _use_pos[reg] = block_pos;
4990 }
4991 }
4992 }
4993 inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) {
4994 assert(i->assigned_reg() != any_reg, "interval has no register assigned");
4995 if (block_pos != -1) {
4996 set_block_pos(i->assigned_reg(), i, block_pos);
4997 set_block_pos(i->assigned_regHi(), i, block_pos);
4998 }
4999 }
5000
5001
5002 void LinearScanWalker::free_exclude_active_fixed() {
5003 Interval* list = active_first(fixedKind);
5004 while (list != Interval::end()) {
5005 assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned");
5006 exclude_from_use(list);
5007 list = list->next();
5008 }
5009 }
5010
5011 void LinearScanWalker::free_exclude_active_any() {
5012 Interval* list = active_first(anyKind);
5013 while (list != Interval::end()) {
5014 exclude_from_use(list);
5015 list = list->next();
5016 }
5017 }
5018
5019 void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) {
5020 Interval* list = inactive_first(fixedKind);
5021 while (list != Interval::end()) {
5022 if (cur->to() <= list->current_from()) {
5023 assert(list->current_intersects_at(cur) == -1, "must not intersect");
5024 set_use_pos(list, list->current_from(), true);
5025 } else {
5026 set_use_pos(list, list->current_intersects_at(cur), true);
5027 }
5028 list = list->next();
5029 }
5030 }
5031
5032 void LinearScanWalker::free_collect_inactive_any(Interval* cur) {
5033 Interval* list = inactive_first(anyKind);
5034 while (list != Interval::end()) {
5035 set_use_pos(list, list->current_intersects_at(cur), true);
5036 list = list->next();
5037 }
5038 }
5039
5040 void LinearScanWalker::spill_exclude_active_fixed() {
5041 Interval* list = active_first(fixedKind);
5042 while (list != Interval::end()) {
5043 exclude_from_use(list);
5044 list = list->next();
5045 }
5046 }
5047
5048 void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) {
5049 Interval* list = inactive_first(fixedKind);
5050 while (list != Interval::end()) {
5051 if (cur->to() > list->current_from()) {
5052 set_block_pos(list, list->current_intersects_at(cur));
5053 } else {
5054 assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect");
5055 }
5056
5057 list = list->next();
5058 }
5059 }
5060
5061 void LinearScanWalker::spill_collect_active_any() {
5062 Interval* list = active_first(anyKind);
5063 while (list != Interval::end()) {
5064 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false);
5065 list = list->next();
5066 }
5067 }
5068
5069 void LinearScanWalker::spill_collect_inactive_any(Interval* cur) {
5070 Interval* list = inactive_first(anyKind);
5071 while (list != Interval::end()) {
5072 if (list->current_intersects(cur)) {
5073 set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false);
5074 }
5075 list = list->next();
5076 }
5077 }
5078
5079
5080 void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) {
5081 // output all moves here. When source and target are equal, the move is
5082 // optimized away later in assign_reg_nums
5083
5084 op_id = (op_id + 1) & ~1;
5085 BlockBegin* op_block = allocator()->block_of_op_with_id(op_id);
5086 assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary");
5087
5088 // calculate index of instruction inside instruction list of current block
5089 // the minimal index (for a block with no spill moves) can be calculated because the
5090 // numbering of instructions is known.
5091 // When the block already contains spill moves, the index must be increased until the
5092 // correct index is reached.
5093 LIR_OpList* list = op_block->lir()->instructions_list();
5094 int index = (op_id - list->at(0)->id()) / 2;
5095 assert(list->at(index)->id() <= op_id, "error in calculation");
5096
5097 while (list->at(index)->id() != op_id) {
5098 index++;
5099 assert(0 <= index && index < list->length(), "index out of bounds");
5100 }
5101 assert(1 <= index && index < list->length(), "index out of bounds");
5102 assert(list->at(index)->id() == op_id, "error in calculation");
5103
5104 // insert new instruction before instruction at position index
5105 _move_resolver.move_insert_position(op_block->lir(), index - 1);
5106 _move_resolver.add_mapping(src_it, dst_it);
5107 }
5108
5109
5110 int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) {
5111 int from_block_nr = min_block->linear_scan_number();
5112 int to_block_nr = max_block->linear_scan_number();
5113
5114 assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range");
5115 assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range");
5116 assert(from_block_nr < to_block_nr, "must cross block boundary");
5117
5118 // Try to split at end of max_block. If this would be after
5119 // max_split_pos, then use the begin of max_block
5120 int optimal_split_pos = max_block->last_lir_instruction_id() + 2;
5121 if (optimal_split_pos > max_split_pos) {
5122 optimal_split_pos = max_block->first_lir_instruction_id();
5123 }
5124
5125 int min_loop_depth = max_block->loop_depth();
5126 for (int i = to_block_nr - 1; i >= from_block_nr; i--) {
5127 BlockBegin* cur = block_at(i);
5128
5129 if (cur->loop_depth() < min_loop_depth) {
5130 // block with lower loop-depth found -> split at the end of this block
5131 min_loop_depth = cur->loop_depth();
5132 optimal_split_pos = cur->last_lir_instruction_id() + 2;
5133 }
5134 }
5135 assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary");
5136
5137 return optimal_split_pos;
5138 }
5139
5140
5141 int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) {
5142 int optimal_split_pos = -1;
5143 if (min_split_pos == max_split_pos) {
5144 // trivial case, no optimization of split position possible
5145 TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible"));
5146 optimal_split_pos = min_split_pos;
5147
5148 } else {
5149 assert(min_split_pos < max_split_pos, "must be true then");
5150 assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise");
5151
5152 // reason for using min_split_pos - 1: when the minimal split pos is exactly at the
5153 // beginning of a block, then min_split_pos is also a possible split position.
5154 // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos
5155 BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1);
5156
5157 // reason for using max_split_pos - 1: otherwise there would be an assertion failure
5158 // when an interval ends at the end of the last block of the method
5159 // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no
5160 // block at this op_id)
5161 BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1);
5162
5163 assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order");
5164 if (min_block == max_block) {
5165 // split position cannot be moved to block boundary, so split as late as possible
5166 TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block"));
5167 optimal_split_pos = max_split_pos;
5168
5169 } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) {
5170 // Do not move split position if the interval has a hole before max_split_pos.
5171 // Intervals resulting from Phi-Functions have more than one definition (marked
5172 // as mustHaveRegister) with a hole before each definition. When the register is needed
5173 // for the second definition, an earlier reloading is unnecessary.
5174 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos"));
5175 optimal_split_pos = max_split_pos;
5176
5177 } else {
5178 // search optimal block boundary between min_split_pos and max_split_pos
5179 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()));
5180
5181 if (do_loop_optimization) {
5182 // Loop optimization: if a loop-end marker is found between min- and max-position,
5183 // then split before this loop
5184 int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2);
5185 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d", loop_end_pos));
5186
5187 assert(loop_end_pos > min_split_pos, "invalid order");
5188 if (loop_end_pos < max_split_pos) {
5189 // loop-end marker found between min- and max-position
5190 // if it is not the end marker for the same loop as the min-position, then move
5191 // the max-position to this loop block.
5192 // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading
5193 // of the interval (normally, only mustHaveRegister causes a reloading)
5194 BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos);
5195
5196 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()));
5197 assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between");
5198
5199 optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2);
5200 if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) {
5201 optimal_split_pos = -1;
5202 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary"));
5203 } else {
5204 TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful"));
5205 }
5206 }
5207 }
5208
5209 if (optimal_split_pos == -1) {
5210 // not calculated by loop optimization
5211 optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos);
5212 }
5213 }
5214 }
5215 TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d", optimal_split_pos));
5216
5217 return optimal_split_pos;
5218 }
5219
5220
5221 /*
5222 split an interval at the optimal position between min_split_pos and
5223 max_split_pos in two parts:
5224 1) the left part has already a location assigned
5225 2) the right part is sorted into to the unhandled-list
5226 */
5227 void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) {
5228 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: "); it->print());
5229 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos));
5230
5231 assert(it->from() < min_split_pos, "cannot split at start of interval");
5232 assert(current_position() < min_split_pos, "cannot split before current position");
5233 assert(min_split_pos <= max_split_pos, "invalid order");
5234 assert(max_split_pos <= it->to(), "cannot split after end of interval");
5235
5236 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true);
5237
5238 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range");
5239 assert(optimal_split_pos <= it->to(), "cannot split after end of interval");
5240 assert(optimal_split_pos > it->from(), "cannot split at start of interval");
5241
5242 if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) {
5243 // the split position would be just before the end of the interval
5244 // -> no split at all necessary
5245 TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval"));
5246 return;
5247 }
5248
5249 // must calculate this before the actual split is performed and before split position is moved to odd op_id
5250 bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos);
5251
5252 if (!allocator()->is_block_begin(optimal_split_pos)) {
5253 // move position before actual instruction (odd op_id)
5254 optimal_split_pos = (optimal_split_pos - 1) | 1;
5255 }
5256
5257 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos));
5258 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary");
5259 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary");
5260
5261 Interval* split_part = it->split(optimal_split_pos);
5262
5263 allocator()->append_interval(split_part);
5264 allocator()->copy_register_flags(it, split_part);
5265 split_part->set_insert_move_when_activated(move_necessary);
5266 append_to_unhandled(unhandled_first_addr(anyKind), split_part);
5267
5268 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)", move_necessary));
5269 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print());
5270 TRACE_LINEAR_SCAN(2, tty->print (" "); split_part->print());
5271 }
5272
5273 /*
5274 split an interval at the optimal position between min_split_pos and
5275 max_split_pos in two parts:
5276 1) the left part has already a location assigned
5277 2) the right part is always on the stack and therefore ignored in further processing
5278 */
5279 void LinearScanWalker::split_for_spilling(Interval* it) {
5280 // calculate allowed range of splitting position
5281 int max_split_pos = current_position();
5282 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from());
5283
5284 TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: "); it->print());
5285 TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d", min_split_pos, max_split_pos));
5286
5287 assert(it->state() == activeState, "why spill interval that is not active?");
5288 assert(it->from() <= min_split_pos, "cannot split before start of interval");
5289 assert(min_split_pos <= max_split_pos, "invalid order");
5290 assert(max_split_pos < it->to(), "cannot split at end end of interval");
5291 assert(current_position() < it->to(), "interval must not end before current position");
5292
5293 if (min_split_pos == it->from()) {
5294 // the whole interval is never used, so spill it entirely to memory
5295 TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval"));
5296 assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position");
5297
5298 allocator()->assign_spill_slot(it);
5299 allocator()->change_spill_state(it, min_split_pos);
5300
5301 // Also kick parent intervals out of register to memory when they have no use
5302 // position. This avoids short interval in register surrounded by intervals in
5303 // memory -> avoid useless moves from memory to register and back
5304 Interval* parent = it;
5305 while (parent != nullptr && parent->is_split_child()) {
5306 parent = parent->split_child_before_op_id(parent->from());
5307
5308 if (parent->assigned_reg() < LinearScan::nof_regs) {
5309 if (parent->first_usage(shouldHaveRegister) == max_jint) {
5310 // parent is never used, so kick it out of its assigned register
5311 TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used", parent->reg_num()));
5312 allocator()->assign_spill_slot(parent);
5313 } else {
5314 // do not go further back because the register is actually used by the interval
5315 parent = nullptr;
5316 }
5317 }
5318 }
5319
5320 } else {
5321 // search optimal split pos, split interval and spill only the right hand part
5322 int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false);
5323
5324 assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range");
5325 assert(optimal_split_pos < it->to(), "cannot split at end of interval");
5326 assert(optimal_split_pos >= it->from(), "cannot split before start of interval");
5327
5328 if (!allocator()->is_block_begin(optimal_split_pos)) {
5329 // move position before actual instruction (odd op_id)
5330 optimal_split_pos = (optimal_split_pos - 1) | 1;
5331 }
5332
5333 TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d", optimal_split_pos));
5334 assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary");
5335 assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary");
5336
5337 Interval* spilled_part = it->split(optimal_split_pos);
5338 allocator()->append_interval(spilled_part);
5339 allocator()->assign_spill_slot(spilled_part);
5340 allocator()->change_spill_state(spilled_part, optimal_split_pos);
5341
5342 if (!allocator()->is_block_begin(optimal_split_pos)) {
5343 TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d", it->reg_num(), spilled_part->reg_num()));
5344 insert_move(optimal_split_pos, it, spilled_part);
5345 }
5346
5347 // the current_split_child is needed later when moves are inserted for reloading
5348 assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child");
5349 spilled_part->make_current_split_child();
5350
5351 TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts"));
5352 TRACE_LINEAR_SCAN(2, tty->print (" "); it->print());
5353 TRACE_LINEAR_SCAN(2, tty->print (" "); spilled_part->print());
5354 }
5355 }
5356
5357
5358 void LinearScanWalker::split_stack_interval(Interval* it) {
5359 int min_split_pos = current_position() + 1;
5360 int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to());
5361
5362 split_before_usage(it, min_split_pos, max_split_pos);
5363 }
5364
5365 void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) {
5366 int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1);
5367 int max_split_pos = register_available_until;
5368
5369 split_before_usage(it, min_split_pos, max_split_pos);
5370 }
5371
5372 void LinearScanWalker::split_and_spill_interval(Interval* it) {
5373 assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed");
5374
5375 int current_pos = current_position();
5376 if (it->state() == inactiveState) {
5377 // the interval is currently inactive, so no spill slot is needed for now.
5378 // when the split part is activated, the interval has a new chance to get a register,
5379 // so in the best case no stack slot is necessary
5380 assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise");
5381 split_before_usage(it, current_pos + 1, current_pos + 1);
5382
5383 } else {
5384 // search the position where the interval must have a register and split
5385 // at the optimal position before.
5386 // The new created part is added to the unhandled list and will get a register
5387 // when it is activated
5388 int min_split_pos = current_pos + 1;
5389 int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to());
5390
5391 split_before_usage(it, min_split_pos, max_split_pos);
5392
5393 assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register");
5394 split_for_spilling(it);
5395 }
5396 }
5397
5398 int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) {
5399 int min_full_reg = any_reg;
5400 int max_partial_reg = any_reg;
5401
5402 for (int i = _first_reg; i <= _last_reg; i++) {
5403 if (i == ignore_reg) {
5404 // this register must be ignored
5405
5406 } else if (_use_pos[i] >= interval_to) {
5407 // this register is free for the full interval
5408 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) {
5409 min_full_reg = i;
5410 }
5411 } else if (_use_pos[i] > reg_needed_until) {
5412 // this register is at least free until reg_needed_until
5413 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) {
5414 max_partial_reg = i;
5415 }
5416 }
5417 }
5418
5419 if (min_full_reg != any_reg) {
5420 return min_full_reg;
5421 } else if (max_partial_reg != any_reg) {
5422 *need_split = true;
5423 return max_partial_reg;
5424 } else {
5425 return any_reg;
5426 }
5427 }
5428
5429 int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) {
5430 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm");
5431
5432 int min_full_reg = any_reg;
5433 int max_partial_reg = any_reg;
5434
5435 for (int i = _first_reg; i < _last_reg; i+=2) {
5436 if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) {
5437 // this register is free for the full interval
5438 if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) {
5439 min_full_reg = i;
5440 }
5441 } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) {
5442 // this register is at least free until reg_needed_until
5443 if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) {
5444 max_partial_reg = i;
5445 }
5446 }
5447 }
5448
5449 if (min_full_reg != any_reg) {
5450 return min_full_reg;
5451 } else if (max_partial_reg != any_reg) {
5452 *need_split = true;
5453 return max_partial_reg;
5454 } else {
5455 return any_reg;
5456 }
5457 }
5458
5459 bool LinearScanWalker::alloc_free_reg(Interval* cur) {
5460 TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for "); cur->print());
5461
5462 init_use_lists(true);
5463 free_exclude_active_fixed();
5464 free_exclude_active_any();
5465 free_collect_inactive_fixed(cur);
5466 free_collect_inactive_any(cur);
5467 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0");
5468
5469 // _use_pos contains the start of the next interval that has this register assigned
5470 // (either as a fixed register or a normal allocated register in the past)
5471 // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely
5472 #ifdef ASSERT
5473 if (TraceLinearScanLevel >= 4) {
5474 tty->print_cr(" state of registers:");
5475 for (int i = _first_reg; i <= _last_reg; i++) {
5476 tty->print(" reg %d (", i);
5477 LinearScan::print_reg_num(i);
5478 tty->print_cr("): use_pos: %d", _use_pos[i]);
5479 }
5480 }
5481 #endif
5482
5483 int hint_reg, hint_regHi;
5484 Interval* register_hint = cur->register_hint();
5485 if (register_hint != nullptr) {
5486 hint_reg = register_hint->assigned_reg();
5487 hint_regHi = register_hint->assigned_regHi();
5488
5489 if (_num_phys_regs == 2 && allocator()->is_precolored_cpu_interval(register_hint)) {
5490 assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals");
5491 hint_regHi = hint_reg + 1; // connect e.g. eax-edx
5492 }
5493 #ifdef ASSERT
5494 if (TraceLinearScanLevel >= 4) {
5495 tty->print(" hint registers %d (", hint_reg);
5496 LinearScan::print_reg_num(hint_reg);
5497 tty->print("), %d (", hint_regHi);
5498 LinearScan::print_reg_num(hint_regHi);
5499 tty->print(") from interval ");
5500 register_hint->print();
5501 }
5502 #endif
5503 } else {
5504 hint_reg = any_reg;
5505 hint_regHi = any_reg;
5506 }
5507 assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal");
5508 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval");
5509
5510 // the register must be free at least until this position
5511 int reg_needed_until = cur->from() + 1;
5512 int interval_to = cur->to();
5513
5514 bool need_split = false;
5515 int split_pos;
5516 int reg;
5517 int regHi = any_reg;
5518
5519 if (_adjacent_regs) {
5520 reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split);
5521 regHi = reg + 1;
5522 if (reg == any_reg) {
5523 return false;
5524 }
5525 split_pos = MIN2(_use_pos[reg], _use_pos[regHi]);
5526
5527 } else {
5528 reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split);
5529 if (reg == any_reg) {
5530 return false;
5531 }
5532 split_pos = _use_pos[reg];
5533
5534 if (_num_phys_regs == 2) {
5535 regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split);
5536
5537 if (_use_pos[reg] < interval_to && regHi == any_reg) {
5538 // do not split interval if only one register can be assigned until the split pos
5539 // (when one register is found for the whole interval, split&spill is only
5540 // performed for the hi register)
5541 return false;
5542
5543 } else if (regHi != any_reg) {
5544 split_pos = MIN2(split_pos, _use_pos[regHi]);
5545
5546 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax
5547 if (reg > regHi) {
5548 int temp = reg;
5549 reg = regHi;
5550 regHi = temp;
5551 }
5552 }
5553 }
5554 }
5555
5556 cur->assign_reg(reg, regHi);
5557 #ifdef ASSERT
5558 if (TraceLinearScanLevel >= 2) {
5559 tty->print(" selected registers %d (", reg);
5560 LinearScan::print_reg_num(reg);
5561 tty->print("), %d (", regHi);
5562 LinearScan::print_reg_num(regHi);
5563 tty->print_cr(")");
5564 }
5565 #endif
5566 assert(split_pos > 0, "invalid split_pos");
5567 if (need_split) {
5568 // register not available for full interval, so split it
5569 split_when_partial_register_available(cur, split_pos);
5570 }
5571
5572 // only return true if interval is completely assigned
5573 return _num_phys_regs == 1 || regHi != any_reg;
5574 }
5575
5576
5577 int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int ignore_reg, bool* need_split) {
5578 int max_reg = any_reg;
5579
5580 for (int i = _first_reg; i <= _last_reg; i++) {
5581 if (i == ignore_reg) {
5582 // this register must be ignored
5583
5584 } else if (_use_pos[i] > reg_needed_until) {
5585 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) {
5586 max_reg = i;
5587 }
5588 }
5589 }
5590
5591 if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) {
5592 *need_split = true;
5593 }
5594
5595 return max_reg;
5596 }
5597
5598 int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, bool* need_split) {
5599 assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm");
5600
5601 int max_reg = any_reg;
5602
5603 for (int i = _first_reg; i < _last_reg; i+=2) {
5604 if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) {
5605 if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) {
5606 max_reg = i;
5607 }
5608 }
5609 }
5610
5611 if (max_reg != any_reg &&
5612 (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to)) {
5613 *need_split = true;
5614 }
5615
5616 return max_reg;
5617 }
5618
5619 void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) {
5620 assert(reg != any_reg, "no register assigned");
5621
5622 for (int i = 0; i < _spill_intervals[reg]->length(); i++) {
5623 Interval* it = _spill_intervals[reg]->at(i);
5624 remove_from_list(it);
5625 split_and_spill_interval(it);
5626 }
5627
5628 if (regHi != any_reg) {
5629 IntervalList* processed = _spill_intervals[reg];
5630 for (int i = 0; i < _spill_intervals[regHi]->length(); i++) {
5631 Interval* it = _spill_intervals[regHi]->at(i);
5632 if (processed->find(it) == -1) {
5633 remove_from_list(it);
5634 split_and_spill_interval(it);
5635 }
5636 }
5637 }
5638 }
5639
5640
5641 // Split an Interval and spill it to memory so that cur can be placed in a register
5642 void LinearScanWalker::alloc_locked_reg(Interval* cur) {
5643 TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for "); cur->print());
5644
5645 // collect current usage of registers
5646 init_use_lists(false);
5647 spill_exclude_active_fixed();
5648 assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0");
5649 spill_block_inactive_fixed(cur);
5650 spill_collect_active_any();
5651 spill_collect_inactive_any(cur);
5652
5653 #ifdef ASSERT
5654 if (TraceLinearScanLevel >= 4) {
5655 tty->print_cr(" state of registers:");
5656 for (int i = _first_reg; i <= _last_reg; i++) {
5657 tty->print(" reg %d(", i);
5658 LinearScan::print_reg_num(i);
5659 tty->print("): use_pos: %d, block_pos: %d, intervals: ", _use_pos[i], _block_pos[i]);
5660 for (int j = 0; j < _spill_intervals[i]->length(); j++) {
5661 tty->print("%d ", _spill_intervals[i]->at(j)->reg_num());
5662 }
5663 tty->cr();
5664 }
5665 }
5666 #endif
5667
5668 // the register must be free at least until this position
5669 int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1);
5670 int interval_to = cur->to();
5671 assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use");
5672
5673 int split_pos = 0;
5674 int use_pos = 0;
5675 bool need_split = false;
5676 int reg, regHi;
5677
5678 if (_adjacent_regs) {
5679 reg = find_locked_double_reg(reg_needed_until, interval_to, &need_split);
5680 regHi = reg + 1;
5681
5682 if (reg != any_reg) {
5683 use_pos = MIN2(_use_pos[reg], _use_pos[regHi]);
5684 split_pos = MIN2(_block_pos[reg], _block_pos[regHi]);
5685 }
5686 } else {
5687 reg = find_locked_reg(reg_needed_until, interval_to, cur->assigned_reg(), &need_split);
5688 regHi = any_reg;
5689
5690 if (reg != any_reg) {
5691 use_pos = _use_pos[reg];
5692 split_pos = _block_pos[reg];
5693
5694 if (_num_phys_regs == 2) {
5695 if (cur->assigned_reg() != any_reg) {
5696 regHi = reg;
5697 reg = cur->assigned_reg();
5698 } else {
5699 regHi = find_locked_reg(reg_needed_until, interval_to, reg, &need_split);
5700 if (regHi != any_reg) {
5701 use_pos = MIN2(use_pos, _use_pos[regHi]);
5702 split_pos = MIN2(split_pos, _block_pos[regHi]);
5703 }
5704 }
5705
5706 if (regHi != any_reg && reg > regHi) {
5707 // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax
5708 int temp = reg;
5709 reg = regHi;
5710 regHi = temp;
5711 }
5712 }
5713 }
5714 }
5715
5716 if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) {
5717 // the first use of cur is later than the spilling position -> spill cur
5718 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));
5719
5720 if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) {
5721 assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)");
5722 // assign a reasonable register and do a bailout in product mode to avoid errors
5723 allocator()->assign_spill_slot(cur);
5724 BAILOUT("LinearScan: no register found");
5725 }
5726
5727 split_and_spill_interval(cur);
5728 } else {
5729 #ifdef ASSERT
5730 if (TraceLinearScanLevel >= 4) {
5731 tty->print("decided to use register %d (", reg);
5732 LinearScan::print_reg_num(reg);
5733 tty->print("), %d (", regHi);
5734 LinearScan::print_reg_num(regHi);
5735 tty->print_cr(")");
5736 }
5737 #endif
5738 assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found");
5739 assert(split_pos > 0, "invalid split_pos");
5740 assert(need_split == false || split_pos > cur->from(), "splitting interval at from");
5741
5742 cur->assign_reg(reg, regHi);
5743 if (need_split) {
5744 // register not available for full interval, so split it
5745 split_when_partial_register_available(cur, split_pos);
5746 }
5747
5748 // perform splitting and spilling for all affected intervals
5749 split_and_spill_intersecting_intervals(reg, regHi);
5750 }
5751 }
5752
5753 bool LinearScanWalker::no_allocation_possible(Interval* cur) {
5754 #ifdef X86
5755 // fast calculation of intervals that can never get a register because the
5756 // the next instruction is a call that blocks all registers
5757 // Note: this does not work if callee-saved registers are available (e.g. on Sparc)
5758
5759 // check if this interval is the result of a split operation
5760 // (an interval got a register until this position)
5761 int pos = cur->from();
5762 if ((pos & 1) == 1) {
5763 // the current instruction is a call that blocks all registers
5764 if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) {
5765 TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call"));
5766
5767 // safety check that there is really no register available
5768 assert(alloc_free_reg(cur) == false, "found a register for this interval");
5769 return true;
5770 }
5771
5772 }
5773 #endif
5774 return false;
5775 }
5776
5777 void LinearScanWalker::init_vars_for_alloc(Interval* cur) {
5778 BasicType type = cur->type();
5779 _num_phys_regs = LinearScan::num_physical_regs(type);
5780 _adjacent_regs = LinearScan::requires_adjacent_regs(type);
5781
5782 if (pd_init_regs_for_alloc(cur)) {
5783 // the appropriate register range was selected.
5784 } else if (type == T_FLOAT || type == T_DOUBLE) {
5785 _first_reg = pd_first_fpu_reg;
5786 _last_reg = pd_last_fpu_reg;
5787 } else {
5788 _first_reg = pd_first_cpu_reg;
5789 _last_reg = FrameMap::last_cpu_reg();
5790 }
5791
5792 assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range");
5793 assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range");
5794 }
5795
5796
5797 bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) {
5798 if (op->code() != lir_move) {
5799 return false;
5800 }
5801 assert(op->as_Op1() != nullptr, "move must be LIR_Op1");
5802
5803 LIR_Opr in = ((LIR_Op1*)op)->in_opr();
5804 LIR_Opr res = ((LIR_Op1*)op)->result_opr();
5805 return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num();
5806 }
5807
5808 // optimization (especially for phi functions of nested loops):
5809 // assign same spill slot to non-intersecting intervals
5810 void LinearScanWalker::combine_spilled_intervals(Interval* cur) {
5811 if (cur->is_split_child()) {
5812 // optimization is only suitable for split parents
5813 return;
5814 }
5815
5816 Interval* register_hint = cur->register_hint(false);
5817 if (register_hint == nullptr) {
5818 // cur is not the target of a move, otherwise register_hint would be set
5819 return;
5820 }
5821 assert(register_hint->is_split_parent(), "register hint must be split parent");
5822
5823 if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) {
5824 // combining the stack slots for intervals where spill move optimization is applied
5825 // is not benefitial and would cause problems
5826 return;
5827 }
5828
5829 int begin_pos = cur->from();
5830 int end_pos = cur->to();
5831 if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) {
5832 // safety check that lir_op_with_id is allowed
5833 return;
5834 }
5835
5836 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)) {
5837 // cur and register_hint are not connected with two moves
5838 return;
5839 }
5840
5841 Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode);
5842 Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode);
5843 if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) {
5844 // register_hint must be split, otherwise the re-writing of use positions does not work
5845 return;
5846 }
5847
5848 assert(begin_hint->assigned_reg() != any_reg, "must have register assigned");
5849 assert(end_hint->assigned_reg() == any_reg, "must not have register assigned");
5850 assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move");
5851 assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move");
5852
5853 if (begin_hint->assigned_reg() < LinearScan::nof_regs) {
5854 // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur
5855 return;
5856 }
5857 assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled");
5858 assert(!cur->intersects(register_hint), "cur should not intersect register_hint");
5859
5860 if (cur->intersects_any_children_of(register_hint)) {
5861 // Bail out if cur intersects any split children of register_hint, which have the same spill slot as their parent. An overlap of two intervals with
5862 // the same spill slot could result in a situation where both intervals are spilled at the same time to the same stack location which is not correct.
5863 return;
5864 }
5865
5866 // modify intervals such that cur gets the same stack slot as register_hint
5867 // delete use positions to prevent the intervals to get a register at beginning
5868 cur->set_canonical_spill_slot(register_hint->canonical_spill_slot());
5869 cur->remove_first_use_pos();
5870 end_hint->remove_first_use_pos();
5871 }
5872
5873
5874 // allocate a physical register or memory location to an interval
5875 bool LinearScanWalker::activate_current() {
5876 Interval* cur = current();
5877 bool result = true;
5878
5879 TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval "); cur->print());
5880 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()));
5881
5882 if (cur->assigned_reg() >= LinearScan::nof_regs) {
5883 // activating an interval that has a stack slot assigned -> split it at first use position
5884 // used for method parameters
5885 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use"));
5886
5887 split_stack_interval(cur);
5888 result = false;
5889
5890 } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) {
5891 // activating an interval that must start in a stack slot, but may get a register later
5892 // used for lir_roundfp: rounding is done by store to stack and reload later
5893 TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use"));
5894 assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned");
5895
5896 allocator()->assign_spill_slot(cur);
5897 split_stack_interval(cur);
5898 result = false;
5899
5900 } else if (cur->assigned_reg() == any_reg) {
5901 // interval has not assigned register -> normal allocation
5902 // (this is the normal case for most intervals)
5903 TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register"));
5904
5905 // assign same spill slot to non-intersecting intervals
5906 combine_spilled_intervals(cur);
5907
5908 init_vars_for_alloc(cur);
5909 if (no_allocation_possible(cur) || !alloc_free_reg(cur)) {
5910 // no empty register available.
5911 // split and spill another interval so that this interval gets a register
5912 alloc_locked_reg(cur);
5913 }
5914
5915 // spilled intervals need not be move to active-list
5916 if (cur->assigned_reg() >= LinearScan::nof_regs) {
5917 result = false;
5918 }
5919 }
5920
5921 // load spilled values that become active from stack slot to register
5922 if (cur->insert_move_when_activated()) {
5923 assert(cur->is_split_child(), "must be");
5924 assert(cur->current_split_child() != nullptr, "must be");
5925 assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval");
5926 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()));
5927
5928 insert_move(cur->from(), cur->current_split_child(), cur);
5929 }
5930 cur->make_current_split_child();
5931
5932 return result; // true = interval is moved to active list
5933 }
5934
5935
5936 // Implementation of EdgeMoveOptimizer
5937
5938 EdgeMoveOptimizer::EdgeMoveOptimizer() :
5939 _edge_instructions(4),
5940 _edge_instructions_idx(4)
5941 {
5942 }
5943
5944 void EdgeMoveOptimizer::optimize(BlockList* code) {
5945 EdgeMoveOptimizer optimizer = EdgeMoveOptimizer();
5946
5947 // ignore the first block in the list (index 0 is not processed)
5948 for (int i = code->length() - 1; i >= 1; i--) {
5949 BlockBegin* block = code->at(i);
5950
5951 if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) {
5952 optimizer.optimize_moves_at_block_end(block);
5953 }
5954 if (block->number_of_sux() == 2) {
5955 optimizer.optimize_moves_at_block_begin(block);
5956 }
5957 }
5958 }
5959
5960
5961 // clear all internal data structures
5962 void EdgeMoveOptimizer::init_instructions() {
5963 _edge_instructions.clear();
5964 _edge_instructions_idx.clear();
5965 }
5966
5967 // append a lir-instruction-list and the index of the current operation in to the list
5968 void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) {
5969 _edge_instructions.append(instructions);
5970 _edge_instructions_idx.append(instructions_idx);
5971 }
5972
5973 // return the current operation of the given edge (predecessor or successor)
5974 LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) {
5975 LIR_OpList* instructions = _edge_instructions.at(edge);
5976 int idx = _edge_instructions_idx.at(edge);
5977
5978 if (idx < instructions->length()) {
5979 return instructions->at(idx);
5980 } else {
5981 return nullptr;
5982 }
5983 }
5984
5985 // removes the current operation of the given edge (predecessor or successor)
5986 void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) {
5987 LIR_OpList* instructions = _edge_instructions.at(edge);
5988 int idx = _edge_instructions_idx.at(edge);
5989 instructions->remove_at(idx);
5990
5991 if (decrement_index) {
5992 _edge_instructions_idx.at_put(edge, idx - 1);
5993 }
5994 }
5995
5996
5997 bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) {
5998 if (op1 == nullptr || op2 == nullptr) {
5999 // at least one block is already empty -> no optimization possible
6000 return true;
6001 }
6002
6003 if (op1->code() == lir_move && op2->code() == lir_move) {
6004 assert(op1->as_Op1() != nullptr, "move must be LIR_Op1");
6005 assert(op2->as_Op1() != nullptr, "move must be LIR_Op1");
6006 LIR_Op1* move1 = (LIR_Op1*)op1;
6007 LIR_Op1* move2 = (LIR_Op1*)op2;
6008 if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) {
6009 // these moves are exactly equal and can be optimized
6010 return false;
6011 }
6012
6013 } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) {
6014 assert(op1->as_Op1() != nullptr, "fxch must be LIR_Op1");
6015 assert(op2->as_Op1() != nullptr, "fxch must be LIR_Op1");
6016 LIR_Op1* fxch1 = (LIR_Op1*)op1;
6017 LIR_Op1* fxch2 = (LIR_Op1*)op2;
6018 if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) {
6019 // equal FPU stack operations can be optimized
6020 return false;
6021 }
6022
6023 } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) {
6024 // equal FPU stack operations can be optimized
6025 return false;
6026 }
6027
6028 // no optimization possible
6029 return true;
6030 }
6031
6032 void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) {
6033 TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d", block->block_id()));
6034
6035 if (block->is_predecessor(block)) {
6036 // currently we can't handle this correctly.
6037 return;
6038 }
6039
6040 init_instructions();
6041 int num_preds = block->number_of_preds();
6042 assert(num_preds > 1, "do not call otherwise");
6043 assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed");
6044
6045 // setup a list with the lir-instructions of all predecessors
6046 int i;
6047 for (i = 0; i < num_preds; i++) {
6048 BlockBegin* pred = block->pred_at(i);
6049 LIR_OpList* pred_instructions = pred->lir()->instructions_list();
6050
6051 if (pred->number_of_sux() != 1) {
6052 // this can happen with switch-statements where multiple edges are between
6053 // the same blocks.
6054 return;
6055 }
6056
6057 assert(pred->number_of_sux() == 1, "can handle only one successor");
6058 assert(pred->sux_at(0) == block, "invalid control flow");
6059 assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch");
6060 assert(pred_instructions->last()->as_OpBranch() != nullptr, "branch must be LIR_OpBranch");
6061 assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch");
6062
6063 if (pred_instructions->last()->info() != nullptr) {
6064 // can not optimize instructions when debug info is needed
6065 return;
6066 }
6067
6068 // ignore the unconditional branch at the end of the block
6069 append_instructions(pred_instructions, pred_instructions->length() - 2);
6070 }
6071
6072
6073 // process lir-instructions while all predecessors end with the same instruction
6074 while (true) {
6075 LIR_Op* op = instruction_at(0);
6076 for (i = 1; i < num_preds; i++) {
6077 if (operations_different(op, instruction_at(i))) {
6078 // these instructions are different and cannot be optimized ->
6079 // no further optimization possible
6080 return;
6081 }
6082 }
6083
6084 TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: ", num_preds); op->print());
6085
6086 // insert the instruction at the beginning of the current block
6087 block->lir()->insert_before(1, op);
6088
6089 // delete the instruction at the end of all predecessors
6090 for (i = 0; i < num_preds; i++) {
6091 remove_cur_instruction(i, true);
6092 }
6093 }
6094 }
6095
6096
6097 void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) {
6098 TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d", block->block_id()));
6099
6100 init_instructions();
6101 int num_sux = block->number_of_sux();
6102
6103 LIR_OpList* cur_instructions = block->lir()->instructions_list();
6104
6105 assert(num_sux == 2, "method should not be called otherwise");
6106 assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch");
6107 assert(cur_instructions->last()->as_OpBranch() != nullptr, "branch must be LIR_OpBranch");
6108 assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch");
6109
6110 if (cur_instructions->last()->info() != nullptr) {
6111 // can no optimize instructions when debug info is needed
6112 return;
6113 }
6114
6115 LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2);
6116 if (branch->info() != nullptr || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) {
6117 // not a valid case for optimization
6118 // currently, only blocks that end with two branches (conditional branch followed
6119 // by unconditional branch) are optimized
6120 return;
6121 }
6122
6123 // now it is guaranteed that the block ends with two branch instructions.
6124 // the instructions are inserted at the end of the block before these two branches
6125 int insert_idx = cur_instructions->length() - 2;
6126
6127 int i;
6128 #ifdef ASSERT
6129 for (i = insert_idx - 1; i >= 0; i--) {
6130 LIR_Op* op = cur_instructions->at(i);
6131 if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != nullptr) {
6132 assert(false, "block with two successors can have only two branch instructions");
6133 }
6134 }
6135 #endif
6136
6137 // setup a list with the lir-instructions of all successors
6138 for (i = 0; i < num_sux; i++) {
6139 BlockBegin* sux = block->sux_at(i);
6140 LIR_OpList* sux_instructions = sux->lir()->instructions_list();
6141
6142 assert(sux_instructions->at(0)->code() == lir_label, "block must start with label");
6143
6144 if (sux->number_of_preds() != 1) {
6145 // this can happen with switch-statements where multiple edges are between
6146 // the same blocks.
6147 return;
6148 }
6149 assert(sux->pred_at(0) == block, "invalid control flow");
6150 assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed");
6151
6152 // ignore the label at the beginning of the block
6153 append_instructions(sux_instructions, 1);
6154 }
6155
6156 // process lir-instructions while all successors begin with the same instruction
6157 while (true) {
6158 LIR_Op* op = instruction_at(0);
6159 for (i = 1; i < num_sux; i++) {
6160 if (operations_different(op, instruction_at(i))) {
6161 // these instructions are different and cannot be optimized ->
6162 // no further optimization possible
6163 return;
6164 }
6165 }
6166
6167 TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: ", num_sux); op->print());
6168
6169 // insert instruction at end of current block
6170 block->lir()->insert_before(insert_idx, op);
6171 insert_idx++;
6172
6173 // delete the instructions at the beginning of all successors
6174 for (i = 0; i < num_sux; i++) {
6175 remove_cur_instruction(i, false);
6176 }
6177 }
6178 }
6179
6180
6181 // Implementation of ControlFlowOptimizer
6182
6183 ControlFlowOptimizer::ControlFlowOptimizer() :
6184 _original_preds(4)
6185 {
6186 }
6187
6188 void ControlFlowOptimizer::optimize(BlockList* code) {
6189 ControlFlowOptimizer optimizer = ControlFlowOptimizer();
6190
6191 // push the OSR entry block to the end so that we're not jumping over it.
6192 BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry();
6193 if (osr_entry) {
6194 int index = osr_entry->linear_scan_number();
6195 assert(code->at(index) == osr_entry, "wrong index");
6196 code->remove_at(index);
6197 code->append(osr_entry);
6198 }
6199
6200 optimizer.reorder_short_loops(code);
6201 optimizer.delete_empty_blocks(code);
6202 optimizer.delete_unnecessary_jumps(code);
6203 optimizer.delete_jumps_to_return(code);
6204 }
6205
6206 void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* header_block, int header_idx) {
6207 int i = header_idx + 1;
6208 int max_end = MIN2(header_idx + ShortLoopSize, code->length());
6209 while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) {
6210 i++;
6211 }
6212
6213 if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) {
6214 int end_idx = i - 1;
6215 BlockBegin* end_block = code->at(end_idx);
6216
6217 if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) {
6218 // short loop from header_idx to end_idx found -> reorder blocks such that
6219 // the header_block is the last block instead of the first block of the loop
6220 TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d",
6221 end_idx - header_idx + 1,
6222 header_block->block_id(), end_block->block_id()));
6223
6224 for (int j = header_idx; j < end_idx; j++) {
6225 code->at_put(j, code->at(j + 1));
6226 }
6227 code->at_put(end_idx, header_block);
6228
6229 // correct the flags so that any loop alignment occurs in the right place.
6230 assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target");
6231 code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag);
6232 code->at(header_idx)->set(BlockBegin::backward_branch_target_flag);
6233 }
6234 }
6235 }
6236
6237 void ControlFlowOptimizer::reorder_short_loops(BlockList* code) {
6238 for (int i = code->length() - 1; i >= 0; i--) {
6239 BlockBegin* block = code->at(i);
6240
6241 if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) {
6242 reorder_short_loop(code, block, i);
6243 }
6244 }
6245
6246 DEBUG_ONLY(verify(code));
6247 }
6248
6249 // only blocks with exactly one successor can be deleted. Such blocks
6250 // must always end with an unconditional branch to this successor
6251 bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) {
6252 if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) {
6253 return false;
6254 }
6255
6256 LIR_OpList* instructions = block->lir()->instructions_list();
6257
6258 assert(instructions->length() >= 2, "block must have label and branch");
6259 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label");
6260 assert(instructions->last()->as_OpBranch() != nullptr, "last instruction must always be a branch");
6261 assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional");
6262 assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor");
6263
6264 // block must have exactly one successor
6265
6266 if (instructions->length() == 2 && instructions->last()->info() == nullptr) {
6267 return true;
6268 }
6269 return false;
6270 }
6271
6272 // substitute branch targets in all branch-instructions of this blocks
6273 void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) {
6274 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()));
6275
6276 LIR_OpList* instructions = block->lir()->instructions_list();
6277
6278 assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label");
6279 for (int i = instructions->length() - 1; i >= 1; i--) {
6280 LIR_Op* op = instructions->at(i);
6281
6282 if (op->code() == lir_branch || op->code() == lir_cond_float_branch) {
6283 assert(op->as_OpBranch() != nullptr, "branch must be of type LIR_OpBranch");
6284 LIR_OpBranch* branch = (LIR_OpBranch*)op;
6285
6286 if (branch->block() == target_from) {
6287 branch->change_block(target_to);
6288 }
6289 if (branch->ublock() == target_from) {
6290 branch->change_ublock(target_to);
6291 }
6292 }
6293 }
6294 }
6295
6296 void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) {
6297 int old_pos = 0;
6298 int new_pos = 0;
6299 int num_blocks = code->length();
6300
6301 while (old_pos < num_blocks) {
6302 BlockBegin* block = code->at(old_pos);
6303
6304 if (can_delete_block(block)) {
6305 BlockBegin* new_target = block->sux_at(0);
6306
6307 // propagate backward branch target flag for correct code alignment
6308 if (block->is_set(BlockBegin::backward_branch_target_flag)) {
6309 new_target->set(BlockBegin::backward_branch_target_flag);
6310 }
6311
6312 // collect a list with all predecessors that contains each predecessor only once
6313 // the predecessors of cur are changed during the substitution, so a copy of the
6314 // predecessor list is necessary
6315 int j;
6316 _original_preds.clear();
6317 for (j = block->number_of_preds() - 1; j >= 0; j--) {
6318 BlockBegin* pred = block->pred_at(j);
6319 if (_original_preds.find(pred) == -1) {
6320 _original_preds.append(pred);
6321 }
6322 }
6323
6324 for (j = _original_preds.length() - 1; j >= 0; j--) {
6325 BlockBegin* pred = _original_preds.at(j);
6326 substitute_branch_target(pred, block, new_target);
6327 pred->substitute_sux(block, new_target);
6328 }
6329 } else {
6330 // adjust position of this block in the block list if blocks before
6331 // have been deleted
6332 if (new_pos != old_pos) {
6333 code->at_put(new_pos, code->at(old_pos));
6334 }
6335 new_pos++;
6336 }
6337 old_pos++;
6338 }
6339 code->trunc_to(new_pos);
6340
6341 DEBUG_ONLY(verify(code));
6342 }
6343
6344 void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) {
6345 // skip the last block because there a branch is always necessary
6346 for (int i = code->length() - 2; i >= 0; i--) {
6347 BlockBegin* block = code->at(i);
6348 LIR_OpList* instructions = block->lir()->instructions_list();
6349
6350 LIR_Op* last_op = instructions->last();
6351 if (last_op->code() == lir_branch) {
6352 assert(last_op->as_OpBranch() != nullptr, "branch must be of type LIR_OpBranch");
6353 LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op;
6354
6355 assert(last_branch->block() != nullptr, "last branch must always have a block as target");
6356 assert(last_branch->label() == last_branch->block()->label(), "must be equal");
6357
6358 if (last_branch->info() == nullptr) {
6359 if (last_branch->block() == code->at(i + 1)) {
6360
6361 TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d", block->block_id()));
6362
6363 // delete last branch instruction
6364 instructions->trunc_to(instructions->length() - 1);
6365
6366 } else {
6367 LIR_Op* prev_op = instructions->at(instructions->length() - 2);
6368 if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) {
6369 assert(prev_op->as_OpBranch() != nullptr, "branch must be of type LIR_OpBranch");
6370 LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op;
6371
6372 if (prev_branch->stub() == nullptr) {
6373
6374 LIR_Op2* prev_cmp = nullptr;
6375 // There might be a cmove inserted for profiling which depends on the same
6376 // compare. If we change the condition of the respective compare, we have
6377 // to take care of this cmove as well.
6378 LIR_Op4* prev_cmove = nullptr;
6379
6380 for(int j = instructions->length() - 3; j >= 0 && prev_cmp == nullptr; j--) {
6381 prev_op = instructions->at(j);
6382 // check for the cmove
6383 if (prev_op->code() == lir_cmove) {
6384 assert(prev_op->as_Op4() != nullptr, "cmove must be of type LIR_Op4");
6385 prev_cmove = (LIR_Op4*)prev_op;
6386 assert(prev_branch->cond() == prev_cmove->condition(), "should be the same");
6387 }
6388 if (prev_op->code() == lir_cmp) {
6389 assert(prev_op->as_Op2() != nullptr, "branch must be of type LIR_Op2");
6390 prev_cmp = (LIR_Op2*)prev_op;
6391 assert(prev_branch->cond() == prev_cmp->condition(), "should be the same");
6392 }
6393 }
6394 // Guarantee because it is dereferenced below.
6395 guarantee(prev_cmp != nullptr, "should have found comp instruction for branch");
6396 if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == nullptr) {
6397
6398 TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d", block->block_id()));
6399
6400 // eliminate a conditional branch to the immediate successor
6401 prev_branch->change_block(last_branch->block());
6402 prev_branch->negate_cond();
6403 prev_cmp->set_condition(prev_branch->cond());
6404 instructions->trunc_to(instructions->length() - 1);
6405 // if we do change the condition, we have to change the cmove as well
6406 if (prev_cmove != nullptr) {
6407 prev_cmove->set_condition(prev_branch->cond());
6408 LIR_Opr t = prev_cmove->in_opr1();
6409 prev_cmove->set_in_opr1(prev_cmove->in_opr2());
6410 prev_cmove->set_in_opr2(t);
6411 }
6412 }
6413 }
6414 }
6415 }
6416 }
6417 }
6418 }
6419
6420 DEBUG_ONLY(verify(code));
6421 }
6422
6423 void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) {
6424 #ifdef ASSERT
6425 ResourceBitMap return_converted(BlockBegin::number_of_blocks());
6426 #endif
6427
6428 for (int i = code->length() - 1; i >= 0; i--) {
6429 BlockBegin* block = code->at(i);
6430 LIR_OpList* cur_instructions = block->lir()->instructions_list();
6431 LIR_Op* cur_last_op = cur_instructions->last();
6432
6433 assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label");
6434 if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) {
6435 // the block contains only a label and a return
6436 // if a predecessor ends with an unconditional jump to this block, then the jump
6437 // can be replaced with a return instruction
6438 //
6439 // Note: the original block with only a return statement cannot be deleted completely
6440 // because the predecessors might have other (conditional) jumps to this block
6441 // -> this may lead to unnecessary return instructions in the final code
6442
6443 assert(cur_last_op->info() == nullptr, "return instructions do not have debug information");
6444 assert(block->number_of_sux() == 0 ||
6445 (return_converted.at(block->block_id()) && block->number_of_sux() == 1),
6446 "blocks that end with return must not have successors");
6447
6448 assert(cur_last_op->as_Op1() != nullptr, "return must be LIR_Op1");
6449 LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr();
6450
6451 for (int j = block->number_of_preds() - 1; j >= 0; j--) {
6452 BlockBegin* pred = block->pred_at(j);
6453 LIR_OpList* pred_instructions = pred->lir()->instructions_list();
6454 LIR_Op* pred_last_op = pred_instructions->last();
6455
6456 if (pred_last_op->code() == lir_branch) {
6457 assert(pred_last_op->as_OpBranch() != nullptr, "branch must be LIR_OpBranch");
6458 LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op;
6459
6460 if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == nullptr) {
6461 // replace the jump to a return with a direct return
6462 // Note: currently the edge between the blocks is not deleted
6463 pred_instructions->at_put(pred_instructions->length() - 1, new LIR_OpReturn(return_opr));
6464 #ifdef ASSERT
6465 return_converted.set_bit(pred->block_id());
6466 #endif
6467 }
6468 }
6469 }
6470 }
6471 }
6472 }
6473
6474
6475 #ifdef ASSERT
6476 void ControlFlowOptimizer::verify(BlockList* code) {
6477 for (int i = 0; i < code->length(); i++) {
6478 BlockBegin* block = code->at(i);
6479 LIR_OpList* instructions = block->lir()->instructions_list();
6480
6481 int j;
6482 for (j = 0; j < instructions->length(); j++) {
6483 LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch();
6484
6485 if (op_branch != nullptr) {
6486 assert(op_branch->block() == nullptr || code->find(op_branch->block()) != -1, "branch target not valid");
6487 assert(op_branch->ublock() == nullptr || code->find(op_branch->ublock()) != -1, "branch target not valid");
6488 }
6489 }
6490
6491 for (j = 0; j < block->number_of_sux() - 1; j++) {
6492 BlockBegin* sux = block->sux_at(j);
6493 assert(code->find(sux) != -1, "successor not valid");
6494 }
6495
6496 for (j = 0; j < block->number_of_preds() - 1; j++) {
6497 BlockBegin* pred = block->pred_at(j);
6498 assert(code->find(pred) != -1, "successor not valid");
6499 }
6500 }
6501 }
6502 #endif
6503
6504
6505 #ifndef PRODUCT
6506
6507 // Implementation of LinearStatistic
6508
6509 const char* LinearScanStatistic::counter_name(int counter_idx) {
6510 switch (counter_idx) {
6511 case counter_method: return "compiled methods";
6512 case counter_fpu_method: return "methods using fpu";
6513 case counter_loop_method: return "methods with loops";
6514 case counter_exception_method:return "methods with xhandler";
6515
6516 case counter_loop: return "loops";
6517 case counter_block: return "blocks";
6518 case counter_loop_block: return "blocks inside loop";
6519 case counter_exception_block: return "exception handler entries";
6520 case counter_interval: return "intervals";
6521 case counter_fixed_interval: return "fixed intervals";
6522 case counter_range: return "ranges";
6523 case counter_fixed_range: return "fixed ranges";
6524 case counter_use_pos: return "use positions";
6525 case counter_fixed_use_pos: return "fixed use positions";
6526 case counter_spill_slots: return "spill slots";
6527
6528 // counter for classes of lir instructions
6529 case counter_instruction: return "total instructions";
6530 case counter_label: return "labels";
6531 case counter_entry: return "method entries";
6532 case counter_return: return "method returns";
6533 case counter_call: return "method calls";
6534 case counter_move: return "moves";
6535 case counter_cmp: return "compare";
6536 case counter_cond_branch: return "conditional branches";
6537 case counter_uncond_branch: return "unconditional branches";
6538 case counter_stub_branch: return "branches to stub";
6539 case counter_alu: return "artithmetic + logic";
6540 case counter_alloc: return "allocations";
6541 case counter_sync: return "synchronisation";
6542 case counter_throw: return "throw";
6543 case counter_unwind: return "unwind";
6544 case counter_typecheck: return "type+null-checks";
6545 case counter_fpu_stack: return "fpu-stack";
6546 case counter_misc_inst: return "other instructions";
6547 case counter_other_inst: return "misc. instructions";
6548
6549 // counter for different types of moves
6550 case counter_move_total: return "total moves";
6551 case counter_move_reg_reg: return "register->register";
6552 case counter_move_reg_stack: return "register->stack";
6553 case counter_move_stack_reg: return "stack->register";
6554 case counter_move_stack_stack:return "stack->stack";
6555 case counter_move_reg_mem: return "register->memory";
6556 case counter_move_mem_reg: return "memory->register";
6557 case counter_move_const_any: return "constant->any";
6558
6559 case blank_line_1: return "";
6560 case blank_line_2: return "";
6561
6562 default: ShouldNotReachHere(); return "";
6563 }
6564 }
6565
6566 LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) {
6567 if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) {
6568 return counter_method;
6569 } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) {
6570 return counter_block;
6571 } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) {
6572 return counter_instruction;
6573 } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) {
6574 return counter_move_total;
6575 }
6576 return invalid_counter;
6577 }
6578
6579 LinearScanStatistic::LinearScanStatistic() {
6580 for (int i = 0; i < number_of_counters; i++) {
6581 _counters_sum[i] = 0;
6582 _counters_max[i] = -1;
6583 }
6584
6585 }
6586
6587 // add the method-local numbers to the total sum
6588 void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) {
6589 for (int i = 0; i < number_of_counters; i++) {
6590 _counters_sum[i] += method_statistic._counters_sum[i];
6591 _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]);
6592 }
6593 }
6594
6595 void LinearScanStatistic::print(const char* title) {
6596 if (CountLinearScan || TraceLinearScanLevel > 0) {
6597 tty->cr();
6598 tty->print_cr("***** LinearScan statistic - %s *****", title);
6599
6600 for (int i = 0; i < number_of_counters; i++) {
6601 if (_counters_sum[i] > 0 || _counters_max[i] >= 0) {
6602 tty->print("%25s: %8d", counter_name(i), _counters_sum[i]);
6603
6604 LinearScanStatistic::Counter cntr = base_counter(i);
6605 if (cntr != invalid_counter) {
6606 tty->print(" (%5.1f%%) ", _counters_sum[i] * 100.0 / _counters_sum[cntr]);
6607 } else {
6608 tty->print(" ");
6609 }
6610
6611 if (_counters_max[i] >= 0) {
6612 tty->print("%8d", _counters_max[i]);
6613 }
6614 }
6615 tty->cr();
6616 }
6617 }
6618 }
6619
6620 void LinearScanStatistic::collect(LinearScan* allocator) {
6621 inc_counter(counter_method);
6622 if (allocator->has_fpu_registers()) {
6623 inc_counter(counter_fpu_method);
6624 }
6625 if (allocator->num_loops() > 0) {
6626 inc_counter(counter_loop_method);
6627 }
6628 inc_counter(counter_loop, allocator->num_loops());
6629 inc_counter(counter_spill_slots, allocator->max_spills());
6630
6631 int i;
6632 for (i = 0; i < allocator->interval_count(); i++) {
6633 Interval* cur = allocator->interval_at(i);
6634
6635 if (cur != nullptr) {
6636 inc_counter(counter_interval);
6637 inc_counter(counter_use_pos, cur->num_use_positions());
6638 if (LinearScan::is_precolored_interval(cur)) {
6639 inc_counter(counter_fixed_interval);
6640 inc_counter(counter_fixed_use_pos, cur->num_use_positions());
6641 }
6642
6643 Range* range = cur->first();
6644 while (range != Range::end()) {
6645 inc_counter(counter_range);
6646 if (LinearScan::is_precolored_interval(cur)) {
6647 inc_counter(counter_fixed_range);
6648 }
6649 range = range->next();
6650 }
6651 }
6652 }
6653
6654 bool has_xhandlers = false;
6655 // Note: only count blocks that are in code-emit order
6656 for (i = 0; i < allocator->ir()->code()->length(); i++) {
6657 BlockBegin* cur = allocator->ir()->code()->at(i);
6658
6659 inc_counter(counter_block);
6660 if (cur->loop_depth() > 0) {
6661 inc_counter(counter_loop_block);
6662 }
6663 if (cur->is_set(BlockBegin::exception_entry_flag)) {
6664 inc_counter(counter_exception_block);
6665 has_xhandlers = true;
6666 }
6667
6668 LIR_OpList* instructions = cur->lir()->instructions_list();
6669 for (int j = 0; j < instructions->length(); j++) {
6670 LIR_Op* op = instructions->at(j);
6671
6672 inc_counter(counter_instruction);
6673
6674 switch (op->code()) {
6675 case lir_label: inc_counter(counter_label); break;
6676 case lir_std_entry:
6677 case lir_osr_entry: inc_counter(counter_entry); break;
6678 case lir_return: inc_counter(counter_return); break;
6679
6680 case lir_rtcall:
6681 case lir_static_call:
6682 case lir_optvirtual_call: inc_counter(counter_call); break;
6683
6684 case lir_move: {
6685 inc_counter(counter_move);
6686 inc_counter(counter_move_total);
6687
6688 LIR_Opr in = op->as_Op1()->in_opr();
6689 LIR_Opr res = op->as_Op1()->result_opr();
6690 if (in->is_register()) {
6691 if (res->is_register()) {
6692 inc_counter(counter_move_reg_reg);
6693 } else if (res->is_stack()) {
6694 inc_counter(counter_move_reg_stack);
6695 } else if (res->is_address()) {
6696 inc_counter(counter_move_reg_mem);
6697 } else {
6698 ShouldNotReachHere();
6699 }
6700 } else if (in->is_stack()) {
6701 if (res->is_register()) {
6702 inc_counter(counter_move_stack_reg);
6703 } else {
6704 inc_counter(counter_move_stack_stack);
6705 }
6706 } else if (in->is_address()) {
6707 assert(res->is_register(), "must be");
6708 inc_counter(counter_move_mem_reg);
6709 } else if (in->is_constant()) {
6710 inc_counter(counter_move_const_any);
6711 } else {
6712 ShouldNotReachHere();
6713 }
6714 break;
6715 }
6716
6717 case lir_cmp: inc_counter(counter_cmp); break;
6718
6719 case lir_branch:
6720 case lir_cond_float_branch: {
6721 LIR_OpBranch* branch = op->as_OpBranch();
6722 if (branch->block() == nullptr) {
6723 inc_counter(counter_stub_branch);
6724 } else if (branch->cond() == lir_cond_always) {
6725 inc_counter(counter_uncond_branch);
6726 } else {
6727 inc_counter(counter_cond_branch);
6728 }
6729 break;
6730 }
6731
6732 case lir_neg:
6733 case lir_add:
6734 case lir_sub:
6735 case lir_mul:
6736 case lir_div:
6737 case lir_rem:
6738 case lir_sqrt:
6739 case lir_abs:
6740 case lir_f2hf:
6741 case lir_hf2f:
6742 case lir_log10:
6743 case lir_logic_and:
6744 case lir_logic_or:
6745 case lir_logic_xor:
6746 case lir_shl:
6747 case lir_shr:
6748 case lir_ushr: inc_counter(counter_alu); break;
6749
6750 case lir_alloc_object:
6751 case lir_alloc_array: inc_counter(counter_alloc); break;
6752
6753 case lir_monaddr:
6754 case lir_lock:
6755 case lir_unlock: inc_counter(counter_sync); break;
6756
6757 case lir_throw: inc_counter(counter_throw); break;
6758
6759 case lir_unwind: inc_counter(counter_unwind); break;
6760
6761 case lir_null_check:
6762 case lir_leal:
6763 case lir_instanceof:
6764 case lir_checkcast:
6765 case lir_store_check: inc_counter(counter_typecheck); break;
6766
6767 case lir_fpop_raw:
6768 case lir_fxch:
6769 case lir_fld: inc_counter(counter_fpu_stack); break;
6770
6771 case lir_nop:
6772 case lir_push:
6773 case lir_pop:
6774 case lir_convert:
6775 case lir_roundfp:
6776 case lir_cmove: inc_counter(counter_misc_inst); break;
6777
6778 default: inc_counter(counter_other_inst); break;
6779 }
6780 }
6781 }
6782
6783 if (has_xhandlers) {
6784 inc_counter(counter_exception_method);
6785 }
6786 }
6787
6788 void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) {
6789 if (CountLinearScan || TraceLinearScanLevel > 0) {
6790
6791 LinearScanStatistic local_statistic = LinearScanStatistic();
6792
6793 local_statistic.collect(allocator);
6794 global_statistic.sum_up(local_statistic);
6795
6796 if (TraceLinearScanLevel > 2) {
6797 local_statistic.print("current local statistic");
6798 }
6799 }
6800 }
6801
6802
6803 // Implementation of LinearTimers
6804
6805 LinearScanTimers::LinearScanTimers() {
6806 for (int i = 0; i < number_of_timers; i++) {
6807 timer(i)->reset();
6808 }
6809 }
6810
6811 const char* LinearScanTimers::timer_name(int idx) {
6812 switch (idx) {
6813 case timer_do_nothing: return "Nothing (Time Check)";
6814 case timer_number_instructions: return "Number Instructions";
6815 case timer_compute_local_live_sets: return "Local Live Sets";
6816 case timer_compute_global_live_sets: return "Global Live Sets";
6817 case timer_build_intervals: return "Build Intervals";
6818 case timer_sort_intervals_before: return "Sort Intervals Before";
6819 case timer_allocate_registers: return "Allocate Registers";
6820 case timer_resolve_data_flow: return "Resolve Data Flow";
6821 case timer_sort_intervals_after: return "Sort Intervals After";
6822 case timer_eliminate_spill_moves: return "Spill optimization";
6823 case timer_assign_reg_num: return "Assign Reg Num";
6824 case timer_allocate_fpu_stack: return "Allocate FPU Stack";
6825 case timer_optimize_lir: return "Optimize LIR";
6826 default: ShouldNotReachHere(); return "";
6827 }
6828 }
6829
6830 void LinearScanTimers::begin_method() {
6831 if (TimeEachLinearScan) {
6832 // reset all timers to measure only current method
6833 for (int i = 0; i < number_of_timers; i++) {
6834 timer(i)->reset();
6835 }
6836 }
6837 }
6838
6839 void LinearScanTimers::end_method(LinearScan* allocator) {
6840 if (TimeEachLinearScan) {
6841
6842 double c = timer(timer_do_nothing)->seconds();
6843 double total = 0;
6844 for (int i = 1; i < number_of_timers; i++) {
6845 total += timer(i)->seconds() - c;
6846 }
6847
6848 if (total >= 0.0005) {
6849 // print all information in one line for automatic processing
6850 tty->print("@"); allocator->compilation()->method()->print_name();
6851
6852 tty->print("@ %d ", allocator->compilation()->method()->code_size());
6853 tty->print("@ %d ", allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2);
6854 tty->print("@ %d ", allocator->block_count());
6855 tty->print("@ %d ", allocator->num_virtual_regs());
6856 tty->print("@ %d ", allocator->interval_count());
6857 tty->print("@ %d ", allocator->_num_calls);
6858 tty->print("@ %d ", allocator->num_loops());
6859
6860 tty->print("@ %6.6f ", total);
6861 for (int i = 1; i < number_of_timers; i++) {
6862 tty->print("@ %4.1f ", ((timer(i)->seconds() - c) / total) * 100);
6863 }
6864 tty->cr();
6865 }
6866 }
6867 }
6868
6869 void LinearScanTimers::print(double total_time) {
6870 if (TimeLinearScan) {
6871 // correction value: sum of dummy-timer that only measures the time that
6872 // is necessary to start and stop itself
6873 double c = timer(timer_do_nothing)->seconds();
6874
6875 for (int i = 0; i < number_of_timers; i++) {
6876 double t = timer(i)->seconds();
6877 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);
6878 }
6879 }
6880 }
6881
6882 #endif // #ifndef PRODUCT